Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Spreads' native dependencies and low-level IL methods.


Native compression

Spreads.Native.Compression class exposes methods from Blosc:

  • SIMD-optimized shuffle/unshuffle.
  • Compression: LZ4, Zstd, Zlib/GZip/Deflate compression/decompression. Currently works on Windows x64/x86 and Linux x64 (tested on WSL & Docker Ubuntu). Targets netstandard2.0.


Full mimalloc API in .NET.

Cpu.GetCurrentCoreId method

Equivalent of Thread.GetCurrentProcessorId method that works on .NET Standard 2.0 and guarantees that the returned value could be used directly as an index in arrays with Cpu.CoreCount length. This allows to avoid expensive modulo operation in the most common use cases of per-core data structures.



UnsafeEx class contains unsafe IL helper methods that we cannot implement in C#.

Constrained generic calls without constraints

Generic methods ending with Constrained emit a constrained call to instance methods of known interfaces on instances of a generic type T without a type constraint where T : IKnownInterface<T>.

For example, calling the IComparable<T>.CompareTo method is implemented like this:

  .method public hidebysig static int32 CompareToConstrained<T>(!!T& left, !!T& right) cil managed aggressiveinlining
        .custom instance void System.Runtime.Versioning.NonVersionableAttribute::.ctor() = ( 01 00 00 00 )
        .maxstack 8
        ldobj !!T
        constrained. !!T
        callvirt instance int32 class [System.Runtime]System.IComparable`1<!!T>::CompareTo(!0)
  } // end of method Unsafe::CompareToConstrained

In addition to the IComparable<T> interface there are IEquatable<T> and the following custom ones in Spreads namespace:


public interface IDelta<T>
    T AddDelta(T delta);
    T GetDelta(T other);


public interface IInt64Diffable<T> : IComparable<T>
    T Add(long diff);
    long Diff(T other);


The main use case and sample usage is KeyComparer<T>. A benchmark shows that the unsafe CompareToConstrained method and the KeyComparer<T> that uses it are c.2x faster than the Comparer<T>.Default when called via the IComparer<T> interface and are c.1.6x faster when the default comparer is called directly as a class.


Case MOPS Elapsed GC0 GC1 GC2 Memory
Unsafe 403.23 248 ms 0.0 0.0 0.0 0.000 MB
KeyComparer* 396.83 252 ms 0.0 0.0 0.0 0.000 MB
Default 255.75 391 ms 0.0 0.0 0.0 0.000 MB
Interface 211.42 473 ms 0.0 0.0 0.0 0.000 MB

* KeyComparer<T> uses the JIT compile-time constant optimization for known types and falls back to the Unsafe.CompareToConstrained method for types that implement IComparable<T> interface. On .NET 4.6.1 there is no visible difference with and without the special cases: Unsafe.CompareToConstrained performs as fast as the if (typeof(T) == typeof(Int32)) { ... } pattern. See the discussion here and implementation with comments here explaining why the special cases could be needed on some platforms.

Unsafe methods could only be called on instances of a generic type T when the type implements a relevant interface. KeyComparer<T> has a static readonly field that (in theory) allows to use the same JIT optimization mentioned above:

private static readonly bool IsIComparable = typeof(IComparable<T>).GetTypeInfo().IsAssignableFrom(typeof(T));

public int Compare(T x, T y)
    if (IsIComparable) // JIT compile-time constant 
    return Unsafe.CompareToConstrained(ref x, ref y);

But even if such optimization breaks in this particular case (see the linked discussion) then checking a static bool field is still much cheaper than a virtual call, especially given that its value is constant for the lifetime of a program and branch prediction should be 100% effective.


Another use case is FastDictionary<TKey,TValue> that uses unsafe methods via KeyEqualityComparer<T>, which is very similar to KeyComparer<T> above. FastDictionay is a rewrite of S.C.G.Dictionary<TKey,TValue> that avoids virtual calls to an equality comparer.

A benchmark for <int,int> types shows that FastDictionary<int,int> is c.70% faster than S.C.G.Dictionary<int,int>:


Case MOPS Elapsed GC0 GC1 GC2 Memory
FastDictionary 120.48 415 ms 0.0 0.0 0.0 0.000 MB
Dictionary 71.63 698 ms 0.0 0.0 0.0 0.000 MB

Such implementation is much simpler than one with an additoinal generic parameter for a comparer, as recently discussed in this blog post. It is also more flexible than constraining TKey to where TKey : IEquatable<TKey> and gives the same performance.

Another benchmark with a key as a custom 16-bytes Symbol struct shows c.50% performance gain:


Case MOPS Elapsed GC0 GC1 GC2 Memory
FastDictionary 63.69 157 ms 0.0 0.0 0.0 0.000 MB
Dictionary 43.29 231 ms 0.0 0.0 0.0 0.000 MB


MPL 2.0. See the license file and third-party licenses.