Replaces the watchdog-bump approach with an automated check. The next 5× (or
worse) substrate regression will trip the alarm at build time instead of
hiding behind a deadline bump and only being noticed weeks later.
Components:
* lib/perf-smoke.sx — four micro-benchmarks chosen for distinct substrate
failure modes: function-call dispatch (fib), env construction (let-chain),
HO-form dispatch + lambda creation (map-sq), TCO + primitive dispatch
(tail-loop). Warm-up pass populates JIT cache before the timed pass so we
measure the steady state.
* scripts/perf-smoke.sh — pipes lib/perf-smoke.sx to sx_server.exe, parses
per-bench wall-time, asserts each is within FACTOR× of the recorded
reference (default 5×). `--update` rewrites the reference in-place.
* scripts/sx-build-all.sh — perf-smoke wired in as a post-step after JS
tests. Hard fail if any benchmark regressed beyond budget.
Reference numbers: minimum across 6 back-to-back runs on this dev machine
under typical concurrent-loop contention (load ~9, 2 vCPU, 7.6 GiB RAM,
OCaml 5.2.0, architecture @ 92f6f187). Documented in
plans/jit-perf-regression.md including how to update them.
The 5× factor is chosen so contention noise (~1–2× variance) doesn't trigger
false alarms but a real ≥5× substrate regression — the kind that motivated
this whole investigation — fails the build immediately.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Conflict in lib/tcl/test.sh: architecture had bumped `timeout 2400 → 7200`,
this branch had restored it to `timeout 300` based on the Phase 1
quiet-machine measurement (376/376 in 57.8s wall, 16.3s user). Resolved by
keeping `timeout 300` — the 7200s bump was preemptive against contention,
not against an actual substrate regression. Phase 1 confirms the original
180s deadline is comfortable; 300s gives 5× headroom for moderate noise.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 1 of the jit-perf-regression plan reproduced and quantified the alleged
30× substrate slowdown across 5 guests (tcl, lua, erlang, prolog, haskell). On
a quiet machine all five suites pass cleanly:
tcl test.sh 57.8s wall, 16.3s user, 376/376 ✓
lua test.sh 27.3s wall, 4.2s user, 185/185 ✓
erlang conformance 3m25s wall, 36.8s user, 530/530 ✓ (needs ≥600s budget)
prolog conformance 3m54s wall, 1m08s user, 590/590 ✓
haskell conformance 6m59s wall, 2m37s user, 156/156 ✓
Per-test user-time at architecture HEAD vs pre-substrate-merge baseline
(83dbb595) is essentially flat (tcl 0.83×, lua 1.4×, prolog 0.82×). The
symptoms reported in the plan (test timeouts, OOMs, 30-min hangs) were heavy
CPU contention from concurrent loops + one undersized internal `timeout 120`
in erlang's conformance script. There is no substrate regression to bisect.
Changes:
* lib/tcl/test.sh: `timeout 2400` → `timeout 300`. The original 180s deadline
is comfortable on a quiet machine (3.1× headroom); 300s gives some safety
margin for moderate contention without masking real regressions.
* lib/erlang/conformance.sh: `timeout 120` → `timeout 600`. The 120s budget
was actually too tight for the full 9-suite chain even before this work.
* lib/erlang/scoreboard.{json,md}: 0/0 → 530/530 — populated by a successful
conformance run with the new deadline. The previous 0/0 was a stale
artefact of the run timing out before parsing any markers.
* plans/jit-perf-regression.md: full Phase 1 progress log including
per-guest perf table, quiet-machine re-measurement, and conclusion.
Phases 2–4 (bisect, diagnose, fix) skipped — there is no substrate regression
to find. Phase 6 (perf-regression alarm) still planned to catch the next
quadratic blow-up early instead of via watchdog bumps.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
vwait used frame-lookup which doesn't honor `::` global routing. So
`vwait ::done` after `set ::done fired` (where set routes to root frame)
never saw the var change in the local frame, looping forever.
Added tcl-vwait-lookup helper that mirrors tcl-var-get's `::` routing
but returns nil instead of erroring on missing vars.
