perf(gc): the minor collector still does a full-region sweep walk over the old generation (#6181 sub-cost a)

[proggeramlug]

The one unfixed sub-cost from #6181, refiled on its own so it stops sharing a ticket with two that are done.

Fixed and closed with #6181: (b) the whole-heap remembered-set rebuild and (c) the per-cycle page-meta iterations, via #6213 / #6277 / #6568.

Still live: sub-cost (a) — the minor collector performs a full-region sweep walk over the old generation. crates/perry-runtime/src/gc/oldgen.rs::sweep (line ~476) is reached on the minor path, so minor-GC cost scales with old-gen size rather than with nursery survivors — which is the property a generational collector exists to avoid.

Why this is worth doing and what to watch

The remembered set already exists precisely so a minor need not retrace the old generation, so the walk is redundant work on the hot path. But two things make this less mechanical than it looks:

• The malloc sweep is a separate path (sweep_malloc_objects, copied_minor_malloc_sweep_due) and is gated by trigger kind. Do not conflate the two — a change that skips one and not the other will look like it works.
• A guard alone can livelock. P0: silent wrong answer under the moving collector — stale heap-string deref in generated code (realistic interpreter workload) #7682's fix is the precedent: adding the immobility guard by itself deadlocked, because the nursery cap was keyed on from-space bytes that a non-moving minor never frees. The complete fix needed cap-gating as well (fix(gc): restore evacuation at precise safepoints — the pacing half of #7682 #7690). Expect the same shape here — removing work from the minor path changes what the trigger arithmetic sees.

Before coding

Re-measure. #6181's figures predate #6213/#6277/#6568, so the remaining share attributable to (a) alone has never been isolated. Use PERRY_GC_DIAG=1 and confirm the cycle KINDS you are measuring are the ones you think — census counters cover copying minors ONLY, and a run whose cycles are full mark-sweeps says nothing about the minor path.

Bench host per the standing rule: the quiet M1 mini.

[proggeramlug]

Re-measured first, as this issue instructs. Sub-cost (a) is 0.0 ms on the shipped default path. The premise — "minor-GC cost scales with old-gen size rather than with nursery survivors" — no longer holds on main @ db44b31b7.

Method: release build, PERRY_GC_TRACE=1, per-cycle phase_us split by collection_kind, on three workloads with very different old-gen sizes. Cycle kinds were checked rather than assumed, per the warning in this issue.

Default configuration (loop polls ON since #7721 — copying minors)

workload	minors	minor pause	sweep share of minor pause
retain_live_big (large live old gen)	15	1,739.0 ms	0.0 ms (0.0%)
churn	105	21.1 ms	0.0 ms (0.0%)
tree	0 minors (40 fulls)	—	—

Where minor pause actually goes on retain_live_big:

copying_nursery      1553.9 ms
in_place_promotion    129.4 ms
sweep                   0.0 ms

The sweep key is not missing from the accounting — the same runs attribute 68.0 ms of sweep to their full cycles, and tree's fulls are 53.7% sweep. Minors simply do not spend time there.

The reason is structural: the minor path is now the copying minor, which evacuates the nursery and never walks the old region. That became the default via #7690 / #7721, after #6181 was written — which is precisely why this issue told the next person to re-measure.

The one place (a) still exists

Forcing the non-moving minor path with PERRY_GC_MOVING_LOOP_POLLS=0 on the same workload:

	polls ON (default)	polls OFF (non-moving minors)
minors	15, all copying	13, none copying
total minor pause	1,739 ms	21,989 ms
sweep	0.0 ms (0.0%)	1,443.1 ms (6.6%)
atomic_finalize	—	13,715.8 ms (62.4%)
reference_rewrite	—	12,045.9 ms (54.8%)

So the old-gen sweep walk is reachable only on the non-moving minor path, where it is 6.6% of a path that is 12.6× slower in total for reasons that dwarf it. Optimising it would be optimising a configuration that is off by default and already pathological, and the two hazards this issue warns about — the separate malloc-sweep path, and the #7682/#7690 livelock shape where removing work from the minor path changes what the trigger arithmetic sees — would all still have to be paid to get there.

Closing as measured-moot. What would reopen it: a default-path workload whose minors show non-zero sweep in phase_us, or a decision to make the non-moving minor a supported steady-state configuration again rather than a bisection tool.

One caveat stated plainly: this measures phase attribution, not call reachability. oldgen::sweep may still be entered on a copying minor and return immediately; what the numbers refute is the cost claim, which is what the issue is about.