System.Memory

[adamsitnik]

No description provided.

[adamsitnik]

Why is this manually unrolled? Can we use a loop instead?

It's due to the nature of Span - it's stack only, so I can't store it in the field of a class. So I need to create it as part of every benchmark. It's not a problem when it comes to time consuming benchmarks like span.CopyTo or span.Fill but span.Slice is super fast and if I write benchmark like this:

[Benchmark]
public Span<T> Slice() => new Span<T>(_arrayField).Slice(0, 1)

the Span constructor will be a huge part of the benchmark itself (up to 50%)

this is why I called the ctor once and used OperationsPerInvoke to amortize the cost of span creation. (the result is (1x span ctor + 16x slice) / 16) not 1x span ctor + 1x slice

Can we use a loop instead?

No, because the loop would be an overhead for this particular nano-benchmark

[adamsitnik]

result is a local variable and loop independent. Does the loop get optimized?

No, our compiler is not there yet.

[adamsitnik]

ok, I have covered entire Span, ReadOnlySpan, Memory, ReadOnlyMemory and MemoryMarshaller API with some good benchmarks. I believe the PR is ready for final review

[adamsitnik]

@jorive this is the last one ;D

[jorive]

        return result;;

nit: ;;

---

Refers to: src/benchmarks/corefx/System.Collections/Iterate/IterateFor.cs:75 in 68587cd. [](commit_id = 68587cd, deletion_comment = False)

[jorive]

[adamsitnik]

the port is complete now!