Make it easier to add built-in sections

[davidlattimore]

When adding new built-in sections, or sections that we need to reference from linker code, there's currently quite a few places that need to be changed. The tests help a bit, since they generally tell you if you miss one of the places that needs changing. It's still not ideal though.

There are a few things that make fixing this complicated.

• Sections that come from input files, we bucket by alignment, whereas for sections that we generate internally, we don't.
• Some internally generated sections have two parts that we track separately, but then merge into a single section for output. For example, .symtab is split into symtab_locals and symtab_globals, since all the locals needs to come before all the globals. A similar split occurs with .rela.dyn which has all relative relocations before all others.

Since this is a pretty large change, it'd be good to try to do it in a few stages, although it's entirely possible that once the thread is pulled, it'll all unravel and end up as one massive change.

I think probably the key will be to introduce SectionPartId or even just PartId. A section can then be composed of multiple Parts. Each Part would have a single alignment.

Change OutputSectionPartMap to do all access via PartId rather than via direct references to named fields. AlignmentMap would go away and be flattened into OutputSectionPartMap.

We might, at least temporarily, need a way to map between PartId and OutputSectionId. This is probably easier if we put all the 1-1 parts together, then the weird 2-1 parts (mentioned above) then all the 16-1 parts, which are categorised by alignment.

A careful eye needs to be kept on performance when doing this to make sure we don't regress. I think that should be entirely doable, so long as things that are currently together in a single heap allocation remain in a single heap allocation, we continue to use constant offsets into Vecs etc. I'm hopeful that we might even see a slight performance improvement due to the code being a bit simpler with data being treated in a more homogenous way.

[davidlattimore]

I think this is done, at least for now. There's still a few steps to adding a new section, but they should all be relatively straightforward. I think further reduction would require some kind of code generation. e.g. macro, proc-macro or a build script. I'm not sure it's worthwhile doing that - especially since those often have costs around how easy it is to understand what's going on in the code.