Repository navigation
Native register ABI regresses generated-code throughput ~5-10x (MKDD 51 -> 10 fps) #24
Description
Activity
Correction: the headline numbers in this issue do not hold up. Closing.
The original measurements were single first-runs on freshly generated modules, which include shader-compilation warmup, on a machine that shows large run-to-run variance. Re-measured warm and interleaved today:
title / scene current main PR-era build verdict MKDD bench.sav(interleaved 2+2)45.8, 47.1 fps 39.1, 52.3 fps no signal; box variance dominates Colosseum Phenac, warm 24.2, 27.1 fps 25-29 fps equal Luigi's Mansion foyer, uncapped 45.0 fps (1.51x) 44.1 fps (1.50x) equal The "51 -> 10" and "28 -> 2.4" figures were cold-run artifacts, not the native ABI. Apologies for the noise -- the ABI gate suggestion stands only as a nice-to-have for future A/B work, not as a regression need.
Method note for anyone benchmarking these builds: never take the first run on a newly generated module as a measurement; it compiles shaders.
1bec355("improve llvm backend") and0137523("Add native register ABI") incorporate the changes from #21, #22 and #23 -- thanks for folding those in. Measuring currentmainagainst the PR branches on the same runtime build and pinned savestates, though, shows a large performance regression:bench.sav, cappedCounters for the Colosseum runs show the same per-cycle hook-fallback rate on both builds, but guest throughput collapses ~12x on main (1.09G cycles retired vs 13.3G in an equal window), so the cost is in the generated code itself rather than in fallback traffic. The native ABI applies automatically per function via
nativeABIFlags()with no opt-out flag, so there is no way to A/B it from a built module -- an env or CLI gate would make this easy to bisect.Setup notes for reproducing: Windows x86-64, clang toolchain modules, runtime with ExpansionPak/RecompCore#18 applied (that PR is the delivery half of #23's fix -- without it the mtmsr exit in
emitStateWritecreates the boundary but pending external interrupts still wait for slice granularity, and Colosseum's retrace wait starves back to ~0.25 fps).