## 概要 LLVM backend のCall処理、PHI wiring、Plugin loader を強化。 次のリファクタリング(箱化モジュール化)のための準備も含む。 ## 変更内容 ### LLVM Call処理強化 - `mir_call/__init__.py`: Call ルーティングロジック改善 - `mir_call/global_call.py`: print処理の marshal強化 - `mir_call/method_call.py`: メソッド呼び出し処理改善 - `boxcall.py`: BoxCall処理改善 ### PHI処理強化 - `phi_manager.py`: PHI管理改善 - `phi_wiring/wiring.py`: PHI配線ロジック強化(+17行) - `phi_wiring/tagging.py`: Type tagging改善 - `resolver.py`: Value解決ロジック強化(+34行) ### Copy伝播 - `copy.py`: Copy命令のType tag伝播追加(+10行) ### Plugin loader強化 - `library.rs`: エラー出力改善、[plugin/missing]ログ追加(+34行) - fail-fast強化 ### テスト - `phase97_json_loader_escape_llvm_exe.sh`: Phase 97 E2Eテスト追加 - `phase97_next_non_ws_llvm_exe.sh`: Phase 97 E2Eテスト追加 ### その他 - `nyash_kernel/lib.rs`: Kernel側の改善(+23行) ## 統計 - 14ファイル変更 - +256行 / -53行 = +203 net ## 次のリファクタリング準備 以下の箇所がリファクタリング対象として識別済み: 1. Call ルーティング箱の明文化 2. print の marshal 箱 3. TypeFacts/Tagging 箱の一本化 4. PHI Snapshot 契約のSSOT 5. Plugin loader のエラー出力統合 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Nyash Kernel
Minimal runtime kernel for Nyash language - Plugin-First Architecture
Generated: 2025-09-24 Architecture: Phase 2.4 NyRT→NyKernel Revolution Complete
Overview
The Nyash Kernel (nyash_kernel) is the minimal runtime core that replaced the legacy NyRT system. This represents a 42% reduction in runtime complexity by moving from VM-dependent architecture to a unified Plugin-First system.
Architecture Revolution
✅ From NyRT to NyKernel (Phase 2.4 Complete)
Before (Legacy NyRT):
- Mixed VM/Plugin dependencies
with_legacy_vm_argsscattered throughout codebase- 58% essential + 42% deletable functions
- Complex shim layer for LLVM integration
After (NyKernel):
- Pure Plugin-First architecture
- Zero legacy VM dependencies
- Only essential kernel functions remain
- Clean C ABI for LLVM integration
🏗️ Core Components
Essential Kernel Functions (58% - Kept)
- GC Management: Safepoints, write barriers, memory management
- Handle Registry: Object handle management for AOT/JIT
- Plugin Host: Unified plugin loading and method resolution
- Process Entry: Main entry point and runtime initialization
Removed Shim Functions (42% - Deleted)
with_legacy_vm_args- 11 locations completely removed- Legacy VM argument processing
- String/Box operation shims
- VM-specific encoding functions
Build Output
Target: libnyash_kernel.a (static library)
Status: Clean build (0 errors, 0 warnings)
Integration: LLVM + VM unified
Implementation Details
Deleted Legacy Functions
| File | Locations | Status |
|---|---|---|
encode.rs |
1 | ✅ Removed |
birth.rs |
1 | ✅ Removed |
future.rs |
2 | ✅ Removed |
invoke.rs |
6 | ✅ Removed |
invoke_core.rs |
1 | ✅ Removed |
| Total | 11 | ✅ Complete |
Plugin-First Integration
All Box operations now route through the unified plugin system:
// Before: VM-dependent
with_legacy_vm_args(|args| { ... })
// After: Plugin-First
let host = get_global_plugin_host().read()?;
host.create_box(type_name, &args)?
🔥 ExternCall Print修正 (codex技術力)
Phase 2.4で解決した重大問題: LLVM EXEでprint()出力されない
問題の詳細
- 症状: VM実行は正常、LLVM EXEは無音
- 根本原因:
src/llvm_py/instructions/externcall.pyの引数変換バグ - 技術詳細: 文字列ハンドル→ポインタ変換後にnull上書き
修正内容
# src/llvm_py/instructions/externcall.py:152-154
else:
# used_string_h2p was true: keep the resolved pointer (do not null it)
pass
検証結果
/tmp/direct_python_test_fixed
# 出力:
# 🎉 ExternCall print修正テスト!
# codex先生の名前解決修正確認
# Result: 0
Usage
For LLVM Backend
# Build with LLVM integration
cargo build --release -p nyash_kernel
# Output: crates/nyash_kernel/target/release/libnyash_kernel.a
For VM Backend
# Runtime integration (automatic)
./target/release/hakorune program.hako
Design Philosophy
"Everything is Plugin" - The kernel provides only the essential infrastructure for plugin management, leaving all Box implementations to the plugin system.
Core Principles
- Minimal Surface: Only GC, handles, plugins, and process entry
- Plugin-First: All Box operations through unified plugin host
- C ABI Clean: Stable interface for LLVM/VM integration
- Zero Legacy: Complete removal of VM-dependent code paths
ChatGPT5 × codex × Claude Collaboration
This kernel represents a historic achievement in AI-assisted architecture design:
- Design: ChatGPT5 Pro architectural analysis (42% reduction strategy)
- Implementation: Claude systematic implementation (11 locations)
- Debugging: codex root cause analysis (ExternCall print fix)
- Result: 100% successful architecture revolution + critical bug resolution
Integration
The Nyash Kernel integrates seamlessly with:
- LLVM Backend: Static linking via libnyash_kernel.a
- VM Backend: Dynamic plugin loading
- Build System: tools/build_llvm.sh integration complete
Part of Phase 15 Nyash Self-hosting Revolution Documentation: ChatGPT5 NyRT→NyKernel Design