`performance.timeOrigin` was being set from when JS started executing,
but `op_now` measures from an `std::time::Instant` stored in `OpState`,
which is created at a completely different time. This caused
`performance.timeOrigin` to be very incorrect. This PR corrects the
origin and also cleans up some of the timer code.
Compared to `Date.now()`, `performance`'s time origin is now
consistently within 5us (0.005ms) of system time.
Remove `--allow-hrtime` and `--deny-hrtime`. We are doing this because
it is already possible to get access to high resolution timers through
workers and SharedArrayBuffer.
Co-authored-by: Bartek Iwańczuk <biwanczuk@gmail.com>
Also removes permissions being passed in for node resolution. It was
completely useless because we only checked it for reading package.json
files, but Deno reading package.json files for resolution is perfectly
fine.
My guess is this is also a perf improvement because Deno is doing less
work.
Improves #19100
Fixes #20356
Replaces #20428
Changes made in deno_core to support this:
- [x] Errors must be handled in setTimeout callbacks
- [x] Microtask ordering is not-quite-right
- [x] Timer cancellation must be checked right before dispatch
- [x] Timer sanitizer
- [x] Move high-res timer to deno_core
- [x] Timers need opcall tracing
Splitting the sleep and interval ops allows us to detect an interval
timer. We also remove the use of the `op_async_void_deferred` call.
A future PR will be able to split the op sanitizer messages for timers
and intervals.
Rewrites 3 ops that used "op(deferred)" to use "op2(async(lazy))"
instead.
This will allow us to remove codepath for handling "deferred" ops in
"deno_core".
If a timer is requested with <=100ms resolution, request the high-res
timer. Since the default Windows timer period is 15ms, this means a
100ms timer could fire at 115ms (15% late). We assume that timers longer
than 100ms are a reasonable cutoff here.
The high-res timers on Windows are still limited. Unfortuntely this
means that our shortest duration 4ms timers can still be 25% late, but
without a more complex timer system or spinning on the clock itself,
we're somewhat bounded by the OS' scheduler itself.
