Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Inconsistent and counterintuitive codegen behavior of + operator for Vector512<T> depending on the number and order of operands generated by compare instructions, unlike for Vector256<T> #92261

Closed
MineCake147E opened this issue Sep 19, 2023 · 3 comments · Fixed by #92282
Closed

Inconsistent and counterintuitive codegen behavior of + operator for Vector512<T> depending on the number and order of operands generated by compare instructions, unlike for Vector256<T> #92261

MineCake147E opened this issue Sep 19, 2023 · 3 comments · Fixed by #92282
Labels
area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI avx512 Related to the AVX-512 architecture

Comments

@MineCake147E
Copy link
Contributor

Description

When I was playing with the Vector512<T>, I found some inconsistent behavior of + operator for Vector512<byte>, depending on one or more arguments being generated by Vector512.Equals or not.

Both the .NET 7 and .NET 8 RC 1 are doing the right thing for Vector256<byte>:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector256<byte> Test1(Vector256<byte> ymm0, Vector256<byte> ymm1, Vector256<byte> ymm2, Vector256<byte> ymm3)
    => Avx2.BlendVariable(ymm0, ymm2, Vector256.Equals(ymm0, ymm1) + Vector256.Equals(ymm2, ymm3));

In .NET 7, it generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovupd     ymm0,ymmword ptr [rdx]  
vmovupd     ymm1,ymmword ptr [r9]  
vpcmpeqb    ymm2,ymm0,ymmword ptr [r8]  
vpcmpeqb    ymm3,ymm1,ymmword ptr [rax]  
vpaddb      ymm2,ymm2,ymm3  
vpblendvb   ymm0,ymm0,ymm1,ymm2  
vmovupd     ymmword ptr [rcx],ymm0  
mov         rax,rcx  
vzeroupper  
ret

In .NET 8 RC 1, it generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     ymm0,ymmword ptr [rdx]  
vmovups     ymm1,ymmword ptr [r9]  
vpcmpeqb    ymm2,ymm0,ymmword ptr [r8]  
vpcmpeqb    ymm3,ymm1,ymmword ptr [rax]  
vpaddb      ymm2,ymm2,ymm3  
vpblendvb   ymm0,ymm0,ymm1,ymm2  
vmovups     ymmword ptr [rcx],ymm0  
mov         rax,rcx  
vzeroupper  
ret

But when it comes to Vector512<byte>, it somehow goes wrong.

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));

In .NET 8 RC 1, it generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpcmpeqb    k2,zmm1,zmmword ptr [rax]  
kaddq       k1,k1,k2                 ; WRONG INSTRUCTION!
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

While what I expect it to generate looks like:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

And even weirder, if I add Vector512<byte>.Zero in advance of adding second operand, like:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

It generates what I expect it to generate from Test512Blend:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Adding the Vector512<T>.Zero MUST NOT alter the result of the whole integer + operation.

Reproduction Steps

  1. Obtain a JIT assembly for the codes below with .NET 8 RC 1:
[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector256<byte> Test1(Vector256<byte> ymm0, Vector256<byte> ymm1, Vector256<byte> ymm2, Vector256<byte> ymm3)
    => Avx2.BlendVariable(ymm0, ymm2, Vector256.Equals(ymm0, ymm1) + Vector256.Equals(ymm2, ymm3));
[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));
[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

Expected behavior

The code:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));

and

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

For both functions, it should generate something like:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Actual behavior

For the code:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));

It instead generates something like:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpcmpeqb    k2,zmm1,zmmword ptr [rax]  
kaddq       k1,k1,k2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Regression?

