* Fix four bugs surfaced by the commonmark-app example
- regex: bounded quantifiers {n,m} no longer expand exponentially. The
desugaring inlined the continuation per optional level, doubling it each
time, so {0,61} built a PEG the compiler expanded to ~2^61 nodes and hung.
Each level is now its own grammar rule referenced by name (jolt-3xur).
- strings are a seqable of chars for vec/set/into, matching seq. They went
through realize-for-iteration, which had no string case, so into/set got
raw bytes (code points) and set threw; vec used string/from-bytes (1-char
strings). realize-for-iteration now maps make-char over the bytes, and
core-vec matches (jolt-dl4s).
- clojure.string/split takes Java Pattern.split limit semantics: negative
keeps trailing empties, 0/omitted drops them (a no-match result stays
[input]), positive caps the part count with the remainder unsplit. It used
to delegate the limit to take, so a negative limit returned [] and the
2-arg form never trimmed trailing empties (jolt-ik3a).
- System/exit is registered (maps to os/exit), so (System/exit n) works
(jolt-w2wf).
* regex: single-digit backreferences \1..\9 (jolt-xtss)
\1..\9 now match the text captured by the corresponding group, so
patterns like ([-*_])\1\1 or (\w+) \1 work. The parser records which
groups are referenced; a referenced group additionally captures its text
under a tag and the backref compiles to a PEG (backmatch). Only referenced
groups change — they match possessively (the CPS-over-possessive-PEG engine
can't backtrack into a tagged capture), so backtracking back into a
backreferenced group isn't supported (rare). Unreferenced groups keep full
backtracking and position-based result capture unchanged.
---------
Co-authored-by: Yogthos <yogthos@gmail.com>
canon-key canonicalizes a collection key by re-keying a native Janet
table by the canonical form of each element/entry. canon-key returns nil
for nil, and a Janet struct can't hold a nil key or value, so a nil set
member / nil map key / nil map value was dropped during canonicalization
— #{nil 1} canonicalized like #{1} and collided as a map key. So
(count {#{nil 1} :a, #{1} :b}) was 1 and (get {#{nil 1} :a} #{1}) was :a.
Box a nested nil before it goes into the table. The marker has to be
value-hashable, not the identity-hashed mutable-table sentinel the
transients use: the canonical struct becomes a long-lived phm key, and
its hash must survive the marshal/snapshot/fork that init-cached relies
on — an unmarshalled table gets a fresh address, so its hash isn't
preserved and the map can't find its own key. An interned keyword hashes
by content. Collision risk is only a real value equal to that exact
keyword, the same negligible class as canon-key's existing set/map struct
aliasing.
The transient sentinel stays a mutable table (it's built and consumed
within one op, never crossing a marshal boundary, so identity hashing is
stable there).
Co-authored-by: Yogthos <yogthos@gmail.com>
Two paths dropped a nil set member. phs-seq read members via
phm-to-struct, whose Janet struct can't hold a nil key, so the nil was
lost on seq/print and on into-an-existing-set even though the backing phm
counted it (count/contains? then disagreed). Re-attach the nil from the
phm's has-nil slot, keeping struct-key order for the rest so set printing
stays stable.
The transient set keyed its native table by canon-key and stored the
member as the value; canon-key returns nil for nil and a nil value also
drops the entry, so conj!/disj!/contains?/persistent! lost a nil member
outright. Key by tbl-key (the same nil sentinel the transient map uses)
and box a nil value through tbl-nil-key, unboxing on persistent!.
A nil-containing set used as a map key still collides with the nil-free
one (canon-key drops nil during key canonicalization) — separate latent
bug, filed as jolt-zcm9.
Co-authored-by: Yogthos <yogthos@gmail.com>
The map build already used a transient map, but each bucket was rebuilt with
a persistent (conj (get ret k []) x) per element — an O(log n) trie path
rebuild + alloc each. A coarse grouping (few large buckets) was bound on that
conj, not the map build. Buckets are now native arrays (transient vectors,
O(1) push) frozen once; distinct keys are tracked in a side vector so the
buckets freeze in place with no second map rebuild. A bucket's first element
stays a cheap persistent [x] and only promotes to a transient on the second,
so an all-singletons grouping pays no transient alloc.
coarse (10/100 buckets, 50k): ~313ms -> ~125ms (~2.5x)
2 buckets (50k): ~322ms -> ~129ms (~2.5x)
all-unique (50k): ~949ms -> ~892ms (no regression)
Surfaced a latent bug: canon-key returns nil for a nil key and Janet tables
drop a nil key, so the canon-keyed transient map silently lost a nil-key
entry — group-by/frequencies/assoc!/into{} dropped the whole nil bucket
((group-by identity [nil nil 1]) gave {1 [1]}, not {nil [nil nil], 1 [1]}).
Route nil through a sentinel (tbl-key) at the transient-map keying sites;
persistent!/count/dissoc! work unchanged since the real [nil v] pair is kept
as the stored value, and phm already has its own has-nil slot. The transient
set has the analogous bug (needs phs nil support) — filed separately.
Co-authored-by: Yogthos <yogthos@gmail.com>
into {}, frequencies, group-by, set, into #{} and persistent! all built
their result by folding an immutable assoc/conj per element — each call
rebuilt the O(log32 n) trie path and allocated a fresh wrapper. Add a
one-pass bottom-up HAMT builder (phm-from-pairs) and route the builders
through it, the map/set analog of the pvec bulk build in #153.
phm-from-pairs partitions entries by hash and constructs the bin/array/
collision nodes directly, with the same bin<=16 / array-node>=17 promotion
the incremental path uses — so the trie is byte-identical to one built by
phm-assoc (validated across the size and branching boundaries, including
hash collisions, duplicate keys and the nil key). persistent! map/set and
the set constructor bulk-build; into {} keeps the small-scalar-map-stays-a
-struct rule via bulk-map-from-pairs; frequencies/group-by switch to the
canonical transient form and ride the fast persistent!.
