JIT: Unroll more Equals/StartsWith/EndsWith on arm64

[EgorBo]

Seems like it's quite beneficial even on large sizes, at least on arm64:

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

public class Bench
{
    string s = "https://pkgs.dev.azure.com/dnceng/public/packaging/dotnet10/";

    [Benchmark]
    public bool Bestcase() => 
        s.StartsWith("https://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.Ordinal);
    
    [Benchmark]
    public bool Bestcase_IgnoreCase() => 
        s.StartsWith("https://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.OrdinalIgnoreCase);

    // Worst case: differs at 1st char:

    [Benchmark]
    public bool Worstcase() => 
        s.StartsWith("Xttps://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.Ordinal);
    
    [Benchmark]
    public bool Worstcase_IgnoreCase() => 
        s.StartsWith("Xttps://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.OrdinalIgnoreCase);
}

BenchmarkDotNet v0.14.0, Ubuntu 24.04 LTS (Noble Numbat)
Arm64
  Job-NEWBNL : .NET 10.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD
  Job-LSVZLO : .NET 10.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD
EnvironmentVariables=DOTNET_JitDisasm=Bestcase_IgnoreCase

Method	Toolchain	Mean	Error	Ratio
Bestcase	Main	4.1458 ns	0.0071 ns	1.00
Bestcase	PR	0.6151 ns	0.0007 ns	0.15
Bestcase_IgnoreCase	Main	5.9848 ns	0.1685 ns	1.00
Bestcase_IgnoreCase	PR	1.5489 ns	0.0044 ns	0.26
Worstcase	Main	1.7136 ns	0.0003 ns	1.00
Worstcase	PR	0.6572 ns	0.0013 ns	0.38
Worstcase_IgnoreCase	Main	2.5230 ns	0.0015 ns	1.00
Worstcase_IgnoreCase	PR	1.2213 ns	0.0016 ns	0.48

[EgorBo]

@MihuBot

[EgorBo]

@EgorBot -arm64 --envvars DOTNET_JitDisasm:Bestcase_IgnoreCase

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

public class Bench
{
    string s = "https://pkgs.dev.azure.com/dnceng/public/packaging/dotnet10/";

    [Benchmark]
    public bool Bestcase() => 
        s.StartsWith("https://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.Ordinal);
    
    [Benchmark]
    public bool Bestcase_IgnoreCase() => 
        s.StartsWith("https://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.OrdinalIgnoreCase);

    // Worst case: differs at 1st char:

    [Benchmark]
    public bool Worstcase() => 
        s.StartsWith("Xttps://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.Ordinal);
    
    [Benchmark]
    public bool Worstcase_IgnoreCase() => 
        s.StartsWith("Xttps://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.OrdinalIgnoreCase);
}

[EgorBo]

@MihuBot -arm -dependsOn 108838

[EgorBo]

PTAL @jakobbotsch @dotnet/jit-contrib this improves perf on arm64 (even in the worst case when the first char is different). Seems like makes no sense on x64 judging by benchmarks.

[tannergooding]

Seems like makes no sense on x64 judging by benchmarks.

Do you have a link to the perf numbers on x64?

I would expect this to be similar perf on most x64, but likely better on CPUs like e-cores or other scenarios where call overhead is greater

[EgorBo]

@EgorBot -amd -intel --envvars DOTNET_JitDisasm:Bestcase_IgnoreCase

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

public class Bench
{
    string s = "https://pkgs.dev.azure.com/dnceng/public/packaging/dotnet10/";

    [Benchmark]
    public bool Bestcase() => 
        s.StartsWith("https://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.Ordinal);
    
    [Benchmark]
    public bool Bestcase_IgnoreCase() => 
        s.StartsWith("https://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.OrdinalIgnoreCase);

    // Worst case: differs at 1st char:

    [Benchmark]
    public bool Worstcase() => 
        s.StartsWith("Xttps://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.Ordinal);
    
    [Benchmark]
    public bool Worstcase_IgnoreCase() => 
        s.StartsWith("Xttps://pkgs.dev.azure.com/dnceng/public/pack", StringComparison.OrdinalIgnoreCase);
}

[EgorBo]

@tannergooding you can find x64 numbers here EgorBot/runtime-utils#162, Basically, almost no wins and clear regressions unlike arm64

[tannergooding]

you can find x64 numbers here EgorBot/runtime-utils#162, Basically, almost no wins and clear regressions unlike arm64

Thanks! That roughly fits my expectations for that hardware. I don't think we have any e-cores to see if there is a contrast for those devices though, unfortunately.

I have no concerns with it being merged, just wanted to see some of the numbers since it was called out