Optimize Vector128/256.Equals via TestZ

[EgorBo]

This PR optimizes Vector128/256.Equals via the following improvements:

• Substitutes Vector_.Zero during inlining
• Optimizes Xor(x, zero) to just x in morph
• Adds Avx2/Sse41 TestZ path in Vector_.Equals

// Compare two vectorsboolTest128(Vector128<int>v1,Vector128<int>v2)=>v1.Equals(v2);boolTest256(Vector256<int>v1,Vector256<int>v2)=>v1.Equals(v2);// Compare against Zero:boolTest128(Vector128<int>v1)=>v1.Equals(Vector128<int>.Zero);boolTest256(Vector256<int>v1)=>v1.Equals(Vector256<int>.Zero);

Codegen diff: https://www.diffchecker.com/zJhfk9yq
According to benchmarks, it makes it 10-15% faster

jit-diff is quite small:

Total bytes of base: 61261912
Total bytes of diff: 61261864
Total bytes of delta: -48 (-0.00% of base)
Total relative delta: -0.19
diff is an improvement.
relative diff is an improvement.
Top file improvements (bytes):
-48 : System.Private.CoreLib.dasm (-0.00% of base)
1 total files with Code Size differences (1 improved, 0 regressed), 272 unchanged.
Top method improvements (bytes):
-16 (-10.74% of base) : System.Private.CoreLib.dasm - Vector128`1:Equals(Vector128`1):bool:this (6 methods)
-16 (-2.48% of base) : System.Private.CoreLib.dasm - Vector128`1:Equals(Object):bool:this (6 methods)
-8 (-5.06% of base) : System.Private.CoreLib.dasm - Vector256`1:Equals(Vector256`1):bool:this (6 methods)
-8 (-1.18% of base) : System.Private.CoreLib.dasm - Vector256`1:Equals(Object):bool:this (6 methods)
Top method improvements (percentages):
-16 (-10.74% of base) : System.Private.CoreLib.dasm - Vector128`1:Equals(Vector128`1):bool:this (6 methods)
-8 (-5.06% of base) : System.Private.CoreLib.dasm - Vector256`1:Equals(Vector256`1):bool:this (6 methods)
-16 (-2.48% of base) : System.Private.CoreLib.dasm - Vector128`1:Equals(Object):bool:this (6 methods)
-8 (-1.18% of base) : System.Private.CoreLib.dasm - Vector256`1:Equals(Object):bool:this (6 methods)

Superpmi:

asm.coreclr_tests.pmi.windows.x64.checked
Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 4987
Total bytes of diff: 4976
Total bytes of delta: -11 (-0.22% of base)
Total relative delta: -0.01
diff is an improvement.
relative diff is an improvement.
Top file improvements (bytes):
-11 : 251366.dasm (-0.76% of base)
1 total files with Code Size differences (1 improved, 0 regressed), 4 unchanged.
Top method improvements (bytes):
-11 (-0.76% of base) : 251366.dasm - GitHub_18144:Main(System.String[]):int
Top method improvements (percentages):
-11 (-0.76% of base) : 251366.dasm - GitHub_18144:Main(System.String[]):int
1 total methods with Code Size differences (1 improved, 0 regressed), 4 unchanged.
asm.libraries_tests.pmi.windows.x64.checked
Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 1558
Total bytes of diff: 1522
Total bytes of delta: -36 (-2.31% of base)
Total relative delta: -0.20
diff is an improvement.
relative diff is an improvement.
Top file improvements (bytes):
-12 : 309432.dasm (-7.02% of base)
-12 : 309515.dasm (-6.56% of base)
-12 : 309609.dasm (-6.56% of base)
3 total files with Code Size differences (3 improved, 0 regressed), 1 unchanged.
Top method improvements (bytes):
-12 (-7.02% of base) : 309432.dasm - System.Numerics.Tests.Vector2Tests:Vector2ZeroTest():this
-12 (-6.56% of base) : 309515.dasm - System.Numerics.Tests.Vector3Tests:Vector3ZeroTest():this
-12 (-6.56% of base) : 309609.dasm - System.Numerics.Tests.Vector4Tests:Vector4ZeroTest():this
Top method improvements (percentages):
-12 (-7.02% of base) : 309432.dasm - System.Numerics.Tests.Vector2Tests:Vector2ZeroTest():this
-12 (-6.56% of base) : 309515.dasm - System.Numerics.Tests.Vector3Tests:Vector3ZeroTest():this
-12 (-6.56% of base) : 309609.dasm - System.Numerics.Tests.Vector4Tests:Vector4ZeroTest():this

[MichalPetryka]

Partially fixes #55343 since that also talks about optimizing AllBitsSet to a TestC.

[EgorBo]

@tannergooding @dotnet/jit-contrib does it look good?

[EgorBo]

@dotnet/jit-contrib PTAL, a small improvement for Vector128/256.Create

Failed CI job is unrelated (broken gcc build)

[MichalPetryka]

I think that IsIntegralConstVector will only work for integer vectors, so all floating point vectors will be rejected here.

[EgorBo]

Xor is not used in Equals for floats and doubles so it's not a big deal

but it currently handles this (Xor for floats) just fine:

staticVector128<float>Foo(Vector128<float>v)=>Sse.Xor(v,Vector128.Create(0).AsSingle());

[EgorBo]

for zero float it might be tricky as e.g. -0.0 can't be used in this optimization so is not worth the effort

[tannergooding]

Why would it be tricky? You can account for -0.0 by checking the bitwise (rather than floating-point) value is 0.

Speaking of 0.0 vs -0.0, it looks like there might be existing bugs in IsFPZero and IsSIMDZero since they don't take this into account.

[BruceForstall]

LGTM. But would like @tannergooding to also approve.

[EgorBo]

@tannergooding PTAL

[tannergooding]

Do we have an existing issue for this TODO?

[EgorBo]

That's why I want to merge it - to start experimenting with it 🙂 the issue is #40381

[tannergooding]

Thanks. I just wanted to make sure we had some github issue corresponding to the TODO, so its not just a comment we'll forget about

[tannergooding]

Changes LGTM. Would be good to ensure floating-point is equally well handled.