50k A/B: into {} 704->270ms, frequencies 582->160, set 615->241,
into #{} 702->240, group-by 1358->919 (bound on persistent vector conj).
Gate: conformance x3, full suite (4718 >= baseline), new maps/sets bulk
boundary specs.
Co-authored-by: Yogthos <yogthos@gmail.com>
pv-from-indexed (behind vec/mapv/filterv/into-a-vector) built a pvec by calling
pv-conj once per element — each call allocated a fresh pvec wrapper and copied
the up-to-32 tail tuple, so building an n-vector was O(n) allocations + tail
copies. Replace it with a single bottom-up trie construction: chunk the elements
into 32-wide value leaves, group nodes 32-wide up to the root, split off the
tail. The structure is identical to the incremental one — tail-offset(n) =
((n-1)>>5)<<5 is exactly the trie/tail boundary, so nth/conj/assoc/seq read it
unchanged (validated against the old builder across the size boundaries).
into-a-vector likewise stops doing a persistent pv-conj per element: it
accumulates into a native array and bulk-builds once (the transient-style path).
Measured (50k): vec 211 -> 6 ms (~36x), into [] 197 -> 15 ms (~13x). mapv is
unchanged here — it's bottlenecked on lazy map realization, not the build.
The map/set builders (into {}, frequencies, group-by, set — all HAMT-backed)
need the same bulk treatment and are a separate follow-up. Gate: conformance x3,
full suite, new bulk-boundary rows in vectors-spec.
Co-authored-by: Yogthos <yogthos@gmail.com>
Add a PNG writer so the demos produce actual images. Two pieces:
- src/jolt/png.janet — the encoder (8-bit RGB, filter None, stored/uncompressed
DEFLATE so no compressor is needed; correct CRC32 + Adler32). It lives in Janet
because per-byte work in the overlay is far too slow (a byte-array aset loop is
~30s for 360k bytes, and CRC32 over even a tiny image would be worse). Janet's
bit ops are 32-bit signed, so the 32-bit-unsigned arithmetic is done with plain
number ops (doubles hold 2^32) plus byte-level bxor. Exposed to the overlay as
janet.png/* by importing it into eval_base's module-load-env.
- src/jolt/jolt/png.clj — the jolt.png overlay wrapper: image / put! / write. The
overlay only produces pixels; the host encodes them in one pass.
mandelbrot gets a `render` subcommand (jolt -m mandelbrot render out.png [size])
that colours count-point's escape counts; the numeric-arg bench path is untouched.
Verified end to end: macOS `sips` accepts the output (so CRC/zlib are valid).
png-test covers the encoder structure (signature/IHDR/IEND) and the overlay
round-trip.
Co-authored-by: Yogthos <yogthos@gmail.com>
emit-loop compiled every loop/recur to a self-recursive local closure called once
per iteration — relying on Janet TCO for stack safety but paying a fn frame + arg
bind each iteration. The jolt-5vsp spike localized the whole ~1.43x
jolt-over-hand-Janet gap on compute loops to exactly this.
Lower instead to a Janet `while` + state vars: the loop bindings become vars
carried across iterations, a recur writes them and raises a continue flag, and a
non-recur tail value falls out through a result var. recur-name routing in
emit-recur picks the while-set lowering for loops and leaves the fn-arity self-call
path untouched.
The one subtlety is closure capture: Janet closures capture vars BY REFERENCE, so
a closure built in the body over a shared mutable loop var would see the final
value ([3 3 3]) instead of its iteration's ([0 1 2]). Each iteration rebinds the
loop names into a fresh immutable `let` before running the body, which restores
per-iteration capture. recur reads those immutable bindings and writes the state
vars, so cross-referencing args (swap, fib) need no temps.
mandelbrot 218 -> 164 ms (~11.2x JVM, from 15x). fib is unaffected — it's
fn-arity recursion, not a loop. Regression spec in control-flow-spec covers
closure capture, no-clobber recur, nested loops, sequential init, recur through
let, and that fn-arity recur still works. Gate green (conformance x3, full suite).
Note: validating this after a rebuild needs JOLT_NO_DEPS_CACHE=1 — the deps-image
cache keys on the version string, not build identity, so it served stale codegen
(filed separately).
Co-authored-by: Yogthos <yogthos@gmail.com>
Fuse an app's native-compiled numeric-leaf fns plus its source into one
static executable: no sidecar .so, no toolchain on the target. The AOT path
(#148) already produced a prebuilt module + manifest; this links them into
the jpm-built exe so the app ships as a single file.
`jolt cgen-build -m NS -o OUT` stages a build dir (src/jolt-core symlinks
into the jolt tree, a generated cg.c of the hot fns, an uberscript bundle of
the app, and an entry that bakes the runtime, installs the native fns as var
roots, and runs -main), then runs `jpm build` there — declare-native builds
cg.a and declare-executable static-links it (jpm's create-executable marshals
the module cfns and calls its static entry at startup).
Build needs cc + jpm; the result needs neither. Mechanics that bit, codified
in cgen_build.janet: stdlib_embed slurps .clj cwd-relative so the build runs
in a repo-mirroring dir; jpm hardcodes ./project.janet and sets syspath=modpath;
the executable's dofile imports cg and static-links cg.a, neither ordered nor
release-built by default, so deps are wired explicitly; cleanup must lstat (the
tree symlinks must not be followed); the inner build runs --workers=1 so it
doesn't starve siblings in the parallel gate.
test/integration/cgen-build-test.janet builds the mandelbrot fixture, runs it
from a clean dir with no src/ and no cg.so, and checks the total at native
speed. Closes jolt-a7ds.
