Repository navigation
perf: four benchmarks remain 1.4-2.2x Node with no owning issue (pipeline, churn, tree_wide, cycles) #8289
Description
The 2026-08-17 four-engine sweep leaves four rows where Perry is meaningfully behind Node and no issue owns them. interp/iso_miss are #8202 and fib40 is #8175; these are the remainder.
| benchmark | Perry | Node | P/Node | Perry instr | Node instr |
|---|---|---|---|---|---|
pipeline |
0.2043 | 0.0950 | 2.15× | 3.893 B | 1.220 B |
churn |
0.3300 | 0.1666 | 1.98× | 5.006 B | 2.681 B |
tree_wide |
1.2965 | 0.9015 | 1.44× | 26.808 B | 20.262 B |
cycles |
0.1035 | 0.0736 | 1.41× | 2.132 B | 0.669 B |
All four are instruction-bound rather than scheduling-bound — Perry executes 1.3–3.2× the instructions Node does — so the work is to emit less, not to schedule better.
Notes per row, from the sweep:
pipelineis the worst uncovered gap. It was 3.3585 s at 08-14 and 0.2002 s at 08-15, so the catastrophic phase is long fixed (perf: pipeline's +3.95% at #8084 is unattributed — it is NOT the callee rooting (IR is byte-identical) #8243 closed); this is the steady-state residue. Perry beats scriptc here 0.2043 vs 0.4461, and Porffor cannot run it at all, so the gap is specifically against V8.churnalso carries the perf(gc): #8277's object-start bitmap costs 8-10% wall on allocation-heavy workloads #8288 allocation-bitmap regression (+8.2% since 08-15). Fixing perf(gc): #8277's object-start bitmap costs 8-10% wall on allocation-heavy workloads #8288 should be sequenced first, then this re-measured — the two are additive and it would be easy to double-count.cycleshas the largest instruction ratio of the four at 3.19× Node for 1.41× wall, which suggests the emitted work is cheap per instruction. Possibly the most tractable.tree_wideis the mildest ratio but the largest absolute time in the corpus at 1.30 s.
Not a regression report — all four improved or held since 08-15 (tree_wide −24.8%, cycles +1.4%, pipeline +2.0%, churn +8.2% from #8288). Filed so the remaining engine gap is tracked rather than implicit, now that the campaign has cleared the catastrophic rows.
Suggested order: #8288 first (it contaminates churn), then cycles as the cheapest probe into where the extra instructions come from, then pipeline.
Sweep evidence: gc-handoff/current-sweep-2026-08-17/ (RESULTS.md, results-clean.json, report.html).
Activity
proggeramlug commentedon Aug 17, 2026
Progress: #8301 has merged. On the checked-in
cyclesreproducer, memoizing typed layouts by immutableShapeIdreduced the five-run median from 2.177B to 1.804B retired instructions (-17.1%) with unchanged output.Keeping this umbrella issue open:
pipeline,churn, andtree_wideremain, andcyclesstill needs a clean four-engine re-sweep to establish its new wall-time and Node ratios.proggeramlug commentedon Aug 18, 2026
Progress: #8351 has merged. On the checked-in
pipelinereproducer, the four-entry polymorphic closure-dispatch cache reduced the seven-run median from 3.526B to 3.412B retired instructions (-3.23%) with unchanged output.Keeping this umbrella issue open:
churnandtree_widestill have no merged owning fix, andcyclesstill needs the clean four-engine re-sweep noted above.proggeramlug commentedon Aug 18, 2026
All four rows are now within 1.3x of Node — closing
Re-measured on the 2026-08-18 four-engine sweep at
1b61941bc(102 commits past the14468dcbcsweep this issue was filed against). Same corpus, same protocol, verdictCLEAN.pipelinechurntree_widecyclestree_wideno longer just clears the bar, it beats Node outright — and does it on60.6 MiB peak RSS against Node's 641.9 MiB, retiring 7.61 G instructions against 20.28 G.
Sweep-wide, this moved Perry past Node overall for the first time: geomean Perry/Node
0.804x (was 1.076x), 12/19 cells beating Node, 16/19 within 1.3x.
The four rows named here no longer need an owning issue. The one cell that went the
other way in this window is
asyncpipe(+55.6%), which is tracked separately as #8384rather than folded in here, since it is a regression rather than an unclosed gap.
Full sweep:
gc-handoff/current-sweep-2026-08-18-RESULTS.md.