Profiling jolt-i5if showed <=60-bit arithmetic is already native-fast; the real
general overhead in the run/-e/-m path is var resolution. Every var reference
compiled to (var-deref ns name), which builds + hashes a fresh "ns/name" string
and does a hashtable lookup per access (~45ns). The var cell is interned and
def-var! mutates it in place, so caching the resolved cell is sound under
redefinition.
Generalize the devirt per-site cache-cell mechanism to var value references: a
ref inside a fn resolves its cell once into the def's closure, then reads it via
var-cell-deref (a field read after the first). var-cell-deref is the cell-based
var-deref — binding-aware (dynamic vars + *ns* still resolve) and lenient on an
unbound root (a forward-declared var doesn't throw, unlike jolt-var-get).
Gated by a runtime flag: ON for runtime-compiled code (compile-eval.ss), OFF for
the seed mint and AOT build (emit-image.ss) so the seed stays a byte-fixpoint --
prelude.ss is unchanged, only image.ss picks up the new backend. ~5x on a
var-ref-heavy loop (1058ms->205ms); ~1.2x on test.check (its generators are more
deftype/dispatch-bound than var-deref-bound). No C/FFI.
Corpus rows pin redefinition / dynamic binding / forward ref through a cached
ref. make test + shakesmoke green, selfhost holds, SCI 211/218, certify 0-new.
The reader dropped the namespace on ::kw (read ::foo as :foo), so auto-resolved
keywords never matched their qualified form — code that round-trips them (spec
keys, aero's :aero.core/* expansion keys) silently broke. Resolve ::name against
the current ns and ::alias/name through the alias table, as Clojure does. The
runtime loader reads form-by-form with the ns set after the ns form; the
cross-compile reads all forms up front, so ei-emit-ns*/ei-emit-ns-records set the
ns before reading.
clojure.edn/read over a reader discarded its opts map — :readers/:default/:eof
were ignored, so a custom :default never saw the tag. Route the reader arity
through read-string so opts apply, and pass the tag to :default as a symbol (not
the internal :#name keyword), matching Clojure.
Seed re-minted (the ::halt transducer key in clojure.core now reads as
:clojure.core/halt). Corpus gains ::-keyword rows; the unit case that asserted the
old ns-dropping behavior now asserts the qualified result.
types.clj held the inferencer, the success-type checker, and the driver in one
716-line namespace. Move the self-contained checker into jolt.passes.types.check:
the error-domain predicates (not-number?/not-seqable?/not-callable?), the op
tables, type-name, check-invoke, and the user-fn registry. These are pure over
inferred types and the run's env cells, with no inference, so a check-rule edit
can no longer perturb the inferencer.
The infer-coupled probes stay in types.clj — isolated-diag-count and
check-user-call re-run inference, so moving them would make check depend on the
inferencer and reintroduce the cycle. Verbatim move; new ns wired into
ei-compiler-ns-files; seed re-minted to the byte-fixpoint.
Tree-shaking was split across emit-image.ss (the dce-* helpers + record producer)
and build.ss (bld-shake-all + the manifest splice), with the DCE record accessed by
raw (vector-ref r 0..3) in ~10 places and the manifest splice/drop driven by
substring-matching (load "…prelude.ss").
- New host/chez/dce.ss owns the whole DCE concern: a named record API
(dce-rec/-keep?/-fqn/-refs/-str — no more positional vector indexing), the ref
extraction + ref-set constants, dce-blob-records, and dce-shake decomposed into
dce-build-graph / dce-reachable / dce-bail-scan / dce-partition (was one 50-line
bld-shake-all doing five jobs with shared mutable state).
- emit-image.ss keeps only ei-emit-ns-records (it drives the ei-* compiler) and uses
the dce-rec constructor.
- The runtime manifest is now tagged ('prelude/'image/'compile-eval); bld-emit-runtime
dispatches on the tag instead of substring-matching file paths, so the core-splice
and compiler-drop can't silently break on a rename/reorder.
Behaviour-preserving (runtime .ss, no re-mint): build-app shakes identically
(56/460, 8.12MB), make test green, make shakesmoke green (4 git-lib apps
byte-identical, same sizes).
--tree-shake now shakes the clojure.core/stdlib prelude in the same reachability
graph as the app + libraries — only core fns actually reached from -main ship.
dce-blob-records reads prelude.ss with Chez `read`, unwraps each
(guard ... (def-var! "ns" "name" V)) and builds the core->core call graph from the
(var-deref/jolt-var "ns" "name") refs in V — the real emitted edges, no
re-analysis. bld-shake-all merges core + app records into one BFS; roots are -main,
side-effecting forms, and the clojure.core fns the runtime .ss shims reference by
name (enumerated in dce-runtime-core-roots). The shaken core is spliced where the
prelude.ss blob was, in original (tier) order, so load-time deps are preserved.
Bail (reachable runtime resolve) keeps the prelude whole.
Soundness follows Stalin's rule (any reference, value or call, keeps a def live):
dce-collect-refs counts :var and :the-var; non-def registration forms are always
kept (covers protocol/multimethod dispatch). Validated by make shakesmoke: the four
deps.edn git-lib apps build byte-identical output shaken vs not.
