firezone / firezone / 41410528226

80%

push

github

test(connlib): count instructions with mimalloc (#16168)

The Client and the Gateway allocate with mimalloc, but the Callgrind
benchmarks ran on the system allocator. glibc's malloc costs several
times more instructions per allocation, so allocations looked much more
expensive in the instruction counts than in the throughput benchmark:
skipping the STUN parser's allocating error path saved about 185
instructions per packet in the counts but only a few nanoseconds on the
bare-metal runner. With the same allocator as production, the counts
weigh allocations the way the real event loop pays for them.

Related: #16165

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

1370 of 2287 branches covered (59.9%)

Branch coverage included in aggregate %.

94533 of 117440 relevant lines covered (80.49%)

369854.99 hits per line