SuspendingClock on Windows does not suspend #63225
Conversation
grynspan
force-pushed
the
jgrynspan/63224-suspending-clock-windows
branch
from
399945e to
27c684f
Compare
|
@swift-ci please test |
stdlib/public/Concurrency/Clock.cpp
Outdated
| *seconds = 0; | ||
| *nanoseconds = 1000; | ||
| *nanoseconds = 1000000000 / freq.QuadPart; |
There was a problem hiding this comment.
QueryPerformanceFrequency() never reports zero post-Windows XP. If we're concerned about that, we can check.
stdlib/public/Concurrency/Clock.cpp
Outdated
| @@ -159,7 +153,7 @@ switch (clock_id) { | |||
| *nanoseconds = suspending.tv_nsec; | |||
| #elif defined(_WIN32) | |||
| *seconds = 0; | |||
| *nanoseconds = 1000; | |||
| *nanoseconds = 100; | |||
There was a problem hiding this comment.
QueryUnbiasedInterruptTimePrecise() has a unit of 100ns, so its highest possible resolution is 100ns. Neat.
|
@swift-ci please test |
… be dynamically looked up. Also, fix the math after calling it, and avoid floating-point math on the continuous clock.
|
@swift-ci please test |
- Use 128-bit math with QueryPerformanceCounter() to avoid overflowing. - Use swift::fatalError() instead of abort() for bad clock IDs. - Use digit separators (C++14 feature, makes large integers easier to read.)
|
@swift-ci please test |
stdlib/public/Concurrency/Clock.cpp
Outdated
| // count in nanoseconds. Use 128-bit math to avoid overflowing. | ||
| DWORD64 hi = 0; | ||
| DWORD64 lo = _umul128(count.QuadPart, 1'000'000'000, &hi); | ||
| DWORD64 ns = _udiv128(hi, lo, freq.QuadPart, nullptr); |
There was a problem hiding this comment.
_udiv128() generates a GPF if the result can't fit in a 64-bit integer, but if that happens that means the computer's been up for more than 2^64 nanoseconds… that's 584 years. If anybody runs into that problem, please let my great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-great-grandkid know.
|
@swift-ci please test windows |
|
@swift-ci please test windows |
|
@swift-ci please test |
|
@swift-ci please test |
|
@swift-ci please test linux |
|
@swift-ci please test |
Always happy to help! (Danger: I might noodle on this code a bit more in the future if I see ways to improve it further.) |
|
@swift-ci please test |
|
@swift-ci please test windows |
1 similar comment
|
@swift-ci please test windows |
|
The reported Windows failure is not related to this PR. |
|
@swift-ci please test windows |
This PR replaces the implementation of
SuspendingClockon Windows with one that usesQueryUnbiasedInterruptTimePrecise(), which does suspend.I also updated
swift_get_clock_res()to produce accurate values on Windows instead of a hard-coded 1µs resolution.Raymond Chen gives a little bit of info about the various timer APIs here.
Resolves #63224.