JonasKruckenberg/k23

We've been running benchmarks for a while now, and cache them in CI but haven't made any use of the data yet. Goal A PR that regresses a critical benchmark by >threshold fails CI. Everything else posts a PR comment. Performance bugs are bugs. Plan Step 1 — Label benchmarks critical in build/bench.bzl def rust_benchmark(name, critical = False, ...): _rust_benchmark_runner( ... labels = ["critical"] if critical else [], ) Critical benchmarks gate merges; others comment-only. Extractable via: _q_critical_benches(q) := f"attrfilter(labels, critical, ({_q_benchmarks(q)}))" Step 2 — just benchmark-compare recipe benchmark-compare baseline="main": just benchmark critcmp --threshold 5 {{baseline}} new Local regression check without pushing to CI. Step 3 — Extend .github/workflows/benchmark.yml - name: collect criterion output run: | find bench -name 'estimates.json' -print0 \ | xargs -0 cat | jq -s '...' > .bench-out/criterion.json - name: co

Problem CI runs one QEMU config (//sys:k23-qemu-riscv64-tests, 8 harts, 256M, TCG). This misses: Single-hart scheduler / lib/spin bugs TCG-vs-silicon divergences (weak memory, interrupt timing) aarch64 and x86_64 never actually run anything Goal just selftests and CI cover {hart count} × {mem size}. Plan Step 1 — QEMU constraints (build/constraints/BUCK) constraint(name = "qemu-harts", values = ["1", "4", "8", "16"], default = "8") constraint(name = "qemu-mem", values = ["128M", "256M", "1G", "4G"], default = "256M") constraint(name = "qemu-accel", values = ["tcg", "kvm"], default = "tcg") Step 2 — Toolchain reads constraints (build/toolchains/BUCK) Replace hard-coded -smp/-m args with select({...}) over the new constraints. NUMA derived from hart count (1 node ≤4 harts, 2 nodes otherwise). Step 3 — Matrix via modifiers Use Buck2 modifiers (-m constraints//:qemu-harts[1]) rather than multiplying platforms. Lighter weight, same effect. Step 5 — CI matrix (.github/workflow