Skip to content

log.md

Latest commit

 

History

History
103 lines (73 loc) · 5.74 KB

log.md

File metadata and controls

103 lines (73 loc) · 5.74 KB
Raw

Wiki Log

Append-only chronological record of wiki operations. Every ingest, refresh, query, and lint gets an entry. Entries start with ## [YYYY-MM-DD] <op> | <title> so grep '^## \[' wiki/log.md | tail -20 yields recent activity.

Operations:

  • bootstrap — initial scaffolding.
  • ingest — a new source landed, wiki updated.
  • refresh — code drifted, wiki brought in sync.
  • query — a user question was filed back into the wiki.
  • lint — health-check findings.

[2026-08-11] refresh | AOT multi-block resuming — Phase 2 (continuation live-ins) landed

Implemented cross-block continuation live-in threading: a resuming perform's continuation reads values produced in an earlier block (e.g. a mutable accumulator set by one perform, read after a later perform) by threading them through the handler's Resume dispatch. continuation_live_ins (backward liveness over normal successors), resuming_threading (per-site extras + uniform width), threaded slots allocated even for single-site bodies (fixing a loop-carried regression in the single-continuation Resume path). Commit c51df8e. The AOT backend is now functionally complete for non-distributed programs.

Touched: [[queries/aot-multi-block-resuming]], src/aot/codegen.rs

[2026-08-11] refresh | AOT multi-block resuming — general SSA phi fix

While scoping Phase 2 (cross-block prior-result reads) found the representative pattern hits a GENERAL AOT bug: a var assigned in one branch and read after a merge failed with no effects involved. Fixed compute_liveins to phi values live into a merge that are defined in a predecessor, plus fixed the gen/kill terminator liveness (branch-cond wrongly live into a loop back-edge). Commit 8d0c286. Narrowed Phase 2 to continuation live-in threading (still open, rare pattern).

Touched: [[queries/aot-multi-block-resuming]], src/aot/codegen.rs

[2026-08-11] refresh | AOT multi-block resuming handler — Phase 1 landed

Implemented Phase 1 of the AOT multi-block resuming handler design: a resuming effect handler performed from different MIR blocks now compiles (was rejected). Uniform per-body threaded-slot width + compute_liveins exclusion of handler-body blocks + a gen/kill liveness guard that rejects genuine cross-block prior-result reads. Commit 758ce4d. Tests: exclusive-branch if/else, discarded-first-result, cross-read rejection.

Touched: [[queries/aot-multi-block-resuming]], src/aot/codegen.rs

[2026-08-11] query | AOT multi-block resuming handler — scoping

Scoped the last reachable correctness gap in the AOT native backend: a resuming effect handler (| E.op(x) resume => ...) performed from different MIR blocks is rejected ("multi-perform resuming handler with perform sites across multiple blocks is not yet supported"). Confirmed the failing pattern (if cond then perform E.run(1) else perform E.run(2)) on --backend native.

Root-caused the same-block model: cont_thread is per-block while handler_threaded_dsts is per-body (global), so cross-block perform sites make the jump-arg count ≠ handler param count. Proposed a phased design: Phase 1 = uniform threaded-slot width (max per-block, dummy-padded) covering exclusive-branch patterns without cross-block reads; Phase 2 = thread prior perform results through compute_liveins for loop-carried/accumulated results. After this, the only remaining AOT rejects are distribution (remote spawn/ask) and dead defensive arms.

Touched: [[queries/aot-multi-block-resuming]]

Implemented the "Immediate" tier of [[queries/performance-assessment]] and recorded an ## Implementation status section on that page.

Touched: [[queries/performance-assessment]] Key updates: Added #[repr(align(64))] to Scheduler/SchedulerStatsInternal/Mailbox; opt-in core pinning (NULANG_PIN_CORES, Linux sched_setaffinity) wired into the shard spawn; swapped the JIT's per-instruction hot-path maps to rustc_hash::{FxHashMap,FxHashSet}; added VM::new_without_jit() + an interp criterion bench group. Full token-threading deferred (risky core rewrite; hot-counter hashing captured the largest safe slice). Verified: 1650 lib tests pass, interp bench runs, gate default/no-default-features warning-clean.

[2026-08-09] query | Performance proposal assessment

Filed the assessment of PERFORMANCE_ANALYSIS.md's 28 proposals (plus beyond-catalog techniques: threaded dispatch, JIT OSR, NUMA awareness, cache-line padding, non-temporal stores, Auto-SoA) against the current tree.

Touched: [[queries/performance-assessment]] Key updates: Ground-truth verified all claims via read/grep — 14/28 proposals shipped, 4 partial, 7 deferred; the 4 legacy explorer performance findings (Box'd frames, .leak() strings, RGA/MVRegister Vec churn) are all fixed in the current tree. Ranked 5 actionable gaps: interpreter dispatch throughput, rkyv wire serialization, cache-line/NUMA locality, DST harness, JIT OSR.

[2026-08-02] bootstrap | wiki scaffolding

Established the wiki structure per the [[wiki-updater skill|../.claude/skills/wiki-updater/SKILL.md]], instantiating Andrej Karpathy's llm-wiki.md pattern for the Nulang repository.

Created:

  • [[index]] — content catalog.
  • [[log]] — this file.
  • [[overview/architecture-overview]] — top-level architecture map, sourced from AGENTS.md.
  • [[overview/compiler-pipeline]] — compiler stages and backends, sourced from AGENTS.md and src/.

Not seeded (per skill's bootstrap rule: "wiki grows by ingest, not by front-loading"):

  • Subsystem pages (wiki/subsystems/) — created on first ingest per subsystem.
  • Concept pages (wiki/concepts/) — created on first ingest per concept.

Next: user directs which subsystem or concept to expand first via the wiki-updater skill.