MW3 profile after uop merges

Historical source last updated, 2026-10-12. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

Measured frozen commit 837f0a74 on a dedicated ASCII box, 2026-09-30. The result is still about 22 FPS in the stationary cockpit route. Guest execution remains the largest frame component; this is not an isolated A/B measurement of a merge's performance effect.

The follow-up guest and Chrome native disassembly uses this exact module and browser version to explain the hot interpreter paths.

Runtime and method

Box bx_4r5uzdwv, AMD Ryzen 9 9950X, 4 assigned vCPUs, 8 GB, Chrome 151.0.7922.108. CDP confirms ANGLE SwiftShader, not physical GPU rasterization. Both the earlier report and this one used software-backed WebGL.

The repository was exported at the specified commit and built in isolation. The fixture adds diagnostic tools and untracked MW3 assets, MSVCRT and MFC42. One host correction is necessary: the committed lib/guest-worker.js puts the lazy-enable if inside the send({ready,...}) object, a syntax error. The isolated fixture moves that if before send, matching the correction already present in the shared working tree, without its unrelated APC edits. No production WASM, shared source, or canonical build was changed here.

Three 20-second windows follow verified cockpit entry and warmup. Window 0 has frame/wait/transfer timers but no CPU sampling. Windows 1 and 2 sample the page and all workers at 500 microseconds. No screenshots or polling occur inside a measured window. The single surviving guest worker supplies presents; there are no thread lifetime changes in the windows.

Frame breakdown

Window Presents FPS p95 interval Execute excluding waits Wait Other
0 460 22.80 55.64 ms 27.71 ms 13.61 ms 2.54 ms
1 476 23.61 50.54 ms 28.25 ms 11.87 ms 2.23 ms
2 435 21.50 57.26 ms 31.20 ms 12.63 ms 2.67 ms

The three time columns partition guest wall time, including scheduling and host imports; they are not hardware CPU-cycle measurements. Geometry varies from 1,978 to 2,230 triangles/present. One-minute load at the boundaries is 2.74–3.07. All windows have zero GPU-command errors, software-path fallbacks, and dirty audit misses. Here “zero fallbacks” describes the emulator's renderer route; Chrome's WebGL backend itself is SwiftShader. The final screenshot was inspected and retains textured terrain, cockpit, sky and HUD.

An independent unprofiled control (postmerge-control1, two 20-second windows, neither whole-profile nor transfer-profile enabled) records 23.13/24.19 FPS, p95 51.77/48.79 ms, zero renderer errors and dirty-audit misses. Geometry is 2,232/1,847 triangles per present. The first window has 120 renderer fallback events, the second zero. This corroborates the frame-rate scale but cannot isolate instrumentation overhead because the workloads differ.

Sampled hotspots

Self time estimates, normalized to the corresponding measured present rate:

Guest function Window 1 Window 2
uop_fast 3.88 ms 4.23 ms
x87_island_fast 3.27 ms 3.73 ms
branch_end_at 2.40 ms 2.67 ms
th_load32_rop 1.19 ms 1.31 ms
fpu_exec_mem 0.98 ms 1.03 ms
th_store32_rop 0.74 ms 0.77 ms
th_uop_enter 0.63 ms 0.70 ms

WASM accounts for 25.01/27.66 sampled ms per present; host JS outside waits accounts for 3.38/4.05 ms. These sampled estimates overlap the explicit wall partition above and must not be added to it.

Render-worker readPixels spans are 5.00/5.58 ms per present and color-update spans 2.09/2.32 ms, with more than 99.5% of each overlapping guest waits. Renderer getParameter still samples at 2.26/2.22 ms and getError at 2.10/2.27 ms; the latter is already inside color-update timing.

Comparison limits and next work

The historical profile was 21–22 FPS with guest execution 29–30 ms and waits 13–14 ms. This new profile has the same broad bottleneck distribution. Different launch geometry and a restored machine prevent a causal speedup claim.

A same-host historical-WASM control was attempted but failed to link: host.queue_user_apc is required by that frozen binary and absent from the clean committed host. The old measurement included uncommitted APC support. That failed control and two fixture-setup failures are excluded from results.

The x87 island predecode optimization was already in the historical baseline. The next guest opportunity remains mixed x87/integer loop coverage described in the loop investigation. The repeated renderer-name query remains an independent bounded caching experiment.

Command:

DISPLAY=:0 CHROME=/usr/bin/google-chrome node tools/bench-lazy-games.js \
  --app=mw3 --shipped-default --whole-profile --transfer-profile \
  --seconds=20 --samples=3 --wasm=build/postmerge.wasm \
  --out=build/lazy-games/postmerge-profile3