mirror of
https://github.com/software-mansion/argent.git
synced 2026-09-14 19:27:14 +08:00
Ignacy Łątka
1f69f5fedd
## Summary Found while exercising the new Android (Perfetto) native profiler end-to-end on an emulator: the **measurement** is solid, but the **interpretation** layer didn't account for the emulator environment or its own signals, so it surfaced misleading findings/advice. This makes the Android analysis honest: - **CPU hotspots:** on the emulator the hottest frames are the goldfish/QEMU GPU-transport pipe and Linux kernel syscalls (`goldfish_pipe_read_write`, `gup_fast_fallback`, `do_syscall_64`), not app code. They were presented as actionable app hotspots with `reduce view hierarchy depth` advice. Now they're classified and labelled as emulator/kernel overhead, with advice that says so (and to re-profile on a real device); real app frames get actionable advice. - **UI hangs:** advice was always `move heavy work to background queue`, even when the printed main-thread state breakdown showed the thread was **sleeping** (state `S`) for the whole hang — i.e. a GPU/vsync wait, not CPU work. Advice is now derived from the state breakdown (blocked / runnable / executing). - **`hang_stacks`:** an off-CPU (sleeping) hang has no on-CPU stack, so the section was silently empty. It now explains the empty result and points at the state breakdown. ## Why The advice directly contradicted the tool's own state-breakdown data, and the CPU section was dominated by emulator noise — both erode trust in the report. ## Test plan - New `test/android-perfetto/analysis-advice.test.ts` and `hang-stacks-offcpu.test.ts`. - `npm test -w @argent/tool-server` — all pass; `tsc --build` + `typecheck:tests` clean. > Part of a set of native-profiler fixes found in the same session. Touches the shared `render.ts` advice paths and the Android `hang_stacks` renderer.