Performance optimisations: Using more value-types and structs

[Thorium]

This commit reduces the memory usage and speeds up all the serialization.

I used also the benchmark-branch all the test that did have result for FSharp.Json, to run test between the current version and this commit, both run with latest FSharp.Core. Here are the results:

BenchmarkDotNet=v0.11.5, OS=Windows 10.0.22621
13th Gen Intel Core i9-13900H, 1 CPU, 20 logical and 14 physical cores
.NET Core SDK=8.0.100-rc.1.23463.5
  [Host]     : .NET Core 2.0.9 (CoreCLR 4.6.26614.01, CoreFX 4.6.26614.01), 64bit RyuJIT DEBUG
  DefaultJob : .NET Core 2.0.9 (CoreCLR 4.6.26614.01, CoreFX 4.6.26614.01), 64bit RyuJIT

1. BoxedArrayRoundtrip

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated	Version
Newtonsoft.Json	3.680 us	0.0302 us	0.0282 us	2.2736	0.0076	-	14 KB	13.0.3
FSharp.Json	12.753 us	0.1331 us	0.1180 us	1.9531	0.0153	-	12.12 KB	0.4.1
FSharp.Json	12.187 us	0.1455 us	0.1361 us	1.8311	0.0153	-	11.41 KB	This PR

5. FSharpBinaryRoundtrip

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated	Version
Newtonsoft.Json	21.00 ms	0.3417 ms	0.3196 ms	3156.25	968.75	375.0000	18.47 MB	13.0.3
FSharp.Json	120.64 ms	2.3876 ms	2.7495 ms	6600.00	200.00	-	40.88 MB	0.4.1
FSharp.Json	107.32 ms	1.5377 ms	1.4384 ms	6000.00	200.00	-	37.29 MB	This PR

6. FSharpListRoundtrip

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated	Version
Newtonsoft.Json	111.8 us	1.560 us	1.459 us	15.7471	0.7324	-	97.62 KB	13.0.3
FSharp.Json	1,012.4 us	12.436 us	11.632 us	138.6719	21.4844	-	869.35 KB	0.4.1
FSharp.Json	997.4 us	11.905 us	10.553 us	126.9531	19.5313	-	798.89 KB	This PR

11. LargeTupleRoundtrip

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated	Version
Newtonsoft.Json	7.808 us	0.0958 us	0.0896 us	2.6703	0.0305	-	16.59 KB	13.0.3
FSharp.Json	27.109 us	0.3354 us	0.3137 us	3.3569	0.0305	-	20.82 KB	0.4.1
FSharp.Json	26.810 us	0.5355 us	0.9656 us	3.1433	0.0305	-	19.59 KB	This PR

[Thorium]

The figures are still so bad that people should prefer System.Text.Json and Newtonsoft.Json over this library.

[bartelink]

Yes, far too many moving parts If going NSJ (which is a bad idea for new systems as it's a dead end), FsCodec.NewtonsoftJson includes required converters for things like options (FsCodec.SystemTextJson has shims too, but those are a lot less necessary than they are for NSJ) Also if you want lots of features and dot mind taking a dep on a larger lib, look at FSharp.SystemTextJson

[Thorium]

I asked @eiriktsarpalis the best method for custom parsing json with System.Text.Json, he said:

just use Utf8JsonReader under wraps. That way make sure it's compliant and also take advantage of all the vectorization code it's using under the hood.

So I tried a bit: https://gist.github.com/Thorium/09a67ce08adee4fcc02ec7f0048e6962 That could work here as well.

[bartelink]

I should mention though - while GC and perf for STJ is in another league, ultimately the perf of JSON serialization is not make or break for 99% of systems. i.e. my motivations for mentioning other libs is not primarily based on perf, more:

• for new systems, you get to make a choice. For older systems, deciding to shift serializers should not be taken lightly
• depending on less semi-maintained code with bespoke behaviors (vs FSharp.Json and FSharp.SystemTextJson)
• STJ and NSJ by having more traffic/being more mainstream is easier to search up answers for on SO and troubleshoot
• NSJ obviously has broad usage and probably most features, but is also extremely a clear dead end - just scan the issue tracker

[Thorium]

Fair point, however often JSON is used in data communications which often means there is either active user looking progress-bar, or there is a huge batch process going on, so speed is somewhat a property.

Beside the options you mentioned there is FSharp.Data.JsonProvider which is extremely convenient to use, having its own parser.

Right now I feel that F# eco-system has had this problem of everyone copy&pasting their own serialization implementations (and AI-tools will make this even easier in the future). And because everything is OSS, people also referencing single serialization-implementation files directly (e.g. paket files). That's why fixing old code is important even in old less-maintained sources.