The closed-world fixpoint (#226) flowed record types across fn boundaries; this
adds a numeric refinement so a hintless fn whose every call site passes a flonum
has its param unboxed to fl-ops, no ^double hint needed.
Lattice gains :double, a flonum refinement of :num: two doubles join to :double,
a double joined with anything else widens to :num — so a param is :double only
when every contributing value is a flonum, which is what makes the fl-op sound.
infer types a flonum literal and flonum arithmetic (+ - * / min max inc dec over
double/int-literal operands) as :double, and the fixpoint joins those across call
sites and return types like any other lattice value.
The bridge to the existing hint-directed pass is a synthetic [param :double]
nhint: wp-infer! stashes the :double params separately from the structural seeds,
and run-passes injects them as nhints before numeric/annotate, so the fl-op
emission and the exact->inexact entry coercion (a no-op on a proven flonum) apply
unchanged.
Sound subset only: :double, never :long — an untyped integer can be a bignum and
fx-ops would overflow/diverge from jolt's arbitrary precision. So an integer
caller leaves a param generic; an escaped fn (unknown callers) keeps :any.
run-numwp.ss gate: cross-fn :double propagation incl. through a flonum-returning
helper, the integer-caller and escape negatives, and the full run-passes path
emitting fl* + entry coercion. make test / shakesmoke green, selfhost holds, 0
new divergences.
Co-authored-by: Yogthos <yogthos@gmail.com>
Re-derive each app fn's param types from its call sites under --opt, so a
record type flows across fn boundaries: a ctor's return reaches a callee
param, and a typed vector's element reaches a HOF closure's param. The back
end can then bare-index field reads and devirtualize protocol calls at those
sites (it reads the resulting :hint/:devirt annotations; consuming them is
separate work).
This rebuilds the inter-procedural driver the Janet host had — the API
(infer-body/reinfer-def) survived the rehost but nothing drove it, and the
record-shapes/protocol-methods registries were empty stubs.
- records.ss: populate record-shapes (ctor key -> fields/tags/type, resolving
nested record field tags) and protocol-methods (method var -> [proto method])
registries at deftype/defprotocol load time; jolt.host accessors materialize
them.
- passes/types.clj: wp-infer! runs a closed-world fixpoint joining call-site
arg types into callee params; reinfer-def re-seeds each def at emit. Self-
recursive calls and fn-level recur are collected so a recursive fn's params
are constrained by its recursion, not just external callers — else a param
the recursion widens (e.g. binary-trees check-tree, whose untagged child can
be nil) would be unsoundly typed non-nil. A fn used in value position keeps
:any params (callers unknown). Megamorphic sites join to :any.
- build.ss: analyze all app forms and run the fixpoint before per-form emit.
- run-wp.ss: gate (cross-fn propagation, escape soundness, self-recursion).
make test / shakesmoke green, 0 new divergences, selfhost holds.
Co-authored-by: Yogthos <yogthos@gmail.com>
jolt.passes.inline was fully written but dormant — it fetched bodies via the
inline-ir host hook, which was a stub returning nil. Wire it up: run-passes stashes
each inline-eligible defn (single fixed arity) as its form is optimized, and
inline-ir hands the body back at call sites under --opt.
The catch was the ^double/^long coercion: an inlined fn drops its param-entry and
return coercion, so (work 3 4) on a ^double fn would return 25 instead of 25.0. New
:coerce IR node carries the coercion inside the spliced body — the inline pass wraps
a hinted param's arg and the return in :coerce, the back end lowers it
(exact->inexact / jolt->fx), and jolt.passes.numeric reads its :kind. So an inlined
call matches the called one and the body's fl*/fx* fast path still fires.
Only under --opt (closed world); the seed mint and -e don't inline, so selfhost and
the corpus are unaffected. test/chez/inline-test.ss 12/12 (make inline); full make
test green, 0 new corpus divergences.
Bench (hot loop, body is a ^double helper call): direct-link 500ms -> --opt
(inlined) 184ms = 2.7x, by eliminating the call + coercion wrappers and letting Chez
fuse the fl-ops unboxed. ~26x over the default dispatched build.
A ^double/^long param hint (or a float literal) now drives Chez flonum/fixnum
ops instead of generic arithmetic — JVM-style primitive hints, available in every
build and at -e (not gated on direct-linking or whole-program inference).
New pass jolt.passes.numeric: a local forward type-flow seeded from ^double/^long
fn-param hints (analyzer attaches :nhints per arity) and float literals,
propagated through let inits / arithmetic / if / do. It tags an arithmetic invoke
:num-kind :double|:long when every operand is that kind (an integer literal is a
wildcard, coerced to a flonum in a double op). The back end lowers a tagged node
to fl+/fl-/fl*/fl//fl<?/... or fx+/fx*/fx1+/fxquotient/... (unchecked-add etc.
join the fixnum path; == too). Runs last in run-passes, both branches.