Was the deadlock that hung the full test suite past test 32.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Brings the architecture branch (559 commits ahead — R7RS step 4-6, JIT
expansion, host_error wrapping, bytecode compiler, etc.) into the
loops/haskell line of work. Conflict in lib/haskell/conformance.sh:
architecture replaced the inline driver with a thin wrapper delegating
to lib/guest/conformance.sh + a config file. Resolved by taking the
wrapper and extending lib/haskell/conformance.conf with all programs
added under loops/haskell (caesar, runlength-str, showadt, showio,
partial, statistics, newton, wordfreq, mapgraph, uniquewords, setops,
shapes, person, config, counter, accumulate, safediv, trycatch) plus
adding map.sx and set.sx to PRELOADS.
plans/haskell-completeness.md gains three new follow-up phases:
- Phase 17 — parser polish (`(x :: Int)` annotations, mid-file imports)
- Phase 18 — one ambitious conformance program (lambda-calc / Dijkstra /
JSON parser candidate list)
- Phase 19 — conformance speed (batch all suites in one sx_server
process to compress the 25-min run to single-digit minutes)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
hk-bind-exceptions! in eval.sx registers throwIO, throw, evaluate, catch,
try, handle, displayException. SomeException constructor pre-registered
in runtime.sx (arity 1, type SomeException).
throwIO and the existing error primitive both raise via SX `raise` with a
uniform "hk-error: msg" string. catch/try/handle parse it back into a
SomeException via hk-exception-of, which strips nested
'Unhandled exception: "..."' host wraps (CEK's host_error formatter) and
the "hk-error: " prefix.
catch and handle evaluate the handler outside the guard scope (build an
"ok"/"exn" outcome tag inside guard, then dispatch outside) so that a
re-throw from the handler propagates past this catch — matching Haskell
semantics rather than infinite-looping in the same guard.
14 unit tests in tests/exceptions.sx (catch success, catch error, try
Right/Left, handle, throwIO + catch/try, evaluate, nested catch, do-bind
through catch, branch on try result, IORef-mutating handler).
Conformance: safediv.hs (8/8) and trycatch.hs (8/8). Scoreboard now
285/285 tests, 36/36 programs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Trivial wrapper: apl-run-file = apl-run ∘ file-read, where
file-read is built-in to OCaml SX.
Tests verify primes.apl, life.apl, quicksort.apl all parse
end-to-end (their last form is a :dfn AST). Source-then-call
test confirms the loaded file's defined fn is callable, even
when the algorithm itself can't fully execute (primes' inline
⍵ rebinding still missing — :glyph-token, not :name-token).
Parser: :name clause now detects 'name ← rhs' patterns inside
expressions. When seen, consumes the remaining tokens as RHS,
parses recursively, and emits a (:assign-expr name parsed-rhs)
value segment.
Eval-ast :dyad and :monad: when the right operand is an
:assign-expr node, capture the binding into env before
evaluating the left operand. This realises the primes idiom:
apl-run "(2 = +⌿ 0 = a ∘.| a) / a ← ⍳ 30"
→ 2 3 5 7 11 13 17 19 23 29
Also: top-level x←5 now evaluates to scalar 5 (apl-eval-ast
:assign just unwraps to its RHS value).
Caveat: ⍵-rebinding (the original primes.apl uses
'⍵←⍳⍵') is a :glyph-token; only :name-tokens are handled.
A regular variable name (like 'a') works.
Haskell strings are [Char]. Calling reverse / head / length on a SX raw
string transparently produces a cons-list of char codes (via hk-str-head /
hk-str-tail in runtime.sx), but (==) then compared the original raw string
against the char-code cons-list and always returned False — so
"racecar" == reverse "racecar" was False.
Added hk-try-charlist-to-string and hk-normalize-for-eq in eval.sx; routed
== and /= through hk-normalize-for-eq so a string compares equal to any
cons-list whose elements are valid Unicode code points spelling the same
characters, and "[]" ↔ "".
palindrome.hs lifts from 9/12 → 12/12; conformance 33/34 → 34/34 programs,
266/269 → 269/269 tests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ships the algebra for HM-style type inference, riding on
lib/guest/match.sx (terms + unify) and ast.sx (canonical AST):
• Type constructors: hm-tv, hm-arrow, hm-con, hm-int, hm-bool, hm-string
• Schemes: hm-scheme / hm-monotype + accessors
• Free type-vars: hm-ftv, hm-ftv-scheme, hm-ftv-env
• Substitution: hm-apply, hm-apply-scheme, hm-apply-env, hm-compose
• Generalize / Instantiate (with shared fresh-tv counter)
• hm-fresh-tv (counter is a (list N) the caller threads)
• hm-infer-literal (the only fully-closed inference rule)
24 self-tests in lib/guest/tests/hm.sx covering every function above.
