Perf regression: golang/rescan wall +34% (1d523d30, #1403) #1424

[agent] Bench: golang/rescan wall time is +34% and golang/hosted +23% on today's main (17793173), compared with f3c6313a (the previous bench run, 2026-10-09). Request counts are unchanged (64).

Culprit: 1d523d30, #1403 ("Fix Go replace for unselected versions (#391, #392, #509)"). Profiling found it: the hot function, GoSumEditor::zip_versions, was introduced by that commit. It was not bisected with builds.

Same-machine interleaved compare

Base f3c6313a → head 17793173, on a 4 vCPU Intel Xeon @ 2.10GHz cloud sandbox. An A/A run on the same machine was clean (golang/rescan +2.6%, ≈).

Hot spot

Callgrind on the head binary, golang/rescan (1200 modules, 30 patched), inclusive instructions:

68.3%  patch::redirect::rewriter_groups::{{closure}}
30.4%    vendor::go_sum_edit::GoSumEditor::zip_versions
29.0%      <SplitWhitespace as Iterator>::next
15.9%    vendor::go_mod_edit::scan_hosted_replace
10.3%      go_mod_edit::strip_comment

The golang rewriter (patch/redirect/mod.rs, around line 8536) calls go_sum.zip_versions(&fname) once per patched dep, and each call splits every line of go.sum. That makes it O(patched deps × go.sum lines). go_mod_edit::prunes_module_graph(&go_mod) is also evaluated per dep. One pass over go.sum that builds a module → zip versions map, plus a single prunes_module_graph read hoisted out of the loop, should bring it back.

Repro

export CARGO_PROFILE_PERF_INHERITS=release CARGO_PROFILE_PERF_LTO=thin CARGO_PROFILE_PERF_STRIP=none
git worktree add /tmp/base f3c6313a && (cd /tmp/base && CARGO_TARGET_DIR=/tmp/tb cargo build --locked --profile perf -p socket-patch-cli)
CARGO_TARGET_DIR=/tmp/th cargo build --locked --profile perf -p socket-patch-cli -p socket-patch-bench
/tmp/th/perf/socket-patch-bench compare --base /tmp/tb/perf/socket-patch --head /tmp/th/perf/socket-patch -f '^golang/' --runs 20

Runner: nproc 4, Intel(R) Xeon(R) Processor @ 2.10GHz.

Generated by Claude Code