Soundness: :long is seeded ONLY from an explicit ^long hint, never a bare integer
literal, so un-hinted integer code keeps jolt's arbitrary-precision numbers — no
fixnum-overflow surprise, no corpus divergence. :double comes from ^double hints
and float literals (flonum arithmetic is always flonum, matching the generic
result). A ^long hint is a promise the value is a fixnum: fx+ raises on overflow,
like a JVM fixed-width long.
Numeric-hinted params coerce at fn entry (exact->inexact / jolt->fx), the way the
JVM coerces a primitive parameter — so the body's fl*/fx* ops can rely on the
type even when a caller passes an exact int (e.g. Chez's (* 0 1.0) => exact 0).
Round 1 specializes hinted straight-line / fn-body arithmetic. fl-ops are ~4x
generic in a tight Chez loop, but realizing that on loop-carried accumulators
needs loop-var typing — round 2. Sound foundation, gated by test/chez/numeric-test.ss.
- take-last / drop-last return seqs, not vectors: take-last wraps in seq; drop-last
is the JVM (map (fn [x _] x) coll (drop n coll)) form (lazy, () when empty).
- cycle is lazy ((lazy-seq (concat coll (cycle coll)))) so it no longer counts its
argument and terminates on a lazy/infinite input.
- fold's foldable-call catch uses :default, matching the rest of jolt-core and
also catching a raw host condition from a folding primitive.
- alts! rejects non-channel ports with a clear error (put specs / :default are
unsupported) instead of crashing inside ac-poll!.
- Misc: drop the unreachable second getCause clause; jolt-nth on a string raises
'nth "index out of bounds" like the vector branch; name the inline fixpoint cap;
bld-sh-capture rejoins lines with newlines; clarify a couple of comments.
Rename src/jolt -> stdlib (the runtime-loaded layer; jolt-core stays the
seed-baked layer) and update the loader / emit-image / doc paths. Drop dead
code: the spike/ experiments, the duplicate clojuredocs-export.edn (json moves
to tools/), the Janet-era jolt.http binding, and the orphaned
persistent_vector.clj whose ns/path didn't even match.
Strip porting residue from comments and docstrings across host/chez, jolt-core,
stdlib, tests, and docs: internal issue ids, "Phase N" markers, and the "vs
Janet" historical exposition, leaving present-tense descriptions and the real
JVM-Clojure semantic contrasts. Same pass over the corpus suite labels. The seed
is unchanged (docstrings/comments aren't emitted), so the self-host fixpoint and
corpus are untouched.
Port tools/spec_coverage.py off the dead janet probe to bin/joltc and regenerate
coverage.md; drop the dead :host/janet rule from certify.clj and regenerate the
conformance profile. Add docs/host-interop.md (the JVM shims and how to register
your own host class from a library) and a writing-style note in CLAUDE.md.
Stabilize the four racy concurrency corpus cases (future-cancel and agent
send/send-off): give the future a sleeping body and the agent a slow action, so
cancel reliably catches an in-flight future and deref reliably reads the
pre-update snapshot. They certify deterministically now, so drop their :flaky
allowlist entries and the orphaned legend.
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>
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>
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>
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>
passes.clj was a 1486-line grab-bag mixing three weakly-coupled concerns. Split
along the clusters the review mapped (only run-passes + the dirty flag were
shared):
jolt.passes.fold const-fold + the shared scalar-const? predicate (base)
jolt.passes.inline inline + flatten-lets + scalar-replace
jolt.passes.types collection-type inference + success checker + driver API
jolt.passes façade: run-passes + :refer re-exports of the driver fns
the back end looks up by name
scalar-const? was used by both the inline pass and the inference walk, so it
moves to fold (the base layer) and both refer it. The check-mode state stays
private to jolt.passes.types behind a new run-inference fn; run-passes calls it.
build-compiler! loads the three in dependency order before the façade, mirroring
the existing jolt.ir -> jolt.analyzer bootstrap. No behavior change. Also fixed
the stale ns docstring that listed four passes and omitted the type system.
Gate green: conformance 355x3, clojure-test-suite 4718 pass (>= 4695 baseline),
full jpm test exit 0.
* Add architecture refactor plan
Synthesizes a six-part architectural review into phased, gate-validated cleanup
work. Targets LLM-maintainability: one home per feature, no god-files, explicit
checked contracts, no copy-paste dispatch. No code changes yet — the plan only.
