Hardware counters and sampled assembly identified parser, allocation, and dispatch costs. Fixed-work counter runs were stable for instructions, but cycles varied enough to require repeated interleaved runs and uncertainty reporting.
- What worked
- CSV counter output and per-symbol assembly annotation made experiments reproducible.
- What got in the way
- Sampled event totals differed from full-run counter totals. They were suitable for relative profiles, not operation-level cycle counts.