[mono] Simplify runtime assembly representation/remove MonoImage/MonoAssembly split

[CoffeeFlux]

In Mono, we have historically used two high-level structs on the native side to store assembly information—MonoImage and MonoAssembly. The split existed primarily due to MonoImage sharing across domains and multi-module assemblies, neither of which exist in .NET Core and will never be brought back.

For ALC support, we already ensured that on netcore we have a 1:1 relationship between a MonoImage and a MonoAssembly, and each MonoAssembly is tied to a single ALC. For netcore, we could viably get rid of MonoImage entirely and probably also stop refcounting the MonoAssembly. We might still want to move some of the information previous in MonoImage or MonoAssembly into a struct below the main one, but we should have a single top level struct to represent an assembly on the native side and a consistent load process to produce it.

The three main things preventing this are:

• Internal MonoImage * usage
• Embedding MonoImage * usage
• mono/mono support

I think for both MonoImage * cases, the solution would be to eliminate usage of the MonoImage struct within the loader itself and then just have an empty MonoImage * that ties back to the MonoAssembly *. Internally, we access some of the MonoImage fields directly and those cases would probably need to be switched to use an accessor function. For embedders, this should be fairly opaque, though we would probably end up creating more external-only functions taking in a MonoImage * as we switch over to MonoAssembly * everywhere internally.

mono/mono support is a little harder. We already have a lot of fairly distinct loader paths for netcore and legacy, so I personally think it's viable to just split this up further. If we stop mirroring at some point, this becomes much easier.

Tagging subscribers to this area: @CoffeeFlux
See info in area-owners.md if you want to be subscribed.

[naricc]

stop refcounting the MonoAssembly.

I'm not sure what you mean by this; what is the ref counting for? Ensuring that each MonoAssembly is tied to a single ALC?

[CoffeeFlux]

Currently the ref_count on a Monoassembly is incremented when stored in the ALC assemblies list, when stored in the domain assemblies list, and when resolved as a reference on a MonoImage. I don't think this level of granularity makes much sense in the context of a single domain and the netcore unloading algorithm, but maybe it's an orthogonal change to the unification? I'll write up more on how all this works with the unloading documentation.

[lambdageek]

I'm not sure this is a good idea (or at least - it's a complicated idea). The reason is that there are more kinds of ECMA 335 PE images than just assemblies. For example portable PDB files, or EnC delta files. They are all currently represented as a MonoImage.

What is confusing is that for images that do represent assemblies some of their attributes are stored in MonoImage fields.

Any refactoring should probably split out a low-level type that just knows about MonoImage as a container for tables and heaps (and the associate metadata parsing functions) and continue using it for assemblies, PDBs, etc. The rest of the stuff in MonoImage that is only relevant for assemblies can move to MonoAssembly.