* Refactor phase 0: dead code + isolated bugs
Pure cleanup ahead of the structural phases (docs/architecture-refactor-plan.md).
No behavior change except the two bug fixes, which are covered by a regression row.
Dead code (all verified zero-reference or overridden):
- core-resolve / core-satisfies? / core-type->str seed stubs + bindings —
resolve and satisfies? are interned by install-stateful-fns! (the seed copies
were shadowed); type->str was an inert SCI stub with no callers.
- find defined twice in 20-coll.clj; the dead copy returned a plain vector
(wrong — the live def at :787 returns a real map-entry) with a comment that
contradicted it.
- mark-hint (passes.clj), phs-to-struct (phm), shape-vals / ns-imports-fn
(types) — unreferenced.
- redundant local pad2 in javatime (module-level one already in scope).
Bugs:
- File.toURL stored :url but every :jolt/url method reads :spec, so a URL from
(.toURL file) returned nil from all its methods. Now stores :spec (+ spec row).
- pl-rest had a no-op (if (plist? r) r r); collapsed to r.
- :map-shapes? was missing from the deps-image cache key — two runs differing
only in map-shapes could reuse each other's image.
Also dropped read-quote's unused pos param. Full gate green.
---------
Co-authored-by: Yogthos <yogthos@gmail.com>
A ^RecordType hint only resolved against the current namespace's ctor key, so a
hint naming a record defined in another namespace degraded to :any. That made a
decomposed multi-namespace program much slower than the monolith: per-namespace
inference can't see a record param's callers in other namespaces, and the
declared hint that could have typed it was dropped.
Resolution now works cross-namespace, for both record FIELD hints (defrecord)
and fn PARAM hints, in both spellings — ^Vec3 where the type is referred and
^v/Vec3 where the namespace is aliased:
- reader keeps a tag's namespace qualifier (^t/Ray -> "t/Ray", was "Ray").
- make-deftype-ctor-impl indexes each ctor closure by value; record-hint-ctor-key
resolves a hint name against the COMPILE ns (referred names live there; aliases
resolve through it) and maps the type var's root back to its home ctor key.
Using the ctor value, not the var's :ns, is what makes :refer work — :refer
re-interns a fresh var whose :ns is the referring ns.
- the analyzer captures record param hints as arity :phints [name ctor-key];
reinfer-def seeds those param types, so a record param is typed even with no
inferred caller — the open-world / cross-ns case.
Effect on the multi-namespace ray tracer: per-ns compile 30.4s -> 7.9s with
param hints, matching whole-program (8.1s) and the single-ns monolith (8.3s).
cross-ns-hints-test covers field + param hints, refer + as, and the reader tag.
A record field can carry a type hint — ^Vec3 (a defined record type) or ^:num —
and the inference now resolves it so reading the field back yields that exact type
instead of :any. A Vec3 stored in a Ray field reads out as Vec3, so the vec ops on
field-read values prove their reads (bare-index). This is Stalin's per-slot type
sets, but DECLARED rather than inferred: the exact shape is known up front.
- deftype captures each field's :tag / :num metadata (was stripped) and passes it
to make-deftype-ctor; the ctor registers per-field tags, resolving a record-type
hint to its ctor-key (same-ns) so the inference can look it up directly.
- call-ret-type builds a record's struct type with field types resolved from the
hints, recursing into nested record types (depth-bounded for self/cyclic types).
Measured: a nested-record read loop (:r (:origin ray)) runs 1.3s with ^Vec3 hints
vs 7.1s without — 5.5x. This is the lever the ray tracer needed (vecs flow through
container fields); records without it read back as :any and stay unproven.
A protocol method call compiles to (protocol-dispatch proto method this rest) — a
runtime registry walk (type-tag -> proto -> method) on every call, ~19x a direct
call. When the inference proves the receiver (arg 0) is a known record type, the
call now resolves to a DIRECT method call at compile time, skipping the registry.
- defprotocol registers each method's var-key 'ns/method' -> [proto method] (a
ctx-capturing register-protocol-methods! emitted into the do-block); infer-unit!
feeds it to the inference via a box (like record-shapes).
- the record-ctor return type carries :type (the record tag) so the inference
knows the receiver type; the :else invoke case annotates a protocol call whose
arg0 has a known :type with :devirt-{type,proto,method}.
- emit-invoke resolves the impl via find-protocol-method at emit time and emits a
direct call to the embedded impl fn value. Unknown/polymorphic receivers (no
proven :type) fall back to the dispatch path unchanged.
Measured: removes the dispatch overhead (14.7s -> 9.3s on a 10M-call loop); the
remaining cost is the method body itself (non-inlined, unproven reads) — inlining
the resolved method is the follow-up (jolt-t6r) toward direct-call speed.