The lambda / app / let inference rules — the substitution-threading
core of Algorithm W — intentionally live in HOST CODE rather than the
kit, because each host's AST shape and substitution-threading idiom
differ subtly enough that forcing one shared assembly here proved
brittle in practice (an earlier inline-assembled hm-infer faulted with
"Not callable: nil" only when defined in the kit, despite working when
inline-eval'd or in a separate file — a load/closure interaction not
worth chasing inside this step's budget). The host gets the algebra
plus a spec; assembly stays close to the AST it reasons over.
PARTIAL — algebra + literal rule shipped; full Algorithm W deferred
to host consumers (haskell/infer.sx, lib/ocaml/types.sx when
OCaml-on-SX Phase 5 lands per the brief's sequencing note). Haskell
infer.sx untouched; haskell scoreboard still 156/156 baseline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
apl-permutations was doing (append acc <new-perms>) which is
O(|acc|) and acc grows ~N! big — total cost O(N!²).
Swapped to (append <new-perms> acc) — append is O(|first|)
so cost is O((n+1)·N!_prev) per layer, total O(N!). q(7)
went from 32s to 12s; q(8)=92 now finishes well within the
300s timeout, so the queens(8) test is restored.
497/497. Phase 8 complete.
Configurable layout pass that inserts virtual open / close / separator
tokens based on indentation. Supports both styles the brief calls out:
• Haskell-flavour: layout opens AFTER a reserved keyword
(let/where/do/of) and resolves to the next token's column. Module
prelude wraps the whole input in an implicit block. Explicit `{`
after the keyword suppresses virtual layout.
• Python-flavour: layout opens via an :open-trailing-fn predicate
fired AFTER the trigger token (e.g. trailing `:`) — and resolves
to the column of the next token, which in real source is on a
fresh line. No module prelude.
Public entry: (layout-pass cfg tokens). Token shape: dict with at
least :type :value :line :col; everything else passes through. Newline
filler tokens are NOT used — line-break detection is via :line.
lib/guest/tests/layout.sx — 6 tests covering both flavours:
haskell-do-block / haskell-explicit-brace / haskell-do-inline /
haskell-module-prelude / python-if-block / python-nested.
Per the brief's gotcha note ("Don't ship lib/guest/layout.sx unless
the haskell scoreboard equals baseline") — haskell/layout.sx is left
UNTOUCHED. The kit isn't yet a drop-in replacement for the full
Haskell 98 algorithm (Note 5, multi-stage pre-pass, etc.) and forcing
a port would risk the 156 currently passing programs. Haskell
scoreboard remains at 156/156 baseline because no haskell file
changed. The synthetic Python-ish fixture is the second consumer per
the brief's wording.
PARTIAL — kit + synthetic fixture shipped; haskell port deferred until
the kit grows the missing Haskell-98 wrinkles.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
hk-bind-data-ioref! registers newIORef / readIORef / writeIORef /
modifyIORef / modifyIORef' under the import alias (default IORef).
Representation: dict {"hk-ioref" true "hk-value" v} allocated inside IO.
modifyIORef' uses hk-deep-force on the new value before write.
Side-effect: fixed pre-existing bug in import handler — modname was
reading (nth d 1) (the qualified flag) instead of (nth d 2). All
'import qualified … as Foo' paths were silently no-ops; map.sx unit
suite jumps from 22→26 passing.
Conformance now 33/34 programs, 266/269 tests (only pre-existing
palindrome.hs 9/12 still failing on string-as-list reversal, present
on prior commit).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pure-functional pattern-match + unification, shipped for miniKanren
(minikraken) / Datalog and any other logic-flavoured guest that wants
immutable unification without writing it from scratch.
Canonical wire format (config callbacks let other shapes plug in):
var (:var NAME)
constructor (:ctor HEAD ARGS)
literal number / string / boolean / nil
Public API:
empty-subst walk walk* extend occurs?
unify (symmetric, with occurs check)
unify-with (cfg-driven for non-canonical term shapes)
match-pat (asymmetric pattern→value, vars only in pattern)
match-pat-with (cfg-driven)
lib/guest/tests/match.sx — 25 tests covering walk chains, occurs,
unify (literal/var/ctor, head + arity mismatch, transitive vars),
match-pat. All passing.
