[B2-devel] Faster audio in emulated mode

[Gwenole B. <gb...@di...>]

Hi,

I have committed a few arrangements to make the JIT generated code 
somewhat reentrant. i.e. this means it is now possible to recursively 
generate and invoke compiled code. The reasoning was as follows:

1) NanoKernel interrupts could already run in a nested context from 
main execution loop. This is because we could run the interrupt only 
when we exited from a generated block. So, no problem for this one.

2) For nested Execute68k() compilation and invokation, that's trickier. 
Imagine we have the following scenario: ppc execution (toplevel loop) 
-> Execute68k (1) -> Execute68k (2). If Execute68k (2) were to 
invalidate the translation cache and possibly overwrite code from 
Execute68k (1) or toplevel loop return point, you are dead. The trick 
was to handle EmulOps on trampoline code that is not part of the real 
translation cache. That way, even if cache is invalidated, we could 
still run the continuation code at EmlOp return.

In summary, this is a *massive* performance improvement for 
applications that cause a lot of Execute68k(). e.g. audio related 
applications like PlayerPRO, where we end up to around 45% of time 
spent in Execute68k(). It's completely unusable without this 
optimization.

BTW, if you try PlayerPRO with current CVS code, it won't work, it 
needs a special workaround to avoid nested NanoKernel interrupt routine 
call. I still don't know why this would happen. It's not a problem in 
native mode because the whole CPU context is preserved. I could do the 
same with the emulator, but I still think this should not happen.

Bye,
Gwenole.