Sound under the closed-world assumption direct-linking already makes (the impl is
resolved + embedded at compile time). Adds devirt-test (subprocess: dispatched ==
devirtualized across polymorphic dispatch, unknown-receiver fallback, and
heterogeneous collections). Stalin's compile-call/callee-environment is the model.
Records (defrecord/deftype) are now shape-recs in a direct-linking unit by
default — no JOLT_SHAPE flag. A record's shape is DECLARED, so the inference
proves field reads by a lookup, not fragile shape inference, and they bare-index.
Result: ~1.4x faster than the :jolt/deftype table form on a record-heavy loop
(3.9s vs 5.5s), driven by cheaper construction + proven bare-index reads.
Two gates now:
- :shapes? — shape-recs active; records use declared-shape layout + bare
index reads. On with direct-linking (where the inference runs).
- :map-shapes? — also shape generic const-key maps. Opt-in (JOLT_SHAPE), because
shaping maps net-loses on unproven reads (measured). Records win.
- call-ret-type types a record ctor (->Name) as a struct of its declared shape,
fed from a ctx-env registry populated at deftype; field reads on the result
bare-index. (set-record-shapes!/set-map-shapes! wired through infer-unit!.)
- sidx reads the field's position from the :shape vector AS-IS (declared order
for records, str-sorted for map literals) — no re-sort — so any field order
bare-indexes correctly. The map :map case only sets :shape under :map-shapes?.
- record-shape-for interns the descriptor per (type, fields), not per type: a
record redefined with different fields now gets a fresh descriptor instead of a
stale one (fixes redef descriptor staleness; old instances stay valid).
Adds record-declared-shape-test (declared-order reads, incl. non-alphabetical
fields, through fn boundaries + protocol method bodies). Known pre-existing edge
case filed as jolt-wf4 (direct (:f (->R …)) read returns nil after a record is
redefined with different fields; let-bound read works; repros without shapes).
The inference dropped the complete :shape whenever it rebuilt a struct type
(cap) or joined two (join-t/merge-fields), so a vec3 retrieved from a container
or a fn param typed across call sites lost its layout and every field read fell
to the slow descriptor path. Two fixes:
- cap preserves :shape: capping truncates field VALUES below the depth limit but
never the key SET, so the layout is still complete. It also recurses into
fields, so a shaped value nested in a container (a vec3 inside a hit-info)
keeps its own :shape — which is what lets (:r (:normal hit-info)) bare-index.
- join-t preserves :shape when both sides are the SAME complete shape (the
merged struct has the same keys); different shapes still drop it. This carries
the shape through if-joins and the inter-procedural fixpoint's call-site joins.
Result: the ray tracer goes from 22s (R1, correct-but-descriptor-path) to 4.36s
— 2.7x FASTER than the 11.7s no-shape baseline, and ~3x the JVM (was 8.5x), with
byte-identical output. The compounding of cheaper tuple construction plus
bare-index reads across the whole render far exceeds the per-op estimate.
Gate green flag-off, suite 4718, default-path bench even, transparency intact.
Removes the {:r :g :b} hardcoding. ANY constant key set is now a shape:
- inference: a struct type from a map LITERAL carries :shape (its canonical
str-sorted key vector — completeness); joins/access-inferred structs lack
it, so they never get a bare index. The literal node and lookup subjects
carry the shape; the back end derives the index from it.
- backend: emit-map turns any shape-tagged const-key map into a shape tuple;
emit-kw-lookup reads the field by bare index when the complete shape is
proven, else by the value's own descriptor (so a shape-rec whose :shape was
dropped by a join still reads correctly).
- runtime: core-get and core-assoc handle shape-recs.
Status: CORRECT for direct field access, container round-trips, and assoc
(minimal repros pass). NOT yet complete — the full ray tracer still hits an
uncovered path (a shape-rec reaching a map op without coverage: keys/vals/
count/seq/equality/print/jolt-call/dissoc/contains?/the interpreter's
coll-lookup all still need shape-rec branches). And the perf win needs
COMPLETENESS PRESERVATION through joins/containers (merge-fields/cap drop
:shape today, so nested vec3 access falls to the descriptor path, slower than
a struct get) — without it the general version is slower than the vec3
prototype.
All behind JOLT_SHAPE (off by default). Gate green with the flag off, suite
4718. This preserves the general design; the transparency layer + completeness
preservation are the remaining multi-session work.
Validated prototype of the hidden-class object-model change. A vec3-shaped
{:r :g :b} map literal is represented as a cheap Janet tuple [shape vb vg vr]
instead of a struct (~2x cheaper to construct); a lookup on a value the
inference PROVES is the shape reads by bare index with no runtime check.
