The current caching strategy had a flaw in that it tried
to lazy-lock the mutex only if g_currentAssetData was non-zero.
For this to be somewhat reliable, g_currentAssetData would have
to be volatile. But that would still not be enough since
thread-unaware code optimizations might also happen on the CPU level.
Instead of complicating the current logic more, change it to
only do caching per thread. Since QThreadStorage will take ownership
of its data, we can't let it store a pointer to QIOSAssetData directly
since we need to control the life time of QIOSAssetData using deleteLater.
Change-Id: I2c3ffb3257ec2bdec8be71a3d63f666ab33b5277
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@theqtcompany.com>
6eaee855c7