The new Nvidia drivers have a bug where the FastReplicateTo6 function produces a lookup into the REPLICATE_TO_8 table rather than the REPLICATE_TO_6 table.
This seems to be an optimization gone wrong. Combining the logic of the FastReplicate functions seems to address the bug.
- This is used to terminate a thread asynchronously after it has been exited.
- This fixes a crash that can occur in Pokemon Sword/Shield because a thread is incorrectly closed on svcExitThread, then, the thread is destroyed on svcCloseHandle while it is still scheduled.
- Instead, we now wait for the thread to no longer be scheduled on all cores before destroying it from KWorkerTaskManager, which is accurate to HOS behavior.
- PR #7699 attempted to fix CreateServiceThread and ReleaseServiceThread to be thread safe, but inadvertently introduced a possible dead-lock.
- With this PR, we use a worker thread to manage the service thread list, allowing it only to be accessed by a single thread, and guaranteeing threads will not destroy themselves.
- Fixes a rare crash in Pokemon Sword/Shield, I've now run this game for ~12 hours non-stop and am quite confident this is a good solution for this issue.
With the current settings 2p mode in pokemon let's go wasn't showing up. By making the shake more violent we can make it appear without any effort using the keyboard
- CreateServiceThread and ReleaseServiceThread can be accessed by different threads, uses a lock to make this thread safe.
- Fixes a rare crash in Pokemon Sword/Shield that can occur when a new service thread is being created while an old one is being destroyed.
Motion stops working in Mario Tennis in swing mode if the update rate is too fast even when HW it updates at the same speed. 10ms it's the minimum period that the game needs to start working again.
Changes tab initialization to happen after the configuration is loaded,
which means that it no longer happens as member initializers in the
ConfigurePerGame constructor. Removes the cluster of
??_tab->SetConfiguration's that I added earlier to get around this issue
initially.
Fixes a regression in #6774
bunnei