The OP_DIV/numeric-tower work on this branch made the OCaml `/` primitive
return an exact Rational for (/ int int) (e.g. (/ 5 2)=5/2), diverging from
the canonical spec ("/ always returns inexact float"), the test-rationals.sx
header ("in the JS host, (/ int int) returns float — backward-compatible"),
and the JS host itself. That leaked rationals into arithmetic results and
rendered CSS (tw-opacity emitted `opacity:1/20` instead of `0.05`).
Decision (with the user): keep exact rationals as an explicit opt-in
(literals 1/3, make-rational) but bring `/` back into spec/host parity —
the isomorphic SSR↔hydration invariant requires both hosts to agree, and
JS has no native rational type.
- sx_primitives.ml `/`: (/ int int) → integer when exactly divisible, else
inexact float; a Rational operand still yields an exact rational (matches
test-numeric-tower: (/ 6 2)=3, (/ 1 4)=0.25, (/ 5 2)=2.5, (/ 1/2 2)=1/4).
- sx_primitives.ml number? / exact?: recognise the Rational type (real bugs —
test-rationals asserts (number? 1/3) and (exact? 1/3); inexact?/float?
already returned false for Rational, correct).
- sx_vm.ml OP_DIV: comment updated (it delegates to the now-float `/`).
- test-rationals.sx: fix typo in "rational * float = float" — used int 2,
should be 2.0 (1/2 * 2 = 1 exact, not a float; name + siblings use floats).
OCaml conformance 4834→4863 (+29 fixed, zero new failures); rationals,
numeric-tower, arithmetic, tw-opacity suites all 0 failures. Remaining run_tests
failures are the pre-existing environmental hyperscript (host-call-fn) set.
JS host already handles number?/exact? on rationals and float `/`; its
remaining float?/contagion failures are a separate pre-existing limitation
(JS has no distinct float type), out of scope here.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fixes the silent miscompile under SX_SERVING_JIT=1 (http-listen + cek_run_with_io):
a perform inside a native HO-primitive callback (map/filter/reduce/for-each)
unwound the native loop, corrupting the stack so the next CALL_PRIM read garbage
args (map/rest/drop). (A) call_closure_reuse resolves callback IO inline in
serving mode so the loop survives; (A') resume_vm restores _active_vm; (B)
register_jit_hook resolve_loop falls back to CEK on resume error (no 500).
Repro 9/9 (hosts/ocaml/bin/repro_jit_resume.ml); conformance unchanged 4834/1110.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Root cause (found via bin/repro_jit_resume.ml, 9 surgical cases): when a
`perform` (durable kv read) fires inside a native HO-primitive callback
(map/filter/reduce/for-each/some/every?), the VmSuspended unwound through
the primitive's native OCaml loop (List.map etc.), destroying the loop's
iteration state. The remaining elements were dropped and the stack left
misaligned, so the NEXT CALL_PRIM (map/rest/drop) read wrong args —
"map: expected (fn list)", "rest: 1 list arg", "drop: list and number".
Only triggers in the http-listen + cek_run_with_io serving path (epoch
eval has no synchronous resolver, so conformance was 271/271).
(A) lib/sx_vm.ml call_closure_reuse: when a callback suspends AND a
synchronous IO resolver is installed (serving mode), resolve the
callback's IO inline and run it to completion right there, returning its
value to the native loop — so the loop is never unwound. Scoped to the
resolver-set path; the CEK-driven path (flow/reactive/async tests) keeps
its existing reuse_stack behaviour, so nothing else changes. reuse_stack
is isolated across the nested resume.
(A') lib/sx_vm.ml resume_vm: re-assert _active_vm := Some vm for the
duration of the resumed run (mirrors call_closure). call_closure restored
_active_vm to the caller when VmSuspended unwound, so HO callbacks during
a resume could land on the wrong VM. Latent-bug fix.
(B) bin/sx_server.ml register_jit_hook: the resolve_loop runs inside the
VmSuspended handler, so a non-VmSuspended exception from resume_vm escaped
to the http handler (→ 500). Catch it and fall back to CEK for THIS call
(mark jit_failed, return None → interpreter re-runs it). Self-heals on the
first hit, not a retry. Defense-in-depth; with (A) it shouldn't trigger.
