\p{L}/\p{Lu}/\p{Ll}/\p{N}/\p{Z}/\p{Ps}/\p{Pe} (+\P negation) land in
both escape positions of the regex compiler, mapped onto the byte PEGs:
ASCII exact, any high byte (inside a UTF-8 sequence) counts as a LETTER —
so ^\p{L}+$ accepts UTF-8 words while \p{N}/\p{Z} stay ASCII. (?u) was
already a tolerated no-op flag. Unknown property names error at compile.
Chasing the acceptance target (cuerdas via deps-conformance) pulled in the
rest of its clj-compat chain, each a real gap:
- the deps-conformance harness reads libraries under clj-compat reader
features (deps are clj/cljc by definition — without :clj, cuerdas's
#?(:clj (instance? Pattern x)) branches resolved to NIL bodies)
- instance? knows Pattern/java.util.regex.Pattern (regex values) and
Character (cuerdas's rx/regexp? gate on split)
- the java.lang.String method surface: .toLowerCase/.toUpperCase/.trim/
.indexOf(-1 on miss)/.lastIndexOf/.substring/.charAt/.startsWith/
.endsWith/.contains/.replace/.equalsIgnoreCase/... — ASCII case mapping,
unknown methods error (the old path silently returned nil)
- the (.method obj args) SUGAR now desugars to (. obj method args) in the
interpreter — it was never implemented (bare .method heads resolved as
vars, hence 'Cannot call nil')
- Long/MAX_VALUE / MIN_VALUE statics (f64 approximations)
deps-conformance: medley ok, cuerdas ok (was check-error); dependency now
loads its clj branches and fails only on its single-segment ns resolution.
30 new spec rows (11 regex, 19 interop). Gate exit 0.
ifn? (jolt-1vx) is the canonical IFn set in the overlay: fns, keywords,
symbols, maps (sorted included), sets, vectors, and vars — NOT lists. The
seed version said true for lists and false for struct maps and vars.
Mutable-mode caveat documented (vectors and lists share the array repr
there). 13 predicate rows.
Multimethod dispatch (jolt-heo) now collects EVERY isa-matching method key
and picks the dominant one — x dominates y when prefer-method'd over it or
(isa? x y) — and two matches with no dominant is an ambiguity ERROR, as in
Clojure. It used to take whichever key the table yielded first, silently
ignoring prefer-method. The prefers store upgrades to Clojure's
{x -> set-of-dominated} shape, shared between the dispatch closure and
prefer-method-setup via the var; prefers becomes a macro over a setup fn
(the store lives on the VAR — the multifn value can't carry it, so the old
fn read {} forever). 6 multimethod rows + the conformance row updated to
the canonical shape (335x3).
The reader (jolt-ou8) kept the pending KEY when a comment or #_ sits in a
map's VALUE slot: the old code dropped both, desyncing kv pairing — the
real value became the next key and the closing brace landed in value
position ('Unmatched closing brace'). Selmer's deps.edn (a '; for
development (REPL, etc)' comment between key and value) now parses; 6
reader rows incl. nested commented maps.
Gate: jpm exit 0, conformance 335x3, all tests passed.
jolt.passes is the new portable pipeline stage between the analyzer and the
back end: pure IR -> IR rewrites, total over node :ops (unknown ops pass
through with folded children), loaded with the compiler namespaces and
resolved lazily by analyze-form (JOLT_NO_IR_PASSES=1 disables — the same
escape-hatch pattern as the macro oracle). The shape is flatiron's opt.clj
applied to the jolt IR, which is what jolt-2om asked for.
The first pass is constant folding: a call of a foldable numeric SEED fn
(the later tiers don't exist when the compiler loads) whose args are all
constant numbers becomes a constant, and an if with a constant test becomes
the taken branch (dead-branch elimination — the untaken side never even
resolves). Folding computes with the ACTUAL jolt fns, so results match
runtime semantics by construction; a fold that would throw (mod 5 0) is
left for runtime.