Co-authored-by: Yogthos <yogthos@gmail.com>
Lever 1 (native codegen) is built and merged (PRs #143-148): the floor is
localized, cgen translates numeric-leaf fns to C (JOLT_CGEN, 18x on mandelbrot,
cached), and the build-time AOT path deploys native code with no cc. Replaces the
stale START HERE (which still pointed at the now-done spike) with current status
and the open work: jolt-a7ds binary fusion, jolt-v28u while-lowering, jolt-l1l4
grammar widening, jolt-qx70 hot-fn detection.
Co-authored-by: Yogthos <yogthos@gmail.com>
Splits native codegen into a build phase (needs cc) and a deploy phase (none):
- gen-c-module/compile-module compile MANY numeric-leaf fns into ONE native
module (the AOT shape), generalizing the one-fn-per-.so JIT path.
- Backend :cgen-collect? records each numeric-leaf defn's IR while the app loads
as bytecode; cgen/aot-build compiles them into one module and write-manifest
persists {sopath, [{ns name sym}]}.
- Backend :cgen-prebuilt + cgen/load-aot: the deploy run loads the prebuilt .so
(via the native builtin, no cc) and installs each cfunction as the var root
with the same timing as the JIT path, so callers direct-link to native code.
- toolchain-available? no longer crashes when cc is off PATH (os/execute raises
on a missing exe) — a toolchain-less target now gets false.
Proven end-to-end in two processes (spike/native/aot-demo.janet): build with cc,
then deploy with cc removed from PATH -> count-point still native, mandelbrot
3288753 at 12.4ms (full 18x). Test: test/integration/cgen-aot-test.janet. Default
path unchanged; the modes are opt-in. Gate green (118 files).
Remaining for a literal single binary: fuse the .so + manifest into the jpm exe.
Co-authored-by: Yogthos <yogthos@gmail.com>
compile-fn now keys the .so on a hash of the generated C + the Janet ABI + the
platform, in a persistent cache dir (default jolt-cgen under TMPDIR, override
with JOLT_CGEN_CACHE_DIR; JOLT_CGEN_NO_CACHE=1 forces a rebuild). cc runs only on
the first build of a given fn; later runs with the same source reuse the cached
.so, so the per-startup compile cost is paid once.
mandelbrot 100 whole-process wall: cold ~0.71s -> warm ~0.21s (the ~0.5s cc
cost). These cache knobs don't shape output, so they stay out of
ctx-shaping-env-vars (same as the image-cache knobs). Test asserts the .so is
content-addressed and a second compile hits the cache without the source .c.
Co-authored-by: Yogthos <yogthos@gmail.com>
Wires src/jolt/cgen.janet into the backend's :def emit. With JOLT_CGEN=1 (off by
default, needs direct-linking), a plain defn of a numeric-leaf fn is compiled to
C at def time and the cfunction installed as the var root, so direct-linked
callers embed native code. The fn is not inline-stashed when cgen fires —
callers must call the C fn, not inline the bytecode body. ^:redef/^:dynamic stay
bytecode.
The leaf-first rule falls out: run calls count-point (a user var), so run isn't a
numeric leaf and stays bytecode, calling the native count-point over the cheap
forward crossing. mandelbrot 200: 224ms -> 12.4ms (~18x), result unchanged.
Adds JOLT_CGEN to ctx-shaping-env-vars (rides the disk-cache key) and :cgen? to
resolve-run-mode. Default path (cgen off) is a no-op: cgen-root returns nil and
the normal bytecode emit runs. Gate green (117 files). Test:
test/integration/cgen-pipeline-test.janet.
Co-authored-by: Yogthos <yogthos@gmail.com>
First slice of the native-codegen tier. A new standalone module, src/jolt/
cgen.janet, that translates a numeric-leaf fn (numeric in/out, body uses only
native-op arithmetic + loop/recur/if/let/do) to a Janet native C module: params
unboxed to C doubles at entry, loop/recur lowered to a while loop, reboxed at
return. compile-fn runs cc and loads the .so via the native builtin, returning a
cfunction; it returns nil for non-candidates or when the toolchain is absent.
count-point compiles and matches the bytecode fn across the mandelbrot grid
(test/integration/cgen-test.janet, which skips the behavioral leg where cc/janet.h
are missing). Nothing wires this into the default compile path yet — detecting
hot fns and installing the C version onto the var cell is the next step.
See docs/foundational-runtime-lever1-native-codegen.md for the ceiling
(native-C ~18-22x faster than bytecode, edges out JVM) and the leaf-first rule.
Co-authored-by: Yogthos <yogthos@gmail.com>
Probes the ceiling and incremental strategy for compiling hot fns to native C,
the only lever that moves the ~10.8x Janet-VM floor the localization spike found.
Native-C mandelbrot (Janet native module) runs ~10-12ms — faster than JVM
Clojure (14.2ms) and ~18-22x faster than jolt's 219ms. The boundary cost is
asymmetric: a bytecode loop calling a C hot-fn 40k times is nearly free (~11ms),
but a C fn calling back into bytecode via janet_call costs ~3.5us/call (~152ms,
no win). So the strategy is leaf-first / whole-hot-cluster compilation, crossing
only at cold edges. A plain cc-built .so (no jpm) loads at runtime via require at
full speed, so the native tier fits jolt's dynamic compile model.
Adds the spike artifacts under spike/native/ and the writeup. Next step is
jolt-ihdp (IR->C for the numeric subset). No source changes.
Co-authored-by: Yogthos <yogthos@gmail.com>
The jolt-vs-hand-Janet-vs-JVM mandelbrot comparison splits the 15.4x floor
into two layers: a Janet-VM floor (~10.8x JVM, optimal while-loop Janet over
unboxed doubles — only native codegen moves it) plus a ~1.43x jolt loop-
lowering overhead on top. The overhead is entirely the loop/recur -> recursive-
closure-called-per-iteration lowering; hand-Janet written the same way matches
jolt, while a while+var/set version is 1.43x faster. So a cheap backend win
(jolt-v28u) sits above the structural native-codegen lever.
