Avoid adding duplicates to SRCS/OBJS/SOBJS/POBJS

[arichardson]

Summary

This is a change in preparation for stopping to use lorder.sh (D26044) and
instead assume that we have a linker new than ~1990. Without lorder
duplicates end up being passed to the linker when building .so files and this
results in duplicate symbol definition errors otherwise.

There is one minor change: libcompiler_rt.a will no longer provide
gcc_personality_v0 and instead we now only have it in libgcc_eh.a/libgcc_s.so.
This matches GCC's behaviour.

[emaste]

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

Where do the duplicates come from? Perhaps we could just avoid adding the dups in the first place, although perhaps that's awkward and this is fine.

lib/libprocstat/Makefile

61

SOBJS/POBJS/ originally came in from rS221807, maybe just due to a misunderstanding. Normally only OBJS is set, so I'd argue the comment here is unnecessary and is probably more confusing to a future code archaeologist - I think it can just be deleted.

stand/userboot/userboot/Makefile

23

Is this just unused?

[arichardson]

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

The two include lines above seem to add duplicates. I think it might be a bit awkward to remove the duplicates but I can look into that tomorrow.

.include "../libcompiler_rt/Makefile.inc"
.include "../libgcc_eh/Makefile.inc"

lib/libprocstat/Makefile

61

Will fix.

stand/userboot/userboot/Makefile

23

It's added below in the .include "${BOOTSRC}/loader.mk" line.

[emaste]

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

OK - then I think it's reasonable to commit this one for now, and we can later look at avoiding the duplicates and removing this again.

stand/userboot/userboot/Makefile

23

OK

[arichardson]

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

Seems like the problem is gcc_personality_v0 which is included in both the compiler_rt and libgcc_eh makefile.inc.

@dim does it need to be in both?

It seems like on Linux libgcc (i.e. libcompiler_rt) doesn't provide it, only libgcc_eh (and libgcc_s.so.1) include it.

/usr/lib/gcc/x86_64-linux-gnu/7/libgcc.a:morestack.o: 0000000000000000 V DW.ref.__gcc_personality_v0
/usr/lib/gcc/x86_64-linux-gnu/7/libgcc.a:morestack.o:                  U __gcc_personality_v0
/usr/lib/gcc/x86_64-linux-gnu/7/libgcc_eh.a:unwind-c.o: 00000000000001b0 T __gcc_personality_v0

[emaste]

Doing some archaeology, it seems I moved gcc_personality_v0 to lib/libcompiler_rt/Makefile.inc (from lib/libcompiler_rt/Makefile) in rS306377, added to lib/libgcc_eh/Makefile.inc in rS307230 (D8188).

[arichardson]

Remove unncessary comment

[dim]

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

Good question. A very long time ago in rS276851, I added compiler-rt trunk r276851, and at that point added the gcc_personality_v0.c file to what was then lib/libcompiler_rt/Makefile. (This got renamed to Makefile.inc later.)

Then in rS307230, @emaste added the scaffolding Makefiles for libgcc_eh and libgcc_s based on compiler-rt, and this initial version also contained an entry for gcc_personality_v0.c.

I am unsure which library is *supposed* to provide it, though. Upstream compiler-rt places it in its lib/builtins directory, which typically goes in to libcompiler_rt.a.

Note that gcc_personality_v0.c has a comment added by @cem in rS354347, related to this:

* XXX On FreeBSD, this file is compiled into three libraries:
*   - libcompiler_rt
*   - libgcc_eh
*   - libgcc_s
*
* In the former, the include path points to the contrib/libcxxrt/unwind-arm.h
* copy of unwind.h.  In the latter, the include path points to the
* contrib/libunwind/include/unwind.h header (LLVM libunwind).
*
* Neither (seemingly redundant) variant of unwind.h needs the redefinitions
* provided in the "helpful" header below, and libcxxrt's unwind-arm.h provides
* *no* useful distinguishing macros, so just forcibly disable the helper
* header on FreeBSD.
*/

So this may indeed be some historical mistake. I'd say it only belongs in libcompiler_rt proper, that is, libcompiler_rt.a and libgcc_s.so, and not in libgcc_eh.a. But it is possible that it must also be available in libgcc_eh.a for compatibility reasons.

[dim]

Looking at what gcc does, from its build script and Makefiles, I infer that they only put the personality functions in libgcc_eh.a and libgcc_s.so, *not* in libgcc.a.

This is also what I see in actual installed binaries, for example from the gcc9 port:

% grep -R gcc_personality /usr/local/lib/gcc9
Binary file /usr/local/lib/gcc9/libgcc_s.so matches
Binary file /usr/local/lib/gcc9/gcc/x86_64-portbld-freebsd12.1/9.3.0/32/libgcc_eh.a matches
Binary file /usr/local/lib/gcc9/gcc/x86_64-portbld-freebsd12.1/9.3.0/libgcc_eh.a matches
Binary file /usr/local/lib/gcc9/libgcc_s.so.1 matches

E.g., I think we can remove the duplicate copy from libcompiler_rt.a safely, and ensure that gcc_personality_v0.c is only added to libgcc_eh.a and libgcc_s.so.

[cem]

In D26042#578249, @dim wrote:

E.g., I think we can remove the duplicate copy from libcompiler_rt.a safely, and ensure that gcc_personality_v0.c is only added to libgcc_eh.a and libgcc_s.so.

Sounds good to me.

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

If we remove gcc_personality_v0 from libcompiler_rt, can this SRCS deduplication line be removed?

[arichardson]

lib/libgcc_s/Makefile

56 ↗

(On Diff #75737)

If we remove gcc_personality_v0 from libcompiler_rt, can this SRCS deduplication line be removed?

Yes I believe that was the only duplicate. Will check shortly.

[arichardson]

• Remove duplicates in libgcc_s without uniq:
  • Remove gcc_personality from libcompiler_rt
  • move int_util.c from libgcc_eh/Makefile.inc to Makefile

[arichardson]

Fix libgcc_eh

[cem]

Thanks for doing this! I’ve been really happy to see all your bootstrap portability changes streaming in recently.

lib/libgcc_eh/Makefile

11

I’m not sure this needs a comment. Makefile.inc is the weird case, srcs in Makefile is normal. The commit log for this line will justify it for anyone especially curious.