• v0.8.0 a9350ded53

    v0.8.0 Stable

    Hapto released this 2026-11-16 20:34:33 +00:00 | 22 commits to main since this release

    ducielperf 0.8.0

    Changes since 0.5.1. Drop-in replacement: copy ducielperf.dll over the old one. Every new feature can be switched off in ducielperf.cfg (in the WoW folder).

    New

    Faster frame raising: no more freezes when opening panels or popups (0.8.0)

    Showing a "toplevel" frame raises it above the rest of its strata. That covers the character and spellbook panels, StaticPopup dialogs such as group invites, the breath/fatigue timer, and many addon windows. Before raising it, the client closes the gaps between frame levels. To do that it walks its list of every frame ever created (about 17,000 with a large addon set) once per level, so a single raise took 150–975 ms. ducielperf now does it in one walk and gets the same result: same levels, same level count, same order within each level. In game the worst raise is now about 2 ms.

    • Lua: DucielPerf_GetFrameLevelStats() returns calls, fallbacks, frames moved, total ms, worst ms and frames in the list.
    • Disable: framelevel=0

    GC deferral in combat, on by default (0.7.0)

    A full Lua garbage collection takes 95–145 ms with a raid UI, and it caused 23 of 56 boss-fight hitches in a BWL/MC night. While you are in combat, collections are now postponed: the threshold is raised in 2 MB steps, up to 128 MB of extra growth per fight. When combat ends, the next allocation collects. In the ZG test: no hitches in boss fights, 60 fps.

    • Lua: DucielPerf_GCDefer([on]). DucielPerf_GetGCStats(t) now also reports deferOn, deferring, deferred, deferredFights, deferCapHits and deferPeakExtra.
    • Config: gcdefer=0 to disable, gcdefer_mb=<MB> for the cap (default 128).

    Faster SavedVariables saving at logout and /reload (0.6.0)

    The client writes each saved value with its own system call. A logout with large SavedVariables (aux-addon, pfExtend, ~8 MB each) spent about 6 s of an 8.5 s frame writing. Writes are now collected in a 256 KB buffer and written in large chunks. The files come out byte for byte the same.

    • Disable: svwrite=0

    Diagnostics

    • Watchdog line numbers (0.8.0): freeze reports now show the Lua line that is actually running. Before, the line could be stale or :-1.
    • Watchdog in multi-client setups (0.6.1): clients that are not the active window are skipped. They throttle themselves to about 4 fps, which would fill the log at low thresholds.
    • Allocation tracking (0.8.0): a vararg function's arg table is now charged to that function instead of its caller. A wrapper such as a profiler no longer pays for its callees' arg tables. Per-addon numbers from older logs aren't directly comparable.

    Tests

    New host-side tests:

    • test_framelevel: 3,000 random frame setups checked against a literal port of the client's code.
    • test_svwrite: byte-identical saves with about 300× fewer system calls.
    • test_watchdog: now also covers GC deferral and the live line number.
    • bench_alloctrack: now also covers vararg attribution.

    tests/run_tests.sh runs them all.

    Downloads