Linking currently takes a long time because many symbol references (in particular, to statically allocated data) are not resolved until link time, despite many of these references being to data defined in the same file.

So far, this patch removes about half the symbols in compiler-libs, mostly unexported data. (The benefit applies only to Closure builds at the moment).

The major change is to use the following type for symbols in Cmm onwards, instead of string:

type is_global = Global | Local

(* Symbols are marked with whether they are local or global,
   at both definition and use sites.
   Symbols defined as [Local] may only be referenced within the same file,
   and all such references must also be [Local].
   Symbols defined as [Global] may be referenced from other files.
   References from other files must be [Global], but references from the
   same file may be [Local].
   (Marking symbols in this way speeds up linking, as many references can
   then be resolved early) *)
type symbol =
  { sym_name : string;
    sym_global : is_global }

At assembly time, Local symbols are generated as assembler labels rather than ELF symbols, so references can be resolved eagerly. Global symbols are generated as both an ELF symbol and an assembler label, so that Local references within the same file can use the label form.

There is quite a lot of low-hanging fruit here. The last commit here removes another ~8% of symbol references by wrapping caml_call_gc once per file. We should also be able to turn references to functions in the same file into assembler label references, and to improve the encoding of debug info and frame tables to generate fewer relocations.