symbolizer.o should be buildable with a regular cmake rule

[pcc]

Bugzilla Link	30750
Version	unspecified
OS	Linux
Attachments	attempt
CC	@eugenis,@vitalybuka

Extended Description

The internal symbolizer.o object is currently built with the script compiler-rt/lib/sanitizer_common/symbolizer/scripts/build_symbolizer.sh.

We should really be using a simpler approach to building this object
file, and it should be available as a regular cmake rule. Conceptually, we
want to be doing "ld -r" followed by "objcopy -G" to create a relocatable
object file with only our entry points exposed. Attached is an attempt at doing this in cmake. However, this does not work at present due to linker and objcopy bugs:

• the presence of local symbols named the same as global symbols confuses at
  least bfd and gold, so we need to rename local symbols
• ld.gold fails to link symbolizer0.o (fails an assertion)
• ld.bfd succeeds, but seemingly creates an invalid .eh_frame section (segfaults
  in the unwinder)
• ld.lld succeeds but creates multiple .eh_frame sections which confuses
  objcopy into creating an invalid output file (lld refuses to accept it, gold
  creates an invalid .eh_frame, bfd fails assertions)
  To the extent that these are bugs in lld and objcopy, we can probably fix the
  bugs in lld and write our own llvm-objcopy (perhaps with a nicer interface for
  doing this transformation), then call the tools directly from cmake.

[Enna1]

FYI, I tried using ld.lld -r followed by llvm-objcopy -G to create the symbolizer.o, sanitizers tests (ninja check-msan check-tsan check-asan check-hwasan check-lsan check-ubsan check-sanitizer) are passed with this version symbolizer.o.

--no-dependent-libraries is passed to ld.lld to bypass the link error "unable to find library from dependent library specifier: pthread"

But ld.lld -r + llvm-objcopy -G version symbolizer.o is extremely large, compared with symbolizer.o created with current llvm-link + opt -internalize approach. In my case, it is 33M v.s. 3.7M.

cc @vitalybuka