gh-148211: decompose [_POP_TWO/_INSERT_2]_LOAD_CONST_INLINE_BORROW in JIT

[NekoAsakura]

• Issue: Refactor the JIT uops #148211

[NekoAsakura]

Conflicts already 😲 JIT is moving so fast.
Next PR would be: _POP_CALL(_ONE/_TWO)

[Fidget-Spinner]

Awesome, another nice simplification. Thanks!

[Fidget-Spinner]

Conflicts already 😲 JIT is moving so fast. Next PR would be: _POP_CALL(_ONE/_TWO)

Yeah there're like 5 active contributors sending stuff to the JIT almost daily. It's a happy problem :).

[Fidget-Spinner]

@NekoAsakura I would like to add you to the contributors in the What's New page. Would you be alright with that and do you mind if I use your Neko Asakura handle?

[NekoAsakura]

@NekoAsakura I would like to add you to the contributors in the What's New page. Would you be alright with that and do you mind if I use your Neko Asakura handle?

That'd be lovely, thanks! Would it be alright to use a different name though?

[Fidget-Spinner]

That'd be lovely, thanks! Would it be alright to use a different name though?

Yeah of course.

[Fidget-Spinner]

I'll open the PR to add new JIT contributors. Give me a moment. Will tag you in it.

[Fidget-Spinner]

This seems a little inefficient, but I think it should be fine as it's just a register to register move with stack caching in the JIT.

[NekoAsakura]

Are there any better practices for this?

[Fidget-Spinner]

don't think so: no matter what you do, you probably have to shuffle the register anyways. We can remove some movs, but I doubt that will speed up things by that much on modern CPUs considering register renaming is a thing.

[NekoAsakura]

Sound about right.

[Fidget-Spinner]

I verified that for the following:

def constant():
    for _ in range(4096):
        a = 1
        b = 2
        c = a + b

constant()

We emit the following:

  25 OPTIMIZED: _LOAD_CONST_INLINE_BORROW_r23 (0, target=21, operand0=0x5fdb9ad03c88, operand1=0)
  26 OPTIMIZED: _SWAP_3_r33 (3, target=21, operand0=0, operand1=0)
  27 OPTIMIZED: _SWAP_2_r33 (2, target=21, operand0=0, operand1=0)

Which means everything is in registers. So that's great.

[NekoAsakura]

Next PR would be: _POP_CALL(_ONE/_TWO)

Would you like me to decompose _POP_TWO as well? I see op_without_pop already peels it during refcount elimination, so we'd just be making that explicit.

[Fidget-Spinner]

Next PR would be: _POP_CALL(_ONE/_TWO)

Would you like me to decompose _POP_TWO as well? I see op_without_pop already peels it during refcount elimination, so we'd just be making that explicit.

Yes please, the op elimination pass should be as simple as possible. We want this so any optimizations we do in the future are easier and more effective.