Verification: repro 9/9 (incl. host shape: map[cb→interpreted-helper
perform]→drop = (7 8); reduce; nested map). Standard + --full OCaml
conformance unchanged at 4834/1110 (baseline identical — the 1110 are
pre-existing environmental: host-call-fn/browser-platform symbols,
rational display, tw/regex). Host loop to re-verify 271/271 serving and
drop its (jit-exclude! "host/*" "dream-*" "dr/*") band-aid.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The 6 common-lisp opt-in-JIT failures were all condition-system continuation
escape: cl-restart-case/cl-handler-case/cl-handler-bind wrap their body in
call/cc (restarts + non-local handler exit). When an SX function that drives
the condition system (the parse-recover / interactive-debugger fixtures, e.g.
parse-numbers, make-policy-debugger) is JIT-compiled, the call/cc form runs in
a NESTED cek-run where invoking the captured continuation
runs-to-completion-and-returns instead of escaping — so a restart fails to
abort and the body falls through. Observed as result accumulation
(got (1 3 0 3) vs (1 3)) and no-abort (restart returns the 999 sentinel).
These callers are arbitrary user/fixture code, not a fixed namespace, so they
can't be prefix-excluded. New data-driven mechanism:
- jit-exclude-callers-of! registers call/cc-establishing form names in
Sx_types.jit_excluded_caller_names.
- jit_compile_lambda skips any function whose constant pool (recursively,
incl. nested closures) references a registered name — code_refs_escaping_caller.
Guarded by Hashtbl.length > 0 so it's a no-op for every guest that doesn't
register (zero effect outside CL).
- lib/common-lisp/runtime.sx registers the establish side (cl-restart-case,
cl-handler-case, cl-handler-bind) and the invoke side (cl-invoke-restart,
cl-invoke-debugger, cl-signal, cl-error-with-debugger).
Result: CL conformance under SX_SERVING_JIT=1 = 487/0, EXACTLY matching the CEK
baseline (was 484/6 with a +3 double-execution over-count). parse-recover
3/4 -> 6/0, interactive-debugger 7/2 -> 7/0.
Note: the geometry/mop-trace suites report 0/0 on BOTH CEK and JIT — they error
"Undefined symbol: refl-class-chain-depth-with" (the CLOS suites don't preload
lib/guest/reflective/class-chain.sx). Pre-existing conformance-harness gap, not
a JIT issue; left as-is.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The host combined-binary integration test exposed a new JIT-unsafe class:
Dream's error middleware (host/wrap-errors -> dream-catch-with) failed to catch
a thrown error under JIT — it escaped as "Unhandled exception" and truncated the
host middleware suite (7/9 vs 9/9 on CEK).
Root cause: the VM's OP_PUSH_HANDLER (the compiled form of `guard`) only
intercepts a VM-level RAISE (opcode 37); it does NOT catch the OCaml Eval_error
that the `error` primitive throws from a CALL/CALL_PRIM in a callee frame. So a
JIT-compiled `guard` silently fails to catch. dream-catch-with is curried
((fn (on-error) (fn (next) (fn (req) (guard ...))))), so the guard lives in a
NESTED closure — JIT-compiling the outer function mints that inner guard as a
VmClosure with the broken VM handler.
Fix (central, not per-callsite): scan a JIT candidate's bytecode RECURSIVELY —
including nested closure code in the constant pool — for OP_PUSH_HANDLER, and
skip JIT for any handler-installing function. It then runs on the CEK, whose
guard catches correctly. Covers dream-catch-with, host wrap-errors/blog-render,
and every other guard / handler-bind user automatically.
Verified: minimal direct guard and curried cross-frame guard both return the
caught value under JIT (were "Unhandled exception"); the host run's "kaboom"
escapes went 2 -> 0. (Remaining host blog/page failures are "Undefined symbol:
render-page" — the host's native render fn, absent from the standalone
sx_server.exe; identical on CEK, i.e. an environment artifact, not a JIT
regression. The combined host binary has render-page.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Enabling the epoch serving-mode JIT globally regressed continuation-based guest
interpreters (the epoch mode is the shared command channel every loop's
conformance runner uses). Two-part fix:
1. SAFE DEFAULT GATE. register_jit_hook in the persistent server branch is now
opt-in via SX_SERVING_JIT=1 (default OFF). Default behaviour is unchanged
(no JIT in epoch serving) → zero regression for sibling loops. The
content/Smalltalk page server opts in.
