feat(topic28): extend RISC-V instruction selection and f64 support

[ForsakenDelusion]

Summary

DSL / ONNX 标量 i32、f32、f64
  → 保留类型的 IR 和优化
  → RV32IM / F / D 指令选择
  → 整数或浮点寄存器、4/8 字节保存、每函数栈空间
  → 合法汇编、独立调用者、QEMU 执行

增加除法、余数、平方根、min/max、符号操作、访存及比较分支，提供四个 DSL 示例、支持列表和使用说明。R1–R4 的 LLVM 失败处理、函数内部重名、常量编码和分配器复用问题已修复。后续 R5 也已修复：ONNX 浮点 Min/Max 在任意输入为 NaN 时返回 NaN，IR 执行器同步处理，DSL 保持既有规则。

最低 Python 版本保留为 3.8：替换不兼容注解和标准库方法，保留调度器构造行为，统一版本说明，并在现有测试任务中增加独立的 3.8 兼容检查。

Review 报告与修复记录

2026-10-14 工程复审报告：新增 R7（P1）和 R6（P2），目前尚未修复；两项规范/维护建议为 P3。

Evidence

• 2026-10-14 对固定提交 37516a7 再跑完整回归：3826 passed，0 failed，0 skipped。独立探针仍确认 9 个 LLVM 常量错误场景和 36 个 ONNX 符号零结果差异，已有测试尚未覆盖或采用了目标预期。
• Before: 在 77bef84 的生产代码上，新 NaN 专项测试为 156 失败、60 通过；加上纠正 ONNX 预期后的 96 项既有编码测试，共 216 失败、96 通过。
  After: 新增 216 项专项测试通过，既有编码测试继续检查位模式、ABI 和异常标志；R5 修复后的 Python 3.12 完整回归为 3826 passed，0 failed，0 skipped。
• Python 3.8.20：NaN 专项和相关回归 917 passed，0 failed，0 skipped，包含更新后的最低版本 CI 入口 361 项。
• 8 个 NaN 场景的 ONNX 官方参考实现、ONNX Runtime 1.30.0 和 QEMU 执行结果一致。
• Before: 同一份 6 项兼容测试在 e061861 的生产代码和真实 Python 3.8 下全部失败，导入阶段报错。
  After: 6 项全部通过；Python 兼容性修复时最低版本 CI 入口的 145 项测试通过，包含原始课题验收 99 项和真实 RISC-V 执行。
• Python 兼容性修复时，Python 3.8.20 完整回归：3610 passed，0 failed，0 skipped。
• Python 兼容性修复时，Python 3.12 完整回归：3610 passed，0 failed，0 skipped；仅有已有的 TinyFive 弃用提示。
• 使用 Clang/LLD 与 QEMU 汇编、链接、执行；git diff --check 通过。远端 Actions 结果以实际运行记录为准。
• evidence 和 repro 文件只保留在本地，不纳入提交。

Merge Danger

Door: two-way

代码和配置可通过回退提交恢复。

Blast Radius: 编译管线

影响共享前端、IR 优化、RISC-V 选择与分配、栈布局及部分基准工具。范围为 RV32 标量和硬件 F/D；内部 encoder/emulator 仍仅支持 RV32IM。

ONNX Min/Max 的 NaN 传播问题已修复。本轮复审仍有两项待处理：R7（P1），共享 typed const 输入进入 LLVM 后，普通 f32 小数可产生无法解析的文本，f64 常量可能丢精度或负零符号；R6（P2），ONNX 正负零经 Min/Max 后接 Div 的结果与官方参考实现和 ONNX Runtime 符号相反。R6 是已复现的兼容差异，公开 ONNX 算子说明没有明文规定符号零相等时的选择规则。范围仍为 RV32 标量与硬件 F/D。

[github-actions]

🤖 AI Code Review

共审查 10 个变更文件
⚠️ 另有 20 个文件超过上限（最多 10 个）未审查

📁 .github/workflows/ci.yml

Code Review: .github/workflows/ci.yml

🔴 Risk: venv module availability on Python 3.8 — "$MINIMUM_PYTHON" -m venv will fail if python3.8-venv isn't installed. On GitHub-hosted ubuntu runners this is typically pre-installed, but if you ever self-host or change the runner image, this is a silent landmine. Consider adding an explicit install step or a pre-check:

"$MINIMUM_PYTHON" -c "import ensurepip" || sudo apt-get install -y python3.8-venv

🟡 Hardcoded test file list is fragile — Six individual test files listed. If tests are renamed or moved, this breaks silently. Consider a glob or directory-level invocation:

"$COMPAT_PYTHON" -m pytest tests/test_topic28_*.py tests/test_regalloc_result_reuse.py

If the explicit list is intentional for pinning scope, add a comment explaining why — currently it's unclear.

