Redundant nullcheck in ARR_LEN/IND on top of non-null array

[EgorBo]

void Foo()
{
 var arr = new int[10];
 var len = arr.Length;
}

Codegen:

; Method Program:Foo():this
G_M52879_IG01: 
 4883EC28 sub rsp, 40
G_M52879_IG02: 
 48B928E4DFFCF87F0000 mov rcx, 0x7FF8FCDFE428 ; int[]
 BA0A000000 mov edx, 10
 E8384A865F call CORINFO_HELP_NEWARR_1_VC
 8B4008 mov eax, dword ptr [rax+08H] ;; <----
G_M52879_IG03: 
 4883C428 add rsp, 40
 C3 ret 
; Total bytes of code: 32

and other patterns where JIT should be able to detect that ARR_LEN/IND is invoked on something known to be non-null, e.g.:

void Foo(int[] arr)
{
 if (arr != null)
 {
 var _ = arr.Length;
 }
}

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

Issue Details

---

void Foo()
{
 var arr = new int[10];
 var len = arr.Length;
}

Codegen:

; Method Program:Foo():this
G_M52879_IG01: ;; offset=0000H
 4883EC28 sub rsp, 40
G_M52879_IG02: ;; offset=0004H
 48B928E4DFFCF87F0000 mov rcx, 0x7FF8FCDFE428 ; int[]
 BA0A000000 mov edx, 10
 E8384A865F call CORINFO_HELP_NEWARR_1_VC
 8B4008 mov eax, dword ptr [rax+08H] ;; <----
G_M52879_IG03: ;; offset=001BH
 4883C428 add rsp, 40
 C3 ret 
; Total bytes of code: 32

Ideally, the whole thing should be folded to nop

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

[JulieLeeMSFT]

Pushing to Future.

[BruceForstall]

 8B4008 mov eax, dword ptr [rax+08H] ;; <----

This isn't a null-check, per se -- it's loading the asked-for Length -- but I guess you're saying that we should be able to dead code eliminate this even though it appears to have a null-check exception side-effect and the value result is unused?

If so, could we then eliminate the memory allocation, too, since it would be unused?

[EgorBo]

but I guess you're saying that we should be able to dead code eliminate this even though it appears to have a null-check exception side-effect and the value result is unused?

Yes, this load is here only because of GTF_EXCEPT flag, if it's removed - the load is optimized. The allocation will also be removed

[EgorBo]

Basically, the fix is simple - extend optAssertionProp_Ind to handle GT_ARR_LEN