Global dead code elimination on final firmware image

[rmn30]

At the moment we have dead code elimination within a compartment via -ffunction-sections -fdata-sections and -gc-sections but all exported functions are used as roots for GC purposes even if they're never used in the final firmware image. This applies to both compartments and libraries, which is really bad for the final image size. To solve this properly we could use a global pass at the point of firmware linking, but this will probably require extensive changes to the linkage model. Doing it just for library compartments may be easier or we could support explicitly enable and disabling specific entry points when building compartments, although this may be a bit painful to use.

[resistor]

It seems like we could easily narrow the roots set to only compartment export tables, right? Are library export tables allowed to be thread entry points?

[resistor]

Thinking about this further, the obvious way to do it would be with --gc-sections with some customization of the GC root logic, but by the final link step we've already merged all of the function sections away. So this would necessitate retaining un-merged sections until final link, which is at odds with the goal of doing intra-compartment layout at compartment link time.