Two walk lessons paid for in debugging: let/loop bindings are
[name init-ir] PAIRS, not maps (assoc'ing :init into a pair corrupts it);
and a throw inside the interpreted pass unwinds past the interpreter's ns
restores, so analyze-form restores the compile ns after the (protected)
pass call — without that, one pass error left current-ns in jolt.passes and
the rest of the tier compile resolved against the wrong namespace (sort-by
landed on the 2-arg JANET builtin).
ir-passes-test pins folds, conservatism (free vars, throwing folds), and
end-to-end eval. Gate exit 0.
Every walk over a lazy seq created FRESH wrapper tables around the shared
rest-thunks (ls-rest, ls-seq/ls-count, realize-for-iteration, the printers,
reduce — each had its own make-lazy-seq loop), so independent walks re-ran
the thunks: side effects duplicated, and a doall'd seq of futures was
re-spawned serially by the deref walk. Every walker now goes through
ls-rest-cached, which memoizes the rest wrapper on its node — thunks run
exactly once, as in Clojure. Costs ~10% on walk-heavy benches (the per-node
cache get/put — Clojure's LazySeq pays the same); net still -9% vs the
pre-linear-walks baseline. Three regression rows pin once-only effects and
value stability across walks.
On top of that: pmap/pcalls/pvalues (jolt-oeu) over the real-thread futures
— spawn-all-then-deref (the once-only fix is what makes the doall actually
mean that), snapshot semantics documented, multi-coll arity via the
canonical vector-zip. System/currentTimeMillis + nanoTime land as System
statics (the realtime clock — os/time is whole seconds, which quantized
every elapsed measurement to 1000ms). Seven pmap rows incl. a generous-
margin parallelism check (4 x 200ms sleeps under 700ms after warmup).
The native-ops table grows from 9 to 16, each verified for semantic parity
with the jolt fn before inclusion (incl. negative operands): mod is floored
on both sides, rem (janet %) truncates, / is variadic with (/ x) -> 1/x.
quot is deliberately absent — janet div floors where Clojure truncates.
jolt's bit fns are 2-arg (unlike Clojure's variadic), so the bit ops emit
native only at exactly that arity; bit-not is unary. Eight new conformance
rows pin compile=interpret on the new ops at their guarded arities (334x3).
map-read 10.8 -> 9.2 ms (the (mod i 100) in its loop inlines). From the
flatiron review's unchecked-primitive-loops idea.
Reviewing flatiron's morsel batching for applicability turned up something
better hiding under the lazy machinery: every cell over a concrete
collection was produced by slicing the REMAINDER ((tuple/slice c 1) per
element in coll->cells, and rest/next sliced the same way), so any full walk
was O(n^2). mapv over 40k elements took 10.4 SECONDS; a 20k-element
first/rest loop took two.
Cells over indexed collections now walk by INDEX (one shared indexed-cells
helper, O(1) per step), and rest/next of a vector/tuple/list return an O(1)
lazy view from index 1 — which also makes (rest [1 2 3]) a SEQ, as in
Clojure (it was a vector-typed slice; seq?/vector? rows pin the change).
mapv 40k: 10405 -> 182 ms. rest-loop 20k: 2040 -> 31 ms. Whole bench:
seq-pipe -28%, into-vec -24%, str-join -18%, hof -26%, TOTAL 4565 -> 3753
(-18% vs main, back to back). Chunked seqs (jolt-yqc) drop in priority:
the quadratic walks were the actual cost; chunking now only amortizes
per-element closure allocation.
Nine regression rows incl. 20k/50k linear-scaling smoke tests. Gate exit 0.
require-:as wrote the string-keyed :imports table (which resolution reads)
while ns-aliases read the symbol-keyed :aliases table (which nothing wrote)
— so (ns-aliases) was always empty and the alias fn had to write both as a
bridge. :aliases (alias-name string -> ns-name string) is THE store now:
require :as and the alias fn write it, both resolution paths read it first
(falling back to :imports for class imports, which is all that table holds
now), ns-unalias removes one entry, and ns-aliases presents Clojure's
{alias-symbol -> namespace object} shape built from it. ns-resolve's
qualified path goes through the same lookup.
