Reduce size of generated code

[Bergmann89]

Large schemas have the problem, that the generated code base it huge and is not able to being compiled anymore. We need to find a suitable way to reduce the size and the complexity of the generated code.

Here is a list of potential solutions:

• Instead of generics use runtime evaluation (like Box<dyn Any>).
• Use crates instead of modules to be able to compile them in parallel

[rcbevans]

Running the generated code through rustfmt so it's not all on one line seemed to reduce build time from 4 minutes to 45 seconds for the hpxml v4 schema.

This is still slow (though it is I think ~520k lines of generated code)

Splitting generated types across crates for compilation unit parallelism with a root facade to export as a single interface would allow for much more efficient build times for complex schemas

[Bergmann89]

Running the generated code through rustfmt so it's not all on one line seemed to reduce build time from 4 minutes to 45 seconds for the hpxml v4 schema.

OK, this is a really interesting observation. Unfortunately, I’m having compilation issues with some schemas, and rustfmt is also unable to produce suitable output in a reasonable amount of time. So I may need to add some newlines to the generated code to achieve at least some basic formatting. I’ll investigate this further once I find the time.

Splitting generated types across crates for compilation unit parallelism with a root facade to export as a single interface would allow for much more efficient build times for complex schemas

There is already an option to split the generated code into multiple modules, depending on the namespace and/or the schema the element was defined in. I also thought about to have a module per type. So rust should already be able to compile the modules in parallel, but I think the actual problem is the linking.

[rcbevans]

In rust a compilation unit maps to a crate, modules are just logical namespaces for code, if they're in the same crate, they compile within the same compilation unit.

To have rust use multiple workers to build in parallel the generated code needs splitting into multiple crates.

In C/C++ every file is its own translation unit and produces a separate .so which gets linked at the end, that's not so with rust.

Once rustc has run type checking, borrow checking etc and output the IL, llvm can then work concurrently across codegen units to output the compiled code.

[Bergmann89]

Ah yes, sorry. You are right. I'll add this to the list of potential improvements.