- Most turns: finding the code paths. Before the context reset I spent most of the run reading the checker and the emitter: where a parameter is declared, where a call is checked and named, how
T::fields is unrolled, how a copy is named and emitted, and where moves are decided. check_core_pilot.nika and the emit_* files are large, and reading them filled my context until it reset. Two smaller costs came after that:
- One build cycle went to a rustc inference error in the lowered
(bound ?? on).clone().
- Some turns went to
pre-push-check.sh failing on ADR-384, which has nothing to do with my change.
- What the brief could have handed me:
- The emitter side: where a specialised copy’s name is made (
emit_function.nika, walk.specialising), and where a call to a copy is rewritten (emit_core_args.nika, emit_call_special).
- Where the move check is decided:
hands_over_read_at in check_arm_leaves.nika.
- Where declared types are collected, so I knew where to add NK1320:
collect_declared, check_function_frame, and the typed let.
- That
comptime value parameters do not parse at all. The brief said comptime f: meta::Field “means the same”, and that turned out to be impossible inside its allowed files.
- Checks:
- Useful: the old-binary baseline on both tests, the build and
lower-std fixpoint cycle, the 41 regression programs with --both-settings, the NK1320 placement probe, fmt and clippy.
- Mostly wasted: the
nikaia-std tools:: unit tests (11 tests, unrelated to this change); the second fixpoint build, since the diff stat was already the same; the bash -x trace of pre-push-check.sh.
- One change: the brief’s Locations should name the emitter and move-analysis functions as well as the checker’s (one line each, from
scripts/where.nika). It should also say up front that a comptime parameter does not parse yet. That would have saved most of the reading that used up the first context. In builder.md, a known-red pre-push step (like ADR-384) belongs under Red on develop now.
This site is open source.
Improve this page.