Also: the coverage dashboard's last 'resolvable-not-interned' entry was '.'
— which (resolve '.) returns nil for on the JVM too; the tool now classifies
it as the special form it is, and that category reads ZERO.
7 new unified-alias spec rows (require/alias/ns-unalias round-trips through
both the resolution and introspection views); the white-box namespace test
tracks the accessor rename. Gate exit 0.
(recur (inc acc) (rest xs)) re-entered the fn through its varargs collector,
so the rest seq came back wrapped in a fresh 1-element rest list — xs never
emptied and the interpreter hung (jolt-4df; the compile path was already
correct). recur now re-enters through a dedicated entry that binds the LAST
arg directly as the rest param (n-fixed + 1 args, Clojure's contract), in
both the single-arity and multi-arity fn* paths; the shared body runner keeps
the ns-swap/restore in one place, and fixed arities still re-dispatch through
the arity dispatcher exactly as before.
Six spec rows: the original repro, zero-fixed variadic, rest-empties-to-nil,
multi-arity variadic recur, nil rest, and a fixed-arity control.
clojure.edn was nearly complete (sets, #uuid/#inst, :eof all landed earlier);
the :readers opt was ignored and :default missing. Both work now — the
reader stores a tag as a :#name keyword, so the lookup normalizes it to the
symbol Clojure keys :readers with; :default gets (tag value); built-in data
readers stay the fallback. 8 new edn spec rows. This was the last open item
of jolt-0mb (the vendored walk/zip/data/edn battery has been green for a
while: 34/33/61/50, all clean).
Chasing the probe cascade ('Unable to resolve symbol: edn/...' after one
error) found a real evaluator bug: an interpreted fn body runs with
current-ns rebound to its DEFINING ns and restores it with a plain trailing
call — an UNCAUGHT throw skips every restore on the way out, leaving the ctx
stuck in the deepest callee's ns, where alias-qualified lookups then fail
(the same cascade previously seen via sci). The repair lives at the
TOP-LEVEL boundary (loader/eval-toplevel saves the entry ns and restores it
on error before re-raising) — NOT per-call defer/try, which builds a fiber
per frame and blew the C stack on deep interpreted recursion (file-seq)
when tried first. Regression tests cover the cross-eval leak and that
aliases keep resolving.
test/spec/untested-vars-spec.janet adds 143 rows asserting jolt's documented
behavior for the whole implemented-untested category — primed arithmetic,
the array/aset/coercion stubs, unchecked-*, the chunk family, JVM-shape
stubs (class/bean/proxy/memfn as resolve-only or :throws), ns/REPL
machinery, and the misc seqs. tools/spec_coverage.py now checks each var as
a whole TOKEN in the test sources (call-position-only matching missed *1,
+', ., .., /, and bare transducer refs like cat).
Writing rows from probed truth surfaced five real bugs, all fixed:
- comp with a jolt-IFn stage silently returned nil ((comp seq :content)) —
raw Janet keyword application is not jolt invoke. comp is the canonical
overlay defn now (fixed-arity composed fn, so the hot 1-arg path is two
direct calls); the seed keeps a private td-comp only for the transducer
machinery. hof bench +9% vs native, the price of correct IFn dispatch.
- extend (the fn) was a nil-expanding stub MACRO shadowing any definition;
it's a real fn over register-method now, and extends? (a constant-false
stub) is real over extenders
- (.. x f g) hit the 'ClassName.' constructor branch (a name ending in a
dot) and died; .. is the canonical threading macro now
- aclone errored on pvecs; ns-interns/ns-imports returned live host tables
that count/seq reject (now structs)
Thread/sleep + Thread/yield land as Thread statics beside Math/: sleep parks
the WORKER's own event loop (each future thread has one), which makes timed
deref provably fire — futures-spec gains the timeout-fires, sleep-in-body,
and timed-out-future-still-completes rows. The futures impl itself already
ran on real OS threads (ev/spawn-thread + marshalled results); jolt-ejx was
stale.
