Skip to content

performance_audit.md

Latest commit

 

History

History
86 lines (63 loc) · 7.15 KB

performance_audit.md

File metadata and controls

86 lines (63 loc) · 7.15 KB
Raw

Script + Variant perf audit

Audit date: 2026-09-06.

Baseline commit: 0d384c0b; host: Windows x64; Rust: 1.97.1.

Path Cost + result Action
with_state! / with_state_mut! Dense slot + full ID check + cached TypeId; no codec or heap alloc Kp closure borrow; fx active-slot miss aft peer swap-remove
with_node! Arena lookup + exact type dispatch; no node clone Kp typed read path
with_node_mut! Transform/base snapshots + UI payload hash + dirty hooks Kp hooks; group related writes in one closure
params! Borrow stack arr of owned Variant vals Kp macro; arg-list heap alloc = 0; text/container vals may still alloc
variant! Variant::from(value); move owned vals Move unique Arc<T> / Rc<T> payload; clone only w/ shared strong owners
Variant Inline scalar/ID; shared text/bytes; owned arr/object; box large engine structs Kp layout + <=32-byte stride guard; avoid deep container clone on hot paths
#[derive(Variant)] / DeriveVariant Default array struct + u16 enum tag; owned codec + static object keys Kp wire forms; rm scratch heap alloc for small fixed-arr dcod
get_var! Const-ID match for known fields/nested paths; owned result Skip leaf fields in dynamic fallback; add safe erased-state type guard
set_var! Const-ID match + typed dcod; object patch fallback Use owned dcod for non-object vals; kp object patch behavior
call_method! Instance lookup + Rc handle + callback frame + trait call + const-ID match Kp dispatch; reuse callback context; avoid per-call text-name hash via method! / func!
Gen member tables Rust match on const hashes; no runtime map build Kp; no evidence for map replacement
Gen script registry Static constructor slice + binary search; dynamic constructor map Kp; attach/load path, separate from member dispatch
Nested dynamic fallback Encode root + walk fields + hash full paths Reuse one path buffer; skip scalar/empty roots; kp full-path hash format
JSON / scene codec Tree conversion + resource resolver work Kp at data boundaries; avoid in per-frame typed state access

Fix active-state access after peer removal: old dense slot may point elsewhere or fall out of range. Retry full generational ID only on slot mismatch. Kp exact state-type check + stale-ID rejection. Kp scheduler snapshot validation separate.

Fix gen get_var / set_var / scene-state casts: public safe trait accepts dyn Any; old unchecked cast permits wrong-type refs. Add checked downcast -> panic on mismatch. Cost: one Any type check at erased-state boundary; no extra check in cached with_state* path.

Kp strict numeric arg types, missing-arg defaults, unknown-member result, object patch fallback, scene resource resolution, visibility rules + node dirty hooks.

Measure via variant_hotpaths Criterion bench: 1 s warmup, 2 s total measurement, 30 samples; same host + bench profile (opt-level = 3). Use 3 s measurement for final large-arr comparison. Isolate encode setup via iter_batched; no whole-frame/FPS claim. Report time estimates below; round relative gains.

Case Before After Time cut
Fixed [i32; 3] dcod 59.0 ns 18.35 ns ~69%
Fixed [i32; 16] dcod 76.9 ns 61.76 ns ~20%
Fixed [i32; 128] dcod 330 ns 335.5 ns No gain claim; see reference below
Unique Arc<Variant> encode, 128 vals 1.112 us 70.8 ns ~94%
Unique Rc<Variant> encode, 128 vals 1.134 us 57.5 ns ~95%
Shared Arc<Variant> encode, 128 vals 1.023 us 1.027 us No significant change
Nested set, last of 64 fields 12.71 us 2.021 us ~84%
params![i32, f32, bool] 7.90 ns 7.86 ns No significant change

Reject full stack-array dcod: 128-item probe regresses ~22%. Cap stack path at 16 vals + 4 KiB of Option<T> scratch; kp heap path above cap. Treat cap as measured policy for this host, not universal optimum.

Restore original large-arr body after shared-helper regression. Recheck original body beside final code: 337.0 ns reference vs 335.5 ns final; overlapping confidence intervals. Kp reference case in bench for future host-drift checks.

Kp claims scoped to measured cases. No end-to-end call_method, frame-time or node-speed claim. Gen var type guard adds one check; net get/set dispatch cost remains unmeasured.

Run baseline before code edits:

cargo bench -p perro_scripting --bench variant_hotpaths -- --warm-up-time 1 --measurement-time 2 --sample-size 30 --save-baseline audit_before

Run comparison after code edits:

cargo bench -p perro_scripting --bench variant_hotpaths -- 'variant_hotpaths/(array_decode_[0-9]+$|unique_|shared_|nested_|params_)' --warm-up-time 1 --measurement-time 2 --sample-size 30 --baseline audit_before

Run final large-arr + old-body reference together:

cargo bench -p perro_scripting --bench variant_hotpaths -- array_decode_128 --warm-up-time 1 --measurement-time 3 --sample-size 30

Prioritize next gains only w/ workload data:

  1. Add typed nested-field access metadata for imported/dynamic schemas -> skip full-root encode + walk. Kp unknown-key + patch semantics.
  2. Add owned scene dcod contract -> into_parse_scene still calls borrowed from_scene_variant, despite owned input.
  3. Add borrowed typed method args if container-heavy dispatch dominates -> current &[Variant] requires clone for owned Variant / String params. Use raw &[Variant] or &str methods where suitable today.
  4. Split type validation from is_type::<T>() dcod -> current probe may alloc full result + drop it. Use kind / exact accessors for cheap kind checks.
  5. Bench heapify for BinaryHeap dcod + static keys in hand-written engine codecs -> current push-per-item heap build + repeated key alloc.

Kp API/wire layout in this patch. Defer shared-array/object or borrowed-Variant redesign: public enum ownership + mutation + serialization contracts need separate design + workload proof.

Verify: 946 unit tests + 24 integration tests pass; include borrow compile-pass/fail fixtures + derive roundtrips. Pass generated all-type compile + 4 generated-glue runtime tests. Pass Clippy for Variant, scripting + compiler. Pass diff whitespace check.

Trace sources: