Linking FreeBSD with lld: -dc option missing

[Ed Maste via llvm-commits]

I'm now looking in more detail at the final two issues preventing us
from fully using lld for building FreeBSD userland and kernel. The
issues have been discussed in sub-threads in the past, but as we're
very close to building FreeBSD with lld "out of the box" I'm focusing
on final resolutions for them. I need to either have lld gain the last
pieces of missing functionality, or commit to a permanent change in
FreeBSD.

One of the outstanding issues is the -d/-dc/-dp option, which converts
common symbols into allocated .bss symbols with -r linking. The option
is used by FreeBSD's rescue utility, which is a Busybox-like tool that
combines many binaries into one - e.g., cat, ls, sh, etc., and is
probably best explained by the comment in the crunchide tool used in
the build process:

 * crunchide.c - tiptoes through a symbol table, hiding all defined
 *      global symbols.  Allows the user to supply a "keep list" of symbols
 *      that are not to be hidden.  This program relies on the use of the
 *      linker's -dc flag to actually put global bss data into the file's
 *      bss segment (rather than leaving it as undefined "common" data).
 *
 *      The point of all this is to allow multiple programs to be linked
 *      together without getting multiple-defined errors.
 *
 *      For example, consider a program "foo.c".  It can be linked with a
 *      small stub routine, called "foostub.c", eg:
 *          int foo_main(int argc, char **argv){ return main(argc, argv); }
 *      like so:
 *          cc -c foo.c foostub.c
 *          ld -dc -r foo.o foostub.o -o foo.combined.o
 *          crunchide -k _foo_main foo.combined.o
 *      at this point, foo.combined.o can be linked with another program
 *      and invoked with "foo_main(argc, argv)".  foo's main() and any
 *      other globals are hidden and will not conflict with other symbols.

In order to make this work with lld I need to do one of three things:

1) compile the object files with -fno-common
2) give crunchide some simple linking ability, so that it can convert
common symbols into allocated .bss symbols
3) add -dc support to lld

Rafael previously suggested #1. It's feasible, but makes the crunch
build system slightly less portable (in that it couldn't be used with
other software that relies on common symbol behaviour). It also
requires that we build a second set of objects for the rescue build,
if -fno-common isn't applied to the "normal" build of the tools.

#2 seems like more effort than is sensible.

I have a slight preference for #3, and I think will be a relatively
small change in lld. However, a mitigating factor is that I've had
trouble finding any other users of this option. I did find (through
Debian's code search) that vim and emacs include
FORCE_COMMON_ALLOCATION, which is the linker script option
corresponding to -dc. It wasn't immediately clear though if it's
actually used in the build on operating systems of interest.

Rafael/Rui, do you think it's reasonable to add -dc to lld?

-Ed