Result on the ray tracer (direct-link): 12.3s -> 10.7s (~13% faster), with
byte-identical pixel output. The shape value flows transparently through
hit-info/ray/material containers and the colors vector; core-get handles it
(inline check, no fn call) so an unspecialized access is still correct.
Key lessons baked in: the lookup MUST compile to a bare index (a runtime
shape check, even inlined, taxed every field read and made it 2.5-3.4x
SLOWER) — so the inference gained a :shape hint (struct type with keys
exactly {:r :g :b}) that the back end turns into (in m idx). The descriptor
is quoted when embedded (its keys are a parens tuple Janet would otherwise
try to CALL).
All behind JOLT_SHAPE (off by default). Gate green, suite 4718, default-path
bench even. Scoped to the one shape; NOT yet sound in general (assumes every
vec3-shaped value is a shape-rec, true under the flag for the ray tracer) nor
fully transparent (only core-get + the inlined lookup; jolt-call/equality/
print/keys not yet covered). Those are the next steps toward a real feature.
Two provably-wrong cases the inference already has the facts for, closing the
last RFC 0006 open question:
- Calling a non-function. At an :invoke whose callee is provably :num or :str
(the only non-callable types — keywords/maps/vectors/sets are IFn), report
"cannot call a number as a function". Default level (no closed-world: the
callee type is inferred at the call site). Covers (5 1), ("hi" 0),
((+ 1 2) :k), a let-bound number, and a var holding a number (via vtype-box
in direct-link). A union is non-callable only when every member is, so
((if c 1 :k) x) is accepted (:kw is callable). Verified zero false positives
on the ray tracer, which calls maps/keywords/vectors as fns throughout.
- Wrong arity to a user fn. The registered single-fixed-arity sig (jolt-zo1)
makes a mismatched arg count provably throw; reported under the
JOLT_TYPE_CHECK_USER opt-in (same closed-world boundary; ^:redef/variadic
skipped). Caught at compile time before the runtime arity error.
Both fold into the existing infer walk, carry :pos for file:line:col, and keep
no-false-positives. Gate green, suite 4718, conformance 335/335, runtime bench
even (compile-time only).
Checking inherently needs an inference pass (~2.6x compile as a standalone
pass). But direct-link builds ALREADY run one inference pass for
specialization (run-passes' infer-top), so checking can ride along: set a
check-mode flag, turn checking? on during that existing pass, and collect
the diagnostics after — ~2% overhead measured on the ray tracer, vs 2.6x
for the separate pass.
So the checker now defaults to `warn` in direct-link builds (where it's
nearly free) and stays OFF in plain REPL/dev builds (no inference to ride,
no forced cost — opt in with JOLT_TYPE_CHECK there). JOLT_TYPE_CHECK still
overrides in both directions (off to disable, error to escalate).
It checks the POST-optimization IR, which matches what the optimized
program actually evaluates — scalar-replace only drops provably-pure code,
an accepted opt-mode divergence, so no real error is hidden. The loaders
enable position tracking whenever checking will run (env-selected or
direct-link). type-check! (the standalone pass) stays for plain builds;
both paths share report-diags!.
cli-test pins: plain build silent, direct-link warns by default,
JOLT_TYPE_CHECK=off disables. Gate green, suite 4718, runtime bench even.
The checker ran a separate check-walk that re-inferred each argument's
subtree AND recursed into it — quadratic in expression nesting. Fold the
diagnostic emission into `infer` itself (gated by a checking? flag, off
during the optimization fixpoint): one O(n) walk that both types and
checks. Removes check-walk entirely; check-form now drives infer.
This is a cleanup and removes the deep-nesting blowup, but it does NOT make
warn-by-default cheap: on a real 360-line file the checker still adds ~2.6x
compile time (277ms -> 720ms). That cost is the structural inference pass
itself, which checking inherently requires — not redundancy. A cheap
default-on path would need either piggybacking on the inference direct-link
already runs, or a lighter scalar-only checker inference. Gate green,
type-check tests pass.
RFC 0006 error reporting wanted file:line:col but IR nodes carried no
position, so diagnostics read only "type error in <ns>: <msg>". Now:
type error /tmp/scene.clj:5:5: `inc` requires a number, but argument 1 is a string
The reader records each LIST form's absolute start offset in a table keyed
by form identity (lists are fresh arrays, never interned), gated behind a
flag the loaders enable only when JOLT_TYPE_CHECK is on — zero cost off.
Keying by identity makes positions survive macroexpansion exactly when the
user's own sub-form is spliced through, and absent for macro-synthesized
structure: a `(inc :k)` written inside `(when c ...)` reports at its own
line, never at the expansion's generated if/do.
