Codegen optimization of `(a << 4) | (b >> 4)`

[deeprobin]

@deeprobin feel free to file an issue. Related one is #13816 (you can use PR that closed it as a foundation for your PR if you want to contribute 🙂)

Originally posted by @EgorBo in #69884 (comment)

---

See https://godbolt.org/z/T8eqE9chv

LLVM

        shr     sil, 4
        shl     dil, 4
        or      dil, sil
        movzx   eax, dil
        ret

.NET JIT

       movzx    rax, dil
       shl      eax, 4
       movzx    rdi, sil
       sar      edi, 4
       or       eax, edi
       ret      

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

@deeprobin feel free to file an issue. Related one is #13816 (you can use PR that closed it as a foundation for your PR if you want to contribute 🙂)

Originally posted by @EgorBo in #69884 (comment)

---

See https://godbolt.org/z/T8eqE9chv

LLVM

        shr     sil, 4
        shl     dil, 4
        or      dil, sil
        movzx   eax, dil
        ret

.NET JIT

       movzx    rax, dil
       shl      eax, 4
       movzx    rdi, sil
       sar      edi, 4
       or       eax, edi
       ret      

Author:	deeprobin
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[deeprobin]

(feel free to assign me)

[hopperpl]

Assuming rsi/rdi are byte parameters (as in the example) and all high-bits therefore are zero, and additionally that the result is not masked back to byte (as in the example)... use this... otherwise all examples return nonsense.

        shr     rsi, 4                             ; can use shr sil, 4 as well, but that gives a stall
        lea     eax, [rdi * 4 + rsi]
        ret

This will return a masked byte value. Yes, lea is a very nice C = A + B instruction to keep both source registers. And on top, it's free, as the memory address unit is bored.

        shr     sil, 4
        shl     dil, 4
        lea     eax, [rdi + rsi]
        ret

But I think that's overkill for special optimization.

[deeprobin]

@EgorBo
This is my first change that should change something about the JIT logic. How should I start?
Is there any guide for the JIT changes?

[EgorBo]

@deeprobin it turns out that this opt is not as simple for first-time-contributors (and not useful) as I thought so I'll find something better for you. For now you can start from building everything and follow steps in https://github.com/dotnet/runtime/blob/main/docs/workflow/debugging/coreclr/debugging.md

[deeprobin]

@EgorBo Thank you 👍🏼