Unknown

Known Workarounds

By adding Vector512<T>.Zero in advance of adding second operand, like:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

It generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Configuration

  • .NET version: 8.0.100-rc.1.23455.8
  • PC: Custom built PC
    • OS: Windows 11 Home 22H2 (22621.2283)
    • CPU: Intel(R) Xeon(R) w5-2455X 3.19 GHz
    • RAM: 32 GB DDR5 ECC RDIMM
    • Graphics: NVIDIA GeForce GTX1060(6 GB)
    • Storage:
      • OS(C:) : WD_BLACK SN770 1TB PCIe 4.0 NVMe M.2 SSD
      • DATA (D:) : Western Digital WDC WD10EZEX-22MFCA0 1000.2 GB 7200RPM SATA HDD
      • WORKSPACE (E:) : Seagate ST1000DM003-1ER162 1000.2 GB 7200RPM SATA HDD
Visual Studio 2022 installation

Microsoft Visual Studio Community 2022
Version 17.8.0 Preview 2.0
VisualStudio.17.Preview/17.8.0-pre.2.0+34112.27
Microsoft .NET Framework
Version 4.8.09032

インストールされているバージョン:Community

Visual C++ 2022 00482-90000-00000-AA248
Microsoft Visual C++ 2022

ASP.NET and Web Tools 17.8.226.21692
ASP.NET and Web Tools

AvaloniaPackage Extension 1.0
AvaloniaPackage Visual Studio Extension Detailed Info

Azure App Service Tools v3.0.0 17.8.226.21692
Azure App Service Tools v3.0.0

Azure Functions and Web Jobs Tools 17.8.226.21692
Azure Functions and Web Jobs Tools

C# ツール 4.8.0-2.23429.7+44555193fd1135b5d53a2099f76fec91e0d1ebde
IDE で使用する C# コンポーネント。プロジェクトの種類や設定に応じて、異なるバージョンのコンパイラを使用できます。

Code Cleanup On Save 1.0.12
Automatically run one of the Code Clean profiles when saving the document. This ensures your code is always formatted correctly and follows your coding style conventions.

Common Azure Tools 1.10
Provides common services for use by Azure Mobile Services and Microsoft Azure Tools.

Extensibility Message Bus 1.4.39 (main@e8108eb)
Provides common messaging-based MEF services for loosely coupled Visual Studio extension components communication and integration.

File Icons 2.7
Adds icons for files that are not recognized by Solution Explorer

Microsoft JVM Debugger 1.0
Provides support for connecting the Visual Studio debugger to JDWP compatible Java Virtual Machines

Mono Debugging for Visual Studio 17.8.14 (0c9914e)
Support for debugging Mono processes with Visual Studio.

NuGet パッケージ マネージャー 6.8.0
Visual Studio 内の NuGet パッケージ マネージャー。NuGet の詳細については、https://docs.nuget.org/ にアクセスしてください

Razor (ASP.NET Core) 17.8.2.2345506+ade90399d42c1a7bf92191b1c067816c0ae1c311
ASP.NET Core Razor の言語サービスを提供します。

SonarLint for Visual Studio 7.3.0.77872
SonarLint is an extension to your favorite IDE that provides on-the-fly feedback to developers on new bugs and quality issues injected into their code.

SQL Server Data Tools 17.8.64.0
Microsoft SQL Server Data Tools

Tweaks 2022 1.1.143
A collection of minor fixes and tweaks for Visual Studio to reduce the paper cuts and make you a happier developer

TypeScript Tools 17.0.20830.2001
TypeScript Tools for Microsoft Visual Studio

Visual Basic ツール 4.8.0-2.23429.7+44555193fd1135b5d53a2099f76fec91e0d1ebde
IDE で使用する Visual Basic コンポーネント。プロジェクトの種類や設定に応じて、異なるバージョンのコンパイラを使用できます。

Visual F# Tools 17.8.0-beta.23425.10+0d3549fa5b8b6387ade191d76768405cefed8229
Microsoft Visual F# Tools

Visual Studio IntelliCode 2.2
Visual Studio 向けの AI 支援付き開発。

VisualStudio.DeviceLog 1.0
パッケージに関する情報

VisualStudio.Mac 1.0
Mac Extension for Visual Studio

Xamarin 17.8.0.118 (main@35c256f)
Xamarin.iOS と Xamarin.Android の開発を有効にする Visual Studio 拡張機能

Xamarin Designer 17.8.1.11 (remotes/origin/d17-8@13ef934098)
Visual Studio で Xamarin Designer ツールを有効にするための Visual Studio 拡張機能。

Xamarin Templates 17.8.16 (830b56a)
Templates for building iOS, Android, and Windows apps with Xamarin and Xamarin.Forms.

