Instruction Counts (--instr)
asb profile # --instr is the default
asb profile --instrThe default profiling mode. A Binaryen pass injects per-function call and executed-instruction counters at region granularity (function entry, loop bodies, if-arms), then the bench runs once. Counts are exact and fully deterministic — identical totals across runs, machines, and build flags — which makes this the right tool for comparing the effect of source or compile-flag changes.
Weighted vs raw
Raw instruction counts treat every op as equal. That under-ranks expensive operations, so the table also shows a weighted cost from a static table:
| Op class | Weight |
|---|---|
| ALU / const / local | 1 |
| integer multiply | 3 |
| load / store | 3 / 2 |
| float arithmetic | 2 |
| call / indirect call | 5 / 8 |
| division / sqrt | 12–15 |
| float→int truncation | 3 |
| atomics | 10 |
memory.grow | 100 |
Output columns: % | weighted instrs | raw instrs | calls | wt/call | name. The table ranks by weighted cost; both columns are shown.
arith/divide-heavy 1,204 weighted · 612 instructions
66.1% 796 wt 240 instrs 1 call 796 wt/call divideLoop
...Why a static table, not measured times
Per-instruction times aren't additive under superscalar execution, and weights can't see cache behavior — a load is costed as an L1 hit. The table is deliberately a ranking heuristic, not a timing model. When you need real nanoseconds (including cache effects), use --time; when you need an exact, reproducible count, the raw column here is ground truth.
Caveats
- Counts are exact, but the weighted column is a heuristic ranking, not a prediction of wall-clock time.
- Region attribution is approximate around early
return/br: the region's cost is charged on entry, so an early exit can over-count. The raw counts remain deterministic.
Next
- Self Time — real wall-clock, cache-aware.
- Allocations — bytes, not instructions.