Derive the .cmi / .cmt / .ast magic numbers instead of bumping them in every PR
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 35/100
- Issue 类型
- 功能
- 描述清晰度
- 描述清楚
- 活跃度
- 活跃
- 技术栈
- ocaml
- 领域
- compilers
调研方向
Start with compiler/ext/config.ml and inspect how the .cmi, .cmt, and .ast magic numbers are defined, then trace the serialized interfaces and release tooling described in the proposal. The work spans digest generation, release recording and counter updates, and possibly dev build suffixes and prefix changes. Done means release counters change only when serialized layouts change, outside tools can distinguish compatible releases, and the changelog notes the prefix change.
由索引模型根据 Issue 内容生成。
描述
Problem
compiler/ext/config.ml holds hand-numbered magic numbers for .cmi (Caml1999I036), .cmt (Caml1999T043) and marshaled .ast files (ResImpl01315 / ResIntf01315). Every PR that changes a serialized type bumps them, so parallel PRs constantly conflict on the same lines. That's the same problem we had with CHANGELOG.md before fragments. There have been about 20 bumps since August.
The per-PR precision buys nothing:
- Rebuilds after a compiler change are driven by
bsc_hash, which both build systems store, not by the magic numbers. - Outside tools only ever see released versions, so intermediate bumps between two releases are invisible to them.
- A bump forgotten in a PR fails silently: mixing files across a layout change crashes in
Marshalinstead of giving a clean error.
The Caml1999 prefix is a separate problem. OCaml 5.5 uses Caml1999I037 / Caml1999T037, so our I036 is most likely a real OCaml release's .cmi number, and the next bump would collide with OCaml 5.5. An OCaml tool would then accept a ReScript .cmi and misread it, instead of rejecting it.
What outside tools need
Tools that read .cmi/.cmt files may rely on the magic number. They need:
- A stable number for each release.
- A change between releases only when the format may have changed, so they can tell that two releases share a format.
- Ordering, for "built by a newer compiler" checks (
cmi_format.mlalready uses it for its older/newer error).
Proposal
Bump the counters at release time, not in PRs:
- A dune rule computes a digest of the interfaces whose values are serialized. For
.cmithat istypes,path,ident,parsetree, …; for.cmtaddtypedtree; for.astit isparsetree. Prefer whole directories or dune's cmi digests over a hand-picked file list, which would go stale like manual bumps do. - A checked-in file records the counter and digest from the last release. PRs never touch it.
- The release script, or a CI check on release PRs, compares the current digest with the recorded one. If they differ, it bumps the counter and records the new digest. Comment-only edits cause harmless extra bumps.
- Optional: dev builds append the digest (e.g.
…037+3fa9c1d2), so mixing two dev builds that share a counter but differ in layout gives a clean error instead of a crash. Releases ship the bare counter. - Switch to our own prefixes (e.g.
ResCmi…/ResCmt…, like the.astnumbers already did), so OCaml tools reject our files. Note this in the changelog because outside tools will see it.
Not in scope: .cmj has no magic number yet, only a digest of its own content (js_cmj_format.ml has a TODO). It could get a counter under the same scheme. The frozen Parsetree0 PPX numbers (Caml1999M022 / N022) must never change and stay as they are.
Alternatives considered
- The compiler version as the magic number. Stable per release and never touched by PRs. However, it changes on every release even when the format didn't, it has no meaningful ordering across prereleases, and dev builds sharing a version can't detect a layout mismatch.
- A pure content digest. Exact and never conflicts, but outside tools can't see that two releases share a format, and there's no ordering.
- 主要语言
- OCaml
- 星标
- 7.5k
- 派生
- 483
- 平均合并
- 21 小时 8 分钟
- 30 天内合并 PR
- 183
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
rescript-lang/rescript 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 72/100
rescript-lang/rescript#8996 ·
维护者通常 1 天内回复
-
难度 3/5 1-2 天 新手友好度 70/100
rescript-lang/rescript#8997 ·
维护者通常 1 天内回复
-
难度 4/5 1-2 天 新手友好度 55/100
rescript-lang/rescript#8995 ·
维护者通常 1 天内回复
-
难度 3/5 1-2 天 新手友好度 68/100
rescript-lang/rescript#8994 ·
维护者通常 1 天内回复
-
Issues from differential semantic drift analysis (a7721303 to b123b7c3)可能已有人在做 @cknitt 今天认领。 未关闭
rescript-lang/rescript#8980 · 已指派 1 人 ·
维护者通常 1 天内回复
查看 rescript-lang/rescript 的全部 Issue
相似的 Issue
-
bug P2
难度 2/5 1-3 小时 新手友好度 74/100
维护者通常 1 天内回复
-
duplicate
难度 2/5 1-3 小时 新手友好度 72/100
维护者通常 1 天内回复
-
REPL and --show-tokens abort with "Token conversion not implemented yet" on .xor., byte and in out未关闭bug low priority tokenizer
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
难度 1/5 1 小时以内 新手友好度 76/100
WebAssembly/spec#2300 ·
维护者通常 1 天内回复
-
crash generated by fuzzer mlir
难度 2/5 1-3 小时 新手友好度 72/100
llvm/llvm-project#234636 · 1 条评论 ·
维护者通常 1 天内回复