Wins (binary): build-app 9.84MB -> 8.12MB (dropped 403/457 core defs); malli-app
10.0MB -> 8.1MB; markdown 9.9 -> 8.3; commonmark 9.8 -> 8.1 — all output-identical.
build-smoke asserts an unused core fn (group-by) is dropped; full make test green.
Two things, both from studying Stalin's closed-world DCE:
1. Soundness fix: dce-collect-refs now counts a :the-var (#'x / (var x)) reference,
not just a :var. Stalin's rule is that ANY reference — value position, not only a
direct call — keeps the target live; a var referenced as #'x would otherwise be
wrongly dropped. (My :var collection already covered value-position refs; this
closes the the-var hole.)
2. host/chez/tree-shake-smoke.sh (make shakesmoke): builds example apps default vs
--tree-shake and requires identical output — the real risk is shaking a binary
that pulled libraries via deps.edn. Covers markdown/malli/commonmark/hiccup
(git-lib apps). All produce byte-identical output shaken vs not, and drop
~0.8-1MB (malli 10.0MB -> 9.0MB) from the compiler-drop. Slow; not in the default
gate. Skips without the examples repo.
An AOT-compiled app only needs the analyzer/back end at runtime to compile from
source — eval / load-string / load-file. Macros are expanded at build time and a
require of a baked namespace no-ops, so a closed app that never compiles at runtime
carries the compiler image (~0.8MB) as dead weight.
Under --tree-shake, when reachability is trustworthy (no bail) and no reachable code
references eval/load-string/load-file/load-reader/load, omit host/chez/seed/image.ss
+ compile-eval.ss from the runtime manifest. bld-tree-shake returns the flag
alongside the shaken forms; bld-emit-runtime filters the manifest.
Measured: build-app 9.84MB -> 9.05MB, still runs. Safety verified: an app that evals
keeps the compiler (eval is a bail + compile ref) and eval works at runtime.
build-smoke asserts the compiler is gone in the no-eval app; full make test green.
`jolt build --tree-shake` (or deps.edn :jolt/build {:tree-shake true}) does
reachability DCE over the re-emitted app + library namespaces: keep -main, every
side-effecting (non-def) top-level form, and every def reachable from those; drop
the rest. A macro (expanded at AOT, never called at runtime) is prunable too.
Sound: bails (keeps everything) if REACHABLE code resolves vars by name at runtime
(eval/resolve/ns-resolve/requiring-resolve/find-var/intern/load-string/...), which a
static call graph can't follow. Unreached eval-using library code is simply shaken
away and never triggers the bail. clojure.core and the compiler image stay baked
(prelude + image blobs), so only re-emitted namespaces are shaken for now.
The reachability machinery is in emit-image.ss (records: keep?/fqn/refs/str via
reduce-ir-children) + build.ss (BFS + bail check). build-smoke covers it (drops the
unreachable `twice` macro, output unchanged). Opt-in; default builds are untouched.
full make test green.
Scope note: this shakes the re-emitted app/lib code only. Measurement shows jolt's
compiled code is ~5.8MB of a ~9.8MB binary, dominated by the clojure.core prelude
(~1.5-2MB) and the compiler image (~0.8MB) — both baked blobs this pass doesn't
touch. Those (shake-core, drop-compiler-when-no-eval) are the larger footprint wins,
filed as follow-ups.
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.
A release/optimized `jolt build` is a closed world: every app def is final, so
an app->app call can bind to the def's Scheme binding directly instead of going
through (jolt-invoke (var-deref ns name)).
The emitter gains a direct-link mode (off for the seed mint, runtime -e/repl, and
dev builds). With it on, a top-level app def also emits a binding jv$<ns>$<name>
that def-var! aliases; an app->app call or value-ref to a name already emitted in
the unit lowers to that binding, skipping both the var-table lookup and the
generic IFn dispatch. ^:dynamic/^:redef defs and nested defs (a defonce's inner
def) opt out and stay indirect. Off direct-link mode, emit-top-form is exactly
emit, so the seed and runtime eval are byte-unchanged (selfhost holds).
build.ss turns it on for release + optimized; the defined-set accumulates across
the dependency-ordered namespaces so a dep's defs are linkable by the time the
entry that calls them is emitted. App->core calls stay indirect for now (core is
the baked seed); that's a later stage.
~1.74x on a hot cross-namespace call loop (26.5s -> 15.2s).
types.clj was 852 lines mixing the pure structural-type algebra with the
inference engine, checker, and driver. Move the lattice — scalar/struct/vec/set/
union types, join-t, depth-cap, shape, and the numeric/vector return-fn sets —
into jolt.passes.types.lattice (no inference state, no requires). types.clj
requires it; the engine is now ~720 lines. Compiled into the image before
jolt.passes.types. Re-minted seed differs only by gensym label renumbering.