Adds the spike artifacts under bench/ and the results writeup; marks the spike
done in the handoff. No source changes.
Co-authored-by: Yogthos <yogthos@gmail.com>
The targeted-specialization work (jolt-ffn) concluded that the constant-factor
gap vs JVM is structural, not per-form: three targeted passes (field-read,
inline cache, ctor descriptor-bake) all came back flat. mandelbrot (pure
compute) is ~15x off JVM and that's the floor — Janet bytecode VM + mark-sweep
GC + indirect calls.
This doc hands off the successor epic (jolt-5vsp): the foundational levers
(native codegen, GC-pressure reduction, deeper devirt+inline) and, importantly,
the spike to run first — localize the 15x floor by comparing jolt-compiled vs
hand-written-Janet vs JVM mandelbrot before committing to any big lever. Also
records what not to repeat.
Co-authored-by: Yogthos <yogthos@gmail.com>
A protocol method reads its fields through the generic guarded keyword lookup
because the method's `this` param is untyped. defrecord now hints `this` with
the record type, the per-form inference seeds ^Record-hinted params (the
:fn branch previously typed all params :any — only the whole-program path
seeded phints), and run-passes feeds the inference the record shapes. So a
hinted param's field reads bare-index instead of going through the :jolt/type
tag guard.
This needed a with-meta fix: (with-meta sym ..) returned a proto'd table, so
symbol? was false and the macro-attached hint broke fn destructuring. Symbols
now carry metadata in-place in their struct (matching how the reader attaches
^hint), keeping symbol? true, as in Clojure.
Modest on dispatch (~3-5%): the field read is a small fraction of a dispatch;
the machinery (record-tag + protocol lookup + wrapper) dominates, which is the
inline-cache target (jolt-ez5h). But it's a correctness fix and lets any
^Record-hinted code — not just methods — drop the field-read guard per-form,
not only under whole-program.
Co-authored-by: Yogthos <yogthos@gmail.com>
The ray tracer is compute-bound and the three existing benches only cover
alloc / megamorphic-dispatch / collections. Add three axes the epic needs to
judge itself holistically:
- mono-dispatch: monomorphic protocol dispatch. Its jolt/JVM ratio (~110x) is
*worse* than megamorphic (~76x) — the JVM inline-caches a runtime-monomorphic
call site to near-free while jolt does a full registry dispatch (devirt only
fires on statically-proven receivers). Points at the call-site inline cache.
- mandelbrot: pure float compute, no alloc/dispatch. The floor at ~15x — native
arith already gets close to the JVM.
- fib: recursion, call + integer-arith overhead.
run.sh gains JVM=1, which runs each bench on JVM Clojure too and prints the
jolt/JVM ratio. collections sized up now that the map is a HAMT (jolt-684u).
README documents the axes and the current scorecard.
Co-authored-by: Yogthos <yogthos@gmail.com>
run-tests.janet runs the same file set as `jpm test` across a pool of worker
processes (one `janet FILE` each, ev-based). The full gate goes from ~790s
serial to ~98s here (8x), and more on CI where the heavy files don't thrash on
swap. CI and the docs point at it; `jpm test` still works serially.
Three things dominated the wall:
- Nine integration tests cold-built a compile ctx (~8s each); switch them to
api/init-cached so they share the prebuilt image. The cache key already
fingerprints the ctx-shaping env vars, so the direct-link ones share one DL
image and the rest share the plain one.
- core-bench's main ran on every gate (~35s of benchmark loops that assert
nothing); gate it behind JOLT_BENCH=1.
- cli-test spawned `janet src/jolt/main.janet` ~20 times at ~8s cold each
(340s under parallel load, and it was the whole wall); prefer build/jolt
(~20ms baked ctx) when present, fall back to from-source for an unbuilt tree.
type-check-test stays on cold init: a snapshot-loaded ctx loses the success
checker's op/msg detail (jolt-vley). jolt-pria tracks caching from-source
startup generally, which would let cli-test drop the build/jolt preference.
Co-authored-by: Yogthos <yogthos@gmail.com>
* Bind *command-line-args* after the deps-image cache swap (jolt-4mui)
Under whole-program (deps-image cache active), `jolt -m NS ARG` dropped ARG:
run-main set *command-line-args* on the current ctx, but a cache HIT then
replaced ctx with the saved image (via `set ctx cached`), whose *command-line-
args* was whatever got baked when the image was saved. The stale binding won at
`(apply NS/-main *command-line-args*)`, so -main ran with the wrong (usually
default) args — silently, for any optimized -m program.
Move set-command-line-args to AFTER the cache swap so it binds on the final ctx.
Repro/regression in deps-cache-args-test.janet: first run builds the image
(arg "first"), second run (cache hit) must echo "second", not the baked "first".
* docs: RFC 0003 — phm is a HAMT, sorted colls a red-black tree
The transients RFC described phm as "bucket-based copy-on-write" and mused about
"if it ever becomes a HAMT" — it is one now (jolt-684u), and sorted maps/sets are
a red-black tree (jolt-0hbr). Update the deviation/future-work notes accordingly.
---------
Co-authored-by: Yogthos <yogthos@gmail.com>
Sorted collections were a sorted VECTOR — insert-at = (into (conj (subvec es
0 i) x) (subvec es i)) is O(n) per assoc with a large constant, so building was
O(n^2): 2000 entries took 55.6s.