Dashboard: implemented+tested 433 -> 564 of 694; implemented-untested and
missing-portable are both EMPTY. Gate: jpm exit 0, all tests passed.
The dashboard's missing-portable category is now EMPTY (was 35 when the issue
was filed; this session's io/leaf work had already landed most of them).
The final seven:
- extenders — ctx-capturing clojure.core fn over the protocol type-registry:
the type-tags implementing a protocol, as symbols; nil when none
- find-keyword — keyword: jolt keywords have no intern table, so it always
finds (babashka makes the same call)
- inst-ms* — the raw Inst method; one inst representation, so = inst-ms
- read+string — over the 50-io readers, which now expose :buf and :fill-fn;
returns [form exact-text-consumed], EOF throws or yields [eof-value ""]
with the 3-arity, works for string AND stdin readers
- with-local-vars — fresh free-standing var cells (__local-var seam) bound as
locals; var-get/var-set work on any cell
- with-open — canonical recursive expansion closing through the __close seam:
a map-like value's :close fn or a host file (no .close interop here);
nested closes run inner-first, finally runs on throw
- with-precision — body evaluates with precision/:rounding accepted and
ignored (doubles, no BigDecimal context) — documented divergence
30 new spec rows (test/spec/missing-vars-spec.janet); coverage.md
regenerated: implemented+tested 426 -> 433, missing-portable 7 -> 0.
Gate: jpm exit 0, all tests passed.
The phm had a FIXED 8 buckets, so a 100-entry map was a ~12-entry linear
scan per lookup — and phm-get walked the bucket twice (contains? then find).
This went mostly unnoticed until the canonical zipmap (batch 2) started
returning phms where kvs->map had built structs for scalar keys, regressing
the map-read bench ~7x (jolt-s3y).
phm-assoc now rehashes into a doubled bucket array when the count passes 2
entries/bucket (done on the fresh copy, so persistence is untouched);
phm-get is single-pass with a presence flag (nil values still distinguish
from missing); key= tries identity/scalar equality before paying for
canonicalization; the bucket count is derived from (length (m :buckets)),
not a constant, so any already-marshaled map keeps working. core-contains?'s
phm branch goes through phm-contains? instead of poking buckets directly.
map-read 48.5 -> 10.9 ms (the residual vs the pre-batch-2 6.7 is the
canonicalizing-representation constant); map-build steady; bench TOTAL 4457
vs 4565 on main back-to-back. New unit case crosses the resize boundary at
500 entries: every key found, nil values present, collection keys canonical,
dissoc + persistence intact. Gate: jpm exit 0, conformance 326x3.
nrepl-test was CI's only failure: the server subprocess ran main.janet from
source, paying the full compile-mode init, which outran the 5s connect poll
on slow runners (locally it always won the race). The test now prefers
build/jolt — its ctx is baked at build time, so it accepts in ~20ms and CI
builds it anyway — falls back to source, polls up to 60s under a 90s
watchdog, and dumps the server's stderr when startup fails so the next CI
failure is diagnosable.
CLAUDE.md's placeholder sections become real: build/test commands with the
run-the-gate-with-a-real-exit-code protocol (a piped gate once shipped masked
spec failures), the seed/overlay/tier architecture sketch, and the porting
gotchas that have each bitten at least once (leaf verification, stub-breaks-
self-recursion, tier macro ordering, ref-get vs get on attached-ops wrappers,
:jolt/type map keys, expander-called fns, canonical-port policy) — previously
only in local bd memories.