The 1123-line collection tier is the largest source file. Cut it at two existing
section banners into 20-coll (predicates, printing, hierarchies, pure-over-core
leaves), 21-coll (rand/sort seams, the test runner, fn combinators), and 22-coll
(canonical Clojure ports, transduce/into, JVM-shape stubs). No macros in this tier,
so order is the only constraint; the emit-image manifest lists the three in
sequence. Re-minted seed is identical apart from gensym label renumbering.
ei-emit-ns (emit-image) and bld-emit-ns (build) were near-verbatim copies that had
drifted: the minter guard-wraps and skips failing forms, the build is strict, and
since the passes were wired the build also runs run-passes. Fold both into
ei-emit-ns* with optimize?/guard? flags; ei-emit-ns and bld-emit-ns become one-line
callers. Output is byte-identical (selfhost fixpoint and build smoke stay green).
The fold/inline/types passes and the jolt.passes façade were baked into neither
seed half and never invoked: compile-eval and build went analyze -> emit directly,
and `jolt build --opt` flipped an optimize flag that nothing consumed.
- Compile the passes into the image (emit-image manifest): fold, inline, types,
then the jolt.passes façade, after jolt.ir.
- compile-eval and build.ss now run jolt.passes/run-passes between analyze and
emit. Off the direct-link path it is a pure const-fold; `jolt build --opt`
turns on inline + flatten + scalar-replace + type inference (it sets
hc-optimize?, which inline-enabled? reads).
- The seed minter (emit-image) stays analyze -> emit, so the seed is built
un-optimized and the self-host fixpoint is unaffected.
build-smoke already exercised --opt; it now actually optimizes and still matches
the release binary's output. Corpus floor and the fixpoint are green.
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.
Rephrase comments that pointed at deleted Janet files (emit.janet, the seed
sources, 'the Janet back end punts ...') to present-tense descriptions of the
Chez behavior. Comment/docstring-only; the self-host fixpoint is unchanged
(comments don't affect the compiled seed).
Delete five files that were Janet-host shims with no Chez path: clojure.java.io
(provided natively by host/chez/io.ss), and jolt.{nrepl,png,interop,shell}
(the janet.* bridge, os/shell, janet.net — none exist on Chez).
jolt-cf1q.6
The compiler image is already resident at runtime on the Chez spine, so eval
and load-string are just wiring: make them clojure.core functions instead of
analyzer special forms.
- eval / load-string are now functions, not special forms. Dropped "eval" from
the host-contract special-symbol lists so it resolves as an ordinary var, and
def-var! both in compile-eval.ss. eval takes an already-read form (e.g. from
quote/list) and compiles+evals it in the current ns; load-string reads every
form from a source string and evals each, returning the last.
- Runtime defmacro: jolt-compile-eval-form intercepts a (defmacro ...) form
before analysis, defs the expander fn + mark-macro!s the var, exactly as
emit-image.ss does at build time. The two helpers (macro-form? / defmacro->fn)
move to compile-eval.ss and emit-image.ss reuses them.
- Top-level (do ...) is now unrolled form-by-form, like Clojure, so a defmacro
or def in an earlier subform is visible (macro flag set / var interned) before
a later subform is analyzed. This is what makes multi-form -e with a macro work.
Seed is byte-identical (no source references eval), so no re-mint; bootstrap-test
still passes. Zero-Janet corpus 2534 -> 2544 (eval/load-string cases now run),
0 new divergences; floor raised. Prelude corpus, JVM cert, full Janet gate green.
jolt-r8ku
Extends the fixpoint beyond the compiler image to the whole emitted system.
emit-image.ss now handles macros (defmacro -> bare expander fn + def-var! +
mark-macro!) and re-emits the clojure.core prelude (all tiers + stdlib) on Chez
via jolt-emit-prelude; driver's emit-image-on-chez takes an emit-fn arg.
The prelude converges at pstage3==pstage4 (one stage later than the compiler's
stage2==stage3) because macro expanders bake an auto-gensym id at emit time, so a
Janet-emitted macro carries a different id than a Chez-emitted one — only once
both stages load a Chez-emitted prelude does it stabilize.
fixpoint-test now proves: compiler stage2==stage3, prelude pstage3==pstage4, and
the fully Chez-emitted system (Chez prelude + Chez image, no Janet artifact in the
loop) compiles+runs real cases. 10/10.
The zero-Janet spine proves the on-Chez analyzer/emitter compile arbitrary
Clojure faithfully. This proves the stronger property: the on-Chez compiler
reproduces itself. emit-image.ss re-emits the compiler sources (jolt.ir +
jolt.analyzer + jolt.backend-scheme) ON CHEZ via the loaded image; feeding it
stage1 (the Janet cross-compile) yields stage2, feeding stage2 yields stage3.
stage2 and stage3 are byte-for-byte identical, and stage2 is a working compiler
(real cases compile+run through it). stage1 differs from stage2 only in gensym
numbering, so the fixpoint is stage2==stage3.
driver: emit-image-on-chez / program-emit-image spawn a fresh chez per stage
(clean gensym state). test/chez/fixpoint-test.janet gates it (skips without chez).
