ZJIT: Stop running trampolines and stubs through full LIR backend

[tekknolagi]

We have two different ways of generating code:

• Ruby->HIR->LIR->machine code
• LIR->machine code

They have very different needs. The former needs a register allocator, linearization, etc. The latter is often just writing one or two instructions into a buffer.

We should not pay for the full LIR pipeline cost if all we are doing is writing out a jump instruction somewhere. Additionally, we should have finer-grained control over when and where we do mprotect and icache coherence stuff.

ruby#16563 is a partial split into a new interface for sketching purposes but it's possible we should have a Compiler layer over the existing Assembler and just let Assembler be lower-level.

Thoughts?

[k0kubun]

Having a mode in Assembler that turns off some LIR passes seems fine to me. Assembler::new_with_scratch_reg is what we do today for trampolines, and it's kinda similar in the sense that it changes the backend behavior only for trampolines. I'm totally fine if that mode is replaced by something that turns off particular LIR passes like the register allocator / linearization.

[XrXr]

In theory we can do this with just CodeBlock and a tag that describes the specific instruction encoding form to pick and completely bypass Assembler. Basically, we want linker relocations. But it's harder to implement than paring down Assembler.