dangernoodle-io / TaipanMiner / 26537435963

90%

push

github

perf(mining): cache-isolate hot-loop data to protect against layout shifts (#435)

The per-nonce inner loop in mine_nonce_range reads hw_ctx->midstate_hw[]
and block2_words[] (via pointer to block2[]) on every iteration. Both
addresses depend on mining_task's stack base, which the FreeRTOS
allocator places relative to cumulative BSS/heap state. Each new feature
that adds BSS shifts the stack base, putting these hot reads on different
D-cache lines that compete with the new feature's static data.

This caused the +112 ns/nonce regression bisected to TM #413 (per-pool
stats + block-found detection) on tdongle-s3 — see TA-392 for the full
diagnostic trail. Without protection, every future feature addition will
keep eating ~1-3% off mining throughput.

fix: make block2 in mine_nonce_range and hw_ctx in mining_task `static
__attribute__((aligned(32)))`. Aligned to ESP32-S3's cache line size (32B)
so they occupy their own cache lines with no false sharing, and pinned
to fixed BSS addresses immune to stack-base shifts.

verified on tdongle-s3 .78: 297,894 H/s → 307,770 H/s steady (+3.3%
recovery, matching v0.23.0 baseline of 309,650). Per-Tier-2 inner-loop
timing drops from 3520 ms to 3408 ms per 1M nonces.

scope:
- mine_nonce_range's `block2[64]` (stack-local → static cache-aligned)
- mining_task's `hw_backend_ctx_t hw_ctx` (stack-local → static cache-aligned)
- both functions only run on the single mining task; static storage is
  semantically equivalent to a stack local that outlives the function

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

2549 of 3217 branches covered (79.24%)

Branch coverage included in aggregate %.

5158 of 5332 relevant lines covered (96.74%)

154876.83 hits per line