This means the shared memory struct is just 255 bytes longer than a few
commits ago instead of 32K.
Also introduce a function for putting path strings in the shared memory
as there was too much copied and pasted code and it was error-prone.
Also free some handles once we're done with them so they don't leak.
Well... unless we fail to get the user profile directory.
Also put freeze dumps in a more appropriately-named file.
Discussed in https://gitlab.com/OpenMW/openmw/-/issues/7455
Manual rewrite of c21695c951
As well as that, I found a bug where the crash dump path was being converted from a std::filesystem::path to a utf8-encoded std::string, then passed to the crash catcher, which converted it to a std::filesystem::path implicitly, but using the system eight-bit code page, which wouldn't usually be UTF-8. That's now fixed by passing the function that expects a path a path instead of a string.
high_resolution_clock may not give real time. MSVC implements it as
steady_clock that is basically CPU time which is usually desynchronized with
real time.
When the crash catcher catch a signal it forks to collect data about its
parent. In the process the child reinitialize the log file, which ends
up empty.
This should fix the issue where Windows Release builds (compiled as
/SUBSYSTEM:WINDOWS instead of /SUBSYSTEM:CONSOLE) can't have their
output redirected.
Basically, a console application creates a console if not given one, so
you get a console window behind OpenMW while it's running. It was
decided that this was ugly, so we set Release builds to be windows
applications, which don't get an automatic console and don't
automatically connect to a console if given one anyway.
Of course, we still wanted to actually be able to print to a console if
given one, so we manually attach to the parent process' console if it
exists, then reopen the standard streams connected to CON, the Windows
pseudo-file representing the current console.
This is a little like connecting a second wire into a dumb terminal in
that you're pumping characters into the display rather than onto a
pipeline, so output can't be redirected.
It turns out, though, that if a /SUBSYSTEM:WINDOWS application has its
standard streams redirected by the calling process, it still gets its
handles as normal, so everything starts off connected just how we want
it and we were clobbering this good setup with the straight-to-console
fix.
All we need to do to fix that is check if we've got valid standard
handles and that they go somewhere useful, and if so, avoid reopening
them once the console is attached. Simples.