An indirect global reference emitted ((var-get 'cell) ...) — a function call
per deref, whose body is a binding-stack check plus a root read. The emitter
now inlines that: (if (in 'cell :dynamic) (var-get 'cell) (in 'cell :root)).
Non-dynamic vars — the vast majority of references — pay two native table
ops and a branch instead of a call; dynamic vars take the full var-get
(thread-binding walk). Redefinition stays live (:root is read per call) and
binding semantics are exact: the :dynamic check is PER CALL, not at emit,
because a (def ^:dynamic x) in the same compiled unit marks the cell dynamic
only when the def runs — the same reason JVM Clojure's Var.deref() checks
the thread-bound bit every call (an emit-time variant was 1.7x faster still
on fib but failed conformance exactly there).
fib 130 -> 74 ms (1.75x); bench TOTAL 4564 -> 4437 back-to-back. This
displaces the gen-counter inline-cache design from jolt-8sq: with var-get's
existing fast path, resolution was never the cost — the call was. A
gen-guarded cache would add state per site to save nothing further, and
couldn't skip the dynamic check anyway.
Found while benching: map-read regressed ~7x back in batch 2 (canonical
zipmap builds a phm where kvs->map built a struct) — filed as jolt-s3y.
Gate: jpm test exit 0, conformance 326x3, suite >= baseline.
sort-by, rand-int, shuffle, random-uuid, char-escape-string, and
char-name-string move to 20-coll over the two host seams that stay (rand and
sort — they ARE the randomness/ordering primitives). Canonical upgrades ride
along: sort-by defaults its comparator to compare, so nil sorts FIRST (the
kernel fn used host ordering and put nil last); rand-int truncates toward
zero via int (the kernel fn floored, wrong for negative n); shuffle is a
pure-functional Fisher-Yates over vector assoc and rejects non-collections
(a string is seqable but not shuffleable, as on the JVM — the honest gate
caught that one); random-uuid builds over rand-int and validates through
parse-uuid; the char tables are char-keyed Clojure maps (Clojure's shape —
the seed keeps its private code-keyed copies for pr-render).
22 new spec rows. Gate: jpm test exit 0 verified, suite 4698 >= 4660, bench
parity with main back-to-back (4733 vs 4817).
recompile-defns! is the defn analog of recompile-macros!: pre/at-kernel
overlay defns (00-syntax's destructure and friends; the kernel tier too in
interpret mode) load as interpreted closures, the evaluator stashes their fn
source on the var (:defn-src, scoped by a flag only api/load-core-overlay!
sets), and the end-of-init pass compiles them and swaps the var root. With
that in place, keys/vals/empty? — the fns the 00-syntax expanders call at
expansion time — move to the top of 00-syntax as raw fn* defs (canonical:
keys/vals project (seq m), so sorted maps come back in comparator order and
(keys {}) is nil; empty? keeps O(1) count dispatch with seq's cell check only
for the lazy/list fallback). The sorted tier drops its now-dead :keys/:vals
ops.
Correctness fixes that surfaced once the gate was run with a REAL exit code
(the previous 'jpm test | grep' gates reported grep's exit and masked spec
failures across #48-#50):
- map conj is strict again: a non-nil/non-map arg must be a 2-element vector
('Vector arg to map conj must be a pair'), and merge inherits it — the
batch-2 canonical merge had silently dropped the validation
- conj onto a lazy seq prepends (it fell into the MAP fallback); upstream
clojure.data/diff relies on (conj seq x) via set/union over keys, so diff
now matches Clojure exactly
- (seq {}) / (seq #{}) / empty phm are nil, not ()
- key/val are strict (a plain vector is not an entry); find mints a REAL
entry as the first entry of a one-entry map, nil values intact
- the sci avoid-method-too-large stub passes its registry map through
instead of returning a raw host table (strict conj rejected it; sci's
clojure-core registry is also no longer discarded)
Test updates: lazy-infinite pins take-nth realization at 5 (was 7 — the
canonical lazy impl realizes fewer); self-host asserts the analyzer IS loaded
in interpret mode (compiled expanders, PR #50) and is NOT in the
:compile-macros? false oracle. 18 new maps-spec rows.
