IndexOf and LastIndexOf are for some test cases few times slower on Linux

[adamsitnik]

IndexOf and LastIndexOf are for some test cases few times slower on Linux.

Slower	Lin/Win	Win Median (ns)	Lin Median (ns)	Modality
System.Memory.ReadOnlySpan.IndexOfString(input: "だ", value: "た", comparisonType: InvariantCulture)	34.21	107.43	3675.43
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (, None, True))	7.66	2952.92	22631.17
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (en-US, IgnoreCase, True))	7.62	2954.37	22515.44
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (en-US, None, True))	5.75	2963.99	17052.76
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (, IgnoreCase, True))	5.55	2958.29	16412.98
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (en-US, IgnoreSymbols, False))	4.87	3222.91	15680.86
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (en-US, OrdinalIgnoreCase, False))	3.47	596.96	2072.40
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (en-US, IgnoreSymbols, False))	3.17	3183.26	10084.88
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (en-US, OrdinalIgnoreCase, False))	3.12	678.58	2118.31
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (en-US, IgnoreCase, True))	2.73	3713.38	10119.90
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (, None, True))	2.67	3850.16	10264.14
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (en-US, None, True))	2.64	3935.11	10406.63
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (, IgnoreCase, True))	2.64	3707.07	9803.11
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (en-US, Ordinal, False))	2.36	131.66	311.09
System.Memory.ReadOnlySpan.IndexOfString(input: "AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAX	2.35	235.63	553.18
System.Globalization.Tests.StringSearch.LastIndexOf_Word_NotFound(Options: (pl-PL, None, False))	2.17	8300.73	18011.26
System.Memory.ReadOnlySpan.IndexOfString(input: "Hello WorldbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbareallyreallylongHello World	1.72	45.05	77.69
System.Memory.ReadOnlySpan.IndexOfString(input: "More Test's", value: "Tests", comparisonType: OrdinalIgnoreCase)	1.61	53.91	86.82
System.Globalization.Tests.StringSearch.IndexOf_Word_NotFound(Options: (pl-PL, None, False))	1.37	9237.59	12681.21
System.Memory.ReadOnlySpan.IndexOfString(input: "StrIng", value: "string", comparisonType: OrdinalIgnoreCase)	1.30	37.58	48.82

How to run the benchmarks:

git clone https://github.com/dotnet/performance.git
python3 ./performance/scripts/benchmarks_ci.py -f netcoreapp5.0 --filter 'System.Globalization*IndexOf*'  '*IndexOfString'

Recommended profilers are PerfCollect and VTune which works great on Linux.

Some of the differences come from ICU (on Windows we are using WinAPI), some might come from the managed hot path. The person who is willing to work on this issue should use VTune/PerfCollect, identify the problem and solve it. It might require tuning the ICU itself ;)

[adamsitnik]

I've used VTune to profile the biggest difference just to make sure we don't do anything stupid:

static int Main(string[] args)
{
    int result = 0;
    
    for (int i = 0; i < 100_000; i++)
        result ^= IndexOfString("\u3060", "\u305F", StringComparison.InvariantCulture);

    return result;
}

[MethodImpl(MethodImplOptions.NoInlining)]
static int IndexOfString(string input, string value, StringComparison comparisonType)
{
    ReadOnlySpan<char> inputSpan = input.AsSpan();
    ReadOnlySpan<char> valueSpan = value.AsSpan();

    int result = 0;
    
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);
    result ^= inputSpan.IndexOf(valueSpan, comparisonType); result ^= inputSpan.IndexOf(valueSpan, comparisonType);

    return result;
}

80% of time is spent in usearch_openFromCollator (ICU):