Repository navigation
perf(coverage): the DEBUG trap re-mangles the file path twice per executed line #1102
Description
Problem
With the report phase down to ~1s (#1088, #1090) and the buffer gone (#1101),
capture is what a --coverage run costs. record_line runs per executed line
and does two cache lookups — the tracked decision and the normalized path — and
each one rebuilds its key with
_BASHUNIT_COVERAGE_LOOKUP_KEY_OUT="$1${2//[^a-zA-Z0-9]/_}"That pattern substitution is the whole cost, and it is not linear in path
length. Measured on Bash 3.2 arm64, 5000 mangles:
| path length | time |
|---|---|
| 7 chars | 39 ms |
| 59 chars | 653 ms |
| 121 chars | 2746 ms |
So record_line costs 402 µs per line for a real path
(/Users/…/src/coverage/engine.sh), of which 328 µs is the two lookups and
only 29 µs the actual append. A project whose sources sit deeper pays
quadratically more, for nothing it can control.
Proposal
A one-entry memo for the last file seen. Consecutive executed lines come from
the same file almost always — that is what a function body is — so a single
string compare replaces both key rebuilds and both indirect reads:
| 10,000 iterations | |
|---|---|
| mangle per call | 1393 ms |
| same-file memo | 56 ms |
Expected: ~402 µs to ~75 µs per line, and the path-length sensitivity mostly
gone. The existing per-path caches stay for the miss path.
coverage::init has to clear the memo along with the other in-memory caches,
or a second run in the same shell inherits a stale decision.
Metadata
Metadata
Assignees
Labels
Type
Projects
- Status
Activity
Chemaclass commentedon Aug 13, 2026
Measured the proposal properly and the premise above is wrong. Recording
the corrected picture.
The memo barely helps. Implemented, it hits 75% of the time (4291 of 5718
record_linecalls) and buys ~4%: 7019ms → 6751ms. Not worth the state.Why the estimate was wrong. The 402 µs/line came from a micro-benchmark
using a 59-character absolute path. The trap actually sees what bash sourced —
./src/config/env.sh, about 20 characters — where the mangle is far cheaper.Path length still matters super-linearly, but not at the lengths that occur.
Where the time really goes. For
tests/unit/assert/basic_test.shagainst--coverage-paths src:srctrackedSo trap dispatch costs 46 ms — #1060 already fixed that — and ~5.1s is
recording. Instrumenting it:
normalize_pathis called 494 times per run,and its body is
which is four forks — roughly 2000 forks per run. It recurs because the
caches live in shell variables and the test body runs in a subshell, so every
test re-pays for the files it touches.
The fix that pays. Same semantics, one fork:
Identical output on relative, absolute and
/tmp/../tmp/./x.sh— thecd+pwdstays, so symlink resolution is unchanged. Pure-bash string normalisation is
not an option for exactly that reason: it would report
/tmp/...where thecurrent code reports
/private/tmp/....Retitling this to the normalization fork; the memo is dropped.
Chemaclass commentedon Aug 13, 2026
Done in #1105 (merged). The memo the issue originally proposed was measured and dropped; the fork reduction is what landed.