Lower similar-list decomposition threshold

AI-assisted change (OpenAI Codex).

Author: Wilfred

PERF_RESEARCH_LOG.md

@@ -1263,6 +1263,37 @@ Both changes are within the roughly 0.1% cross-binary layout noise. The pass
 mostly performs required hash-table counting work, so eliminating only its
 tree-walk control flow has little leverage. The source was fully reverted.
 
+### exp39: lower the similar-list decomposition gate — KEPT at 45,000
+
+Exp35 only attempted similar-list pairing when the product of the two section
+sizes reached one million. A temporary trace showed that `slow` entered through
+one 1.69M top-level section, while `typing` recursively paired sections from
+1.06M to 5.10M but left plausible 0.40M and 0.48M sections untouched. A
+compile-time sensitivity sweep lowered this single gate:
+
+| graph-size gate | `slow` change | `typing` change | 107-pair output oracle |
+| ---: | ---: | ---: | --- |
+| 250,000 | flat | -0.103% | identical |
+| 100,000 | flat | -0.145% | identical |
+| 50,000 | -17.62% | -0.143% | identical |
+| 45,000 | **-19.33%** | **-0.144%** | identical |
+| 40,000 | -19.33% | -0.144% | `slider_1.rs` changed |
+| 25,000 | -31.40% | -0.140% | `load_1.js` and `slider_1.rs` changed |
+
+The cliff is discrete: lowering the threshold enables another recursive split,
+not merely a little more preprocessing. The lowest tested exact-output setting,
+45,000, was retained. Its final ten-run means against a copied exp37 control
+were 642,765,789 to 518,537,235 instructions on `slow` and 2,828,226,154 to
+2,824,159,538 on `typing`.
+
+On `slow`, the number of searches rises from 8 to 17, but their total visited
+vertices fall from 224,960 to 144,260 (**-35.87%**) and the largest search falls
+from 63,774 to 42,265 vertices (**-33.73%**). This confirms that many smaller
+searches are substantially cheaper than a few medium dense ones. Output is
+byte-identical in side-by-side, inline, and JSON modes on all 107 available
+sample pairs. `cargo test` passes (163 passed, one ignored), and the changed
+file passes `rustfmt --check`.
+
 ## Where this leaves things
 
 | workload | earlier reference | now | change |
@@ -1271,8 +1302,8 @@ tree-walk control flow has little leverage. The source was fully reverted.
 | trivial Rust diff, master through query work | 444,668,795 | ~206,000,000 | **-54%** |
 | `huge_cpp`, before exp17 through exp22 | 14,858,905,910 | 9,285,036,692 | **-37.5%** |
 | `huge_cpp` peak RSS, exp21 through exp22 | 696 MB | 491 MB | **-29%** |
-| focused `typing`, exp22 through exp37 | 3,115,140,742 | 2,828,276,712 | **-9.21%** |
-| focused `slow`, exp22 through exp37 | 1,881,539,982 | 642,755,405 | **-65.84%** |
+| focused `typing`, exp22 through exp39 | 3,115,140,742 | 2,824,159,538 | **-9.34%** |
+| focused `slow`, exp22 through exp39 | 1,881,539,982 | 518,537,235 | **-72.44%** |
 
 The large-input pass found a different class of wins from the original suite:
 quadratic display loops (exp13-17), redundant offset-to-line searches (exp19),
@@ -1281,7 +1312,7 @@ The post-exp22 `huge_cpp` profile is now led by line splitting, Imara histogram
 diff construction, allocator traffic, and changed-region line conversion; the
 previous hunk-end and opposite-line hash hotspots are gone.
 
-The focused exp23-37 pass first found small, composable wins in slider range
+The focused exp23-39 pass first found small, composable wins in slider range
 collection and parent-stack dispatch, then a much larger win by decomposing
 oversized graphs at similar sibling lists. It also ruled out three tempting
 graph directions on the exact target inputs: regenerating cached neighbours,

plan.md

@@ -241,10 +241,10 @@ new incremental experiment, use the latest accepted branch tip as the control.
 