Xamarin.Android SDK 13.2.1.2 (d17-5/a8a26c7)
Xamarin.Android Reference Assemblies and MSBuild support.
Mono: d9a6e87
Java.Interop: xamarin/java.interop/d17-5@149d70fe
SQLite: xamarin/sqlite@68c69d8
Xamarin.Android Tools: xamarin/xamarin-android-tools/d17-5@ca1552d

Xamarin.iOS and Xamarin.Mac SDK 16.4.0.16 (b5972410d)
Xamarin.iOS and Xamarin.Mac Reference Assemblies and MSBuild support.

Other information

No response
@dotnet-issue-labeler dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Sep 19, 2023
@ghost ghost added the untriaged New issue has not been triaged by the area owner label Sep 19, 2023
@ghost
Copy link

ghost commented Sep 19, 2023

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

Issue Details

Description

When I was playing with the Vector512<T>, I found some inconsistent behavior of + operator for Vector512<byte>, depending on one or more arguments being generated by Vector512.Equals or not.

Both the .NET 7 and .NET 8 RC 1 are doing the right thing for Vector256<byte>:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector256<byte> Test1(Vector256<byte> ymm0, Vector256<byte> ymm1, Vector256<byte> ymm2, Vector256<byte> ymm3)
    => Avx2.BlendVariable(ymm0, ymm2, Vector256.Equals(ymm0, ymm1) + Vector256.Equals(ymm2, ymm3));

In .NET 7, it generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovupd     ymm0,ymmword ptr [rdx]  
vmovupd     ymm1,ymmword ptr [r9]  
vpcmpeqb    ymm2,ymm0,ymmword ptr [r8]  
vpcmpeqb    ymm3,ymm1,ymmword ptr [rax]  
vpaddb      ymm2,ymm2,ymm3  
vpblendvb   ymm0,ymm0,ymm1,ymm2  
vmovupd     ymmword ptr [rcx],ymm0  
mov         rax,rcx  
vzeroupper  
ret

In .NET 8 RC 1, it generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     ymm0,ymmword ptr [rdx]  
vmovups     ymm1,ymmword ptr [r9]  
vpcmpeqb    ymm2,ymm0,ymmword ptr [r8]  
vpcmpeqb    ymm3,ymm1,ymmword ptr [rax]  
vpaddb      ymm2,ymm2,ymm3  
vpblendvb   ymm0,ymm0,ymm1,ymm2  
vmovups     ymmword ptr [rcx],ymm0  
mov         rax,rcx  
vzeroupper  
ret

But when it comes to Vector512<byte>, it somehow goes wrong.

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));

In .NET 8 RC 1, it generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpcmpeqb    k2,zmm1,zmmword ptr [rax]  
kaddq       k1,k1,k2                 ; WRONG INSTRUCTION!
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

While what I expect it to generate looks like:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

And even weirder, if I add Vector512<byte>.Zero in advance of adding second operand, like:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

It generates what I expect it to generate from Test512Blend:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Adding the Vector512<T>.Zero MUST NOT alter the result of the whole integer + operation.

Reproduction Steps

  1. Obtain a JIT assembly for the codes below with .NET 8 RC 1:
[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector256<byte> Test1(Vector256<byte> ymm0, Vector256<byte> ymm1, Vector256<byte> ymm2, Vector256<byte> ymm3)
    => Avx2.BlendVariable(ymm0, ymm2, Vector256.Equals(ymm0, ymm1) + Vector256.Equals(ymm2, ymm3));
[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));
[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

Expected behavior

The code:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));

and

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

For both functions, it should generate something like:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Actual behavior

For the code:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512.Equals(zmm2, zmm3));

It instead generates something like:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpcmpeqb    k2,zmm1,zmmword ptr [rax]  
kaddq       k1,k1,k2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Regression?