Replace the rep with a red-black tree (assoc/dissoc/get/contains O(log n)),
ported from the ClojureScript PersistentTreeMap (cljs.core: tree-map-add /
balance-left / balance-right / tree-map-append / balance-*-del). This tier (25)
loads before 30-macros so deftype isn't available; a node is a plain vector
[color k v left right] and cljs's BlackNode/RedNode methods become functions —
the algorithm is unchanged. A sorted-set stores elements as keys with a nil
value; its ops project the key.
The seed read the old :entries vector directly for equality/printing; route
those through a new :entries op that materializes ascending from the tree
(core_types/sorted-entries-arr + main.janet's printer).
2000 sorted-map assocs: 55.6s -> 0.98s (57x); now O(log n) (per-op cost flat
from n=2000 to 10000). Correctness in test/integration/sorted-rbtree-test.janet
(shuffled insert ordering, delete rebalancing, custom comparator, comparator
lookup, subseq, count); sorted specs + full gate green. (key/val on sorted
entries stays a pre-existing gap — entries are pvecs not host tuples; jolt-jk23.)
Co-authored-by: Yogthos <yogthos@gmail.com>
* Add benchmark suite for alloc/dispatch/collection workloads (jolt-1r86)
The ray tracer is float-compute-bound (devirt, alloc removal, type-proving all
measured flat on it), so it can't validate the optimization passes. Add a small
cross-language suite (AWFY + CLBG style, portable Clojure) isolating the axes it
misses:
binary-trees allocation / GC pressure (escaping short-lived records)
dispatch megamorphic protocol dispatch (~1M dispatches/s; WP can't devirt)
collections persistent map/vector churn
bench/run.sh runs them; bench/README.md maps each to the pass it exercises.
collections immediately surfaced jolt-684u: the persistent hash map is O(n) per
assoc (flat copy-on-write bucket array, not a HAMT) — n=4000 assocs take 50s.
Invisible to the ray tracer (no maps).
* Persistent hash map: HAMT instead of O(n) copy-on-write (jolt-684u)
The map was a flat bucket array whose assoc copied the whole array every insert
(O(n)/assoc, O(n^2) to build). Compounding it, small maps are Janet structs that
only promoted to phm for collection keys — never for size — so a scalar-key map
stayed an O(n)-copy struct forever. Building a 4000-entry map took ~50s.
Two fixes, following ClojureScript's design:
- phm.janet is now a HAMT (hash array mapped trie): BitmapIndexedNode /
ArrayNode / HashCollisionNode, 32-way, 5 hash bits per level, structural
sharing — assoc/dissoc/get are O(log32 n). Translated from cljs.core, adapted
to Janet's 32-bit bit-ops (the hash is carried unsigned, the level index is
extracted with arithmetic, and bits are tested with band against 1<<i since
brushift rejects negative bitmaps). The public phm-* API and the value shape
(:jolt/type :jolt/phm, :cnt) are unchanged; transients are a separate rep and
untouched.
- core_coll promotes a struct map to a phm past 8 entries (not only for
collection keys), mirroring cljs PersistentArrayMap -> PersistentHashMap, so
incremental building isn't O(n^2).
20000 raw assocs: 7.1s -> 0.105s. The collections benchmark: 16.7s -> 0.2s.
Correctness covered by test/unit/phm-hamt-test.janet (oracle vs a Janet table,
nil keys, dissoc, a real hash-collision pair, and a sub-linear-assoc guard);
full gate green.
---------
Co-authored-by: Yogthos <yogthos@gmail.com>
The ray tracer is float-compute-bound (devirt, alloc removal, type-proving all
measured flat on it), so it can't validate the optimization passes. Add a small
cross-language suite (AWFY + CLBG style, portable Clojure) isolating the axes it
misses:
binary-trees allocation / GC pressure (escaping short-lived records)
dispatch megamorphic protocol dispatch (~1M dispatches/s; WP can't devirt)
collections persistent map/vector churn
bench/run.sh runs them; bench/README.md maps each to the pass it exercises.
collections immediately surfaced jolt-684u: the persistent hash map is O(n) per
assoc (flat copy-on-write bucket array, not a HAMT) — n=4000 assocs take 50s.
Invisible to the ray tracer (no maps).
Co-authored-by: Yogthos <yogthos@gmail.com>
A ^Record param hint was applied only at the final re-emit (reinfer-def), not
during the inter-procedural fixpoint. So a hinted param with no callers stayed
:any while inference ran, and a field read off it (e.g. (:origin ^Ray r)) never
told a non-inlined callee that its arg is a Vec3 — the callee's params stayed
unproven and its field reads kept the dynamic guard.
Seed declared hints as a param-type floor in the fixpoint: phint-seed (passes/
types) resolves an arity's :phints to positional record types via the
record-shapes registry, and infer-unit! initializes each fn's fresh param slots
from them instead of nil. A fixed declared type can't poison the least-fixpoint
the way an early-iteration :any would, and a hinted param now propagates its
(and its field reads') types to its callees during inference.
Scope: this closes the hinted-propagation gap. It does NOT help the ray tracer,
which uses zero ^-hinted params (only hinted fields) — its remaining type gap is
unhinted record-param inference on recursive/non-inlined hot fns, and per the
jolt-15jq A/B it's allocation-bound regardless (jolt-8flj). Tracked on the bead.
Co-authored-by: Yogthos <yogthos@gmail.com>
Dependency resolution now lives in the `jolt` CLI itself instead of a separate
jolt-deps executable. `jolt` resolves a deps.edn into JOLT_PATH/JOLT_APP_PATHS
in-process and dispatches the deps subcommands:
jolt -M:alias [args] run the alias :main-opts
jolt -A:alias CMD run CMD with the alias paths
jolt run FILE resolve, then run FILE
jolt path | tasks | task NAME
A deps.edn in the working dir is auto-resolved for the runnable commands
(repl/-m/-e/nrepl-server/FILE), so e.g. `jolt -M:nrepl` (or plain
`jolt nrepl-server`) starts an nREPL with the project and its deps loaded.