The analyzer stamps the offset onto :invoke nodes (form-position host
contract fn); the checker carries it into each diagnostic as :pos; the
loaders stash the file's source + path on the env (save/restored across
nested requires); backend/type-check! converts offset -> line:col via the
reader's line-col and renders the RFC format. Falls back to the ns when no
position is available (synthetic forms), so it is never worse than before.
Gate green, conformance 335/335, suite 4718, runtime bench even (positions
are compile-time only; off by default).
The success checker fired only against core-fn error domains (stable, not
redefinable). This adds reporting of a call that passes a provably-wrong
type to a USER fn whose body requires otherwise — e.g. a fn that only does
arithmetic on a param, called with a string.
As check-walk sees defs it registers each non-redefinable single-fixed-arity
user fn's {:params :body} in module state (user-sig-box, accumulating across
forms like rtenv-box — a def must precede its call). At a call site (strict
mode only) the body is re-checked with ONE parameter bound to its concrete
argument type, others :any; if that produces a diagnostic the all-:any body
did not, the argument alone is provably wrong and the call is reported.
Monotonic — binding a concrete type can only add error-domain hits — so still
no false positives. A cycle guard (checking-box) terminates mutual recursion.
Gated behind JOLT_TYPE_CHECK_USER (orthogonal to the warn/error level)
because it rests on the closed-world assumption, weaker than the core-fn
case. check-form gains a strict? arity; the default path is unchanged and
user-fn code runs only when the checker is enabled. ^:redef/^:dynamic and
multi/variadic fns are not registered (their body is no stable requirement).
Gate green, suite 4718, conformance 335/335.
The success checker (RFC 0006) used to lose differing if-branches to :any
and accept the use. (inc (if c "a" :k)) typed the if as :any — sound but
imprecise, since the value is provably {:str | :kw}, every member of which
is in inc's error domain.
Adds {:union #{T...}} to the lattice: join-t forms a scalar union of
differing branches instead of collapsing to :any, capped at 4 distinct
scalars (the member space is the five scalar tags, so the lattice stays
finite and the inter-procedural fixpoint still terminates). The checker's
not-number?/not-seqable? report a union only when EVERY member is in the
error domain — any valid member accepts the call, so still no false
positives. type-name renders "a string or a keyword".
Unions are scalar-only and carry no :struct/:vec/:set key, so every
structural predicate already treats them as opaque — specialization sees
them exactly as :any and codegen is unchanged. Gate green, suite 4718,
conformance 335/335, bench even.
Reuse the structural inference from RFC 0005 as a loose type checker. It reports
a core-fn call only when an argument's inferred type is concrete and lies in
that op's throwing error domain, and accepts everything ambiguous (:any, a
union that joined to :any, :truthy). By construction it never produces a false
positive: a correct program has nothing to report even in error mode.
The curated error-domain table starts with the clearest throwing cases:
arithmetic on a provable non-number, and count/first/rest/next/seq/nth on a
provable non-seqable scalar. Lenient operations like (get 5 :k) and (:k 5),
which return nil rather than throw, are deliberately not listed.
Checking is decoupled from specialization: it runs whenever JOLT_TYPE_CHECK is
warn or error, regardless of :inline?, reading the knob at compile time so no
rebuild is needed. warn prints to stderr, error fails the form's compilation,
off (the default) skips it entirely. Core init stays clean under the flag.
jolt-y3b
Replace the ad-hoc inference lattice (a flat :struct-map tag plus {:vec ELEM})
with one recursive structural type: {:struct {field -> T}}, {:vec T}, {:set T},
scalar tags, and :any. A keyword lookup now returns its field's type, so nested
access like (:r (:direction ray)) is typed end to end and drops its guard. join
is field-wise and element-wise with a depth cap of 4 so the inter-procedural
fixpoint still terminates.
The back end honors a struct hint on any subject node, not just locals, so an
inferred field type on a nested lookup specializes. The orchestrator's fixpoint
joins through the portable join-types so compound types no longer collapse to
:any.
Ray tracer goes 12.8s to 11.0s with no hints, matching the explicit ^:struct
version (10.9s). Render checksum unchanged (1915337), full gate green,
conformance x3 modes pass.
jolt-5uj
The inference now tags a :local it proved to be a vector with :hint :vector, and
the back end specializes (count v) -> pv-count (skipping core-count's dispatch
chain) and the 3-arg (nth v i default) -> pv-nth. The 2-arg nth is deliberately
NOT specialized: pv-nth returns nil out-of-bounds where Clojure nth throws.