Unknown

Known Workarounds

By adding Vector512<T>.Zero in advance of adding second operand, like:

[MethodImpl(MethodImplOptions.AggressiveOptimization)]
private static Vector512<byte> Test512Blend3(Vector512<byte> zmm0, Vector512<byte> zmm1, Vector512<byte> zmm2, Vector512<byte> zmm3)
    => Avx512BW.BlendVariable(zmm0, zmm2, Vector512.Equals(zmm0, zmm1) + Vector512<byte>.Zero + Vector512.Equals(zmm2, zmm3));

It generates:

vzeroupper  
mov         rax,qword ptr [rsp+28h]  
vmovups     zmm0,zmmword ptr [rdx]  
vmovups     zmm1,zmmword ptr [r9]  
vpcmpeqb    k1,zmm0,zmmword ptr [r8]  
vpmovm2b    zmm2,k1  
vpcmpeqb    k1,zmm1,zmmword ptr [rax]  
vpmovm2b    zmm3,k1  
vpaddb      zmm2,zmm2,zmm3  
vpmovb2m    k1,zmm2  
vpblendmb   zmm0{k1},zmm0,zmm1  
vmovups     zmmword ptr [rcx],zmm0  
mov         rax,rcx  
vzeroupper  
ret

Configuration

  • .NET version: 8.0.100-rc.1.23455.8
  • PC: Custom built PC
    • OS: Windows 11 Home 22H2 (22621.2283)
    • CPU: Intel(R) Xeon(R) w5-2455X 3.19 GHz
    • RAM: 32 GB DDR5 ECC RDIMM
    • Graphics: NVIDIA GeForce GTX1060(6 GB)
    • Storage:
      • OS(C:) : WD_BLACK SN770 1TB PCIe 4.0 NVMe M.2 SSD
      • DATA (D:) : Western Digital WDC WD10EZEX-22MFCA0 1000.2 GB 7200RPM SATA HDD
      • WORKSPACE (E:) : Seagate ST1000DM003-1ER162 1000.2 GB 7200RPM SATA HDD
Visual Studio 2022 installation

Microsoft Visual Studio Community 2022
Version 17.8.0 Preview 2.0
VisualStudio.17.Preview/17.8.0-pre.2.0+34112.27
Microsoft .NET Framework
Version 4.8.09032

インストールされているバージョン:Community

Visual C++ 2022 00482-90000-00000-AA248
Microsoft Visual C++ 2022

ASP.NET and Web Tools 17.8.226.21692
ASP.NET and Web Tools

AvaloniaPackage Extension 1.0
AvaloniaPackage Visual Studio Extension Detailed Info

Azure App Service Tools v3.0.0 17.8.226.21692
Azure App Service Tools v3.0.0

Azure Functions and Web Jobs Tools 17.8.226.21692
Azure Functions and Web Jobs Tools

C# ツール 4.8.0-2.23429.7+44555193fd1135b5d53a2099f76fec91e0d1ebde
IDE で使用する C# コンポーネント。プロジェクトの種類や設定に応じて、異なるバージョンのコンパイラを使用できます。

Code Cleanup On Save 1.0.12
Automatically run one of the Code Clean profiles when saving the document. This ensures your code is always formatted correctly and follows your coding style conventions.

Common Azure Tools 1.10
Provides common services for use by Azure Mobile Services and Microsoft Azure Tools.

Extensibility Message Bus 1.4.39 (main@e8108eb)
Provides common messaging-based MEF services for loosely coupled Visual Studio extension components communication and integration.

File Icons 2.7
Adds icons for files that are not recognized by Solution Explorer

Microsoft JVM Debugger 1.0
Provides support for connecting the Visual Studio debugger to JDWP compatible Java Virtual Machines

Mono Debugging for Visual Studio 17.8.14 (0c9914e)
Support for debugging Mono processes with Visual Studio.