2. GENERAL FIXES + per-guest interpret-only declarations:
- callable? (sx_server/run_tests/integration_tests/mcp_tree) now accepts
VmClosure. A JIT-compiled higher-order function returns its inner closure
as a VmClosure; callable? previously rejected it, so scheme-apply's
(callable? proc) guard failed with "not a procedure: <vm:anon>".
- jit-exclude! gains a trailing-"*" namespace-prefix form
(Sx_types.jit_excluded_prefixes), the robust way to mark a whole guest
interpreter interpret-only (a name-list misses functions in extra files —
it left erlang's vm/dispatcher JIT'd and 13 tests short).
- Per-guest exclusions in each guest's runtime.sx:
scheme "scheme-*" "scm-*" erlang "er-*" "erlang-*"
prolog "pl-*" common-lisp "cl-*" "clos-*"
js "js-*" haskell "hk-*"
Verified under opt-in JIT (== CEK, no hang): smalltalk 847/847, scheme/flow
166/166, erlang 530/530, prolog 590/590, apl 152/152, js 147/148. Residual
(documented, protected by the default gate): common-lisp 6 fails in advanced
suites (parser-recovery/debugger/CLOS/MOP). lua (0/16) and tcl (3/4) fail
identically on CEK — pre-existing, not JIT. run_tests --jit/no-jit unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
register_jit_hook is now installed in the persistent (epoch) serving-mode
branch of sx_server.ml, not just --http/cli/site. Smalltalk-on-SX conformance
under JIT is 847/847 — identical to the no-JIT baseline; Datalog 356/356.
run_tests --jit/no-jit are byte-identical before/after (no regression).
Five distinct root causes fixed (not one "miscompile"):
1. Serving mode never loaded lib/compiler.sx, so JIT used the native
Sx_compiler.compile stub (arity-0 bytecode, params as GLOBAL_GET →
"VM undefined: <param>"). Server-mode branch now loads compiler.sx
before registering the hook, matching http/cli/site.
2. compile-cond / compile-case-clauses / compile-guard-clauses only treated
keyword :else and true as the catch-all, not the bare symbol `else` that
the CEK's is-else-clause? accepts → GLOBAL_GET "else". (lib/compiler.sx)
3. OP_DIV produced a float for non-divisible Integer/Integer (1/2 → 0.5)
instead of the exact Rational the "/" primitive returns. Now delegates to
the primitive, matching CEK. (sx_vm.ml)
4. OP_EQ / _fast_eq lacked Rational/ListRef cases that the "=" primitive's
safe_eq has → (= 1/2 1/2) false under JIT. OP_EQ now delegates non-scalars
to the "=" primitive; _fast_eq gained rational + ListRef. (sx_vm.ml,
sx_runtime.ml)
5. Continuation-based control flow (Smalltalk ^expr non-local return, block
escape, exceptions via call/cc) can't run in the stack VM. New data-driven
exclusion set Sx_types.jit_excluded + `jit-exclude!` primitive, consulted in
jit_compile_lambda (covers both the CEK hook and vm_call's tiered path).
lib/smalltalk/eval.sx self-declares its continuation dispatch core
interpret-only; pure helpers still JIT. The SUnit suite-runner test helper
pharo-test-class miscompiles mid-loop and is excluded in tests/tokenize.sx.
Also adds SX_JIT_DENY / SX_JIT_ONLY env-var bisection filters to the serving
hook. Known residual documented in plans/jit-bytecode-correctness.md: the hook
re-runs a failed VM execution via CEK (correct result, possible duplicate side
effects); adopting run_tests' propagate-don't-rerun semantics is deferred to
avoid changing shared VM/CEK behavior under this loop.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
blob/put|get|has? backed by <root>/blobs/<cid>, CIDv1 (raw codec,
sha2-256 via Sx_cid/Sx_sha2). put idempotent; persist stores only the
{:cid :size :mime} ref. persist_durable_test.sh extended (8/8): blob
round-trip + content-address idempotency + bytes/ref surviving real
restart. Mock blob suite 14/0 on worktree binary. Durable-storage
Blocker now CLOSED.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sx_persist_store services every persist/* IO op against on-disk storage
(append-only log + separate monotonic .seq high-water + per-key kv files,
SX-serialized). Wired into the (eval) suspension loop, cek_run_with_io
bridge, and in-process _cek_io_resolver. Data-loss repro now (3 3 3).
