Intrinsify Interlocked.CompareExchange and Interlocked.Exchange for null

[EgorBo]

Closes #69000

object Foo(ref object a)
    => Interlocked.Exchange(ref a, null);

object Foo(ref object a, object b)
    => Interlocked.CompareExchange(ref a, null, b);
}

Was:

; Assembly listing for method Prog:Foo(byref):System.Object:this
       488BCA               mov      rcx, rdx
       33D2                 xor      rdx, rdx
       E93616AC5F           jmp      Exchange(byref,System.Object):System.Object
; Total bytes of code 10


; Assembly listing for method Prog:Foo(byref,System.Object):System.Object:this
       488BCA               mov      rcx, rdx
       498BD0               mov      rdx, r8
       4533C0               xor      r8, r8
       E9A216AC5F           jmp      CompareExchange(byref,System.Object,System.Object):System.Object
; Total bytes of code 14

Now:

; Method Prog:Foo(byref):System.Object:this
       33C0                 xor      rax, rax
       488702               xchg     gword ptr [rdx], rax
       C3                   ret      
; Total bytes of code: 6


; Method Prog:Foo(byref,System.Object):System.Object:this
       33C9                 xor      rcx, rcx
       498BC0               mov      rax, r8
       F0                   lock     
       480FB10A             cmpxchg  gword ptr [rdx], rcx
       C3                   ret         
; Total bytes of code: 11

[ghost]

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

Issue Details

---

Closes #69000

object Foo(ref object a)
    => Interlocked.Exchange(ref a, null);

object Foo(ref object a, object b)
    => Interlocked.CompareExchange(ref a, b, null);
}

Was:

; Assembly listing for method Prog:Foo(byref):System.Object:this
       488BCA               mov      rcx, rdx
       33D2                 xor      rdx, rdx
       E93616AC5F           jmp      Exchange(byref,System.Object):System.Object
; Total bytes of code 10


; Assembly listing for method Prog:Foo(byref,System.Object):System.Object:this
       488BCA               mov      rcx, rdx
       498BD0               mov      rdx, r8
       4533C0               xor      r8, r8
       E9A216AC5F           jmp      CompareExchange(byref,System.Object,System.Object):System.Object
; Total bytes of code 14

Now:

; Method Prog:Foo(byref):System.Object:this
       33C0                 xor      rax, rax
       488702               xchg     gword ptr [rdx], rax
       C3                   ret      
; Total bytes of code: 6


; Method Prog:Foo(byref,System.Object):System.Object:this
       33C0                 xor      rax, rax
       F0                   lock     
       4C0FB102             cmpxchg  gword ptr [rdx], r8
       C3                   ret      
; Total bytes of code: 8

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

[jkotas]

Interlocked.CompareExchange(ref a, b, null);

This needs write barrier for b. I do not think you can optimize it like this.

You can optimize Interlocked.CompareExchange(ref a, null, b). It is much less common though.

[EgorBo]

Interlocked.CompareExchange(ref a, b, null);

This needs write barrier for b. I do not think you can optimize it like this.

You can optimize Interlocked.CompareExchange(ref a, null, b). It is much less common though.

Thanks! Fixed, grep: https://grep.app/search?q=CompareExchange%5C%28ref%20%5Cw%2B%2C%20null%2C%20&regexp=true

[EgorBo]

/azp list

[EgorBo]

/azp run runtime-coreclr outerloop, runtime-coreclr gcstress0x3-gcstress0xc, runtime-extra-platforms

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[EgorBo]

@dotnet/jit-contrib PTAL, simple change

[kunalspathak]

LGTM. Thanks!

[EgorBo]

Wow, quite a few improvements: dotnet/perf-autofiling-issues#10542 (e.g. Regex IsMatch as was expected by @GrabYourPitchforks in #69000 (comment))

[drieseng]

@EgorBo, is it expected that IsMatch_Multithreading went from 56.73 ms to 0 ns?

[stephentoub]

is it expected that IsMatch_Multithreading went from 56.73 ms to 0 ns?

No. That's got to be a test harness / reporting issue, or else some kind of catastrophic failure :)

Wow, quite a few improvements

Interlocked.Exchange(ref location, null) is reasonably common in multi-threaded caches / object pools, used to reserve some pooled object from a shared cache. I'd expect to see wins in such uses (which many of the cited improvements seem to be).