A devirtualized protocol call resolved its impl with devirt-resolve on EVERY call
— but the tag/proto/method are compile-time constants, so the resolved fn is a
runtime constant (closed world). That per-call find-protocol-method (three
hashtable lookups) was the cost: on mono-dispatch, dispatch was ~75% of the time
(ablation: same arithmetic direct-call 166ms vs dispatch 673ms).
Resolve once. When emitting a direct-link def, each devirt site gets a fresh cache
cell, bound to #f in a let wrapping the def (so it persists across calls and is
shared by every invocation); the site resolves into it on first use ((or cell
(let ((_f (devirt-resolve ..))) (set! cell _f) _f))) and reuses it after — the
inline cache the JVM gets for free. First call still passes the real receiver, so
the Object/host-tag fallback (devirt-resolve) is unchanged.
mono-dispatch 673ms -> 214ms (~3.15x), 47.5x -> ~15x JVM, near the 166ms
direct-call floor. run-devirt.ss gains the cached-path checks (cell present, 1st
call caches + 2nd reuses, both == dispatch). make test / shakesmoke green, selfhost
holds, 0 new divergences.
Co-authored-by: Yogthos <yogthos@gmail.com>
The devirtualized protocol call emitted find-protocol-method on the inferred
record tag, but a record can satisfy a protocol via an Object/host-tag default
rather than a direct impl — find-protocol-method on its own tag misses that,
while protocol-resolve walks to the default. So a record relying on
(extend-protocol P Object ...) resolved under ordinary dispatch but applied #f
under devirt and crashed. Closed-world opt builds only; the gate previously
covered just direct inline/extend-type impls so it shipped green.
Emit devirt-resolve, which tries the static tag and falls back to
protocol-resolve on a miss — same fast path, correct regardless of how the
record satisfies the protocol. Mirrors jrec-field-at falling back to jolt-get.
The receiver binds to one temp so it feeds the resolve and the application
without double-evaluating a side-effecting arg 0.
Also widen the whole-program fixpoint to :any on hitting the iteration cap: a
non-converged pre-fixpoint is more specific than the least fixpoint, so seeding
it would be unsound. Not reached in practice (~2 passes); a defensive floor.
run-devirt.ss gains an Object-default case. make test / shakesmoke green,
selfhost holds, 0 new divergences.
Co-authored-by: Yogthos <yogthos@gmail.com>
When the inference proves a protocol call's receiver is one record type, the
back end resolves the impl by that static tag (find-protocol-method) instead of
routing through the protocol var -> jolt-invoke -> protocol-resolve, which
re-derives the tag and walks the type table. Same table lookup, minus the
var-deref, the rest-cons, and the receiver-type computation.
Fires only on a monomorphic site: a megamorphic receiver joins to :any and
carries no :devirt-type, so it keeps ordinary dispatch (the dispatch bench is
unaffected). The annotation comes from the whole-program fixpoint typing a
reduce/HOF element or a ctor return as a specific record.
Modest on the dispatch benchmarks (~6% on mono-dispatch) — float boxing in the
reduce accumulator dominates there, a separate numeric lever — but it removes
the dispatch overhead wherever a typed receiver is known.
run-devirt.ss gate: emitted form uses find-protocol-method, and evaluating it
matches ordinary dispatch for an inline impl, an extend-type impl, and the
non-devirt path. make test / shakesmoke green, 0 new divergences.
Co-authored-by: Yogthos <yogthos@gmail.com>
