Scope for improving Microsoft.AspNetCore.Server.Kestrel.Internal.Http.Frame.TakeStartLine(SocketInput)

[sivarv]

TakeStartLine() is one of the hot methods in TechEmPower PlainText and JSON benchmarks.
Jitted code of this method could be improved in the following way:

This method calls the following methods by passing structs by value

MemoryPoolIteratorExtensions:GetAsciiString(struct,struct)
MemoryPoolIteratorExtensions:GetKnownMethod(struct,byref)
MemoryPoolIteratorExtensions:GetKnownVersion(struct,byref)
MemoryPoolIteratorExtensions:GetUtf8String(struct,struct)

The struct param in question is MemoryPoolIterator, which is a struct with two fields: a ref and an int. On Amd64 a local copy of struct param is made and a reference to local copy is passed by reference. Since there are many such calls, at each call site struct params needs to be copied to local. This struct passing overhead could be avoided if caller TakeStartLine() passes struct params by reference.

[benaadams]

Currently these methods are extensions and C# doesn't allow the first parameter to be byref dotnet/roslyn#165 Interestingly VB.NET does.

GetKnownVersion and GetKnownMethod could probably be aggressively inlined and remain extensions (this would have same effect?) as they are only used in one place.

GetAsciiString and GetUtf8String could be changed to regular methods as they are fairly general?

I have a benchmark that gives consistent results so will measure.

/cc @halter73 @CesarBS @davidfowl

[benaadams]

ref changes benaadams@d3ffc67

Pre

Method	Mean	StdDev	Median	Scaled
ParsePlaintext	843.2055 ns	12.1499 ns	838.2252 ns	1.00
ParsePipelinedPlaintext	624.8637 ns	9.5501 ns	622.0941 ns	0.74
ParseLiveAspNet	3,990.4027 ns	40.8259 ns	3,999.2743 ns	4.73
ParsePipelinedLiveAspNet	3,640.7250 ns	55.9658 ns	3,630.0658 ns	4.32
ParseUnicode	6,957.7864 ns	134.4421 ns	6,943.0541 ns	8.25
ParseUnicodePipelined	6,560.8592 ns	100.9833 ns	6,598.5639 ns	7.78

Post

Method	Mean	StdDev	Median	Scaled
ParsePlaintext	785.7661 ns	15.2615 ns	785.2350 ns	1.00
ParsePipelinedPlaintext	620.8864 ns	6.3832 ns	624.4951 ns	0.79
ParseLiveAspNet	3,911.6923 ns	111.3273 ns	3,887.8120 ns	4.98
ParsePipelinedLiveAspNet	3,566.2522 ns	49.6079 ns	3,582.8719 ns	4.54
ParseUnicode	6,791.3005 ns	158.4868 ns	6,767.7367 ns	8.65
ParseUnicodePipelined	6,391.5840 ns	83.6463 ns	6,399.2505 ns	8.14

[benaadams]

Added commit to #1138