 Fresh five-run `perf stat` means at the start of the focused pass:
 
-| pair | post-exp22 instructions | latest accepted after exp37 | cumulative change |
+| pair | post-exp22 instructions | latest accepted after exp39 | cumulative change |
 | --- | ---: | ---: | ---: |
-| `typing` | 3,115,140,742 | 2,828,276,712 | **-9.21%** |
-| `slow` | 1,881,539,982 | 642,755,405 | **-65.84%** |
+| `typing` | 3,115,140,742 | 2,824,159,538 | **-9.34%** |
+| `slow` | 1,881,539,982 | 518,537,235 | **-72.44%** |
 
 The original focused callgrind files were `/tmp/cg-focused-typing.out` and
 `/tmp/cg-focused-slow.out`; sampled cycle profiles were
@@ -347,6 +347,7 @@ than mixing counts across environments.
 | 36 | initialise sibling and predecessor links in one syntax traversal | -0.16% `typing`, flat on `slow` | kept |
 | 37 | initialise parent, ancestor depth, and preorder ID in one traversal | -0.16% `typing`, -0.10% `slow` | kept |
 | 38 | count content IDs during content-ID assignment | -0.07% `typing`, flat on `slow` | rejected |
+| 39 | lower similar-list graph-size gate from 1,000,000 to 45,000 | -19.33% `slow`, -0.14% `typing` | kept |
 
 Every completed experiment is committed and pushed. `results/` holds labelled
 callgrind tables through `exp13-linear-visible-width`; experiments that only
@@ -355,9 +356,9 @@ affect inline, JSON, or synthetic/very-large shapes use repeated targeted
 `huge_cpp` pair from 14.86B to 9.29B instructions (**-37.5%**), and exp22 also
 reduced its peak RSS from 696 MB to 491 MB. If the compiler, dependencies,
 profiler, or machine changed, record a fresh control binary before comparing.
-The focused exp23-37 pass reduced `typing` from 3.115B to 2.828B (**-9.21%**)
-and `slow` from 1.882B to 0.643B (**-65.84%**). Exp35's similar-list pairing is
-responsible for most of that improvement.
+The focused exp23-39 pass reduced `typing` from 3.115B to 2.824B (**-9.34%**)
+and `slow` from 1.882B to 0.519B (**-72.44%**). Exp35's similar-list pairing and
+exp39's lower decomposition gate are responsible for most of that improvement.
 
 ## Research synthesis: where larger wins can come from
 
@@ -575,9 +576,10 @@ percentages above predate exp25-31.
 
 1. **Characterise and tighten similar-list pairing.** Exp35 kept the pairing +
    descent combination after byte-identical output on all 107 available sample
-   pairs. Measure the one-million product gate and vote thresholds one at a
-   time, including values that disable pairing on `typing` or split `slow` more
-   finely. Do not weaken ambiguity or non-crossing guards merely for speed.
+   pairs. Exp39 swept the graph-size gate and retained 45,000: 40,000 changed
+   `slider_1.rs`, while 25,000 also changed `load_1.js`. Measure the minimum
+   common-unique vote threshold next, one value at a time. Do not weaken
+   ambiguity or non-crossing guards merely for speed.
 2. **Add low-overhead search-shape instrumentation.** Counters should be
    compile-time- or log-gated and excluded from timed release measurements.
    Attribute which edge types create and settle the million-state `slow`
@@ -658,7 +660,7 @@ percentages above predate exp25-31.
    rather than comparing across machines or toolchains.
 5. Choose one hypothesis from the prioritised backlog, state whether it is in
    the exact-output or diff-quality lane, change one thing, measure both pairs,
-   and immediately append exp39 (then exp40, etc.) to
+   and immediately append exp40 (then exp41, etc.) to
    `PERF_RESEARCH_LOG.md` and the state table here.
 6. Fully revert rejected source changes with `apply_patch`, but commit and push
    their log entries. For a kept change, run the wider output oracle and full

src/diff/unchanged.rs

@@ -13,7 +13,7 @@ const MOSTLY_UNCHANGED_MIN_COMMON_CHILDREN: usize = 4;
 
 /// Only attempt to decompose a possibly-changed section further when the graph
 /// size would otherwise be at least this big.
-const OVERSIZED_SECTION_MIN_GRAPH_SIZE: u64 = 1_000_000;
+const OVERSIZED_SECTION_MIN_GRAPH_SIZE: u64 = 45_000;
 /// A list smaller than this only needs to share one unique subtree with another
 /// list to be paired with it.
 const SIMILAR_LIST_MIN_DESCENDANTS: u32 = 20;