New persist_durable_test.sh: durable + monotonic-seq + streams + kv +
real process restart all green (5/5).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
NATIVE-ONLY http-request primitive (bin/sx_server.ml). HTTP/1.1 over
Unix sockets + gethostbyname; inline http:// URL parsing (full
url-parse deferred to Phase K); Connection: close + Host +
Content-Length headers auto-supplied; reads response via
Content-Length or read-to-EOF; chunked transfer-encoding rejected.
Test bin/test_http_client.sh spins a Phase-H echo server and drives
a second sx_server: GET+query, POST+body, 404, custom request
header reflected, non-http scheme rejected, integer status — 6/6.
WASM boot green (prim not in lib); Erlang conformance 530/530.
Adds Sx_vm.bytecode_uses_extension_opcodes — an operand-aware
bytecode scanner that walks past CONST u16, CALL_PRIM u16+u8, and
CLOSURE u16+dynamic upvalue descriptors so operand bytes that happen
to be ≥200 don't false-positive as extension opcodes.
jit_compile_lambda calls the scanner on the inner closure's bytecode.
On hit it returns None — the lambda then runs through CEK
interpretation. The VM's dispatch fallthrough still routes the
extension opcodes themselves through the registry; this change just
prevents the JIT from claiming code it has no plan for.
Tests: 7 new foundation cases — pure core eligible, head/middle/
post-CLOSURE detection, CONST + CALL_PRIM + CLOSURE-descriptor false-
positive avoidance. +7 pass vs Phase D baseline, no regressions
across 11 conformance suites.
Loop complete: acceptance criteria 1-4 met. Hand-off to the Erlang
loop — lib/erlang/vm/dispatcher.sx's Phase 9b stub can now be
replaced with a real hosts/ocaml/lib/extensions/erlang.ml consumer.
lib/extensions/ becomes the new home for VM extensions, wired in via
(include_subdirs unqualified). README documents the registration
pattern, opcode-ID range conventions (200-209 guest_vm, 210-219
inline test, 220-229 test_ext, 230-247 ports), and naming rules.
extensions/test_ext.ml is the canonical worked example — two
operand-less opcodes (220 push 42, 221 double TOS) carrying a per-
extension state slot (TestExtState invocation counter). Test_ext.register
called from run_tests.ml at the start of the Phase D suite, on top of
the inline test_reg from earlier suites (disjoint opcode IDs).
Sx_vm.opcode_name now consults extension_opcode_name_ref (forward ref
in the same style as extension_dispatch_ref), so disassemble shows
extension opcodes by name instead of UNKNOWN_n. Registry maintains
name_of_id_table and installs the lookup at module init.
Tests: 5 new foundation cases — primitive resolves test_ext name,
end-to-end bytecode (push + double + return → 84), disassemble shows
"test_ext.OP_TEST_PUSH_42" / "test_ext.OP_TEST_DOUBLE_TOS",
unregistered ext opcodes still fall back to UNKNOWN_n, invocation
counter records the two dispatches. +5 pass vs Phase C baseline, no
regressions across 11 conformance suites.
Registers extension-opcode-id from sx_vm_extensions.ml module init.
Lives downstream of both sx_primitives and sx_vm to avoid a build
cycle. Accepts a string or symbol; returns Integer id when the opcode
is registered, Nil otherwise.
Compilers (lib/compiler.sx) call this to emit extension opcodes by
name. Returning Nil rather than failing on unknown names lets a port's
optimization opt in per-build — missing extensions degrade to slower
correct execution.
Tests: 5 new foundation cases — registered lookup, unknown → nil,
symbol arg, zero-arg + integer-arg rejection. +5 pass vs Phase B
baseline, no regressions across 11 conformance suites.
sx_vm_extension.ml: handler type, extensible extension_state variant,
EXTENSION first-class module signature.
sx_vm_extensions.ml: register / dispatch / id_of_name /
state_of_extension. install_dispatch () runs at module init,
swapping Phase A's stub for the real registry. Rejects out-of-range
opcode IDs (must be 200-247), duplicate IDs, duplicate names, and
duplicate extension names.
Tests: 9 new foundation cases — lookup hits/misses, end-to-end VM
dispatch including opcode composition, all four rejection paths.
+9 pass vs Phase A baseline, no regressions across 11 conformance
suites.
Adds Invalid_opcode of int exception and extension_dispatch_ref forward
ref (default raises Invalid_opcode op), plus the |op when op >= 200 arm
before the catch-all in the bytecode dispatch loop. Partition comment
documents 1-199 core / 200-247 extensions / 248-255 reserved.