Sound, conformance 335/335 x3 and full jpm test pass; type-infer-phase2-test
pins the specialization and the 2-arg exclusion.
Extends the inference lattice with a parametric vector type {:vec ELEM} and
threads element types through the program:
- vector literals, conj/into, and range produce element-typed vectors;
- reduce/map/mapv/filter/filterv seed their closure's element (and reduce's
accumulator) param, so a lookup inside the closure over a vector-of-structs
specializes (the HOF-element-awareness piece);
- a var reference carries a VALUE type — a fn var is :truthy (non-nil, sealed
root), a def var carries its inferred init type (e.g. a color table is
{:vec :struct-map}); element-returning fns (rand-nth/first/nth/...) yield the
collection's element type. These let the dynamically-built scene's sphere
maps type as structs.
The inter-procedural fixpoint now also infers non-fn def value types, and the
recompile re-emits the WHOLE unit callee-first (reverse-topological) so a
caller re-embeds its recompiled, now-specialized callees and a call site
compiled after the pass links the whole chain.
Result on the ray tracer (no hints): the chain closes — hittables infers to
{:vec :struct-map}, hit-sphere's hittable param to :struct-map — and the render
goes 13.1s -> 12.8s. That is only ~3%, far short of the explicit hint's 1.22x.
The remaining gap is nested field access: a lookup RESULT like (:direction ray)
is :any, so (:r (:direction ray)) stays guarded, and the vec3 fns (called with
such values) can't be typed struct. The hint asserts the vec3 params directly
and propagates through inlining; matching it needs field-shape types
(ray.direction : vec3, vec3.r : number) — a structural extension (Phase 4).
Sound: a seeded full render produces an identical checksum (1915337);
conformance 335/335 x3 and the full jpm test pass; type-infer-phase3-test pins
the element-typing + HOF mechanism. Phase 2 (vector nth/count specialization)
was deprioritized — it is orthogonal to this benchmark.
Closed-world (optimization mode): after a unit loads, infer-unit! runs a
whole-unit fixpoint over the call graph and recompiles. A fn's param types are
the lub of its in-unit call-site arg types; its return type is the lub of its
tail positions; iterated to a least fixpoint. Param types are RECOMPUTED FRESH
each iteration (not accumulated) because :any is the lattice top — joining an
early-iteration :any would poison the result permanently. Closures inherit the
enclosing tenv so captured locals keep their types (their own params shadow to
:any). A fn whose var escapes as a VALUE keeps :any params (its callers aren't
all visible). Each fn is then re-inferred with its param types seeded and
re-emitted; recompiled bodies are semantically identical, so correctness holds
regardless of order. Sound under source distribution + whole-program compile
(the consumer compiles all call sites together).
Plumbing: the portable pass (jolt.passes) gained inter-procedural primitives —
set-rtenv!, infer-body (types a body, collects its call sites), reinfer-def
(seeds param types), and escape tracking. The back end stashes each
single-fixed-arity defn's :def IR (:infer-ir); the evaluator triggers
infer-unit! after a unit loads (via an env hook, opt mode only).
Result and honest finding: the fixpoint correctly types scalar-flowing params
(ray-cast/hit-all/hit-sphere all get the ray param as :struct-map, no hint),
but the ray tracer does NOT speed up — its dominant lookups are on `hittable`,
the element of the `hittables` vector threaded through `reduce`, which stays
:any. Typing it needs collection-element types (vector<struct>) plus HOF-element
awareness (knowing reduce applies the closure to elements), which is beyond
inter-procedural param inference. The explicit ^:struct hint reaches it (it
types the reduce closure param directly), which is why the hinted run is 1.22x.
Verified: conformance 335/335 x3, full jpm test; new type-infer-phase1-test
pins the fixpoint, the escape gate, the seeded re-inference, and correctness.
A forward, soft-typing-style pass (simplified HM: monovariant, never-fails,
lattice top = :any) in jolt.passes, run after the inline/scalar-replace
fixpoint when the optimization mode is on. It types expressions from literals
and arithmetic, flows the type through let bindings, and joins at if-branches.
Where a keyword-lookup subject is PROVEN to be a plain struct map it sets
:hint :struct (the same channel a manual hint uses, so the back end drops the
:jolt/type guard); where the type is :any it leaves the dynamic guard in place.
Sound by construction: a concrete type is assigned only when proven (scalar
keys with non-nil/non-false values for a struct-map), so a wrong bare get can't
happen. This is the foundation; on its own it mostly overlaps Route 1
scalar-replacement (which already eliminates non-escaping let-bound maps), so
its standalone win is small. Phase 1 (inter-procedural) is where escaping
params get typed.