Gate: jpm test exit 0 (verified directly, not through a pipe), conformance
326x3, suite 4698 >= 4660.
Macros are ordinary compiled fns in Clojure's model; compile mode has had
that since the staged bootstrap, but interpret mode — the conformance
battery's default — kept interpreted expanders, so every distinct (and ...)/
(cond ...) call form, and every fresh form produced by a recursive expansion,
ran an interpreted closure. ensure-macros-compiled! now runs in every mode:
interpret-mode init loads the tiers fast-interpreted, then one pass at the
end builds the analyzer (which itself stays interpreted there) and compiles
all stashed expanders; user defmacros after init compile too. The new
:compile-macros? opt (JOLT_INTERPRET_MACROS=1) preserves the fully-
interpreted oracle, and joins the ctx-image cache key.
Battery: 4700 pass / 90 clean files / 7 timeouts, from 4577 / 87 / 9 — two
macro-heavy files stopped timing out and 149 more assertions execute. The
compiled-expander delta proper is +67 passes (oracle mode on the same tree
measures 4633). Baselines raised 4540->4660, clean 86->88. Interpret init
grows 0.12s -> 1.12s for the analyzer build; init-cached amortizes it to ~5ms
per process.
New macro-expansion-test pins: expanders compiled in interpret mode (core +
post-init user defmacros), uncompilable bodies fall back interpreted and
still work, compile mode unchanged, oracle opt-out honored.
Follow-up filed (jolt-4j3): the same staged-recompile treatment for early
overlay DEFNS, which is what still pins keys/vals/empty? to the seed.
empty, assoc-in, and update-in move to 20-coll.clj as the canonical recursive
ports; interpose and take-nth move to the lazy tier WITH their canonical
transducer arities (volatile-based), so the seed's td-interpose/td-take-nth
helpers go too. (empty lazy-seq) is () now — the kernel fn returned a bare
host table for it.
keys/vals/empty? stay put for now: they're expander-coupled — 00-syntax's
when/and/or/cond/destructure expanders call them at expansion time, which
happens during the kernel-tier compile, before any later tier exists. They
move when early defns get the staged-recompile treatment macros already have.
26 new spec rows (incl. transducer arities through sequence/into and laziness
checks against (range)). Gate green: conformance 326x3, suite >= baseline,
full jpm test.
key/val/select-keys/zipmap/merge/merge-with/get-in/memoize/partial/
trampoline/some?/true?/false?/max/min/reverse move to 20-coll.clj as the
canonical Clojure definitions, plus find — which was previously missing from
jolt entirely (select-keys/merge-with/memoize build on it). Two behavior
fixes ride along: memoize now caches nil results (the kernel fn re-computed
them — canonical find-based impl), and conj of nil onto a map is a no-op as
in Clojure (it errored; the canonical merge relies on it). max/min keep the
JVM NaN behavior by construction (pairwise >/<). not= stays: the kernel tier
(subvec) uses it.
One new tier-ordering rule, learned the hard way: a tier may only use macros
from tiers that load BEFORE it — memoize's if-let (30-macros) broke compiled
init while interpret mode passed, because compile expands macros at tier
load and the interpreter expands lazily. Now documented in the migration
workflow note.
MIGRATION.md is gone — task tracking lives in beads (jolt-ded; the per-batch
workflow, tier-order rules, perf wall, and remaining candidates are in bd
memory core-migration-workflow). The doc's candidate lists had gone stale
against the actual seed anyway.
43 new spec rows. Gate green: conformance 326x3, suite >= baseline, full
jpm test, bench at parity with main back-to-back (4851 vs 4831 TOTAL).
