ormastes/simple/compiler-performance
Claude Code and agent skill by ormastes/simple
- community
Checking that the SKILL.md file is still in the repository.
Install
npx vskill@latest install ormastes/simple/compiler-performanceSKILL.md
The same SKILL.md was published earlier as ormastes/simple/compiler-performance.
Compiler Performance Expert
Bounded archive production qualification (2026-10-12)
The genuine production controller and final MIR consumer must be tested through the actual compiled CLI/runner paths with complete current bodies, original warning/error order, logical alias specialization ownership and all real mono counters. Initial activation covers managed full-native streaming AOT with nonshard HIR cache; ambient bootstrap-flat semantics remain separately owned. Cold full-inventory, demand and projected routes keep their existing producers and require separate memory qualification. Source-only controls and a passing Rust support build cannot qualify this path.
Canonical attempt13 terminated before producer admission:1,297 seed files compiled and one MIR HIR-lowering failure, exit1/quiescent1, measured peak 1,251,560 KiB under the5,859,375 KiB enforced cap. This early seed build measurement does not prove the later full CLI/runner memory target. The older admitted compiler emitted, linked and ran Hello, while its full runner build exceeded the cap; both original receipts remain retained. Fresh exact-compiler admission, object/link/run Hello, complete finite tool behavior and process-tree memory receipts remain required after activation.
Scope
Own optimizer activation truth, performance/memory diagnostics, shared MIR facts, CollectionPlan decisions, symbolic summaries, and profile correlation.
Non-negotiable rules
- Inspect requested and effective pipelines separately.
- Never activate a dormant pass mechanically.
- Treat
Unknownalias, effect, escape, trip-count, and lifetime facts as rejection. - Use source warnings for likely application mistakes and remarks for compiler decisions.
- Profiles affect priority and profitability only, never transformation legality.
- Preserve order, errors, traps, numeric semantics, ownership, destruction timing, and ABI.
- Reuse
PerfFacts; do not add private CFG/loop/alias scans to a pass. - Require positive/negative sentinels, verification, idempotence, and differential evidence.
- Run only an admitted pure-Simple binary and retain its stage/hash/provenance.
Primary references
doc/01_research/local/simple_compiler_performance_memory_efficiency_audit.mddoc/04_architecture/simple_compiler_performance_memory_efficiency.mddoc/05_design/simple_compiler_performance_memory_efficiency.mddoc/07_guide/compiler/performance_diagnostics.mddoc/03_plan/sys_test/simple_compiler_performance_memory_efficiency.md
Review questions
- Is the pass operational status honest?
- What exact facts prove legality, and how are they invalidated?
- Is the opportunity a lint, transform, remark, deep analysis, or profile finding?
- Are uncertainty and rejection reasons serialized?
- What positive and adversarial witnesses prove behavior?
- What before/after compile-time, runtime, allocation, copy, and RSS evidence exists?
Phase 3 archive journal (2026-10-12)
driver_hir_archive_journal_v1 records real declaration publication and logical aliases. Its source/history encoding and verification run in owned scopes; only compact results escape. Native qualification uses the journal selector in scripts/bootstrap/tests/hir-body-typecheck-lease-native.shs and requires 13 checks plus the canonical managed tool/runtime handoff. Execution is pending. Journal integrity alone does not admit production order, transformed bodies or the final MIR input. No memory improvement is established by source review.
Immutable mono replay now captures the actual applied root/FIFO selection and reconstructs complete bodies before the shared exhaustive verifier. The verification fixture requires 18 checks. Borrowed post-HIR declaration/body readers return only scalar/layout/RT/join-point DTOs; postfacts requires 26 checks and the managed handoff. Both fixtures remain unrun. Their public pins and editable fact rows need genuine driver prerequisite ownership before final admission.
The reviewed MIR declaration/default stack retains complete expression artifacts, canonical callable/global ownership, nested Throws types and lifted closure/trait metadata. Its nine unit examples and 17 native fixture checks remain unrun; the real final repository and driver activation are still needed. Cold SCV source admission now has a separate owned source-capture-plan factory: it admits the pinned generation/snapshot and returns selected input rows without HIR. Four new ownership/alias/snapshot boundary examples are unrun. This source plan cannot authenticate the later transformed body.
