René's URL Explorer Experiment

Title: Simplify and tighten the @perf_event API by congwang-mk · Pull Request #23 · multikernel/kernelscript · GitHub

Open Graph Title: Simplify and tighten the @perf_event API by congwang-mk · Pull Request #23 · multikernel/kernelscript

X Title: Simplify and tighten the @perf_event API by congwang-mk · Pull Request #23 · multikernel/kernelscript

Description: Summary Reworks the @perf_event user-facing API for safety, consistency, and a smaller compiler footprint. Five focused commits: Fold perf_type + perf_config into one event constant. The two old fields had to agree (every config belongs to exactly one type — cache_misses is always hardware), and mismatching them was a silent footgun. A single perf_event enum now packs the perf_event_attr.type tag in the high 32 bits and the config in the low 32 bits, so naming the event fixes both. // before attach(prog, perf_options { perf_type: perf_type_hardware, perf_config: cache_misses }, 0) // after attach(prog, perf_options { event: cache_misses }, 0) Single-source the encoding in stdlib. The PMU-slot analysis had hard-coded the bit layout (>> 32, type tags) and the enum name in the type checker, duplicating what stdlib defined. The layout now lives only in stdlib (perf_event_pack); the type checker resolves the constant generically. Remove the host-dependent static perf-group PMU-budget check. It rejected groups needing more PMU counters than a limit read from the build host's sysfs (or a guessed default of 4) — the wrong layer for a host-dependent resource check, and wrong under cross-compilation. The kernel already enforces group sizing authoritatively at perf_event_open(2) on the real host, and the generated runtime reports it. This removes ~240 lines and all perf-group/PMU/encoding knowledge from type_checker.ml (perf references: 55 → 11, the survivors being the same uniform per-program-type handling @xdp/@kprobe get). Drop the raw group_fd field from perf_options. It exposed a kernel ABI integer alongside the typed group handle. Since every perf fd comes from attach() as a perf_attachment, the raw form was redundant; grouping is now expressed only via group. group_fd: now fails type checking as an unknown field. snake_case the perf types. PerfAttachment → perf_attachment, PerfRead → perf_read, so the whole perf surface matches perf_options and the rest of the language. Behavioral notes Capability trimmed (deliberately): the unified event enum currently covers the named hardware/software events (everything examples/tests used). The old tracepoint/hw_cache/raw/breakpoint type tags and arbitrary numeric configs are not expressible; they can return as encoded enum entries or a raw_event(type, config) constructor if needed. No safety regression from removing the static check: oversized hardware groups fail at perf_event_open(2); >16-member software groups truncate-with-warning in the bounded group read (no overflow); group cycles are already prevented by use-before-definition; non-leader group references are rejected by the kernel. Verification Full test suite green (dune runtest, exit 0). Both perf examples (perf_cache_miss, perf_page_fault) compile and the generated C builds to working binaries — including the software-event path that crosses the 32-bit encoding boundary. Docs (README, SPEC, BUILTINS) updated throughout. Net: +180 / −633 lines. 🤖 Generated with Claude Code

Open Graph Description: Summary Reworks the @perf_event user-facing API for safety, consistency, and a smaller compiler footprint. Five focused commits: Fold perf_type + perf_config into one event constant. The two old ...

X Description: Summary Reworks the @perf_event user-facing API for safety, consistency, and a smaller compiler footprint. Five focused commits: Fold perf_type + perf_config into one event constant. The two old ...

Opengraph URL: https://github.com/multikernel/kernelscript/pull/23

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width