[Bug]: research: MessagePipe parity

[wallstop]

Description

Our Global -> 1 subscriber is ~2x as slow as MessagePipe, Subscribe/Unsubscribe churn is ~2x as slow, and our struct message with no boxing is ~3x as slow.

Why is there such a performance delta? Can we catch up while maintaining our public contract?

Steps to Reproduce

Run performance comparisons

Expected Behavior

DxMessaging is roughly equivalent to MessagePipe

Actual Behavior

DxMessaging is lacking when compared to MessagePipe

Unity Version

6000.x

Package Version

3.2.2

Platform

• Editor
• Windows
• macOS
• Linux
• iOS
• Android
• WebGL
• Other

Additional Context

No response

[wallstop]

Session 215 / green PR #419 added the attribution rows and produced the first Standalone IL2CPP x64 Release evidence in Performance run 31857267407 (commit bf2469c; 169/169 internal tests passed):

• UntargetedFlood_OneHandler (enabled token / AugmentedScalarFast): 30.770M emits/sec
• UntargetedFlood_OneDirectHandler (same payload, bus, priority, by-ref fast slot, and fan-out): 33.916M emits/sec
• RegistrationAttribution_DirectBus_131072: 55.907 ms
• RegistrationAttribution_DirectHandler_131072: 95.451 ms
• RegistrationAttribution_TokenStage_131072: 11.260 ms (orthogonal disabled-token stage/remove row)
• RegistrationAttribution_TokenActive_131072: 119.032 ms

For registration, the cumulative handler layer adds 39.544 ms over direct bus, while active-token work above the handler layer adds 23.581 ms. For dispatch, the direct twin is about 3.0 ns/emit faster, corresponding to a 9.3% throughput difference from the token-augmented path in this diagnostic run. This narrows the next bounded investigations to (1) MessageHandler registration work and (2) the augmented callback / mutable diagnostics branch, as separate candidates.

These are attribution measurements, not an optimization verdict. Any production change still needs fresh bracketed Standalone IL2CPP Release A/B/A controls and must preserve mutable diagnostics semantics. PR: #419

[wallstop]

Session 216 tested the two costs isolated in #419 and rejected both performance candidates.

• Passing the already-resolved typed slot into MessageHandler registration preserved behavior, but the normalized direct-handler/direct-bus ratio was unchanged: 1.620 control versus 1.619 candidate in fresh Mono assemblies.
• Disabling IL2CPP null checks on all 16 token-owned augmented callbacks produced 32.281M emits/sec. That was only +0.43% over the stronger same-code control, below the 3% keep threshold. The two controls also reversed direct/token ordering (0.907 versus 1.082), so absolute movement was not stable enough to support a performance claim.
• Both optimization candidates were removed.

The safety review found an independent correctness bug: an active token with a null callback could mutate the bus before the handler cache threw, leaving a partial registration until reset. Draft PR #420 now contains only the atomic preflight fix and leak-aware coverage: #420

Before another #414 implementation, the direct/token twin needs stronger resistance to host drift or generated IL2CPP code must identify a concrete removable operation. The next bounded lifecycle experiments are typed deregistration helpers that avoid erased cache calls, then an exact-MessageBus specialization that leaves custom buses on their existing interface path.

[wallstop]

Session 217 / green PR #421 tested and rejected the typed-cache deregistration specialization, while retaining stronger evidence capture for future #414 work.

Fresh Standalone IL2CPP x64 Release runs on Unity 6000.5.2f1 used three matching controls and three byte-identical candidates:

Run	Direct bus	Direct handler	Handler excess
A1 control	7.429 ms	28.230 ms	20.801 ms
B1 candidate	7.425 ms	27.055 ms	19.630 ms
A2 control	7.028 ms	26.876 ms	19.848 ms
B2 candidate	7.065 ms	26.534 ms	19.469 ms
B3 identical candidate	7.134 ms	28.235 ms	21.101 ms
A3 final control	7.054 ms	27.829 ms	20.775 ms

B1/B2 initially improved handler excess by 3.81% against A1/A2. The documentation-only B3 run reproduced the candidate player exactly but reversed the result. Across all three controls and candidates, the mean improvement was only 1.99%, below the 3% keep threshold. The runtime specialization and changelog claim were removed.

PR #421 retains:

• a four-window token/direct dispatch palindrome diagnostic;
• shippable player-size, GameAssembly, and metadata sizes/hashes;
• bounded generated-IL2CPP-C++ capture.

The diagnostic also exposed same-code instability: ratio spread reached 8.403% and 10.395% in two controls, and one identical candidate's token rate drifted more than 5%. No dispatch claim was made.

The next exact-MessageBus deregistration hypothesis must first prove from baseline generated C++ that the interface call remains indirect, add durable paired diagnostics with explicit invalidation thresholds, and require stable repeated controls/candidates plus material end-to-end MessagePipe subscribe/unsubscribe improvement. PR: #421

[wallstop]

Session 218 prerequisite evidence is complete in PR #422.

Final exact-head Performance Numbers run 31900934703 passed at 8678a2ca. Artifact perf-6000.5.2f1-standalone-internal (ID 9251194997) captures the exact call chain and reports:

• typedInterfaceDispatchEvidenceLineCount=1
• typedDirectBusDispatchEvidenceLineCount=0
• typedBusDispatchRecognized=true
• one exact wrapper body repeated byte-identically at three IL2CPP locations
• one shared body whose stored bus call is GenericInterfaceActionInvoker1 through RGCTX

That proves the exact-MessageBus specialization has a real interface-dispatch mechanism. It does not yet justify the runtime candidate. The final handler/bus/bus/handler palindrome measured 10.38% handler drift, 5.37% bus drift, and 14.66% handler-excess spread; all three inclusive 3% gates failed and it emitted interpretable=false. Earlier exact-head runs were also unstable, with handler-excess spreads of 51.48%, 3.61%, and 28.82%.

Verdict: retain the diagnostic and exact-codegen harness; defer the production specialization until repeated fresh Standalone IL2CPP Release palindromes are interpretable, then use an independent control/candidate bracket. Static, devcontainer, performance, and all four Unity-version all-mode legs are green. The final review finding is fixed and resolved, and the adversarial re-audit found zero material or minor issues.