ZJIT: Pull GetEP out of GetLocal (with level > 0)

[tekknolagi]

Use the new GetEP instruction. When issuing GetLocal with level > 0, issue a GetEP and then use that GetEP as an operand to LoadField. This may need some re-organizing of GetLocal into separate instructions. Before putting up a PR, sketch out your ideas in the comments please.

[nozomemein]

Happy to take this on. I’ll share a draft idea later.

[nozomemein]

@tekknolagi
I think we could split local access by responsibility:

• Rename the current GetLocal to GetLocalWC0, and keep it for level=0 only (including the use_sp fast path).
• For level>0, emit GetEP { level } and then read the local via LoadField.

So compiling in iseq_to_hir would look like:

• getlocal_WC_0 -> GetLocalWC0 { ep_offset, use_sp, rest_param }
  • only replace GetLocal Insn. so other fast paths will remain.
• getlocal_WC_1 / getlocal(level>0) -> ep = GetEP { level }; LoadField { recv: ep, offset: local_offset, ... }
• getlocal(level == 0) -> GetLocalWC0 { ep_offset, use_sp, rest_param }

We may need to keep the vm_read_from_parent_iseq_local_count counter for each case, and update LoadField’s read effect to Memory, since it will also include frame access.

[tekknolagi]

That sounds good for level>0/no-SP. Just make sure to keep reading from EP/SP/whatever as is already indicated by the various boolean flags on the instruction. As a first step let's either leave GetLocal level==0+SP as already named or, if you want to do more, add LoadSP instruction that just returns SP register as operand and then just make it LoadSP+LoadField.

I agree about making LoadField read from Memory for now.

A follow-up diff could involve determining the memory effects of the load based on the type of the receiver, or storing an AbstractHeap on the LoadField instruction. This part probably requires a bit more discussion.

[nozomemein]

Thanks for the feedback! Makes sense. I’ll keep the current GetLocal naming as-is for now and proceed with a minimal diff. We can handle LoadSP and finer-grained load effects in follow-up PRs. I’ll put up a PR 👍 .