Hot-idiom census across five apps (2026-09-19)

Question: which interpreter idioms are hot in every app, so that a fixed build-time handler (no runtime codegen) pays across the corpus rather than in one title. Method: test/run.js --handler-hist --handler-hist-thread=0,0,0,0,0 --handler-hist-start=A --handler-hist-stop=B --hist-json=F gives five consecutive windows from one gameplay run (JSON per window, named F.T0.<start>), tools/hot-loop-census.js groups the hot blocks, and the top pairs list (now --handler-hist-pairs=N, default 12) names the unfused handler pairs. Counts are load-immune, so the quiet box and this Mac agree exactly; the windows are byte-deterministic (Diablo's four idle windows retired 95,983,584 block entries each).

Routes: StarCraft and Heroes II are the benchmark routes in docs/re-notes/starcraft-shareware.md / heroes2-demo.md; Diablo the project_diablo_gameplay route to town; Caesar III the retry loop of test/test-caesar3-gameplay.js driven over --control-stdin (its stdout teed, because the control session swallows non-[ctl] lines); SimGolf runs headless only with --headless-gl on this Mac (the box has no GL, the game shows "Can't Create A GL Rendering Context" and its windows are a modal box). Every window was gated with a PNG of the scene.

Windows (window 3 of 5, --batch-size=100000, 240-500 batches)

Where the block entries go (mean of five windows; spread in brackets)

Four of the five spend 20-40% of their block entries in a byte-token RLE sprite/span decoder nest, and every one encodes its tokens differently (Storm's span list, Heroes' bit-6/bit-7 tokens, Diablo's signed byte, Caesar's compare ladder). There is no single nest fold to build; what these nests share is their inner idioms, which is what the pair histogram sees. SimGolf is the odd one out: no loop at all, a hot leaf called through call/ret with three pushes and three pops per visit.

Unfused idiom families (share of dispatches, window 3, top-40 pairs)

• cmp/test + Jcc is hot everywhere. test r,r + Jcc is already one handler (H404, 2-7.5% on its own); cmp r,imm + Jcc (H10->H31x/H320: SimGolf 8.4%, Caesar 3.8%, Diablo 1.8%), cmp r,r + Jcc (H19: StarCraft 2.4%, Heroes 1.7%), cmp r,[mem] + Jcc (H48->H309 Diablo 4.2%), shr r,1 + jnb (H53->H310/H311 Diablo 6.3%), cmp [mem],imm + Jcc (H51: Caesar 3.4%), test r16,r16 / test r8,imm / dec [mem] + Jcc (SimGolf 3.1%, StarCraft 1.2%, Heroes 1.6%) are all still two dispatches with the flags materialised in between.
• Token fetch is the RLE decoders' xor eax,eax / mov al,[esi] / inc esi / test al,al / js: H18->H28 (or ->H24/->H403/->H166) 1-3.5% in four apps, H28->H64 1.3-4.5%, H64->H404 4.1% in Diablo. A movzx-style zero-then-load8 handler and a load8+inc handler cover it.
• push/pop runs matter in the call-heavy apps only: SimGolf 16.6% (three pushes, three pops, pop+ret, push+call), Heroes 5.2%.

What to build, and what not

1. A decoder fusion for cmp/test + Jcc in the forms above, mirroring how $th_test_jcc (H404) was done: removes 4-15% of dispatches in every app.
2. xor r,r ; load8 and load8 ; inc fusions: 3-12% of dispatches in the four sprite-decoding apps.
3. Multi-push / multi-pop(+ret) handlers: worth it for SimGolf and Heroes only; third.

Not: another one-app nest fold before (1)-(2) exist, because the nests do not share an encoding. Estimate before measuring: a dispatch is ~8ns (bench-loops.js), and dispatch replication showed the whole per-dispatch tail is worth ~5% of CPU, so removing 10% of dispatches is a few percent, not ten. Measure each on the quiet box with a null arm (gameplay-ab-flags.js), StarCraft and SimGolf both, before landing.

Raw data: census2/ on the box (~/wa-bench/census2), logs with the 40-row pair lists, the per-window JSON, and the gate PNGs.

Outcome of recommendation (1) — 2026-09-19, same day

Built and measured: cmp r32,r32 / cmp r32,imm32 + Jcc as handler 469. Neutral on the quiet box with a null arm (StarCraft +1.3% in a 1.4% band, Heroes II −0.3% in 0.8%, Diablo −1.6% vs null in 3.3%; retired instructions moved ≤0.5%). A first layout that broke $decode_run's fall-through chains was measurably slower on all three. The fusion, its test and the full write-up are on branch cmpjcc-fusion (docs/cmp-jcc-fusion-2026-09-19.md there); nothing landed on main.

What it changes above: a fusion that only removes a dispatch is below what the box can resolve, so (1)'s remaining forms are closed, and (2) and (3) must be estimated against "does it delete memory traffic or a block transfer" before being built. The oracle also showed StarCraft's 2000-batch headless route is not self-deterministic (base vs base: 5,631 px, ~4k API calls), so use Heroes II, Diablo or SimGolf as the picture oracle.