🟡 Unpinned pip install -e . causes CI flakiness — Each run resolves dependencies fresh against current PyPI. If a transitive dep breaks Python 3.8 compatibility, CI turns red with no diff. Consider committing a lock file or pinning key deps:

"$COMPAT_PYTHON" -m pip install -e . "pytest>=7,<9" -c constraints-py38.txt

🟡 No timeout on the test step — A hung test (e.g., waiting on RISC-V execution) blocks the entire job. Add a step-level timeout:

timeout-minutes: 10

💭 Redundant diagnostic output — clang --version, ld.lld --version, qemu-riscv32 --version dump full multi-line output every CI run. If the intent is just to capture versions for debugging, consider logging to a file and uploading as an artifact only on failure:

{ clang --version; ld.lld --version; qemu-riscv32 --version; } > "$RUNNER_TEMP/toolchain_versions.txt"

💭 PYTHONHASHSEED only set here — If other test steps don't set it, results may differ between runs within the same job. Consider setting it at the job level for consistency.

---

📁 CLAUDE.md

Review: CLAUDE.md Python 版本要求变更

🟡 **EOL 风险提示** — Python 3.8 已于 2024-10 EOL。
  作为项目规范文档，此处未说明 3.8 仅为临时兼容窗口还是有长期维护计划。
  建议补充截止条件（如 "Python 3.8 兼容至 YYYY-MM，届时移除"），
  否则后续开发者无法判断何时可以放弃 3.8 支持。

🟡 **引用依赖未落地** — "版本与注解规则以 CONTRIBUTING.md 为准"
  如果 CONTRIBUTING.md 当前不存在或尚未包含对应规则，
  此引用成为空指针，开发者照此执行会得到矛盾指令。
  建议：要么先确保 CONTRIBUTING.md 就绪，要么在此处内联关键约束。

🟡 **CI 声明需与实际一致** — "CI 主环境使用 3.12，Topic 28 另有 3.8 兼容测试"
  将此作为规范写入前，应确认 CI 配置（GitHub Actions / GitLab CI）
  确实已包含 3.8 矩阵测试。文档先于基础设施落地会导致开发者信任偏差。

💭 **单条信息密度偏高** — 第二行同时包含版本号、验证要求、CI 版本、Topic 编号、
  测试分工。考虑拆为两条（一条说要求，一条说验证方式），降低认知负荷。

---

📁 README.md

🟡 Verify actual minimum Python version — Changing the badge from 3.12+ to 3.8+ is a significant claim. Confirm pyproject.toml / setup.cfg / CI matrix actually supports 3.8. If code uses 3.9+ features (dict|dict merge, match/case, X | None union syntax), a 3.8 badge misleads users and breaks their installs.

💭 If 3.8 support is intentional, consider whether transitive dependencies also support it — the badge alone isn't enough to guarantee compatibility.

---

📁 benchmarks/bench_cnn_schedule.py

无问题。 这是一个合理的重构，将路径解析逻辑从三元表达式提取到 try/except 块中，逻辑等价（.relative_to() 在非相对路径时抛出 ValueError，与原代码的 is_relative_to 判断结果一致），且兼容 Python < 3.9。

仅一个可选建议：

💭 Narrow the exception scope — Line 273: except ValueError is fine, but if you want to be extra defensive (and it's a benchmark file worth keeping robust), you could add a brief comment explaining the intent:

try:
    input_path = model_path.relative_to(ROOT).as_posix()
except ValueError:
    # Fallback: model_path is not under ROOT (e.g. symlink or temp dir)
    input_path = str(model_path)

这有助于未来的读者理解为什么捕获 ValueError 而不是使用 is_relative_to。

---

总结：变更正确、安全、可读性更好。可以合并。

---

📁 benchmarks/bench_dsl_diagnostics.py

🟡 Suggestion — No functional issues; the try/except fallback is correct and idiomatic for pre-3.9 compatibility.

