How is Array.Copy implemented in C#?

Question

I tried to look at the implementation of Array.Copy in C# with ILSpy but it didn\'t show me the implementation itself.

I wrote a simple benchmark, Array

Answer 1

Disassembling the Array class will land you on this declaration:

[MethodImpl(MethodImplOptions.InternalCall), SecurityCritical, ReliabilityContract(Consistency.MayCorruptInstance, Cer.MayFail)]
internal static extern void Copy(Array sourceArray, int sourceIndex, Array destinationArray, int destinationIndex, int length, bool reliable);

The [MethodImpl] attribute tells the JIT compiler that the method is actually implemented in the CLR, written in C++ instead of a managed language. It looks in a table of method names and retrieves a pointer to the C++ function that implements the method and compiles it to a simple CALL instruction.

Getting the source code for the CLR is a bit tricky, but the SSCLI20 version is pretty accurate for methods that have been around for a long time and didn't require tweaking. Array.Copy() certainly qualifies. The table I mentioned is defined in clr\src\vm\ecall.cpp, the section that's relevant to your question looks like this:

FCFuncStart(gArrayFuncs)
    FCFuncElement("Copy", SystemNative::ArrayCopy)
    FCFuncElement("Clear", SystemNative::ArrayClear)
    FCFuncElement("get_Rank", Array_Rank)
    //  etc...

The SystemNative::ArrayCopy() function pointer takes you to clr\src\vm\comsystem.cpp. The actual function is too big to copy here without making your eyes glaze over, there is a lot of error checking going on. It looks for a way to optimize the copy, the happy case is where the elements of the array can simply be copied without being converted. That is done by a function named m_memmove(). You'll find that function in the same file, it is used in the 32-bit version of the CLR.

Which first copies a single byte at a time until the destination address is aligned on a multiple of 4 bytes. Then it copies 16 bytes at a time, 4 times 4, these copies are fast because they are aligned. Then it copies what's left one byte at a time.

You can perhaps now see why it can be faster than your own loop. It can move 4 bytes at a time even if the array element size is not 4 bytes wide. And it can do so while ensuring the copy address is aligned, you can't since the physical address of the array element isn't discoverable.

Answer 2

Same techniques used to write a fast memcpy function:

• loop unrolling
• transfer of aligned data in large chunks (often using SIMD)
• CPU caching hints (SIMD helps here as well)

See also:

• Very fast memcpy for image processing?
• How to increase performance of memcpy
• does rewriting memcpy/memcmp/... with SIMD instructions make sense