Skip covariant checks for ldelema T[] with T exact

[EgorBo]

Closes #65789

public sealed class C
{
    C[] _array;

    C Test() => Volatile.Read(ref _array[0]);
}

Was:

; Assembly listing for method C:Test():C:this
       sub      rsp, 40
       mov      rcx, gword ptr [rcx+08H]
       xor      edx, edx
       mov      r8, 0x7FF85CB4CE28      ; C
       call     CORINFO_HELP_LDELEMA_REF
       mov      rax, gword ptr [rax]
       add      rsp, 40
       ret      
; Total bytes of code 33

Now:

; Assembly listing for method C:Test():C:this
       sub      rsp, 40
       mov      rax, gword ptr [rcx+08H]
       mov      edx, dword ptr [rax+08H]
       test     rdx, rdx
       je       SHORT G_M7456_IG04
       mov      rax, gword ptr [rax+10H]
       add      rsp, 40
       ret      
G_M7456_IG04:
       call     CORINFO_HELP_RNGCHKFAIL
       int3     
; Total bytes of code 31

[ghost]

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

Issue Details

---

Closes #65789

public sealed class C
{
    C[] _array;

    C Test() => Volatile.Read(ref _array[0]);
}

Was:

; Assembly listing for method C:Get1():C:this
       sub      rsp, 40
       mov      rcx, gword ptr [rcx+08H]
       xor      edx, edx
       mov      r8, 0x7FF85CB4CE28      ; C
       call     CORINFO_HELP_LDELEMA_REF
       mov      rax, gword ptr [rax]
       add      rsp, 40
       ret      
; Total bytes of code 33

Now:

; Assembly listing for method C:Get1():C:this
       sub      rsp, 40
       mov      rax, gword ptr [rcx+08H]
       mov      edx, dword ptr [rax+08H]
       test     rdx, rdx
       je       SHORT G_M7456_IG04
       mov      rax, gword ptr [rax+10H]
       add      rsp, 40
       ret      
G_M7456_IG04:
       call     CORINFO_HELP_RNGCHKFAIL
       int3     
; Total bytes of code 31

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

[EgorBo]

A few diffs. Size regressions on arm64 because of "heavier" bound checks, but should be clean perf improvement/

@dotnet/jit-contrib @AndyAyersMS PTAL

[markples]

I'm assuming that unverifiable/unsafe code, or specifically code with runtime behavior that actually deviates from the static types, has undefined behavior. In other words, we don't need to throw the exception on an object[] that has been misrepresented as a string[] and then a ldelem.a string performed on it. (This makes sense since such optimizations would then be impossible; it's just against a strict reading of the specification.)

[AndyAyersMS]

Shouldn't this also be conditional on opts.OptimizationEnabled() ? I see you had a minopts diff...

[EgorBo]

Shouldn't this also be conditional on opts.OptimizationEnabled() ? I see you had a minopts diff...

Addressed

[EgorBo]

In other words, we don't need to throw the exception on an object[] that has been misrepresented as a string[] and then a ldelem.a string performed on it.

I'm not sure what is going to happen (and what is expected) in this case so I'm only optimizing a valid case

[markples]

I believe this would optimize that case. Something like

.locals string[] myvar

newarr obj
stloc myvar // type inconsistency, assigning object[] to string[]
// maybe more code to block any kind of optimization between the stloc and ldloc
ldloc myvar
ldc.whatever
ldelema string

or even worse if unsafe part is in a caller... but, as I said, I assume that this isn't a concern because otherwise all type information is useless but just wanted to bring it up in case.

[EgorBo]

I assume this case should fail on the stloc step

[markples]

Oh, do we reject (at least some kinds of) unverifiable code now? This is just moving an object reference around; it's only if one looks at the so-called verification types does it break down. (I guess I should just try it.. I'd be happy to learn this.)

This definitely could be done with Unsafe.As, but the docs there are clear about improper use being undefined.

[markples]

It looks like the standard pattern here is to goto ARR_LD. It would require the pop to be split to impStackTop/impPopStack(2) but pick up the various cases there (unless you're intentionally skipping those here?)

[EgorBo]

Thanks! Addressed

[EgorBo]

SPMI failures are due to recent JIT/EE guid change