Introduce RuntimeHelpers.GetObjectHeader and Optimize Object.GetHashCode() with it

[EgorBo]

A GC-hole free version of #55273, should slightly speed up object.GetHashCode, e.g.:

Example:

[MethodImpl(MethodImplOptions.NoInlining)]
int Test(Prog o) => o.GetHashCode();

Current codegen:

; Method Prog:Test(Prog):int:this (FullOpts)
       sub      rsp, 40
       cmp      byte  ptr [rdx], dl
       mov      rcx, rdx
       call     System.Runtime.CompilerServices.RuntimeHelpers:GetHashCode(System.Object):int
       nop      
       add      rsp, 40
       ret      
; Total bytes of code: 20

New codegen:

; Method Prog:Test(Prog):int:this (FullOpts)
       sub      rsp, 40
       cmp      byte  ptr [rdx], dl
       mov      eax, dword ptr [rcx-0x04] ;; <---
       mov      edx, eax
       and      edx, 0xC000000
       cmp      edx, 0xC000000
       jne      SHORT G_M29166_IG04
       and      eax, 0x3FFFFFF
       jmp      SHORT G_M8578_IG05
G_M8578_IG04:  ;; offset=0x0020
       mov      rcx, rdx
       call     System.Runtime.CompilerServices.RuntimeHelpers:InternalGetHashCode(System.Object):int
G_M8578_IG05:  ;; offset=0x0028
       nop      
       add      rsp, 40
       ret      
; Total bytes of code: 46

This should allow us to convert Monitor.Enter/Exit to C# as well (thin lock, etc).

[dotnet-policy-service]

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

[huoyaoyuan]

I tried non-intrinsic version ealier and measured a significant regression.

[EgorBo]

I tried non-intrinsic version ealier and measured a significant regression.

with pinning? no wonder then 🙂

[huoyaoyuan]

with pinning? no wonder then 🙂

No pinning. I was not sure about how GC deals with exterior pointer. I reverted it then in #97641

[EgorBo]

with pinning? no wonder then 🙂

No pinning. I was not sure about how GC deals with exterior pointer. I reverted it then in #97641

I don't see why my version could be slower (and it's not according to benchmarks) - the codegen has a small CQ problem (#101000) but I am not even sure it can change the result.

[huoyaoyuan]

I don't see why my version could be slower

Yeah, I realized it may need intrinsic handling like retrieving method table.

[EgorBo]

Looks like this PR is a regression when object has a syncblock and I don't think I can surface PassiveGetSyncBlock into C#

[jkotas]

This should allow us to convert Monitor.Enter/Exit to C# as well (thin lock, etc).

We would be need both reads and interlocked writes for this.

To enable rewriting methods like there in C#, I think it would be better to look into general pinning optimizations (#63397) rather than specialized intrinsics.

[jkotas]

surface PassiveGetSyncBlock into C#

It can be done once we have a good reason to.