The runtime core stays deps-agnostic — it only reads JOLT_PATH. The resolver
(deps.janet) is reached only from the CLI entry and loads jpm lazily, so a run
with no deps.edn never touches it and an app baked from its own jolt/api entry
never links it. resolve-deps-argv only resolves on an explicit deps command or
when a deps.edn is present; help/version never do.
jolt-deps stays as a thin deprecation shim that forwards to `jolt`, so existing
scripts keep working. Docs (README, CLAUDE.md, building-and-deps, tools-deps)
and the help text updated.
Co-authored-by: Yogthos <yogthos@gmail.com>
scalar-replace already folds non-escaping const-key map literals
((:k {:k a ..}) -> a, and drops a let-bound map that doesn't escape).
Extend the same fold to record constructors: a (->Rec a b c) is a
positional struct whose declared field order lives in the record-shapes
registry, so a field read on a non-escaping ctor folds to the matching
positional arg and the allocation disappears.
Direct form (:field (->Rec ..)) and the let-bound form both handled,
threaded through run-passes via a per-unit shape registry (new
jolt.host/record-shapes accessor). Soundness: ctor args must be pure
(duplicated/discarded like map vals), arg count must equal the field
count, and only declared-field reads fold — a record answers the virtual
:jolt/deftype key with its type tag and any other key with nil, neither
of which is a positional arg, so those keep the allocation. pure? now
treats a record ctor of pure args as pure, so nested records (a Ray
holding a Vec3) fold bottom-up.
Allocation-bound microbench (non-escaping record built + field-read in a
hot loop): 69.6s -> 2.4s, landing on the no-record arithmetic baseline.
The ray tracer is unchanged — its vec3 results escape (returned/stored
each op), so they genuinely allocate; that's a separate problem.
Co-authored-by: Yogthos <yogthos@gmail.com>
cap truncates a deep type's field VALUES to :any so the inter-procedural
fixpoint stays finite, but it rebuilt the struct via mk-struct and dropped the
record :type tag along the way. The tag is identity — independent of field
depth — so a record stored in a deep container (a Sphere in a world vector, a
material on a hit) degraded to a plain struct, and devirtualization (jolt-41m)
and record? folding silently stopped firing on it.
Preserve :type alongside :shape when capping. Verified: a protocol call on a
record read out of a vector now devirtualizes (the call node gets :devirt-type,
which needs the receiver's record type). Sound — the tag stays accurate; only
field values below the depth cap are truncated.
No measurable wall-clock change on its own (jolt's protocol dispatch is already
cheap), but it restores the record fast path / devirt / record?-folding on
records-in-containers, and unblocks downstream work that keys off record types.
Co-authored-by: Yogthos <yogthos@gmail.com>
A bundle is closed-world — everything it needs is inlined and nothing is
required afterward — so a user defn unreachable from the entry's reference
graph can be dropped. The bundler now computes reachability from main-ns/-main
plus every non-prunable form and drops dead defn/defn- by exact source span
(formatting and reader macros in the surviving code are untouched).
Conservative and sound: only plain defn/defn- are prunable; a defn is kept if
its bare or ns-qualified name appears in any kept form, the closure runs to a
fixpoint, and any use of resolve/ns-resolve/requiring-resolve/find-var/intern/
eval/load-string disables pruning entirely. A parse failure on any file also
falls back to verbatim bundling, so the command stays as robust as a plain
concatenation. defmethod/defrecord/extend bodies are non-prunable and scanned,
so a fn reached only via dynamic dispatch stays live.
New reader/parse-all-spans returns [form start end] byte offsets so the drop
is a verbatim slice, not a re-print.
30-fn library used by a 3-fn entry: bundle 1114 -> 437 bytes (61% smaller, 27
dead fns dropped), output byte-identical.
Co-authored-by: Yogthos <yogthos@gmail.com>
When the collection-type inference proves an argument's type, number?/
string?/keyword?/record?/nil?/some? fold to a compile-time boolean. A
const-fold now runs after inference so a folded predicate propagates and
collapses any if it gates to the taken branch.
Sound by construction: only a provable answer folds, and only when the
argument is side-effect-free (a const or local) so dropping its evaluation
is a no-op. Unknown types (:any/:truthy) and impure args keep the call.
vector?/set?/map? are left out — the :vec tag conflates a real vector with
a range/seq, so vector? could be wrong.
50M-iter loop, same shape isolated with a carry-only control: number? call+
branch 5080ms, predicate folded 1365ms — matching the 1417ms control floor,
so the 3.7x is entirely the eliminated call+branch.
Co-authored-by: Yogthos <yogthos@gmail.com>
* Reader: #() params survive syntax-quote (auto-gensym names)
#(...) named its synthesized params with bare gensyms, so a #() written inside a
syntax-quote had its params qualified to the current ns by sq-symbol — and a
qualified symbol isn't a valid fn param. hiccup's compiler emits
`(let [sb# ..] (iterate! #(.append sb# %) ..)), which broke with "Unable to
resolve symbol: ns/_NNNN".
Name the params with a trailing # (auto-gensym suffix, like Clojure's p1__N#) so
syntax-quote maps them consistently and leaves them unqualified. Harmless outside
a backtick (just a regular symbol name).
* interop: String/valueOf static + String is a CharSequence
Two interop gaps surfaced bringing up hiccup and malli:
- String/valueOf(Object): hiccup's compiler stringifies attribute values with
(String/valueOf (or arg "")). Added the static — "null" for nil, else core-str.
- (instance? CharSequence s) returned false for a string; String implements
CharSequence, and malli's :re validator gates on it before matching, so :re
schemas always failed. instance-check now answers true for strings.