NuGet パッケージ マネージャー 6.8.0
Visual Studio 内の NuGet パッケージ マネージャー。NuGet の詳細については、https://docs.nuget.org/ にアクセスしてください

Razor (ASP.NET Core) 17.8.2.2345506+ade90399d42c1a7bf92191b1c067816c0ae1c311
ASP.NET Core Razor の言語サービスを提供します。

SonarLint for Visual Studio 7.3.0.77872
SonarLint is an extension to your favorite IDE that provides on-the-fly feedback to developers on new bugs and quality issues injected into their code.

SQL Server Data Tools 17.8.64.0
Microsoft SQL Server Data Tools

Tweaks 2022 1.1.143
A collection of minor fixes and tweaks for Visual Studio to reduce the paper cuts and make you a happier developer

TypeScript Tools 17.0.20830.2001
TypeScript Tools for Microsoft Visual Studio

Visual Basic ツール 4.8.0-2.23429.7+44555193fd1135b5d53a2099f76fec91e0d1ebde
IDE で使用する Visual Basic コンポーネント。プロジェクトの種類や設定に応じて、異なるバージョンのコンパイラを使用できます。

Visual F# Tools 17.8.0-beta.23425.10+0d3549fa5b8b6387ade191d76768405cefed8229
Microsoft Visual F# Tools

Visual Studio IntelliCode 2.2
Visual Studio 向けの AI 支援付き開発。

VisualStudio.DeviceLog 1.0
パッケージに関する情報

VisualStudio.Mac 1.0
Mac Extension for Visual Studio

Xamarin 17.8.0.118 (main@35c256f)
Xamarin.iOS と Xamarin.Android の開発を有効にする Visual Studio 拡張機能

Xamarin Designer 17.8.1.11 (remotes/origin/d17-8@13ef934098)
Visual Studio で Xamarin Designer ツールを有効にするための Visual Studio 拡張機能。

Xamarin Templates 17.8.16 (830b56a)
Templates for building iOS, Android, and Windows apps with Xamarin and Xamarin.Forms.

Xamarin.Android SDK 13.2.1.2 (d17-5/a8a26c7)
Xamarin.Android Reference Assemblies and MSBuild support.
Mono: d9a6e87
Java.Interop: xamarin/java.interop/d17-5@149d70fe
SQLite: xamarin/sqlite@68c69d8
Xamarin.Android Tools: xamarin/xamarin-android-tools/d17-5@ca1552d

Xamarin.iOS and Xamarin.Mac SDK 16.4.0.16 (b5972410d)
Xamarin.iOS and Xamarin.Mac Reference Assemblies and MSBuild support.

Other information

No response

Author: MineCake147E
Assignees: -
Labels:

area-CodeGen-coreclr
Milestone: -

@EgorBo EgorBo added the avx512 Related to the AVX-512 architecture label Sep 19, 2023
@EgorBo
Copy link
Member

EgorBo commented Sep 19, 2023

cc @dotnet/avx512-contrib

@tannergooding
Copy link
Member

Looks like an issue in the kmask logic here.

The simplest thing is to avoid generating kadd for .NET 8. The more common cases will be the trivial bitwise operations anyways.

@tannergooding tannergooding mentioned this issue Sep 19, 2023
Merged
@ghost ghost added the in-pr There is an active PR which will close this issue when it is merged label Sep 19, 2023
@ghost ghost removed in-pr There is an active PR which will close this issue when it is merged untriaged New issue has not been triaged by the area owner labels Sep 19, 2023
@jakobbotsch jakobbotsch mentioned this issue Sep 19, 2023
Merged
@ghost ghost locked as resolved and limited conversation to collaborators Oct 20, 2023
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Assignees
No one assigned
Labels
area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI avx512 Related to the AVX-512 architecture
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

Don't generate AddMask as it requires more explicit consideration of semantics tannergooding/runtime
3 participants
@EgorBo @tannergooding @MineCake147E