SPU LLVM: LS Memory Mirrors (Optimize loads/stores)

[elad335]

By allowing the SPU Local Storage to wrap around on addressing overflows around 256k+- offsets stores/loads are allowed to not mask the immediate value after the calculating the sum of addresses.
This makes it use a native, single x86 memory address calculation instruction (e.g. as a LEA instruction) instead of separating it into different instructions and masking it everytime.

An example for a start of a function which stores a few registers one after one for registers savings to stack from GOW3:
Before:
https://user-images.githubusercontent.com/18193363/87683427-be936880-c789-11ea-9f88-a94ffd2c7138.png
After:
https://cdn.discordapp.com/attachments/485204327800242177/733052633950191626/unknown.png

Notice how ($SP) is masked with 0x3fff0 only once and there are no additional LEA instructions using the new code.

Misc: Also fixes MFC transfers which went of bounds of Local Storage, making them wrap around.

[elad335]

Someone wants to perf test?

[alexloxldlol]

Master

Build

same performance

[Satan86]

no performance difference on TLOU either

[nitroos-ai]

Breaks compatibility with Uncharted Drake's Fortune, make the emulator crash with error ·F 0:00:29.917134 {PPU[0x1000000] Thread (main_thread) [0x014a7d80]} SIG: Thread terminated due to fatal error: Verification failed (e=17): (in file ../rpcs3/Emu/Cell/SPUThread.cpp:1190)
RPCS3.log.gz

[Basmurd]

Breaks compatibility with Uncharted Drake's Fortune, make the emulator crash with error ·F 0:00:29.917134 {PPU[0x1000000] Thread (main_thread) [0x014a7d80]} SIG: Thread terminated due to fatal error: Verification failed (e=17): (in file ../rpcs3/Emu/Cell/SPUThread.cpp:1190)
RPCS3.log.gz

Still ingame with no error

[nitroos-ai]

I tried it in windows 10 and it worked fine but tried the emulator with two games in Ubuntu 20.04 and the emulator crashed in both. I will try later with Ubuntu 20.04 live usb.

[paoros1]

I tried under windows 10 and it’s became unstable a lot of crash while loading game or playing with it.

[elad335]

Fixed Linux issue in #8598

[elad335]

I've succedded to repro Windows issue consistently using these added changes in spu_thread constructor:

-		const auto addr = static_cast<u8*>(utils::memory_reserve(SPU_LS_SIZE * 3));
+		auto addr0 = static_cast<u8*>(utils::memory_reserve(SPU_LS_SIZE * 3));
+		auto addr = static_cast<u8*>(utils::memory_reserve(SPU_LS_SIZE * 3));
+
+		while (addr0 + SPU_LS_SIZE * 3 != addr)
+		{
+			utils::memory_release(addr0, SPU_LS_SIZE * 3),utils::memory_release(addr, SPU_LS_SIZE * 3);
+			addr0 = static_cast<u8*>(utils::memory_reserve(SPU_LS_SIZE * 3));
+			addr = static_cast<u8*>(utils::memory_reserve(SPU_LS_SIZE * 3));
+		}

utils::shm::unmap_critical seems to be the cause for the bug, as it can release(amend) memory not within the target memory reservation, causing the assertation in utils::memory_release to trigger as it cant release memory not of the start of the base.

[VVSRevolution]

When you choose a game in KH 2.5 there is a crash and that is the error message. v0.0.11-10664 works

[elad335]

For all crashes use #8598