Nine more seed leaves move to 20-coll.clj (verified leaf-by-leaf: defn +
core-bindings entry only, no internal callers). fnil is upgraded to Clojure's
canonical 2/3/4-arity — it patches only the first 1-3 arguments; the old
kernel fn patched every position it had a default for, which Clojure does
not. The rest carry their kernel semantics over unchanged (bigdec is a
double, numerator/denominator throw, supers is #{}, munge rewrites dashes).
16 new spec rows incl. the fnil arity-contract cases. Gate green:
conformance 326x3, suite 4577, full jpm test (2:18 — first full run with the
ctx image cache on main).
*in*, read-line, read, with-in-str, and line-seq land as a new overlay IO
tier (core/50-io.clj). *in* is a dynamic var holding a reader — a plain map
of two closures, :read-line-fn (next line, nil at EOF) and :read-fn (next
form, advancing past exactly that form). The default *in* reads real stdin
with a shared leftover buffer, so read and read-line interleave correctly;
with-in-str rebinds *in* to a string reader over one atom-held buffer —
(read) consumes its form, a following (read-line) returns the rest of that
line, as in Clojure. read has the 0/1/3 arities (EOF throws, or returns
eof-value when eof-error? is false).
The Janet seed grows two seams next to read-string: __stdin-read-line (one
line off stdin, newline stripped) and __parse-next (one form off a string ->
[form rest], nil at end of input) — and loses the line-seq stub.
Two traps hit and documented for future tiers: a map LITERAL with :jolt/type
as a key is read as a tagged form (don't tag overlay value maps), and a
leftover seed stub holding the same name breaks direct-linked self-recursion
— the overlay line-seq's recursive call bound to the stub's root, truncating
after one line. The stub's string-splitting behavior is kept as a documented
extension.
20 new io-spec rows (read-line EOF/interleave, read arities + eval round-trip,
line-seq incl. real-stdin paths). Gate green: conformance 326x3, suite 4577,
full jpm test.
sorted-map/sorted-map-by/sorted-set/sorted-set-by/sorted?/sorted-map?/
sorted-set?/subseq/rsubseq now live in their own overlay tier (25-sorted.clj).
A sorted coll is a tagged host table with a comparator-ordered :entries
vector, a 3-way :cmp, and the tier's op implementations ATTACHED to the value
(:ops map): the seed's conj/assoc/get/seq/count/... branches are each a
one-line call through (coll :ops), so the ops travel with the value — correct
across contexts, forks, and AOT images, no module-level hooks to re-wire.
The host surface grows by three minimal value primitives: jolt.host/
tagged-table, ref-put! (already there), and ref-get — a raw field read,
because plain get on a sorted coll IS the comparator lookup and reading
:entries with it recurses.
This fixes a pile of Clojure-correctness gaps the Janet kernel had:
- lookup/membership now go through the COMPARATOR: (contains? (sorted-set 1)
1.0) was a deep= scan, (conj (sorted-set 1) 1.0) and assoc of a
comparator-equal key now no-op/replace as in Clojure
- equality is representation-agnostic: (= (sorted-map :a 1) {:a 1}) and
(= (sorted-set 1 2) #{1 2}) were false
- iteration was broken: (map inc (sorted-set 3 1 2)) errored
(realize-for-iteration and coll->cells had no sorted branches)
- empty?/empty saw the host wrapper, not the collection: (empty? (sorted-map))
was false, (empty sc) returned a bare table; it now keeps the comparator
- sorted colls canonicalize as map keys; comparator fns may be boolean
predicates or 3-way (Clojure's fn->comparator)
- sorted-map throws on odd kv count; conj nil is a no-op
Also fixes jolt-h86 en passant: into-conj had no branch for sets (or sorted
colls) and silently returned the target unchanged — (into #{} [:a :b]) was
#{}. The fallback now folds conj. Regression rows in sets-spec.
sorted-spec grows to 77 rows (comparator-based membership, equality,
empty/rseq/printing, seq-fn interop, subseq/rsubseq on maps). Gate green:
conformance 326x3, suite 4577 (vs 4566 prior — the battery gained rows),
sorted+sets specs, full jpm test, bench at parity with main back-to-back
(4521ms vs 4619ms TOTAL under identical load).