Phase B will install the real registry's dispatch into the ref at module
init, replacing this stub.
Tests: 4 new foundation cases (Invalid_opcode for 200/224/247, Eval_error
for 199 to pin the threshold). +4 pass vs baseline, no regressions.
Adds Sx_vm.bytecode_uses_extension_opcodes — an operand-aware
bytecode scanner that walks past CONST u16, CALL_PRIM u16+u8, and
CLOSURE u16+dynamic upvalue descriptors so operand bytes that happen
to be ≥200 don't false-positive as extension opcodes.
jit_compile_lambda calls the scanner on the inner closure's bytecode.
On hit it returns None — the lambda then runs through CEK
interpretation. The VM's dispatch fallthrough still routes the
extension opcodes themselves through the registry; this change just
prevents the JIT from claiming code it has no plan for.
Tests: 7 new foundation cases — pure core eligible, head/middle/
post-CLOSURE detection, CONST + CALL_PRIM + CLOSURE-descriptor false-
positive avoidance. +7 pass vs Phase D baseline, no regressions
across 11 conformance suites.
Loop complete: acceptance criteria 1-4 met. Hand-off to the Erlang
loop — lib/erlang/vm/dispatcher.sx's Phase 9b stub can now be
replaced with a real hosts/ocaml/lib/extensions/erlang.ml consumer.
lib/extensions/ becomes the new home for VM extensions, wired in via
(include_subdirs unqualified). README documents the registration
pattern, opcode-ID range conventions (200-209 guest_vm, 210-219
inline test, 220-229 test_ext, 230-247 ports), and naming rules.
extensions/test_ext.ml is the canonical worked example — two
operand-less opcodes (220 push 42, 221 double TOS) carrying a per-
extension state slot (TestExtState invocation counter). Test_ext.register
called from run_tests.ml at the start of the Phase D suite, on top of
the inline test_reg from earlier suites (disjoint opcode IDs).
Sx_vm.opcode_name now consults extension_opcode_name_ref (forward ref
in the same style as extension_dispatch_ref), so disassemble shows
extension opcodes by name instead of UNKNOWN_n. Registry maintains
name_of_id_table and installs the lookup at module init.
Tests: 5 new foundation cases — primitive resolves test_ext name,
end-to-end bytecode (push + double + return → 84), disassemble shows
"test_ext.OP_TEST_PUSH_42" / "test_ext.OP_TEST_DOUBLE_TOS",
unregistered ext opcodes still fall back to UNKNOWN_n, invocation
counter records the two dispatches. +5 pass vs Phase C baseline, no
regressions across 11 conformance suites.
Registers extension-opcode-id from sx_vm_extensions.ml module init.
Lives downstream of both sx_primitives and sx_vm to avoid a build
cycle. Accepts a string or symbol; returns Integer id when the opcode
is registered, Nil otherwise.
Compilers (lib/compiler.sx) call this to emit extension opcodes by
name. Returning Nil rather than failing on unknown names lets a port's
optimization opt in per-build — missing extensions degrade to slower
correct execution.
Tests: 5 new foundation cases — registered lookup, unknown → nil,
symbol arg, zero-arg + integer-arg rejection. +5 pass vs Phase B
baseline, no regressions across 11 conformance suites.
sx_vm_extension.ml: handler type, extensible extension_state variant,
EXTENSION first-class module signature.
sx_vm_extensions.ml: register / dispatch / id_of_name /
state_of_extension. install_dispatch () runs at module init,
swapping Phase A's stub for the real registry. Rejects out-of-range
opcode IDs (must be 200-247), duplicate IDs, duplicate names, and
duplicate extension names.
Tests: 9 new foundation cases — lookup hits/misses, end-to-end VM
dispatch including opcode composition, all four rejection paths.
+9 pass vs Phase A baseline, no regressions across 11 conformance
suites.
Adds Invalid_opcode of int exception and extension_dispatch_ref forward
ref (default raises Invalid_opcode op), plus the |op when op >= 200 arm
before the catch-all in the bytecode dispatch loop. Partition comment
documents 1-199 core / 200-247 extensions / 248-255 reserved.
Phase B will install the real registry's dispatch into the ref at module
init, replacing this stub.
Tests: 4 new foundation cases (Invalid_opcode for 200/224/247, Eval_error
for 199 to pin the threshold). +4 pass vs baseline, no regressions.