---------
Co-authored-by: Yogthos <yogthos@gmail.com>
Under JOLT_OPTIMIZE a -m program run inferred + specialized EVERY loaded
namespace, including every transitive dependency. On a dep-heavy app that's
prohibitive: malli-app cold-started in ~2m10s (hundreds of dep namespaces, each
run through the per-form inline + inference passes).
The closed world a whole-program pass reasons over is the APP, not its
libraries. jolt-deps now passes the project's own source roots (its deps.edn
:paths) to the runtime as JOLT_APP_PATHS. A namespace loaded from an app root
gets full optimization (and joins the one whole-program fixpoint); a dependency
namespace compiles at default cost — :inline? off for its load, so the per-form
optimize passes don't run over library code — staying direct-linked but
generically typed (the open-world default). With no app roots declared (a bare
program run, or jolt without jolt-deps) everything counts as app, so behavior is
unchanged.
malli-app JOLT_OPTIMIZE cold start: 2m10s -> 4.5s. Compute-heavy programs whose
hot code is their own namespaces (the typed ray tracer) are unaffected — their
code is app code and still fully optimized (9s/frame render). Applied at runtime
in main for the same baked-at-build-time reason as JOLT_PATH; added to the
ctx-image cache key. Help text corrected: optimization is opt-in, not default.
Co-authored-by: Yogthos <yogthos@gmail.com>
A deftype field tagged ^:unsynchronized-mutable / ^:volatile-mutable is set!-able,
but under direct-link immutable records are shape-rec tuples, so set! errored
("Can't set! field on non-deftype: tuple").
A deftype with any mutable field now opts out of the shape-rec layout and uses
the existing :jolt/deftype table form regardless of :shapes? — set! already
mutates that form and field reads route through the tagged-table path. Such a
type is also not registered as a shape, so the inference never emits a bare-index
read against the table. Immutable deftypes/records keep the fast shape-rec.
deftype extracts per-field mutability from the field metadata and passes it to
make-deftype-ctor, which picks the representation at ctor-build time.
Co-authored-by: Yogthos <yogthos@gmail.com>
* Don't direct-link a var redefined earlier in the same unit (jolt-wf4)
defrecord/deftype expands to (do (def R (make-deftype-ctor ...)) (def ->R R) ...),
so the ->R alias references R within one compiled unit. Under direct-link a var
ref embeds the cell's root as a compile-time constant, but on a redefine R's old
root is still in place when that unit compiles — the (def R new) sibling hasn't
run yet — so ->R sealed to the stale pre-redef ctor. (defrecord R [x a])
(defrecord R [a x]) (:a (->R 10 20)) read the old [x a] layout and returned 20.
Track the vars a unit (re)defines and force their later in-unit references to the
live indirect deref. The cell is registered only after its own init is emitted,
so a recursive self-reference inside the init still direct-links (it runs after
the def completes); only sibling references after the def go indirect.
* Emit Janet's `in` as a value so a user local can't shadow it (jolt-fjb1)
The back end emits `in` to deref var cells ((in cell :root)) and index
shape-recs. It emitted the bare symbol, so a user local named `in` shadowed
Janet's builtin in the surrounding scope and the generated cell-deref called the
user's value as a function — "<table> called with 2 arguments, possibly expected
1". malli's explainer binds [value in acc], so m/explain hit this on every
schema (m/validate was unaffected — its path doesn't bind `in`).
Embed `in`'s function value at the emit sites (as jolt-call/core-get already
are); a value in head position can't be shadowed. Fixes m/explain on malli
(loaded with JOLT_FEATURES=clj so its .cljc reader-conditionals resolve).
---------
Co-authored-by: Yogthos <yogthos@gmail.com>
eval-form treated only a reader LIST (a Janet array) as a call; a runtime-built
list — a plist or lazy-seq from cons/concat/list or ~@ (list?/seq? true but
array? false) — fell through to self-eval. So (eval (cons '+ '(1 2))) returned
the list as data instead of 3, and a macro whose output contained such a subform
left it unevaluated. Add a plist?/lazy-seq? branch that coerces to the element
array via d-realize and dispatches through eval-list; an empty list self-evals.
The analyzer already punts these forms to the interpreter (analyze's :else ->
uncompilable -> interpreter fallback), so this one interpreter branch fixes the
correctness bug across the eval and macro-expansion paths; compiling them
directly (vs punting) would be a separate perf change. Verified: conformance
355/355, syntax-quote ~@ splice, list values unchanged.
A deftype with (Object (toString [_] s)) had its toString ignored: the generic
object-methods "toString" fired in dispatch-member before the record's own
method (the record isn't a tagged shim, so that guard passed), and str rendered
the #Type{...} data repr instead of routing through toString.
- dispatch-member: a record's own method (instance/reified/protocol) now wins
over the generic object-methods table — so .toString/.equals/.hashCode on a
record use the record's definitions; plain records still reach object-methods.
- str: add a late-bound record-tostring-cb (wired per-ctx by
install-print-method-cb!, mirroring print-method-cb) that str-render-one
consults for records — a deftype with a custom toString renders via it, plain
records keep the data repr. pr-str is unchanged.
Needed by hiccup's RawString. Adds deftype-tostring-spec (.toString + str +
concatenation + a regression guard that record-less-toString keeps its repr).
walk only handled vector?/map? and fell through :else for everything else, so
postwalk over a quoted list (a plist) never touched its elements —
postwalk-replace with symbol keys silently no-op'd, which broke
clojure.template/apply-template (found during reitit work). Add list? (rebuild as
a list) and seq? (map over it) branches after the vector/map ones so concrete
collections stay authoritative. Adds walk-spec covering list/seq walking plus a
vector/keywordize regression guard and the apply-template trigger.
