Allow -main argument to override TLS entry point

[timcassell]

There is interest for BenchmarkDotNet to support the new dotnet run app.cs feature in .Net 10. dotnet/BenchmarkDotNet#2754

Currently, BenchmarkDotNet generates a separate project with a generated entry point. But it won't work with app.cs files, because we can't use <ProjectReference> on them, and we can't simply <Compile Include="app.cs"> either, because of all the new # directives (and still the TLS issue).

The best way to do it is using Roslyn Source Generators. However, the problem is we can't generate an entry point if top-level statements are used, because the compiler does not allow -main argument. TLS are much more likely to be used in the new dotnet feature.

So what we need is for the already existing -main compiler argument to simply let us use it to override TLS.

I don't know what was the motivation to block it when TLS was introduced.

[jjonescz]

Allowing -main to be specified for programs with top-level statements is just part of the problem. Then your source generator would need to call into the TLS entrypoint, right? But that currently has an unspeakable name (Program.<Main>$) so you couldn't do that without reflection. Seems straightforward to change that too though (make the name speakable).

I was also thinking about a compiler-only solution via a feature flag like TopLevelStatements=SomeName which would tell the compiler to emit top-level statements into method SomeName but not set them as entrypoint. Then your source generator could emit a normal Main method and call the SomeName method.

[timcassell]

Though I can see how calling the original TLS entry point could be useful for other tools.

[jjonescz]

the user's entry point is never used after that.

Can you clarify this bit? The code in the user's entry point must surely have some effect, right? Is your generated entry point going to be somehow derived from the user's code?

[timcassell]

We generate code from the benchmark classes and config values. The user's entry point just calls into BDN (and whatever else the user wants to do).

[RikkiGibson]

No we don't need to call TLS, we can generate a main method the old, speakable way. It needs to be speakable to pass the main argument anyway afaict. When we generate our entry point, the user's entry point is never used after that.

I don't think I'm following, when is the user's TLS main actually run? Is there a "user" process that runs the "user" main, which ends up spawning a "BDN" process that runs the BDN main, or some such?

[timcassell]

You got it. The user runs what we call the host process, then when they call into BDN to run the benchmarks, we generate the code to spawn child processes that run the actual benchmarks.