Re: [PATCH v3 0/3] MR7256: ntdll: monotonic_counter and NtQuerySystemTime performance improvements on macOS.

[Brendan Shanks (＠bshanks)]

On Tue Feb  4 22:24:37 2025 +0000, Marc-Aurel Zent wrote:
> Indeed, but this commit essentially inlines the call to
> `__commpage_gettimeofday()` inside `gettimeofday()`, wheres the latter
> still incurs of overhead of 23 CPU cycles before calling into
> `__commpage_gettimeofday()`, if I counted correctly. 
> It actually does more than inlining though, `gettimeofday()`, will do a
> syscall if `__commpage_gettimeofday()` fails, and use that value instead.
> This can happen rarely (about 1 in a million in my testing), probably
> when page time values that are updated asynchronously from the kernel
> are read simultaneously.
> As the commit is right now, it effectively tries to read from the
> commpage *twice*, making that chance of doing a syscall astronomically low.
> Also it is still measurably faster to use `__commpage_gettimeofday()`
> directly (benchmarked at 10000000 iterations through Rosetta):
> ```
> clock_gettime() Duration: 0.290506 seconds
> gettimeofday() Duration: 0.233024 seconds
> __commpage_gettimeofday() Duration: 0.217468 seconds
> ```
> IMO, this feels like a micro-optimization that isn't worth the complexity and risk to save 23 cycles per call. If it makes a meaningful difference in a game benchmark it could be worth it, but otherwise it feels better to chase bigger gains.

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/7256#note_93758