[JIT]: Codegen for manual devirtualization

[omariom]

This code (sharplab)..

    public T WithManualDevirt(int index)
    {
        var listStore = _listStore;
        
        if (listStore is SingleItemList<T> singleItem)
            return singleItem.EntryAt(index);
            
        return listStore.EntryAt(index);
    }

generates the following assembly:

       mov rcx, [rcx+8]
       mov rax, rcx
       test rax, rax
       je short L001d
       mov r8, 0x7fff4a8ad9c0
       cmp [rax], r8
       je short L001d
       xor eax, eax
L001d: test rax, rax
       je short L0026
       mov eax, [rax+8]
       ret
L0026: mov rax, [rcx]
       mov rax, [rax+0x40]
       mov rax, [rax+0x20]
       jmp rax

I think it could be simpler, with less tests and jumps

       mov rcx, [rcx+8]
       mov rax, rcx
       test rax, rax
       je short L0026
       mov r8, 0x7fff4a8ad9c0
       cmp [rax], r8
       jne short L0026
       mov eax, [rax+8]
       ret
L0026: mov rax, [rcx]
       mov rax, [rax+0x40]
       mov rax, [rax+0x20]
       jmp rax

sharplab is now on 5.0, but 6.0 rc2 generates the same code.

[ghost]

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

Issue Details

---

This code (sharlab)..

    public T WithManualDevirt(int index)
    {
        var listStore = _listStore;
        
        if (listStore is SingleItemList<T> singleItem)
            return singleItem.EntryAt(index);
            
        return listStore.EntryAt(index);
    }

generates the following assembly:

       mov rcx, [rcx+8]
       mov rax, rcx
       test rax, rax
       je short L001d
       mov r8, 0x7fff4a8ad9c0
       cmp [rax], r8
       je short L001d
       xor eax, eax
L001d: test rax, rax
       je short L0026
       mov eax, [rax+8]
       ret
L0026: mov rax, [rcx]
       mov rax, [rax+0x40]
       mov rax, [rax+0x20]
       jmp rax

I think it could be simpler, with less tests and jumps

       mov rcx, [rcx+8]
       mov rax, rcx
       test rax, rax
       je short L0026
       mov r8, 0x7fff4a8ad9c0
       cmp [rax], r8
       jne short L0026
       mov eax, [rax+8]
       ret
L0026: mov rax, [rcx]
       mov rax, [rax+0x40]
       mov rax, [rax+0x20]
       jmp rax

sharplab is now on 5.0, but 6.0 rc2 generates the same code.

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

[Sergio0694]

Potentially related to #36649?
cc. @EgorBo

[EgorBo]

Yes, there were two solutions there: 1) In some cases our inlined castclass/isinst produces a bit suboptimal IR 2) Jump-threading optimization gives up on GT_PHI nodes currently

[EgorBo]

Thanks for reminding me this issue! But I think we can close it as dup of #36649