Minor note: if this file has a consistent Python version target, a comment (e.g., # Python 3.9+ has is_relative_to; fallback for <3.9) would help future readers understand why the more verbose form is used.

---

📁 benchmarks/compare_schedule_disabled.py

🟡 Exception breadth — Catching bare ValueError is fine here, but relative_to() raises ValueError for any reason the path isn't relative. A comment noting this is intentional (vs. e.g. a real path error) would help future readers distinguish the catch from a typo.

# ValueError raised only when path is not relative to root
try:
    Path(compiler.__file__).resolve().relative_to(root)
except ValueError:
    raise RuntimeError(...) from None

💭 Otherwise the change is correct — semantically equivalent to is_relative_to(), improves Python 3.9 compatibility, and from None cleanly suppresses the inner traceback.

---

📁 benchmarks/generate_models.py

No issues. Clean, zero-risk change — enables PEP 563 (deferred annotation evaluation) without side effects.

---

📁 benchmarks/test_regalloc/bench_cnn.py

💭 No issues — from __future__ import annotations is a safe, idiomatic addition (PEP 563). Enables forward references and X | Y union syntax on older runtimes without runtime cost. Good practice.

Only note: confirm no runtime-evaluated annotations exist in this file (e.g., pydantic.BaseModel, dataclasses with __future__ quirks, or explicit typing.get_type_hints() calls), since those need live values rather than strings. Given this is a benchmark harness, that's unlikely — LGTM.

---

📁 benchmarks/test_regalloc/bench_dense.py

This is a trivial, safe change — no issues found.

from __future__ import annotations is a no-op at runtime (just sets __annotations__ to strings), safe to add any time, and enables forward references and PEP 604-style unions (X | None) on older Python versions. 👍

---

📁 benchmarks/test_regalloc/bench_regalloc_linear.py

No issues found.

from __future__ import annotations (PEP 563) defers annotation evaluation to strings — a safe, additive change with no functional impact for a benchmarking script. 👍

---

---

⚠️ 未审查的文件

• benchmarks/test_regalloc/bench_simple.py
• benchmarks/test_regalloc/bench_utils.py
• benchmarks/topic28/f32_sum.dsl
• benchmarks/topic28/f64_dot.dsl
• benchmarks/topic28/f64_sum.dsl
• benchmarks/topic28/i32_div_rem.dsl
• docs/topics/28-扩展指令选择/28-README.md
• docs/topics/28-扩展指令选择/28-交付与验收.md
• docs/topics/28-扩展指令选择/28-代码Review报告-2026-10-12.md
• docs/topics/28-扩展指令选择/28-复审报告-2026-10-14.md
• docs/topics/28-扩展指令选择/28-扩展指令选择-设计文档.md
• docs/topics/28-扩展指令选择/28-扩展指令选择.md
• memory/memory.md
• scratchv/analysis/ir_verifier.py
• scratchv/backend/abi_frame.py
• scratchv/backend/allocation.py
• scratchv/backend/asm_beautifier.py
• scratchv/backend/asm_emit.py
• scratchv/backend/asm_parser_for_beautifier.py
• scratchv/backend/cycle_estimator.py

[Copilot]

Copilot review overview

🔵 Needs a closer look

The design contains unresolved type-validation and floating-point spill/reload issues, plus configuration and ordering ambiguities.

Review effort: Lite
Findings: None

What changed in this PR

Adds a technical design for Topic 28 instruction selection and links it from the topic page.

Changes:

• Defines typed IR, integer/F/D floating-point lowering, register allocation, ABI, and constant folding.
• Adds CLI integration, testing, acceptance criteria, and phased implementation planning.
• Links the new design document from the topic page.

File	Summary
docs/​topics/​28-扩展指令选择/​28-扩展指令选择.md	Adds the design document link.
docs/​topics/​28-扩展指令选择/​28-扩展指令选择-设计文档.md	Adds the comprehensive instruction-selection design. Findings: unknown ONNX types must fail instead of silently becoming FLOAT32 (critical, 1 vote); FPR spill/reload needs value-width/type propagation (moderate, 1 vote); rounding_mode requires validation (moderate, 1 vote); candidate priority direction should be defined (nit, 1 vote).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[ForsakenDelusion]

@copilot code review