Notifier has one thread per instance because the callbacks must be
asynchronous. Watchdog callbacks can be synchronous, so this overhead
can be done away with via a scheduler thread akin to what the HAL
Notifier does.
Imported from https://github.com/akrzemi1/Optional with minor changes:
- Compiler conditional simplifications (we only use recent versions)
- Move from std::experimental to wpi namespace
- Change tests to integrate with Google Test
Update LLVM use cases.
This allows save and restore of camera settings. The restore is a bit
smarter than the save.
* Fix mime types in mjpeg server
* wpiutil: WPI_LOG: Make sure level is an unsigned int
This is set to the FPGA clock by HAL_Initialize. Note this change means
that libuv loops should not be started until after HAL_Initialize is called (on the Rio).
Non-Rio functionality is unchanged.
Static destruction order is not well defined, so it was possible for outs()
or errs() return value to be destroyed even while other code was running,
resulting in a crash. Instead dynamically allocate the static so the
destructor never runs. While this technically leaks, valgrind generally
supresses such leaks as the data is still "reachable" from the static pointer.
Other handles can only be used within the loop itself, but Async is intended
to be used from another thread. This introduces the possibility of a race
condition between the loop being destroyed and the Async being destroyed.
Change Async to keep a weak reference to a loop and check it before performing
libuv operations.
This allows the called function to pass along the promise to another
asynchronous callback.
To avoid memory allocations, add a home-rolled, simplified, non-allocating
version of std::promise and std::future as wpi::promise and wpi::future.
Make wpi::condition_variable typedef to std::condition_variable_any if
wpi::mutex typedefs to priority_mutex.
priority_condition_variable was originally intended as a copy of
std::condition_variable_any that also returned the internal handle like
std::condition_variable. This was needed because NetComm required a
pthread_cond_t. We no longer use it anywhere.
Its args were specialized for priority_mutex, but
std::condition_variable_any supports this and more through
templatization.
This is a breaking change as it makes Async a template (e.g. Async<> must
be used instead of just Async). When data parameters are provided, an
internal mutex and vector is used to hold the parameter packs until the loop
runs.
This fixes two real bugs:
- TimedRobot had a m_period that was hiding the IterativeRobotBase m_period
and was not getting initialized.
- PDPSim was swapping two parameters to getCurrent()
A hash is stored for each native library with the name libraryName.hash.
If the library is not found on the system search path, it is extracted to a cache directory.
Extracted libraries are named with the hash appended, so the library will not be
re-extracted if one with the same hash already exists.
Hashing without the hash file requires double traversing if the file is not in the cache,
but it is still faster than creating a new file in most cases. This won't be needed
after opencv is updated to provide a hash as well.