Verified: conformance 335/335 x3, full jpm test; new type-infer-test pins the
flow rules and the sound :any fallback (cases force the map to escape so the
test isolates inference from scalar-replacement).
A constant-keyword lookup (:k m) currently emits a guarded form,
(if (get m :jolt/type) (core-get m k) (get m k)), to tell a plain struct
(raw get is correct) from a phm/sorted/transient (needs core-get). On a
struct that guard is a second get, so the lookup costs ~36ns where a bare
get is ~20ns. Profiling the ray tracer (jolt-dad) showed keyword lookups are
~50% of a render and the guard is the only avoidable part, but dropping it
needs to know statically that the subject is a plain struct.
Type hints are exactly that information, and jolt already parses them and
otherwise ignores them. This wires one through: a local hinted ^:struct
asserts a plain struct/record map, so a (:k local) lookup on it skips the
guard and emits a bare get. The hint rides on the binding symbol into the
analyzer, which records it per-local and attaches it to :local IR nodes; the
back end reads it on the lookup subject. It also propagates through inlining:
when the inliner let-binds a non-trivial arg to a fresh local, it carries the
called fn's param hint onto that local, so lookups inside the spliced body
keep the bare path. This is a programmer assertion, like a Clojure type hint
(an inaccurate hint just makes the raw get return the wrong value, the same
contract as a wrong ^String), so it stays opt-in and off by default.
On the ray tracer (with inlining on) this is 13.3s to 10.9s, 1.22x, taking it
to 7.8x JVM from 9.4x after the inline pass. The unhinted path emits identical
code (the fast arm is just factored out), so nothing changes without hints.
Verified: a seeded full render produces an identical checksum hinted vs
unhinted; conformance 335/335 in all three modes and the full jpm test pass;
new test/integration/struct-hint-test.janet pins the guard removal, the
inline propagation, and that an accurate hint is correctness-preserving.
Adds two IR passes to jolt.passes that run when a unit opts into
direct-linking (JOLT_DIRECT_LINK=1, off by default). The inline pass splices
small direct-linked fns at their call sites, copy-propagating trivial args so
that scalar replacement can then see map literals across the call boundary.
Scalar replacement is AOT escape analysis: a map allocation whose only use is
constant-keyword lookup is dropped and each (:k m) is replaced with the value
at :k, both for a literal lookup subject and for a non-escaping let-bound map.
Inlining and scalar replacement iterate to a capped fixpoint, since inlining
exposes literals that scalar replacement then collapses.
The back end stashes the body IR of each single-fixed-arity defn on its var
cell (inline-stash!), and the portable pass reads it through two new jolt.host
contract fns (inline-enabled?, inline-ir). Inlining is gated on :inline?, which
is off for all of init so core and the self-hosted compiler compile exactly as
before (const-fold only); api/init and main re-read JOLT_DIRECT_LINK so the
flag works both for a freshly built context and for the build-time-baked one in
the shipped binary.
Only inline-safe targets are spliced: a single fixed arity, no recur/loop/fn/
try crossing the boundary, within a size budget, a closed body (no free locals
beyond the params, so a self-recursive fn's name reference can't dangle), and
not ^:redef / ^:dynamic. Bodies are fully alpha-renamed so no spliced name can
collide with a caller local.
On the ray tracer this is 15.3s -> 13.0s (1.18x). The ceiling is honest: that
workload's cost is dominated by lookups on maps that genuinely escape (rays,
hits, materials) and by dynamic dispatch (the reduce closure, the :scatter fn),
which escape analysis cannot remove. On allocation-bound code where the
temporaries are local it is far larger: a vec3 reflect+dot loop goes 9.3s ->
0.38s (25x), with the loop body reduced to pure arithmetic.
Verified: full jpm test passes (inline off, no regression); conformance 335/335
in all three modes and the clojure-test-suite both pass with inline on; new
inline-sra-test pins the transform and its semantics.
Three canonical-conformance fixes from the post-shrink batch:
- bit-and/bit-or/bit-xor/bit-and-not get Clojure's variadic arities as
20-coll shells folding the binary host ops (now __bit-* seams). 2-arg call
sites still compile to the native janet op via the backend's native-ops
table. The passes.clj constant-fold table now names the seams — the public
fns are overlay and don't exist when the compiler loads (this briefly broke
every compile-mode init).
- core-set? recognizes the :jolt/sorted-set representation (jolt-dpn):
(set? (sorted-set 1)) was false, and ifn? on sorted sets inherited the bug.
- (if) / (if test) / (if test then else extra) throw in both the analyzer
and the interpreter — spec 03-special-forms X1, now marked verified.
Suite 4704 -> 4706; bench and the greeter example benchmark are flat.
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.