Two related kernel bugs were causing the HS conformance test
"arrays containing objects work" to fail with the misleading message
"Expected ({:a 1} {:b 2}) but got ({:a 1} {:b 2})".
1. sx_primitives.ml safe_eq: Dict/Dict only returned true for DOM-wrapped
dicts (those carrying __host_handle); all other dict pairs returned
false unconditionally. Plain dict literals can never have been =
to each other. Add the structural-equality fallback: when neither
side has a host handle, compare lengths and walk keys.
2. sx_browser.ml deep_equal (the kernel binding for equal?): had a
Number/Number branch but no Integer/Integer or cross-Integer/Number
branches, so since the numeric tower change Integer 1 vs Integer 1
was falling through to the catch-all and returning false. Mirror the
cases from run_tests.ml deep_equal which already had them.
Verified via direct kernel probe:
(= {:a 1} {:a 1}) => true (was false)
(= {:a 1 :b 2} {:b 2 :a 1}) => true (was false)
(equal? 1 1) => true (was false)
(equal? {:a 1} {:a 1}) => true (was false)
(equal? (list {:a 1}) (list {:a 1})) => true (was false)
HS suite arrayLiteral: 7/8 → 8/8.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
sx_types.ml:
- Add l_uid field on lambda (unique identity for cache tracking)
- Add lambda_uid_counter + next_lambda_uid () minted on construction
- Add jit_budget (default 5000) and jit_evicted_count counter
- Add jit_cache_queue : (int * value) Queue.t — FIFO of compiled lambdas
- jit_cache_size () helper for stats
sx_vm.ml:
- On successful JIT compile, push (uid, Lambda l) onto jit_cache_queue
- While queue length exceeds jit_budget, pop head (oldest entry) and
clear that lambda's l_compiled slot — evicted entries fall through
to cek_call_or_suspend on next call (correct, just slower)
- Guard JIT trigger by !jit_budget > 0 (budget=0 disables JIT entirely)
sx_primitives.ml:
Phase 2:
- jit-set-budget! N — change cache budget at runtime
- jit-stats includes budget, cache-size, evicted
Phase 3:
- jit-reset-cache! — clear all compiled VmClosures (hot paths re-JIT
on next threshold crossing)
- jit-reset-counters! also resets evicted counter
run_tests.ml:
- Update test-fixture lambda construction to include l_uid
Effect: cache size bounded regardless of input pattern. The HS test harness
compiles ~3000 distinct one-shot lambdas, but tiered compilation (Phase 1)
keeps most below threshold so they never enter the cache. Steady-state count
stays in single digits for typical workloads. When a misbehaving caller
saturates the cache (eval-hs in a tight loop, REPL-style host), LRU
eviction caps memory at jit_budget compiled closures × ~1KB each.
Verification: 4771 passed, 1111 failed in run_tests — identical to
pre-Phase-2 baseline. No regressions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
OCaml kernel changes:
sx_types.ml:
- Add l_call_count : int field to lambda type — counts how many times
a named lambda has been invoked through the VM dispatch path.
- Add module-level refs jit_threshold (default 4), jit_compiled_count,
jit_skipped_count, jit_threshold_skipped_count for stats.
Refs live here (not sx_vm) so sx_primitives can read them without
creating a sx_primitives → sx_vm dependency cycle.
sx_vm.ml:
- In the Lambda case of cek_call_or_suspend, before triggering the JIT,
increment l.l_call_count. Only call jit_compile_ref if count >= the
runtime-tunable threshold. Below threshold, fall through to the
existing cek_call_or_suspend path (interpreter-style).
sx_primitives.ml:
- Register jit-stats — returns dict {threshold, compiled, compile-failed,
below-threshold}.
- Register jit-set-threshold! N — change threshold at runtime.
- Register jit-reset-counters! — zero the stats counters.
bin/run_tests.ml:
- Add l_call_count = 0 to the test-fixture lambda construction.
Effect: lambdas only get JIT-compiled after the 4th invocation. One-shot
lambdas (test harness wrappers, eval-hs throwaways, REPL inputs) never enter
the JIT cache, eliminating the cumulative slowdown that the batched runner
currently works around. Hot paths (component renders, event handlers) cross
the threshold within a handful of calls and get the full JIT speed.
Phase 2 (LRU eviction) and Phase 3 (jit-reset! / jit-clear-cold!) follow.
Verified: 4771 passed, 1111 failed in OCaml run_tests.exe — identical to
baseline before this change. No regressions; tiered logic is correct.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
