Performance
Latest recorded checksums passed, but full performance acceptance is still open. Use the recorded measurements as history, not a current baseline; keep workloads, routes, and gates unchanged.
Earlier completed work is already in this checkout; do not restore removed TODOs or overwrite newer code with saved agent snapshots. In particular, NYC benchmark pinning and dict-set stability are documented in Recorded measurements.
- [ ]
binaryscalar fallback: the gate now passes with outer vectorization on by default (919 us vs 266 us LLVM, 3.45x, 2026-11-05 full run). WithNYTRIX_OUTER_VECTORIZE=0the scalar loop is still 1.45 ms (5.45x). Remaining scalar ideas, about 5% each: branch on the decrementing IV's flags instead ofcmp i, d, and drop the second scalar IV.
- [ ] FPR/vector residency in
linpackanddgemm:spectral,mandelbrot, andpage-rankare done (see measurements).dgemmis 154 us (LLVM 26 us; 5.97x in the 2026-11-05 targeted run). Three of its four invariant broadcasts now stay in registers and are built with register-formvbroadcastsd. What remains is scalar GPR work in the inner loop:r*Mandk*Mare recomputed every iteration because LICM refuses to hoist address arithmetic (licm_memory_address_component), dead address temporaries are stored to homes, and base pointers are reloaded. LLVM's loop is 13 instructions; ours is about 30.linpack(739 us vs 240 us) needs vectorization of its row update once rowsr != kare proven disjoint.
- [ ] Full-gate acceptance: the 2026-11-05 06:25 serial run (3+1, workers
- passed every per-class gate, but other sessions were loading the host (load average about 14), and
collatz,euler9,knight, andqueenFAIL because their working-tree C sources print noelapsed_ns. That happens with the baseline compiler too. Fix those fixtures, then run serial matched NYR/LLVM/C/NYC comparisons with no concurrent builds; retain checksums, medians/spread, tool identity, compile cost, and code size. A subset is not acceptance. Follow Performance.