ZJIT: Introduce SCRATCH2 for IncrCounter on arm64

[k0kubun]

This PR addresses Max's suggestion at #14357 (comment).

Adding SCRATCH2 allows arm64_emit to emit IncrCounter without relying on arm64_split, which lets compile_side_exits remove the cfg!(target_arch = "aarch64") branch. I don't think taking Opnd::const_ptr in IncrCounter as Max suggested isn't worth it, which will duplicate more code in arm64_emit, so it still takes Opnd::Mem.

It generally doesn't seem like a good idea to waste a register when you can save it with no run-time overhead, especially when it's for --zjit-stats. But at the same time, it seems useful to remove an extra register pressure that the other architecture doesn't have for the same LIR instruction.

x18 is a platform register. Since we control the use of every register in JIT code, it seems safe to use it as an extra caller-saved register. We don't seem to have any more caller-saved registers.

[XrXr]

x18 is a platform register. Since we control the use of every register in JIT code, it seems safe to use it as an extra caller-saved register. We don't seem to have any more caller-saved registers.

It's not a great idea to run afoul macOS's platform rules for this https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms#Respect-the-purpose-of-specific-CPU-registers

It's hard to predict what can go wrong. Maybe they use it in the kernel for something in which case we actually don't have full control over that register because we could be preempted in the region of code that uses the register.

[k0kubun]

https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms#Respect-the-purpose-of-specific-CPU-registers
The platforms reserve register x18. Don’t use this register.

Heh, interesting. Thanks for pointing that out.

If we don't have an additional caller-saved register, I'd rather leave the cfg!(target_arch = "aarch64") branch in compile_side_exits as is. Assembler is not designed to hide abstraction leaks after split/alloc_regs in the first place. We should treat compile_side_exits like part of the emit pass, which happens to have slightly richer API.