The 813-line host_interop.janet (java.lang statics, java.time, the 458-line
java.io/util/net/sql/text install-io!, and collection interop all in one file)
becomes a 19-line aggregator over src/jolt/interop/:
- java_base.janet — java.lang statics (Math/Thread/System/Long) + the java.time
shim + the shared chr/pad2/formatter coercion helpers
- host_io.janet — java.io/util/net/sql/text shims; imports java_base for the
shared helpers it reuses (it constructs java.time values and
formats dates)
- collections.janet— the late-bound .iterator/.nth/.count/.seq interop hooks
host_interop.janet just loads the three and runs install!/install-io!/
install-collections! in order. chr/pad2/formatter become public (they cross the
java_base->host_io boundary now). The registry machinery (class-statics/
tagged-methods/register-*!) stays in the evaluator, which loads first — moving it
out touches the hot dot-dispatch and is left as a separate step.
Pure move, no behavior change. Also fixes a cache footgun the subdir exposed:
source-fingerprint walked src/jolt/ with a non-recursive (os/dir), so an edit to
an interop/ file would not have busted the ctx image cache (stale-image bug) —
made it recurse, keyed by repo-relative path.
Full gate green: host-interop-spec (130+ rows across every JDK area), ctx-image
cold/warm, conformance x3, suite >=4695/88, fixpoint.
init-cached (core image) and the deps-image (main.janet) each hand-wrote the same
fork->slurp->validate-ctx?->rewire load and the snapshot->tmp->atomic-rename save.
Extract load-ctx-image / save-ctx-image (in api.janet, beside snapshot/fork): the
two callers now differ only in the validity predicate they pass — none for the
core image (its source fingerprint is already in the path), a deps-manifest mtime
check for the deps image. The per-process print-method-cb rewiring an image
restore must replay is threaded as a callback so the helpers don't depend on core.
Kept in api.janet rather than a new ctx_image.janet module: Janet's `use` doesn't
transitively re-export, and snapshot/fork already live in api and are consumed via
(use ./api) by main and the test harnesses — a separate module would force every
caller to import it directly. (load-image/save-image collide with Janet builtins,
hence the ctx- prefix.)
Full gate green; ctx-image-test cold/warm + deps tests pass.
core-count/core-seq/core-conj each walked a chain of (and (table? x) (= :jolt/X
(get x :jolt/type))) predicates — re-fetching the type tag per predicate and, for
conj, a 14-deep nested if. Replace with a single (case (get coll :jolt/type) ...)
per op: the type is read once and the arm calls the concrete op directly. Host
values (tuple/array/nil) and tuple-based shape-recs carry no :jolt/type and stay
in a per-op fallback cond (shape-rec? kept before tuple? — shape-recs ARE tuples).
Factor the two map-conj paths (phm vs struct) into one conj-into-map parameterized
by the assoc fn.
Behavior-preserving: the :jolt/type tags are disjoint, so the case is an exact
re-expression of the predicate chain; the fallbacks reproduce the originals
(incl. count erroring on a raw host table, which only seq ever handled).
This is the perf-neutral realization of the planned collection vtable. A shared
:jolt/type->{ops} table was tried first and REGRESSED core-bench ~2-4% (table
lookup + indirect call on the hot path); the inline case instead runs ~1.7%
FASTER than main (4864ms vs ~4950ms baseline) since one type-get + a jump beats
the sequential predicate chain. Full gate green (suite >=4695/88, fixpoint).
analyze-try assoc'd :catch-sym/:catch-body/:finally nil-when-absent, so a try
with no catch (or no finally) carried a nil-valued key — which makes the node a
phm in jolt's map representation and forces the back end to densify it
(norm-node) before reading :op. That's the map-nil-representation trap Phase 2
already cleaned up for def/fn/arity nodes. Add those keys only when the clause is
present, matching the arity :rest discipline; a try node stays a fast struct.
Behavior-invisible: emit-try reads each key with a nil-safe (node :k) and gates
on it, so an absent key and a present-nil key are indistinguishable to every
consumer. Adds ir-try-shape-test asserting the node shape across all four
try/catch/finally combinations plus end-to-end eval.
Note on scope: the plan's "delete the defensive norm-node calls" is NOT done — it
can't be. {:op :const :val nil} (e.g. (def x nil)) and nil map keys are
inherently phm, so the emit-dispatch norm-node guards a real case, not a
present-or-absent artifact. This PR removes a source of gratuitous phm nodes
rather than the densification itself. Full gate green.
Six bottom-up IR rewrites (const-fold, inline-node, subst, flatten-lets,
subst-lookup, scalar-replace) each hand-listed every op's child positions —
~250 lines of identical "recurse children, rebuild" arms that had to be kept in
sync whenever an op was added. Extract one map-ir-children into ir.clj that
knows each op's child layout; each walk keeps only its genuine specials
(const-fold's invoke/if, inline-node's invoke, subst's local/let alpha-rename,
scalar-replace's invoke/let folds) and delegates the rest.
The combinator is total over the op set, so the walks are now total too: a
couple soundly gain coverage they previously skipped (const-fold now folds
inside :try; subst-lookup now recurses :def inits, which fixes a latent dangling
ref where a dropped const-key-map binding was referenced inside a def). These
are sound — all six are result-preserving optimizations — and 3-mode conformance
+ fixpoint confirm identical program behavior.
map-ir-children is shape-preserving for :try (recurses :catch-body/:finally only
when present, never assoc's nil) so it can't turn a struct node into a phm.
Written with cond/get only, matching the passes' tier, so no new load-order dep.
Predicates (body-closed?/pure?/local-escapes?), the type-threading infer, and the
Janet backend emit stay as-is: their conservative :else defaults / [type node]
threading / host language don't fit a node-rebuilding combinator.
Adds ir-passes-test coverage for folding reaching fn/loop/try bodies. Full gate
green (conformance x3, suite >=4695/88, fixpoint stage1==2==3, inline-sra + devirt).
