F0074 perf-2¶
Rule: Hot paths on cores without 64-bit maths avoid it, and float formatting.
| Code | F0074 (stable, never reused) |
| Id | perf-2 |
| Category | perf |
| Source | spirit_level |
| Fix | by hand |
| Needs | the firmware's linked ELF (--elf); skipped without a firmware binary |
Level per tier¶
| Tier | Level | Reason |
|---|---|---|
| small | off | |
| small-workspace | off | |
| medium | off | |
| big | off | |
| ai | off | |
| embedded | warn |
bedrock suggestion, reported without failing the run: 64-bit or i128 maths, float formatting, or % / / by a runtime value per element on a hot path of a core without them (__aeabi_lmul, __aeabi_uldivmod, __multi3) |
| ui | off | |
| egui | off | |
| units-strict | off | |
| units-paranoid | off |
Why¶
Cortex-M0/M0+ has no hardware divider and no 64-bit ALU. Any 64-bit add/multiply, i128 arithmetic, float formatting, or a modulo/division by a value only known at runtime (not a compile-time power of two) compiles down to a libcall — __aeabi_lmul, __aeabi_uldivmod, __multi3 — that costs tens to hundreds of cycles, and on a hot path (a ring-buffer index, a per-sample scale) that cost is paid every time.
Example¶
This repo's own bedrock_led (DEV-12) and bedrock_sync (DEV-13) crates are the positive case, checked in the disassembly on purpose: bedrock_led's breath/blink easing is "u32 math only (no 64-bit libcalls on the M0+)", and bedrock_sync's hot path (shared, run every LED tick) has "no libcall on the Cortex-M0+ at opt-level z and 3" for the common case — 64-bit adds and 32-bit products of chunks up to 32 ms — only falling onto a cold path with two __aeabi_lmul for a gap longer than that, and a 64-bit divide once a second at the steer point, not per tick. The negative case: typec_dongle's ww_si unit types cost "~4 / 3.4 KiB" (z/O3) for i128 math alone (ops::scale_i128, i128::checked_pow, ops::compare) — measured in our internal flash-size investigation notes, which call i128 on a Cortex-M0+ "also slow", not just big.
How to fix¶
Keep the hot path in u32 (or fixed-point chunks sized to stay there, as bedrock_sync does); reserve 64-bit/i128 math for the cold path (once-a-second steer points, not every tick); replace a runtime %// with a compile-time power-of-two mask or a precomputed reciprocal where the divisor is known ahead of time.
How it is checked¶
Check: elf mid, chip: M0+ — grep the disassembly or the symbol table for __aeabi_lmul, __aeabi_[u]ldivmod, __multi3, __udivmoddi4 and attribute each hit to the function that pulled it in; a hit reachable from a function on a tight per-tick or per-byte path is the finding, a hit only on a once-a-second cold path (like bedrock_sync's steer) is not.
Source¶
Ported from embedded_bedrock's firmware check catalogue, PERF-2 (book/checks.md, detail in book/checks/perf.md). Found on: embedded_bedrock DEV-12/13/14; typec_dongle FW-42; internal fw-size measurement.
Configure¶
Allow one finding in the code with // spirit-level: allow(perf-2) <reason> (waivers); the code works there too, as in [lints.F0074].