Optimize message body encoding

[pentp]

• Don't encode PooledResponse<T> wrapper types for normal response messages, only encode T.
• Fix some more expected-type handling corner cases.
• Some small Writer and Field/Tag optimizations.

Microsoft Reviewers: Open in CodeFlow

[ReubenBond]

Thanks for all the PRs! Do you mind annotating some of the changes in GitHub using review comments to include some reasoning, etc behind the changes and adding benchmark results, if you have any?

[gfoidl]

This is not necessary, best codegen is when written as

public void WriteByte(byte value)
{
    int bufferPos = _bufferPos;
    if ((uint)bufferPos < (uint)_currentSpan.Length)
    {
        _currentSpan[bufferPos] = value;
        _bufferPos = bufferPos + 1;
    }
    else
    {
        WriteByteSlow(value);
    }
}

• there's no bound check emitted
• _bufferPos is read and written only once (memory accesses), while with _bufferPos++ it's 2 reads + 1 write (for the inc)

[ReubenBond]

Your input is appreciated, @gfoidl

[pentp]

There's still going to be a bounds check unless _currentSpan is also first copied to a local variable, but that introduces additional overhead.

[gfoidl]

Just curious: on what target platform did you see this? Is this after inlining?
Asking because I checked the codegen and didn't see the bound check -- so the JIT should be taught to eliminate that one.

[pentp]

I tested on x64 and after inlining.

[gfoidl]

Well this seems to be dotnet/runtime#72004 (actually opened by myself, and I forgot about this 😉).

[ReubenBond]

Perhaps we should link to that issue and this comment from code so that it can be re-assessed later

[gfoidl]

Suggested change

	var neededBytes = BitOperations.Log2(value) / 7;
	var neededBytes = (int)((uint)BitOperations.Log2(value) / 7);

More ugly code, but shorter machine code.

[ReubenBond]

Nice work. I wonder if this results in measurable differences.
There are serialization microbenchmarks in the test/Benchmarks directory which can be invoked with
dotnet run -c Release -f net7.0 -- suite --filter *erializeBenchmark to run the serialization comparison benchmarks.

[gfoidl]

Did two runs of the benchmarks (current main, then changed to the uint-division). The runs aren't comparable 1:1, but judging based on Ratio-column it improved a bit, which I expected because of less code and the mul instead of imul.

Current main branch

|            Method |       Mean |    Error |   StdDev | Ratio | RatioSD |   Gen0 | Payload | Allocated | Alloc Ratio |
|------------------ |-----------:|---------:|---------:|------:|--------:|-------:|--------:|----------:|------------:|
|           Orleans |   107.1 ns |  2.12 ns |  3.17 ns |  1.00 |    0.00 |      - |    20 B |         - |          NA |
|          Utf8Json |   169.6 ns |  3.23 ns |  7.50 ns |  1.59 |    0.09 |      - |   154 B |         - |          NA |
|    SystemTextJson |   571.5 ns |  5.31 ns |  4.97 ns |  5.29 |    0.14 | 0.0019 |   154 B |      56 B |          NA |
| MessagePackCSharp |   115.0 ns |  0.84 ns |  0.71 ns |  1.06 |    0.03 | 0.0014 |    10 B |      40 B |          NA |
|       ProtobufNet |   283.3 ns |  3.69 ns |  2.88 ns |  2.61 |    0.06 |      - |    18 B |         - |          NA |
|          Hyperion |   165.0 ns |  3.33 ns |  7.66 ns |  1.53 |    0.06 | 0.0019 |    39 B |      56 B |          NA |
|    NewtonsoftJson | 1,700.0 ns | 33.79 ns | 78.31 ns | 16.21 |    0.87 | 0.0763 |   154 B |    2080 B |          NA |
|          SpanJson |   192.5 ns |  2.09 ns |  2.64 ns |  1.79 |    0.05 | 0.0067 |   154 B |     184 B |          NA |

With unsigned divsion

|            Method |        Mean |     Error |    StdDev | Ratio | RatioSD |   Gen0 | Payload | Allocated | Alloc Ratio |
|------------------ |------------:|----------:|----------:|------:|--------:|-------:|--------:|----------:|------------:|
|           Orleans |    95.51 ns |  0.351 ns |  0.311 ns |  1.00 |    0.00 |      - |    20 B |         - |          NA |
|          Utf8Json |   156.21 ns |  1.141 ns |  0.953 ns |  1.64 |    0.01 |      - |   154 B |         - |          NA |
|    SystemTextJson |   537.95 ns |  4.033 ns |  3.575 ns |  5.63 |    0.04 | 0.0019 |   154 B |      56 B |          NA |
| MessagePackCSharp |   118.38 ns |  2.122 ns |  1.985 ns |  1.24 |    0.02 | 0.0014 |    10 B |      40 B |          NA |
|       ProtobufNet |   273.29 ns |  1.479 ns |  1.311 ns |  2.86 |    0.02 |      - |    18 B |         - |          NA |
|          Hyperion |   152.09 ns |  1.320 ns |  1.170 ns |  1.59 |    0.01 | 0.0019 |    39 B |      56 B |          NA |
|    NewtonsoftJson | 1,594.22 ns | 15.754 ns | 14.736 ns | 16.70 |    0.17 | 0.0763 |   154 B |    2080 B |          NA |
|          SpanJson |   189.82 ns |  1.633 ns |  1.528 ns |  1.99 |    0.02 | 0.0067 |   154 B |     184 B |          NA |

Full results:
Benchmarks.zip

[pentp]

These casts get rid of some signed extension instructions.

[ReubenBond]

(closed by mistake)