Repository navigation
perf/gc: regex scratch counted as external GC pressure forces full collections; growing it wins −16..−31% instr but +13..+28% RSS #11549
Description
Found by the regex performance lane (#11543).
The regex slow path reports its per-call scratch buffers to the GC as external allocations. For patterns needing more than 32 match registers (for example dotenv's line regex, which has 42), that phantom external pressure keeps hot loops in frequent full collections.
Measured: letting the per-thread match scratch grow beyond 32 registers gives dotenv/parse −31% and moment/parse_format −16% instructions. But peak RSS rises +27.6% / +13.4%, because once the phantom pressure is gone the young generation grows to its cap before collecting. So it was not shipped, under the "minimize RSS AND keep best compute" rule.
We want both wins. Directions:
- Stop counting short-lived, immediately-freed scratch as external GC pressure, or count it as transient so it doesn't drive full collections.
- Separately, look at the young-generation pacing that lets RSS climb to the cap when allocation pressure is low. An RSS-aware nursery cap or earlier minor GC would keep RSS flat with the faster regex path.
Measurements are on branch wip/perf-regex-matcher and in #11543's description.
Activity
test(regex): operation scratch and lent-cell searches report no exter…
perf(gc): young-only prune for built-in closure metadata on a minor
perf(gc): a minor skips the small-int / ASCII-char string caches
perf(gc): young-log the intern table, skip unarmed array-tail tables …
perf(gc): young-only prune for built-in closure metadata on a minor
perf(gc): a minor skips the small-int / ASCII-char string caches
proggeramlug commentedon Sep 28, 2026
Progress: #11634 landed. A copying minor's fixed cost drops ~87% (alloc loop: 753k -> 98k instructions per minor) via a young-only intern-table log, skipping empty array-tail tables, a young-only closure-metadata prune and skipping pinned string caches on minors; each skip has a completeness check and a sabotage test. No default changed. Leaving this open for the remaining per-minor cost.
perf(regex): stop charging operation scratch to GC external pressure;…
perf(gc): survival-aware nursery pacing: the ladder starts at a quart…
perf(regex): stop charging operation scratch to GC external pressure;…
perf(gc): survival-aware nursery pacing: the ladder starts at a quart…
perf(regex): stop charging operation scratch to GC external pressure;…
perf(gc): survival-aware nursery pacing: the ladder starts at a quart…
perf(gc): copying drain scans ordinary objects through a straight-lin…
perf(gc): raw words consult the mark memo first; skip the carrier gen…
4 remaining items
perf(gc): raw words consult the mark memo first; skip the carrier gen…
perf(gc): the plain-object scan takes the residual-prototype edge and…
perf(gc): copying drain instantiates the slot walk for its visitor; p…
perf(gc): copying drain scans ordinary objects through a straight-lin…
perf(gc): raw words consult the mark memo first; skip the carrier gen…
perf(gc): the plain-object scan takes the residual-prototype edge and…
perf(regex): stop charging operation scratch to GC external pressure;…
perf(gc): survival-aware nursery pacing: the ladder starts at a quart…
perf(gc): cut the copying minor's per-live-object trace cost (straigh…
perf(regex): stop charging operation scratch to GC external pressure;…
perf(gc): survival-aware nursery pacing: the ladder starts at a quart…
perf(regex): stop charging operation scratch to GC external pressure;…
perf(gc): survival-aware nursery pacing: the ladder starts at a quart…
perf(gc): survival-aware nursery pacing, 4 MB floor on the existing i…
proggeramlug commentedon Sep 30, 2026
Status: the regex scratch is no longer counted as GC pressure. That landed with #11645 (merged 9e29f59, carrying #11612), together with #11634, #11668 and #11676 (the minor's fixed cost and per-object trace cost) and survival-aware nursery pacing. The regressions the owner accepted are tracked in #11699. Keeping this open until the post-fix package matrix confirms the dotenv/moment numbers.