Use cmx reachability to decide locality of symbols

[Gbury]

This builds on top of #1222 .

The first commit adds some useful printing functions, as well as prefix all symbols in the cmm output to indicate whether the symbol is global or local (which is extremely helpful for debugging, as most of the bugs I've come across were because the same symbol was declared and used with different "localities").

The second commit centralizes the handling of symbols in to_cmm. Basically, after this commit, all cmm symbols are now generated in to_cmm_result.ml, and there should not be anywhere that still uses strings for symbols (instead using either an flambda Symbol.t or a Cmm.symbol created by the appropriate function in to_cmm_result.ml).

Finally, the third commit uses the name occurrences that are computed during cmx file preparation to decide which symbols should be local. The exact criterion is to use local cmm symbols for symbols from the current compilation unit, and that do not appear in the reachable names.

[stedolan]

Sounds good!

I haven't read the code in detail, but from the description I wonder whether there's a missed optimisation: it is in fact possible and useful to have a Local reference to a Global symbol. (This occurs when a symbol is exported, but there are also references to that symbol from within the current file).

To see how much impact it's having, I recommend running these commands before and after the change:

$ nm _install/lib/ocaml/compiler-libs/ocamlcommon.a  | wc -l
$ objdump -r _install/lib/ocaml/compiler-libs/ocamlcommon.a  | wc -l
$ objdump -r _install/lib/ocaml/compiler-libs/ocamlcommon.a  | grep -v ' \.' | wc -l

These are (approximately, wc -l includes headers etc) the counts of:

• symbols
• relocations
• relocations with symbol references (as opposed to section references)

I'd expect all three to decrease. Also, it's better to measure after configuring with --disable-function-sections.

[Gbury]

Yeah, I was reviewing #1222 and saw that indeed we could do that. I'll try and see what happens.

[Gbury]

I implemented the use of Local for symbols defined in the same compilation unit (see the branch even_fewer_symbols on my fork).

Here are the results of the commands (configured with --disable-function-section):

	main	this PR	even_fewer_symbols
nm	49238	29221 (-40%)	29218 (-40%)
objdump -r	226643	198975 (-12%)	194792 (-14%)
objdump -r | grep -v ' \.'	151517	91836 (-40%)	72655 (-52%)
ocamlopt.opt	49M	44M (-10%)	44M (-10%)