Intrinsify Interlocked.And and Interlocked.Or on XARCH

[EgorBo]

Closes #32239

These were only implemented on ARM64 (and RISC-V). Unlike other Interlocked APIs, these return original value so it makes them impossible to optimize on XARCH when that original value is used. Looks like all usages of it in BCL don't need return value so we can optimize this case. Example:

runtime/src/libraries/System.Private.CoreLib/src/System/Threading/Tasks/Task.cs

Line 756 in 0cf461b

Interlocked.And(ref m_stateFlags, ~(int)TaskStateFlags.WaitCompletionNotification);

void Test1(ref int x, int y) => Interlocked.Or(ref x, y); // return value is not used int Test2(ref int x, int y) => Interlocked.Or(ref x, y); // return value is used

Was:

; Method Tests:Test1(byref,int):this (FullOpts) G_M7530_IG01:  push rax G_M7530_IG02:  mov eax, dword ptr [rdx]  jmp SHORT G_M7530_IG03  align [0 bytes for IG03] G_M7530_IG03:  mov ecx, eax  or ecx, r8d  mov dword ptr [rsp+0x04], eax  lock   cmpxchg dword ptr [rdx], ecx  mov ecx, dword ptr [rsp+0x04]  cmp eax, ecx  je SHORT G_M7530_IG05 G_M7530_IG04:  mov ecx, eax  mov eax, ecx  jmp SHORT G_M7530_IG03 G_M7530_IG05:  add rsp, 8  ret  ; Total bytes of code: 37 ; Method Tests:Test2(byref,int):int:this (FullOpts) G_M30720_IG01:  push rax G_M30720_IG02:  mov eax, dword ptr [rdx]  jmp SHORT G_M30720_IG03  align [0 bytes for IG03] G_M30720_IG03:  mov ecx, eax  or ecx, r8d  mov dword ptr [rsp+0x04], eax  lock   cmpxchg dword ptr [rdx], ecx  mov ecx, dword ptr [rsp+0x04]  cmp eax, ecx  je SHORT G_M30720_IG05 G_M30720_IG04:  mov ecx, eax  mov eax, ecx  jmp SHORT G_M30720_IG03 G_M30720_IG05:  mov eax, ecx G_M30720_IG06:  add rsp, 8  ret  ; Total bytes of code: 39 

Now:

; Method Tests:Test1(byref,int):this (FullOpts) G_M7530_IG01: G_M7530_IG02:  lock   or dword ptr [rdx], r8d G_M7530_IG03:  ret  ; Total bytes of code: 5 ; Method Tests:Test2(byref,int):int:this (FullOpts) G_M30720_IG01: G_M30720_IG02:  mov eax, dword ptr [rdx] G_M30720_IG03:  mov ecx, eax  or ecx, r8d  lock   cmpxchg dword ptr [rdx], ecx  jne SHORT G_M30720_IG03 G_M30720_IG04:  ret  ; Total bytes of code: 14

[ghost]

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

Issue Details

---

Closes #32239

These were only implemented on ARM64 (and RISC-V). Unlike other Interlocked APIs, these return original value so it makes them impossible to optimize on XARCH when that original value is used. Looks like all usages of it in BCL don't need return value so we can optimize this case.

void Test(ref int x, int y) => Interlocked.Or(ref x, y);

Was:

; Method Program:Test(byref,int):this (FullOpts) G_M55114_IG01: ;; offset=0x0000  push rax ;; size=1 bbWeight=1 PerfScore 1.00 G_M55114_IG02: ;; offset=0x0001  mov eax, dword ptr [rdx]  jmp SHORT G_M55114_IG03  align [0 bytes for IG03] ;; size=4 bbWeight=1 PerfScore 4.00 G_M55114_IG03: ;; offset=0x0005  mov ecx, eax  or ecx, r8d  mov dword ptr [rsp+0x04], eax  lock   cmpxchg dword ptr [rdx], ecx  mov ecx, dword ptr [rsp+0x04]  cmp eax, ecx  je SHORT G_M55114_IG05 ;; size=21 bbWeight=8 PerfScore 174.00 G_M55114_IG04: ;; offset=0x001A  mov ecx, eax  mov eax, ecx  jmp SHORT G_M55114_IG03 ;; size=6 bbWeight=4 PerfScore 10.00 G_M55114_IG05: ;; offset=0x0020  add rsp, 8  ret  ;; size=5 bbWeight=1 PerfScore 1.25 ; Total bytes of code: 37

Now:

; Method Program:Test(byref,int):this (FullOpts) G_M55114_IG01: ;; offset=0x0000 ;; size=0 bbWeight=1 PerfScore 0.00 G_M55114_IG02: ;; offset=0x0000  lock   or dword ptr [rdx], r8d ;; size=4 bbWeight=1 PerfScore 16.00 G_M55114_IG03: ;; offset=0x0004  ret  ;; size=1 bbWeight=1 PerfScore 1.00 ; Total bytes of code: 5

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

[EgorBo]

@dotnet/jit-contrib @BruceForstall PTAL, simple PR, optimizes Interlocked.And/Add the way native compilers do: https://godbolt.org/z/csfd3nG89

Diffs

A few size and TP regressions for the cases when we previously didn't inline Interlocked.And/Or and now it's inlined as intrinsic in both Tier0 (non-debug and non-minopts) and Tier1

[kunalspathak]

Looks good, just a comment in lsra.