Repository navigation
perf: method_missing dispatch 2.6x slower than CRuby (call-site IC) #391
Description
Part of #378. method_missing dispatch is ~2.6× slower than CRuby. Measured on lawrences-mac-studio (M2 Max) with the #387 branch, the standard perf build (stdlib,jit-native,_fiber,_json_native,mimalloc), and CRuby 3.4.1. Numbers are ns/op, best of 7, with the empty loop subtracted.
| shape | rubyrs | CRuby |
|---|---|---|
o.nope → method_missing(name, *args) |
357.6 | 135.0 |
same body called directly (o.direct(:nope)) |
~132 | ~83 |
body only (s.end_with?("=") ? nil : s) |
69 | 69 |
Decomposition: the body is at parity. The base call costs ~63 vs ~14 ns (tracked separately). The miss → method_missing resolution costs ~187 vs ~17 ns, which is 76% of the gap.
The profile of the miss path shows three costs:
- ~23% in the slow
do_callcascade of&*name == "..."arms, walked in full beforetry_method_missingat the tail; - ~24% in the generic
invoke_method_with_block; intern("method_missing")on every call (2.6%), an uncachedmethod_missinglookup (1%), and the argsVecbuilt twice.
Proposal
A. Call-site method_missing inline cache.
- After the first miss at a site, cache "(class, method_gen) → user
method_missing" in the site'sCallCache. - Later calls skip the cascade and invoke the method via the stack-args path (
try_invoke_nfa_method_from_stack), prepending the name symbol. - Invalidation uses
method_gen, as for regular entries. - Expected cost is ~170 ns/op.
B. Small fixes, independent of A:
- pre-intern
sym_method_missingonVm; - use the cached lookup for
method_missing; - build the forwarded args once.
Acceptance
o.nopewith a usermethod_missingat ≤ CRuby (≤135 ns/op on the studio).- The diff suite passes, including the JIT runs; add a fixture for
method_missingredefined or removed after a site warmed up.
Activity
linyiru commentedon Sep 26, 2026
Progress: #396 is merged. It implements proposal A and most of B.
A. Call-site method_missing IC
do_callcascade records(receiver class, method_gen) → user method_missingin the site'sCallCache, under the keyclass_ptr | 2.(class, None)entry, so each class uses one way and a site still holdsIC_WAYSclasses.method_missingstraight from the stack.B.
sym_method_missingis pre-interned, and the forwarded args are built once. B's cached lookup was not needed: with A in place, the uncached lookup only runs on a site's first miss or at megamorphic sites.Studio, perf build, ns/op, best of 7 (the loop is inside a block, so absolute numbers are higher than the table above):
o.nope→method_missing(name, *args)The miss overhead is gone:
o.nopenow costs the same as a direct call. The acceptance target is not met yet. The remaining gap is the base call cost (~63 vs ~14 ns), which is tracked separately.Remaining work under this issue:
nopeinside an instance method) still take the slow path. They need the same cache intry_invoke_self_recv_cached.t2_lite_call_*) misses fall back to a full frame and re-dispatch. The result is correct, but the frameless serve is lost.linyiru commentedon Sep 27, 2026
Status: all three
method_missingIC pieces have landedmethod_missingIC for explicit receivers. The entry also serves as the class's negative lookup entry, so each class uses one way.o.nope490 → 283, vs 289 for a direct call to the same body. Polymorphic sites with 3–5 classes add ~40 over direct (CRuby ~48).nopeinside an instance method).method_missingwithout materializing (explicit, bare-self and fusedLoadLocalCallforms).o.nope157 → 125 (direct 125); bare 153 → 127 (direct 118)(#405 recalibrated the peak-RSS ratchet. It was flaking on runner noise while #402 was in CI. There are no code changes.)
Where this leaves the acceptance target: a
method_missingmiss now costs the same as calling the same body directly, in the interpreter and in tier 2. The ≤135 ns / ≤ CRuby target is still not met, because what remains is the base call cost. For example, barenopeis 147 ns against CRuby's 71, and a direct call is 149 against CRuby's 58. Nothing specific tomethod_missingis left.Not covered, by design:
method_missing(name, *args)still materializes on the tier-2 lite path. So does a rest-shaped direct method; that is a general lite-path limit.public_send/send, and force-primitive dispatch never fill the IC.Pre-existing gaps found along the way, not fixed here:
method_missingraisesNoMethodError; CRuby raisesNameError.NameError#namereturns nil."s" + 1raisesNoMethodError; CRuby raisesTypeError.caller_locations(...).labellacks theObject#prefix.I suggest either closing this issue with the base call cost tracked separately, or keeping it open until that work lands.
linyiru commentedon Sep 27, 2026
Closing. method_missing dispatch is now at parity with a direct call (#396, #399, #402). The rest of the ≤ CRuby target is base call cost, now tracked in #406. Pre-existing gaps found along the way: #407, #408, #409, #410.