Optimize BOX+UNBOX for "T? -> T" and "T -> T?"

[EgorBo]

Remove heap allocations for such box+unbox patterns when nullable is involved.

TTo Test<TFrom, TTo>(TFrom o) => (TTo)(object)o;

Current codegen for Test<int?, int> and Test<int, int?>:

; Assembly listing for Test<int?, int>
G_M61296_IG01:
       push     rbx
       sub      rsp, 48
       mov      ebx, ecx
G_M61296_IG02:
       mov      rcx, 0x7FF8348774D0
       call     CORINFO_HELP_NEWSFAST     <-------------- heap allocation!
       mov      dword ptr [rax+0x08], ebx
       mov      r8, rax
       lea      rcx, [rsp+0x28]
       mov      rdx, 0x7FF834B4B750
       call     CORINFO_HELP_UNBOX_NULLABLE
       mov      rax, qword ptr [rsp+0x28]
G_M61296_IG03:
       add      rsp, 48
       pop      rbx
       ret      
; Total bytes of code 59


; Assembly listing for Test<int, int?>
G_M21085_IG01:
       sub      rsp, 40
       mov      qword ptr [rsp+0x30], rcx
G_M21085_IG02:
       lea      rdx, [rsp+0x30]
       mov      rcx, 0x7FF83614B750
       call     CORINFO_HELP_BOX_NULLABLE     <-------------- heap allocation!
       mov      eax, dword ptr [rax+0x08]
G_M21085_IG03:
       add      rsp, 40
       ret      
; Total bytes of code 37

New codegen for Test<int?, int> and Test<int, int?>:

; Assembly listing for Test<int?, int>
G_M61296_IG01:
       push     rax
G_M61296_IG02:
       mov      byte  ptr [rsp], 1
       mov      dword ptr [rsp+0x04], ecx
       mov      rax, qword ptr [rsp]
G_M61296_IG03:
       add      rsp, 8
       ret      
; Total bytes of code 18


; Assembly listing for Test<int, int?>
G_M21085_IG01:
       sub      rsp, 40
       mov      qword ptr [rsp+0x30], rcx
G_M21085_IG02:
       cmp      byte  ptr [rsp+0x30], 0
       je       SHORT G_M21085_IG04
       mov      eax, dword ptr [rsp+0x34]
G_M21085_IG03:
       add      rsp, 40
       ret      
G_M21085_IG04:
       call     CORINFO_HELP_THROWNULLREF
       int3     
; Total bytes of code 31

[dotnet-policy-service]

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

[EgorBo]

@MihuBot

[EgorBo]

@MihuBot

[EgorBo]

@AndyAyersMS @jakobbotsch @dotnet/jit-contrib PTAL, not too many diffs (SPMI is empty due to missing contexts I add here, but there are hits in tests and a few in libs). As part of the effort to make Nullable a bit better.

I am going to re-use impStoreNullableFields for #104954 as well.

NOTE: I am using raw IND/BLK here instead of LCL_FLD because of Nullable<struct> and for simplicity

[jakobbotsch]

NOTE: I am using raw IND/BLK here instead of LCL_FLD because of Nullable<struct> and for simplicity

I don't really follow -- why is that simpler?

[EgorBo]

NOTE: I am using raw IND/BLK here instead of LCL_FLD because of Nullable<struct> and for simplicity

I don't really follow -- why is that simpler?

because apparently it's not that simple to do LclFld store/load for a struct with a T field where T could be struct? Like when I do gtNewStoreLclFldNode(nullableObjTmp, value->TypeGet(), valueOffset, value); - how do I set layout for this value field?

[EgorBo]

@jakobbotsch thanks for the suggestions! I rewrote to LCL_FLD, I guess lack of ClassLayout arg in gtNewLclFldNode confused me 🙂 Can't apply your suggestions to the new version