spu/arm64: clean up assembly code generation #12608
Conversation
|
Hi, you mention a "leak" but the affected code is called once during program initialization. (they are static variables) |
Neko told me to fix it after the SPU arm port was landed. Also, some of this code is called every time we branch to a new "function". We can run out of the 256MB allocated JIT memory in some games with large SPU code size as a result of the memory bloat (some Arm code here is a lot bigger than x86 versions in code size). |
| @@ -1636,18 +1737,24 @@ void spu_recompiler_base::dispatch(spu_thread& spu, void*, u8* rip) | |||
|
|
|||
| atomic_storage<u64>::release(*reinterpret_cast<u64*>(rip - 8), result); | |||
| #elif defined(ARCH_ARM64) | |||
| auto jump_instrs = build_function_asm<spu_function_t>("", [](native_asm& c, auto& args) | |||
There was a problem hiding this comment.
Oh I see, here is the leak. I think these are the only places where it should be changed. (bytecode is much less readable and is more difficult to work with)
There was a problem hiding this comment.
I see. I've removed the other changes.
There was a problem hiding this comment.
And keep the rebuild_ubertrampoline changes as well.
sguo35
force-pushed
the
arm64
branch
2 times, most recently
from
0f1dca9 to
cd03436
Compare
|
I also readded the |
Clean up asmjit usage so we don't unnecessarily allocate memory anymore for SPURecompiler functions.
Clean up asmjit usage so we don't unnecessarily allocate/leak memory anymore for SPURecompiler functions.
Tested with spurs test and 2 basic games and doesn't appear to cause any regressions.