perf(runner): remove avoidable subshell forks from the run_test hot path by Chemaclass · Pull Request #770 · TypedDevs/bashunit

Chemaclass · 2026-07-11T15:50:33Z

🤔 Background

Related #764

run_test and its helpers forked ~12-14 pure-bash $(...) subshells per test plus an rm per test for temp-file cleanup, the single largest external cost on the framework's reference platform (Bash 3.2 macOS). Most of those forks measure or clean up nothing.

💡 Changes

Converted the per-test hot-path helpers to the existing return-slot convention (runtime-error detection, subshell-output decode, name normalization, test-id generation, retry-count lookup) and built the passing-test line with pure-bash concatenation, removing their capture forks.
Per-test temp-file cleanup now probes the glob in pure bash and only forks rm when the test actually created a temp file (rm 102 -> 2 on a 100-test file; ~33% faster wall-clock on that file). A flag could not be used: temp_file runs in a $(...) subshell, so it never reaches the EXIT-trap cleanup.
Left resolve_test_location (its lone inner subshell deliberately contains an extdebug toggle) and did not pre-filter the declare -F cache in get_functions_to_run (shared by the test and bench prefixes; filtering the one builder would couple it to every consumer and removes no fork).
No behaviour change; console output byte-identical. Repo-wide shfmt v3.13.1 drift left untouched, added lines show none (as in perf(runner): cache data-provider lookup per file #767).

#764) detect_runtime_error was called once per attempt inside the retry loop and again after it with the same input, each call a $(...) fork. Convert it to the return-slot convention (_BASHUNIT_RUNNER_RUNTIME_ERROR_OUT) and reuse the final attempt's value after the loop, removing both the fork and the duplicate computation from the per-test hot path. No behaviour change.

…empty (#764) decode_subshell_output was wrapped in a $(...) capture on every test. Convert it to the return-slot convention (_BASHUNIT_RUNNER_SUBSHELL_OUTPUT_OUT) and short-circuit the empty/_BASHUNIT_EMPTY_ case inline, so a passing test with no captured output does no subshell fork at all. The non-empty path keeps the inherent base64 decode fork. No behaviour change.

…_name (#764) normalize_test_function_name forked one or two nested $(...) subshells per call to read the custom title and interpolated-name state via their accessor functions, which only echo reserved-namespace globals. Read the globals directly on this per-test hot path. Accessors are kept for external callers; no behaviour change.

…re fork (#764) generate_id was already internally fork-free but every caller wrapped it in a $(...) capture. Convert it to the return-slot convention (_BASHUNIT_HELPER_ID_OUT); the per-test export_test_identity path and the per-file script-id path now read the slot instead of forking. resolve_test_location already writes a global slot and its single inner subshell is deliberate (it contains the extdebug toggle), so it is left as-is. No behaviour change.

#764) retry_count was captured via $(...) once per test to read a value that is constant across the whole run. Add resolve_retry_count, which validates BASHUNIT_RETRY into the _BASHUNIT_RETRY_VALIDATED global in-shell (no fork); run_test reads the global. retry_count is kept as a thin echoing wrapper for external callers. No behaviour change.

…tf fork (#764) print_successful_test wrapped a $(printf ...) on every passing test to build its output line, though the format only did %s substitution. Concatenate the line directly (byte-identical, covered by the args and no-args console_results tests). No behaviour change.

…reated (#764) cleanup_testcase_temp_files ran rm -rf on every test's temp-file glob even though most tests create none, forking rm once per test (100/100 on the noop census). Probe the glob in pure bash first (a non-matching glob stays literal with nullglob off, so matches[0] does not exist) and rm only when a file matched. A global flag cannot be used here: temp_file/temp_dir run in a $(...) capture subshell, so a flag set there never reaches the EXIT-trap cleanup; the filesystem is the only reliable signal. Fork census on the 100-noop file: rm 102 -> 2. No behaviour change.

bash 3.0 does not expand a compound array assignment attached to `local`: `local matches=(glob)` stores the literal string "(glob)" instead of the matched paths. The cleanup probe therefore never saw a match, [ -e ] was false, and per-test temp files were left behind, failing the two temp file/dir tests on the Bash 3.0 jobs while every newer bash passed. Declare and assign separately so the glob expands on 3.0 too.

bash 3.0 does not expand a compound array assignment attached to `local`. watch.sh collapsed "$@" into a single literal element "(a b c)", so watch mode forwarded its extra args as one broken argument; console_results.sh stored the literal "()" as element 0 instead of building an empty array. Both are the same defect just fixed in globals.sh, and both are silent on every bash >= 3.2. Declare and assign separately in each, and add a guard test that greps src/ for the pattern, since a behavioural test could only fail under bash 3.0.

Chemaclass · 2026-07-11T15:50:33Z

🤔 Background

Related #764

run_test and its helpers forked ~12-14 pure-bash $(...) subshells per test plus an rm per test for temp-file cleanup, the single largest external cost on the framework's reference platform (Bash 3.2 macOS). Most of those forks measure or clean up nothing.

💡 Changes

Converted the per-test hot-path helpers to the existing return-slot convention (runtime-error detection, subshell-output decode, name normalization, test-id generation, retry-count lookup) and built the passing-test line with pure-bash concatenation, removing their capture forks.
Per-test temp-file cleanup now probes the glob in pure bash and only forks rm when the test actually created a temp file (rm 102 -> 2 on a 100-test file; ~33% faster wall-clock on that file). A flag could not be used: temp_file runs in a $(...) subshell, so it never reaches the EXIT-trap cleanup.
Left resolve_test_location (its lone inner subshell deliberately contains an extdebug toggle) and did not pre-filter the declare -F cache in get_functions_to_run (shared by the test and bench prefixes; filtering the one builder would couple it to every consumer and removes no fork).
No behaviour change; console output byte-identical. Repo-wide shfmt v3.13.1 drift left untouched, added lines show none (as in perf(runner): cache data-provider lookup per file #767).