Pins down what a transient is in Jolt (tagged table over a native Janet
array/table, canonical-keyed for maps/sets), where behavior deviates from the
JVM (O(n) transient/persistent! edges with O(1) native ops between, no
owner-thread check — same as Clojure 1.7+, transient-of-list leniency), and
the three reasons the machinery is seed-resident rather than a migration
candidate: it IS the mutation kernel, it sits under the seed's own dispatch,
and the value layer is declared irreducible. Exists so the kernel-shrink
ladder (jolt-tzo) doesn't revisit transients every round.
init in compile mode is ~2.4 s (tier loading, analyzer self-compile, macro
recompilation), paid by every process that builds a ctx from source — each
jpm-test file, embedders, workers. init-cached marshals the built ctx to a
disk image (same root-env dicts as snapshot/fork) and later processes
unmarshal it in ~5 ms, any process: nothing from the baking process is
needed at load.
The cache key fingerprints the embedded .clj stdlib (which covers jolt-core:
analyzer, IR, core tiers), the .janet seed sources next to the module, the
janet version, the init opts, and the env knobs that shape a ctx (JOLT_PATH/
MUTABLE/AOT_CORE/FEATURES) — any change rebuilds. Corrupt or non-ctx images
fall back to a rebuild (unmarshal of garbage can 'succeed' with a scalar, so
the shape is checked, not just the throw). Writes are atomic (tmp + rename)
so racing cold starts never publish a torn image. JOLT_NO_IMAGE_CACHE=1
opts out; JOLT_IMAGE_CACHE_DIR overrides the location (default TMPDIR).
Test consumers switch to init-cached (harness, suite-worker, conformance,
the behavioral unit/integration tests); tests that validate the bootstrap
itself (bootstrap-fixpoint, staged-bootstrap, aot round-trip, direct-linking)
and the deps tests (tmp-dir :paths would fragment the key) keep real init.
Full jpm test: 2:46 -> 1:58 (~29%). New ctx-image-test covers cold/warm,
cross-process load (subprocess runs defn/redef/macros/protocols/multimethods
off the baked image), per-opts keying, and corrupt-image fallback.
The 1104-line Janet bootstrap compiler existed to build jolt.ir/jolt.analyzer
and the kernel tier before the self-hosted analyzer could exist. It is
replaced by the interpreter + one fixpoint turn:
1. bootstrap-load-source loads the compiler sources INTERPRETED (the
evaluator can run the analyzer — it always could).
2. After the overlay is up, self-compile-compiler! re-runs the kernel tier,
jolt.ir, and jolt.analyzer through the SELF-HOSTED pipeline — the
interpreted analyzer compiles itself, and steady state runs compiled with
no bootstrap compiler involved.
Measured: init {:compile? true} 1093 -> ~2400 ms (the one-time interpreted
pass + self-compilation), but steady-state compilation is 2.8x FASTER
(100 forms: 134 -> 48 ms) — the self-hosted pipeline emits better code than
the bootstrap did. An AOT image for init cost is future work (aot.janet's
machinery is the natural vehicle).
The bootstrap's runtime kernel moves to backend.janet (jolt-runtime-env,
ctx-janet-env, build-map-literal); aot imports it from there. The
uncompilable-error? punt check unwraps the interpreter's exception struct
(the interpreted analyzer's throw arrives wrapped). compile-string/
compile-file (the bootstrap's source-text emitter API, no callers outside
the bootstrap's own unit test) are removed with it, as is compiler-test.
Gate green across everything incl. fixpoint stage1==2==3, AOT round-trip,
uberscript, CLI; conformance 326x3; suite 4566 >= 4540; bench in band.
