Linker removes data from "copy" assemblies

[sbomer]

The "copy" action currently sometimes removes assemblyrefs and redirects references to removed type forwarders. This causes issues when there's non-IL data in an assembly that we want to copy as-is. For example, see https://github.com/dotnet/coreclr/issues/25334#issuecomment-506836765.

If we don't support this, there will be workarounds such as passing "skip" and manually copying over the file, or using '-reference' to specify all inputs and prevent resolving this reference. However this is a usability issue, so we may want to consider doing something more.

[sbomer]

More context: the issue we hit wasn't actually with non-IL data - it was with C++/CLI assemblies relying on the layout of the IL. For cases like this where reading/writing the assembly with cecil will break invariants that consumers rely upon, we should consider supporting a true "copy" mode that keeps the original assembly intact. This would also provide a way for users to tell the linker to keep an assembly as-is (similar to fully-preserved types) while still analyzing it to discover dependencies.

Seems to me that "copy" was originally intended to behave like I described, but it has drifted over time. @mrvoorhe, @marek-safar what do you think of this? Also relevant is the discussion in #628.

[sbomer]

@vatsan-madhavan asked in #658 (comment):

Asking for my edification: Does inspecting an assembly using Mono.Cecil API's rewrite the assembly - or is the rewriting specific to the linker implementation?

Just inspecting with Cecil won't rewrite it, but raising it to the cecil object model and writing it back to disk will. I think the fix will be in the linker - to only analyze using cecil, but not write using cecil. There might be a longer term fix to enable cecil to preserve invariants like the ordering of the IL data, but I think the immediate issue is in the linker. Specifically, the description of the "copy" action behavior is what we need, but isn't currently what actually happens.
https://github.com/mono/linker/blob/6071e0a4c8506db06b7ec1873d0b30e31b0cc870/src/linker/Linker/AssemblyAction.cs#L35

[marek-safar]

I think you can get the intended behaviour with the linker when you specify copy and --keep-facades true. That way the linker does not have to touch copied assemblies when forwarders are removed.

[sbomer]

Won't it still rewrite copy assemblies when it removes references in https://github.com/mono/linker/blob/master/src/linker/Linker.Steps/SweepStep.cs#L225? @marek-safar @tlakollo