JIT: support immediate stores

[FioraAeterna]

This is a bit of a WIP given that it.... turned out to be a lot bigger than I expected it to be, and I still need to bench it, but I'd love comments.

Basically I noticed there were tons of examples of code like this:

mov eax, 0x8
bswap eax, eax
mov [rdx+rbx], eax

This was clearly quite silly, so I decided to add support for immediate stores, a'la:

mov [rdx+rbx], 0x08000000

How hard could this be? Well, after a journey into the backpatcher, I can say "much harder than I originally thought..."

[phire]

Would be nice to fix SafeWriteRegToReg so it doesn't clobber both its argument registers. Looks like it only clobbers source if it needs to do a BSWAP instruction and it only needs to clobber addr if it needs to call the external write function.
Currently it always clobbers addr on the slow path, doing an extra add instruction.

But probably out of scope of this PR.

LGTM.

[FioraAeterna]

I've rewritten this to work with all of comex's latest changes, but I'm not sure what I should do about the register issue. If the fast path fails, we need to allocate a register for the immediate reg_value for the slow path, but it's not clear which register we should pick (RSCRATCH or RSCRATCH2 might already be used by addr, and might be requested as callee-save by the caller?).

This should work with MMU games now.

[comex]

You've already saved registers that are in use (other than RSCRATCHes, but those were never going to be saved across a function call anyway - SAFE_LOADSTORE_CLOBBER_RSCRATCH_INSTEAD_OF_ADDR is badly named - it's more about whether to assume RSCRATCH is free; frankly, I'm an insane programmer). Accordingly, you can use any register that's not going to conflict with something else, but keep in in that the result is going to be wasteful code as ABI_CallFunctionRR copies your register into ABI_PARAM2 - not that it matters for a slow case.

[FioraAeterna]

Comex, is FioraAeterna@df2fafe#diff-9aa48ec8e209a074e590018a52b78eb6R70 bad? should I not be doing that?

[comex]

I think it's fine...

[FioraAeterna]

I thought I should be avoiding using RSCRATCH across function calls though?

[comex]

It'll work if you ask for it to be saved.

[FioraAeterna]

Okay, makes sense!