Faster CodeGenerator if FSHARP6 defined

[Thorium]

One option for #405

I was testing SwaggerProvider performance issue of fsprojects/SwaggerProvider#150 (comment)
and even though still way too slow, this did gain a few seconds faster execution.

[Thorium]

This is helping generative TPs (like FSharp.Data and SwaggerProvider) but not erasing providers (like SQLProvider).

[vzarytovskii]

What are some benchmarks with the change?

It adds a lot of clutter and conditional "codegen", where runtime is generally quite fast with reference objects, which don't generally escape gen0.

Also, I don't really remember us (FSharp.Build or elsewhere) to define language constants. Are you sure it works as you expect it to? Or is it TP SDK which takes care of them?

[Thorium]

Because most of the TPs still refer ProvidedTypes.fs as a paket-dependencies file, not as NuGet package, the DefineConstants can be defined in the project referring the file.

It adds a lot of clutter and conditional

We still need a maintenance path for TPs. The current code is old and could be improved. But here, recent 6 weeks you can see that people still use mostly F# 5 and not the latest one. So maybe we need 2 versions? What is a way to separate them, same file, different file, is a political choice not interesting to me.

"codegen", where runtime is generally quite fast

Yeah, right, try the script, it'll complete in a minute (or multiple if you don't have i9 laptop):

#r "nuget: SwaggerProvider"
let [<Literal>] Schema = "https://raw.githubusercontent.com/stripe/openapi/master/openapi/spec3.sdk.json";;
let t = System.DateTime.Now;; //#time didn't give correct results so let's use the old way...
type Stripe = SwaggerProvider.OpenApiClientProvider<Schema, PreferAsync = true>;;
printfn "%f" (System.DateTime.Now-t).TotalSeconds;;

[cartermp]

I think there's a separate discussion to be had about if the base dependency should be FSharp.Core 6 or not. I lean towards that as a reasonable choice. But it seems preprocessors are out.

[vzarytovskii]

Yeah, right, try the script, it'll complete in a minute (or multiple if you don't have i9 laptop):

How does it compare to the changed one? Also, it does involve a bunch of IO, probably better to compare to the already downloaded json.

With local json it runs for 59-ish seconds on my machine.

[Thorium]

But it seems preprocessors are out.

Would you accept preprocessors at fsproj-level? So that there would be ProvidedTypesOld.fs and ProvidedTypesNew.fs ?

I think lot of F# 4 usage comes from the misunderstanding that the recent F# wouldn't support .NET Framework, and I think .NET Framework is still widely used.

With local json it runs for 59-ish seconds on my machine.

It's all ProvidedTypes.fs that seems to take the time, and there is a lot of work to do here.
It's like 7 megabyte file, handling that with modern computers shouldn't take a minute.

I'm not hoping to get this instant, but if that could be dropped to let's say to 10 seconds, it would also speed-up all the other TPs correspondingly.

This PR was not meant to be the full solution, it's just me testing if PRs are accepted, then I'll continue the work :-)

[Thorium]

By accepting a dependency to System.Memory in .NET Standard 2.0 (only) we'd get access to .AsSpan() and ReadOnlySpan<_> which could make parsing more efficient in at least .NET Standard 2.1 environment like dotnet.exe and VS Code.