[LA64_DYNAREC] Added push/pop rsp merge optimization

[ksco]

No description provided.

[ptitSeb]

I really don't like that RSP_TRACK macro that gets in many places... I'm pretty sure there are other way to do this, using PASS0 or PASS1 macro magic to set the rsp state to none by default and override this value by other macros.

[ksco]

I really don't like that RSP_TRACK macro that gets in many places... I'm pretty sure there are other way to do this, using PASS0 or PASS1 macro magic to set the rsp state to none by default and override this value by other macros.

Sure, but this is a whitelist, so we need to be explicit. A blacklist is quite complex, with not much gain.

[ptitSeb]

I really don't like that RSP_TRACK macro that gets in many places... I'm pretty sure there are other way to do this, using PASS0 or PASS1 macro magic to set the rsp state to none by default and override this value by other macros.

Sure, but this is a whitelist, so we need to be explicit. A blacklist is quite complex, with not much gain.

Exactly! That's why default value should be NOPE.
the PASS1 NEWINST can have dyn->insts[ninst].rsp_class = RSP_CLASS_NOP; added to it for example, while the PUSH1mz macro would set dyn->insts[ninst].rsp_class = RSP_CLASS_PUSH; again in PASS1 only (for example).
If you need to put RSP_TRACK to all the opcode that are not whitelisted, that's not really a whitelist, that's a blacklist imho.

[ksco]

RSP_TRACK is used to tell DynaRec this opcode is not a barrier -- it can sit in between multiple push/pop opcodes.

[ptitSeb]

RSP_TRACK is used to tell DynaRec this opcode is not a barrier -- it can sit in between multiple push/pop opcodes.

I understand. But I still think it whould be the default behaviour of an opcode (so without the need to use that macro), as you are doing a whitelist stuffs. I'm sure you can acheive the exact same result without this macro just by using PASS1 stuffs

[ksco]

I use the manual whitelist because I don't want the NOP to be the default behaviour; I want BARRIER to be the default behaviour. Marking most opcodes as NOP will pull in too many things I need to deal with, and those opcodes are not the typical ones that appear in the prologue/epilogue. I know I can make the RSP_TRACK implicit, but that really just turns this into a much bigger PR because of the need to identify every barrier.