rsx: Rewrite IO mappings (again)

[elad335]

• Move io table to rsx::thread, no longer a global variable.
• Minor optimizations in io translations: remove shift left of base mapping 1m block, obsolete validation check and a compiler hint about it.
• Fix HLE cellGcmGetTimeStampLocation (io offset passed to translation was wrong, see cellGcmGetReportDataAddressLocation).
• Make HLE gcm internal translations rely on HLE table.
• Minor bugfix of rsx::thread::pause(). (wait for previous pause request)
• Fix rsx capture re-capturing frame after a crash occured or emulation was stopped in the middle of it.
• Make cellGcm HLE thread-safe.

[Megamouse]

I thought rsx was frozen XD

[Margen67]

@Megamouse Just let it go :^)

[Xcedf]

Uncharted 2, right before intro fmv
E {RSX [0x0010008]} RSX: FIFO error: possible desync event (last cmd = 0xffffffff)
F {RSX [0x006d060]} RSX: class std::runtime_error thrown: GetAddress(offset=0xffffff, location=0xfeed0001): RSXIO memory not mapped
(in file C:\projects\rpcs3\rpcs3\Emu\RSX\RSXThread.cpp:73)

Uncharted 3
E {RSX [0x0012008]} RSX: FIFO error: possible desync event (last cmd = 0xffffffff) x2
E {RSX [0x006e454]} RSX: FIFO error: possible desync event (last cmd = 0xffffffff) x2
F {PPU[0x1000000] Thread (main_thread)} class std::runtime_error thrown: Trap! (0x731b44)

[kd-11]

Rsx is still in maintenance mode so this will be up for a while.

[elad335]

Possibly fixed.

[Megamouse]

We can merge the error code earlier if you seperate it

[Xcedf]

@elad335 yes, fixed

[elad335]

hmm can it be merged already? Honestly if it may regress anything it would be fixed at lightspeed once it reaches master due to its triviality.

[elad335]

oops

[rajkosto]

well LIKELY and UNLIKELY arent macros anymore so it would have to be transformed into [[unlikely] on the side, thats why i asked

[elad335]

Conflicts solved.

[kd-11]

Why not just check ea == 0xffffffff? It's more readable that way. Also, the decision to have ea = 0 as legal is dubious, I do not think that is legal for anything but the lv1 kernel itself. I don't mind either way but this add is pointless here for the comparison bit.

[elad335]

0 isn't compatoble with the opposite operation (ea->io), but io->ea can still use 0.

[elad335]

Also -1 has a cool trick, notice rsx_io_table::get_addr no longer checks for unmapped memory.
Thats because OR doesnt change -1, unlike 0.

[kd-11]

That's fine, but the comparison for invalid still irks me. It relies on arithmetic to infer a hard-coded constant which is messy. Can this be done in the table class with proper named constants? I think a static constexpr in the map class may work for this. The reason 0 comparison gets a pass is because it is a universal 'false' since forever, hence the special handling even in the language itself.
Alternatively get_addr may return a pair, but that is overkill.