ekp/.phrase/phases/phase-kp-overhaul-20260726/issue_narrow_live_append_latency_20260729.md
Kinneyzhang 665164911f perf: make live structural commits incremental
Reuse prepared paragraph data, safe Elisp DP states, common layout lines, and dirty-island source reconstruction for exact append commits. Simplify C int32 validation and add the frozen task030 evaluator, regressions, and decision records.
2026-07-30 08:24:54 +08:00

4.9 KiB
Raw Blame History

Issue: Narrow Live-Append Latency 2026-07-29

issue018 [ ] Unique live appends miss the frame budget at very narrow widths

  • Status: Exact incremental implementation verified; the locked source-instrumented gate remains open.
  • Summary: Stable transactions removed per-key whole-hard-line planning, and task030 now incrementally extends paragraph preparation, Elisp DP state, and layout lines only at structural row crossings. Byte-compiled production paths are within the 16 ms frame budget; the deliberately source-loaded, fully instrumented evaluator still exceeds it.
  • Environment: Emacs 30.2 on macOS, C backend 1.6, benchmark width fixed at 80 px, gc-cons-threshold bound to most-positive-fixnum so mutator work is measured without GC pauses.
  • Repro:
    1. For the production-shaped check, byte-compile the four production Elisp files into a temporary package root, put that root first on load-path, and run tests/ekp-buffer-live-bench.el once with the C module loaded and once with ekp-use-c-module nil.
    2. Confirm ekp--dp-run-1d is byte code and inspect the append and hard-boundary rows.
    3. For the locked source/instrumentation matrix, run tests/run-live-commit-evaluator.sh and inspect .omx/goals/performance/narrow-live-commit/latest-report.json.
  • Expected vs Actual:
    • Expected: live append p99 remains below the 16 ms interaction budget, including narrow windows.
    • Actual production path: three repeated byte-compiled public-command runs measure append p99 at 1.1581.326 ms for C and 1.4291.438 ms for pure Elisp. Hard-boundary p99 is 1.2511.363 ms and 1.4571.470 ms.
    • Actual locked evaluator: four interleaved source-instrumented rounds measure C at 25.490/25.785 ms p95/p99 and Elisp at 43.860/47.578 ms. These improve 77.78/78.37% and 92.54/92.03% over the frozen baseline but still miss the absolute 16 ms stress target.
  • Investigation:
    • The task029 point-motion change is not the cause. Navigation is now a zero-work path and does not enter planning or projection publication.
    • The old workload produced 291 distinct hard-line strings and 291 plan calls. task031 removes that obsolete control flow: only 15 actual visual-row crossings invoke ekp-layout-plan.
    • The remaining high percentiles coincide with those allowed structural commits and must be profiled as commit latency, not average key latency.
    • Wider measurements are materially faster; the pathological result is width-sensitive rather than a general point-motion regression.
    • The previously recorded 6.399 ms task028 append p99 is not reproducible with the current checked-in benchmark and must not remain the current performance claim.
    • Both backends still miss p99 at structural boundaries, so task030 must evaluate both rather than optimize only the C wrapper.
    • The frozen C wrapper measured only 2.615/2.655 ms p95/p99 and the candidate 0.697/0.701 ms. Paragraph preparation, DP state reconstruction, plan construction, and publication owned the end-to-end cost.
    • Byte compilation removes most of the remaining pure-Elisp interpreter and closure overhead. This explains why the production path passes while the frozen source-instrumented stress gate remains red; neither result is substituted for the other.
    • Rust is not selected. It would use the same emacs_env ABI and cannot remove Elisp-owned font measurement, transaction, or projection work.
  • Required Outcome:
    • Diagnose and reduce unique-state narrow-width append planning cost without changing KP output semantics, the core DP contract, C ABI, or source-clean display model.
    • Keep cache identity exact; do not introduce approximate/stale plan reuse, timers, skipped edits, global GC changes, or point-based invalidation.
    • Preserve the task029 invariant that point-only motion performs zero plan, cache, and text-property work.
  • Fix: task030 retains exact prepared paragraph/context data, extends property-free append tails from the last complete-word boundary, resumes pure-Elisp DP from a safe reachable state, reuses common layout lines, and reconstructs only the live dirty source island. Unsupported contexts take the unchanged full path. C int32 validation now performs one extraction.
  • Verification: Exact baseline/C/Elisp hashes and append-chain equivalence pass across 64/80/96/128/160 px, 2/4/8/16-row fixtures, default/excluded GC, unsafe fallbacks, and randomized chains. Normal and random-order ERT pass 199/199; 300 fuzz cases, warning-clean production/C compilation, 9/9 C tests, release checks, and reviewed temporal GUI evidence pass.
  • User Confirmation: Pending; the formal source-instrumented performance target also remains open.
  • Resolved At: Unresolved.
  • Resolved By: Pending.
  • Commit: Pending.