Retained full-inventory SMF/witness emission now uses a per-module owned scope, sharing its actual implementation with borrowed archive capture. It retains all real emitted payload/file/witness outputs and original error categories; two additional compiled-spec examples remain unrun. Full HIR graph retention and the required cap qualification are still unresolved by this local change.
Actual streaming schedule capture binds physical surfaces, each independent authored alias, real visit order and distinct logical/layout orders. The schedule and schedule_shard selectors each require 13 native checks in separate runtime processes. Shard order is prospective and refuses full publication admission. Explicit aspect claim replay preserves ordered real registry inputs without installing another compilation's claims. These controls are unrun and still need the genuine transformed-stage prerequisite owner. The narrow MIR compatibility-default detach API removes only default/ambiguity roots and preserves return/type/link metadata; two frontend examples are unrun.
Current-body registration now uses a distinct complete post-HIR frame and fresh collector, rereading the body and recollecting its rows before publication (current-registration, 13 native checks pending). The borrowed MIR module consumer requires an active lease and exact complete body/declarations; three lifecycle controls and two real emitted-graph controls are pending. Explicit layer/effect/RT claim replay complements aspect replay; each adds two unrun source controls. These mechanical consumers and editable claim DTOs grant no final authority. Real current-backed mono readers and the driver prerequisite owner must bind the actual validator/weave/cold/quiet sequence before activation.
The shared complete-body frame leaf preserves the existing resolved bytes/key contract. A distinct woven-body descriptor supplies actual cold capture before optional quiet resolution, with strict stage/domain separation and the unchanged 8 MiB raw envelope bound. The weave-cold fixture requires 14 native checks, still unrun; its diagnostic boundary labels grant no production or final trust.
Post-HIR selected fact readers now take captured-policy selection inputs rather than consulting another compilation's ambient claims. Ordinary wrappers keep their original selection. Global/local HAL still enables the full graph, while no-claim effect layouts remain unmeasured. The real managed archive fixture now requires 26 checks (including the separately archived HAL source), all unrun.
The once-only policy capture now pins actual ordered claim rows and pre-consumption RT capabilities to the real source/schedule/journal/provider recipe. Canonical newline and independent authored-alias proof are checked. postpolicy requires 22 native checks, including a real stable graph allocation capability, retained manifest consumption and replay preserving the next grant. All are unrun; immutable capture alone is not accepted full-gate admission.
Source-claim replay now reads the exact frozen physical bytes and runs the selected frontend in a file lease with isolated claim registries. Constant-size snapshots restore the old roots before closing scratch, preserving the next RT capability grant. Five added controls cover actual source-produced claims, registry-reset replay, missing obligations, alignment and foreign ownership. These controls remain unrun. Actual full-gate, weave and final ownership are still required before production admission.
Current-body inputs now flow through the existing root/FIFO execution, complete assembly and exhaustive verification replay. Original-body recipes retain their byte contract; current recipes explicitly carry the strict resolved frame and cache identity. The current-mono selector requires 20 native checks, unrun. The registration facade and extracted schema/identity/seal leaves remain below 800 lines with unchanged APIs and private unchecked installation. Genuine driver prerequisite and final repository admission remain required.
The ordinary strict streaming AOT route now selects the production archive controller after source selection. It publishes complete final HIR archives to the actual MIR consumer, retains specialization provenance for logical aliases, and finishes the private owner after compilation. The source review is complete; fresh compiler admission, native controls and measured memory qualification are still pending. Full-inventory and demand routes retain their existing selection.
Imported generic specialization spelling now uses actual template declaration ownership and its extern/export/global/entry linkage metadata. The compact signature carries that bit and digest domain v2 binds it, so signature receipts invalidate when linkage changes. Free and method calls share the same naming rule. Five regression groups are authored and remain unrun; native link and execution qualification is still required.
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingdoc/00_llm_process/feature_expert/compiler_performance/skill.md
Popularity
- 7-day trend score
- 5
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/ormastes/simple/compiler-performance)