PGO: Profiled SequenceEqual

[EgorBo]

Extends #96311 to cover memcmp idiom (SpanHelpers.SequenceEqual):

bool result = SequenceEqual(dst, src, len);

is optimized to:

bool result;
if (len == 10) // PGO data
    result = SequenceEqual(dst, src, 10); // unrolled
else
    result = SequenceEqual(dst, src, len); // fallback

Benchmark 1:

public class Benchmarks
{
    // mutable fields (to hide from jit)
    string s1 = 10000.ToString();
    string s2 = 10001.ToString();

    [Benchmark]
    public bool Compare() => s1 == s2; // ends up calling SpanHelpers.SequenceEqual under the hood
}

Method	Toolchain	Mean	Ratio
Compare	\Core_Root\corerun.exe	1.8656 ns	1.00
Compare	\Core_Root_PR\corerun.exe	0.7309 ns	0.39

Benchmark 2:

public class Benchmarks
{
    public static IEnumerable<object[]> Data()
    {
        yield return new object[] { new int[2], new int[2] };
        yield return new object[] { new int[10], new int[10] };
        yield return new object[] { new int[16], new int[16] };
        yield return new object[] { new int[32], new int[32] };
    }

    [Benchmark]
    [ArgumentsSource(nameof(Data))]
    public bool Compare(int[] a, int[] b) => a.AsSpan().SequenceEqual(b);
}

Method	Toolchain	a	b	Mean	Ratio
Compare	\Core_Root\corerun.exe	Int32[2]	Int32[2]	1.8737 ns	1.03
Compare	\Core_Root_PR\corerun.exe	Int32[2]	Int32[2]	0.4867 ns	0.27
Compare	\Core_Root\corerun.exe	Int32[10]	Int32[10]	1.8131 ns	1.00
Compare	\Core_Root_PR\corerun.exe	Int32[10]	Int32[10]	0.8914 ns	0.49
Compare	\Core_Root\corerun.exe	Int32[16]	Int32[16]	1.5324 ns	0.85
Compare	\Core_Root_PR\corerun.exe	Int32[16]	Int32[16]	0.8889 ns	0.49
Compare	\Core_Root\corerun.exe	Int32[32]	Int32[32]	1.8264 ns	1.01
Compare	\Core_Root_PR\corerun.exe	Int32[32]	Int32[32]	0.9248 ns	0.51

[ghost]

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

Issue Details

---

Extends #96311 to cover memcmp idiom (SpanHelpers.SequenceEqual):

bool result = SequenceEqual(dst, src, len);

is optimized to:

bool result;
if (len == 10) // PGO data
    result = SequenceEqual(dst, src, 10); // unrolled
else
    result = SequenceEqual(dst, src, len); // fallback

Benchmark:

public static IEnumerable<object[]> Data()
{
    yield return new object[] { new int[2], new int[2] };
    yield return new object[] { new int[10], new int[10] };
    yield return new object[] { new int[16], new int[16] };
    yield return new object[] { new int[32], new int[32] };
}

[Benchmark]
[ArgumentsSource(nameof(Data))]
public bool Compare(int[] a, int[] b) => a.AsSpan().SequenceEqual(b);

Method	Toolchain	a	b	Mean	Ratio
Compare	\Core_Root\corerun.exe	Int32[2]	Int32[2]	1.8737 ns	1.03
Compare	\Core_Root_PR\corerun.exe	Int32[2]	Int32[2]	0.4867 ns	0.27
Compare	\Core_Root\corerun.exe	Int32[10]	Int32[10]	1.8131 ns	1.00
Compare	\Core_Root_PR\corerun.exe	Int32[10]	Int32[10]	0.8914 ns	0.49
Compare	\Core_Root\corerun.exe	Int32[16]	Int32[16]	1.5324 ns	0.85
Compare	\Core_Root_PR\corerun.exe	Int32[16]	Int32[16]	0.8889 ns	0.49
Compare	\Core_Root\corerun.exe	Int32[32]	Int32[32]	1.8264 ns	1.01
Compare	\Core_Root_PR\corerun.exe	Int32[32]	Int32[32]	0.9248 ns	0.51

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

[EgorBo]

/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo, runtime-coreclr pgostress

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[EgorBo]

@AndyAyersMS @dotnet/jit-contrib PTAL

[EgorBo]

Leaving Memset idiom for future, because it's currently not a call, but GT_BLK so it will need a bit more changes + tracking IL offset for it.