CodegenAnalysis

Library for obtaining and analyzing the codegen of JIT for .NET. Can be used for both tests (assertions) and making reports (benchmarks).

MIT License

Stars
22
Committers
1

CodegenAnalysis

Library for analyzing the machine code generated by JIT (codegen). Has API for obtaining the codegen, verifying the characteristics for tests (in a similar manner to xUnit), and generating reports as benchmarks (in a similar manner to BenchmarkDotNet). Supports x86_64 on three major platforms (Windows, MacOS, Linux).

Why should I use it?

First of all, most .NET developers don't need this library. If you want to measure performance of your code, use BenchmarkDotNet.

If you're working on low-level, often HW-related project, where it's crucial if a branch or call or etc. is optimized or not, this library is for you.

Table of contents

CodegenAnalysis

static int AddAndMul(int a, int b) => a + b * a;

...

var ci = CodegenInfo.Obrain(() => AddAndMul(3, 5));
Console.WriteLine(ci);

Output:

00007FFD752E42F0 8BC2                 mov       eax,edx
00007FFD752E42F2 0FAFC1               imul      eax,ecx
00007FFD752E42F5 03C1                 add       eax,ecx
00007FFD752E42F7 C3                   ret

CodegenAnalysis.Assertions

Verifying the size of the codegen

using CodegenAssertions;
using Xunit;

public class CodegenSizeTest
{
    public static int SomeMethod(int a, int b)
        => a + b;

    [Fact]
    public void Test1()
    {
        CodegenInfo.Obtain(() => SomeMethod(4, 5), CompilationTier.Tier1)
            .ShouldBeNotLargerThan(20);
    }
}

Having calls in the codegen

public class Tests
{
    public class A
    {
        public virtual int H => 3;
    }

    public sealed class B : A
    {
        public override int H => 6;
    }

    // this will get devirtualized at tier1, but not at tier0
    static int Twice(B b) => b.H * 2;

    [Fact]
    public void NotDevirtTier0()
    {
        CodegenInfo.Obtain(() => Twice(new B()), CompilationTier.Default)
            .ShouldHaveCalls(c => c >= 1);
    }

    [Fact]
    public void DevirtTier1()
    {
        CodegenInfo.Obtain(() => Twice(new B()), CompilationTier.Tier1)
            .ShouldHaveCalls(0);
    }
}

Testing if we have branches

    private static readonly bool True = true;

    static int SmartThing()
    {
        if (True)
            return 5;
        return 10;
    }

    [Fact]
    public void BranchElimination()
    {
        CodegenInfo.Obtain(() => SmartThing())
            .ShouldHaveBranches(0);
    }

    [MethodImpl(MethodImplOptions.NoOptimization)]
    static int StupidThing()
    {
        if (True)
            return 5;
        return 10;
    }

    [Fact]
    public void NoBranchElimination()
    {
        CodegenInfo.Obtain(() => StupidThing(), CompilationTier.Default)
            .ShouldHaveBranches(b => b > 0);
    }

CodegenAnalysis.Benchmarks

CodegenBenchmarkRunner.Run<A>();

[CAJob(Tier = CompilationTier.Default),
 CAJob(Tier = CompilationTier.Tier1)]

[CAColumn(CAColumn.Branches),
 CAColumn(CAColumn.Calls), 
 CAColumn(CAColumn.CodegenSize), 
 CAColumn(CAColumn.StaticStackAllocations)]

[CAExport(Export.Html),
 CAExport(Export.Md)]
public class A
{
    [CAAnalyze(3.5f)]
    [CAAnalyze(13.5f)]
    public static float Heavy(float a)
    {
        var b = Do1(a);
        var c = Do1(b);
        if (a > 10)
            c += Aaa(a);
        return c + b;
    }

    [CAAnalyze(6f)]
    public static float Square(float a)
    {
        return a * a;
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public static float Do1(float a)
    {
        return a * 2;
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public static float Aaa(float h)
    {
        return h * h * h;
    }
}

See the output.