The brief flagged this as the highest-risk step ("revert and redesign
on any regression"). The two existing engines — haskell/match.sx
(pure asymmetric, lazy, returns env-or-nil) and prolog runtime.sx
pl-unify! (mutating symmetric, trail-based, returns bool) — are
structurally divergent and forcing a shared core under either of their
contracts would risk the 746 tests they currently pass. Both are
untouched; they remain at baseline (haskell 156/156, prolog 590/590)
because none of their source files were modified.
PARTIAL — kit shipped, prolog/haskell ports deferred until a guest
chooses to migrate or until a third consumer (minikraken / datalog)
provides a less risky migration path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
programs-e2e.sx exercises the classic-algorithm shapes from
lib/apl/tests/programs/*.apl via the full pipeline (apl-run on
embedded source strings). Tests include factorial-via-∇,
triangular numbers, sum-of-squares, prime-mask building blocks
(divisor counts via outer mod), named-fn composition,
dyadic max-of-two, and a single Newton sqrt step.
The original one-liners (e.g. primes' inline ⍵←⍳⍵) need parser
features we haven't built (compress-as-fn, inline assign) — the
e2e tests use multi-statement equivalents. No file-reading
primitive in OCaml SX, so source is embedded.
Side-fix: ⌿ (first-axis reduce) and ⍀ (first-axis scan) were
silently skipped by the tokenizer — added to apl-glyph-set
and apl-parse-op-glyphs.
Defines the 10 canonical node kinds called out in the brief — literal,
var, app, lambda, let, letrec, if, match-clause, module, import — plus
predicates, ast-kind dispatch, and per-field accessors. Each node is a
tagged keyword-headed list: (:literal V), (:var N), (:app FN ARGS), …
Also lib/guest/tests/ast.sx — 33 tests exercising every constructor +
predicate + accessor, runnable via (gast-tests-run!) which returns the
{:passed :failed :total} dict the shared conformance driver expects.
PARTIAL — pending real consumers. The brief calls Step 5 "Optional —
guests may keep their own AST" and forcing lua/prolog to switch their
internal AST shape risks regressing 775 passing tests for tooling that
nothing yet calls. Both internal ASTs are untouched; lua still 185/185,
prolog still 590/590. Datalog-on-sx (in flight, see plans/datalog-on-sx.md)
will be the natural first real consumer; lua/prolog converters can land
when a cross-language tool wants them.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Parser: apl-collect-fn-bindings pre-scans stmt-groups for
`name ← { ... }` patterns and populates apl-known-fn-names.
is-fn-tok? consults this list; collect-segments-loop emits
(:fn-name nm) for known names so they parse as functions.
Resolver: apl-resolve-{monadic,dyadic} handle :fn-name by
looking up env, asserting the binding is a dfn, returning
a closure that dispatches to apl-call-dfn{-m,}.
Recursion still works: `fact ← {0=⍵:1 ⋄ ⍵×∇⍵-1} ⋄ fact 5` → 120.
Three small unblockers in one iteration:
- tokenizer: read-digits! now consumes optional ".digits" suffix,
so 3.7 and ¯2.5 are single number tokens.
- tokenizer: ⎕ followed by ← emits a single :name "⎕←" token
(instead of splitting on the assign glyph). Parser registers
⎕← in apl-quad-fn-names; apl-monadic-fn maps to apl-quad-print.
- eval-ast: :str AST nodes evaluate to char arrays. Single-char
strings become rank-0 scalars; multi-char become rank-1 vectors
of single-char strings.
Extracted the data-half of Pratt-style precedence parsing: the operator
table format and lookup. The climbing loop stays per-language because
the two canaries use opposite conventions (lua: higher prec = tighter;
prolog: lower prec = tighter, with xfx/xfy/yfx assoc tags) — forcing
one shared loop adds callback indirection that obscures more than it
shares. The brief's literal ask is "Grammar is a dict, not hardcoded
cond" and that's what gets shared.
Entry shape: (NAME PREC ASSOC). Three accessors: pratt-op-name /
pratt-op-prec / pratt-op-assoc. One traversal: pratt-op-lookup.
Ported lua/parser.sx — replaced 18-clause cond and the
lua-binop-right? hardcoded `or` with a 15-entry lua-op-table, now
queried via pratt-op-lookup. Ported prolog/parser.sx — pl-op-find
(linear walk reimpl) deleted; pl-op-lookup wraps pratt-op-lookup;
pl-token-op simplified to return the entry directly.
Verification:
- lua/test.sh: 185/185 = baseline.
- prolog/conformance.sh: 590/590 = baseline (timestamp-only diff).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
