Re: (solved) Re: undefined symbol: __aarch64_ldadd8_acq_rel since llvm12 (mongodb44)

From: Ronald Klop <ronald-lists_at_klop.ws>
Date: Fri, 3 Sep 2021 13:29:01 +0200 (CEST)
 
Van: Michal Meloun <meloun.michal_at_gmail.com>
Datum: vrijdag, 3 september 2021 12:02
Aan: Dimitry Andric <dim_at_FreeBSD.org>, Ronald Klop <ronald-lists_at_klop.ws>
CC: FreeBSD Ports <freebsd-ports_at_freebsd.org>, Ed Maste <emaste_at_freebsd.org>
Onderwerp: Re: (solved) Re: undefined symbol: __aarch64_ldadd8_acq_rel since llvm12 (mongodb44)
> 
> 
> 
> On 03.09.2021 11:08, Dimitry Andric wrote:
> > On 3 Sep 2021, at 10:31, Ronald Klop <ronald-lists_at_klop.ws> wrote:
> >>
> >> Van: Ronald Klop <ronald-lists_at_klop.ws>
> >> Datum: zondag, 25 juli 2021 08:10
> >> Aan: FreeBSD Ports <freebsd-ports_at_freebsd.org>, Michal Meloun <mmel_at_FreeBSD.org>, Dimitry Andric <dim_at_FreeBSD.org>
> >> Onderwerp: (solved) Re: undefined symbol: __aarch64_ldadd8_acq_rel since llvm12 (mongodb44)
> >>
> >> Van: Dimitry Andric <dim_at_FreeBSD.org>
> >> Datum: 19 juli 2021 11:01
> >> Aan: Ronald Klop <ronald-lists_at_klop.ws>
> >> CC: FreeBSD Ports <freebsd-ports_at_freebsd.org>, Michal Meloun <mmel_at_FreeBSD.org>
> >> Onderwerp: Re: undefined symbol: __aarch64_ldadd8_acq_rel since llvm12 (mongodb44)
> >>
> >>>
> >>>
> >>> On 19 Jul 2021, at 09:29, Ronald Klop <ronald-lists_at_klop.ws> wrote:
> >>>>>> Does anybody have a clue why I get this error since clang 12 on aarch64/arm64?
> >>>> Do other ports have the same issue?
> >>>>>> Regards,
> >>>> Ronald.
> >>>>>> Van: Ronald Klop <ronald-lists_at_klop.ws>
> >>>> Datum: vrijdag, 16 juli 2021 09:43
> >>>> Aan: freebsd-arm_at_freebsd.org
> >>>> Onderwerp: undefined symbol: __aarch64_ldadd8_acq_rel since llvm12 (mongodb44)
> >>>>> Hi,
> >>>>> I'm also maintaining databases/mongodb44 and this gives undefined symbols since llvm12 (I think).
> >>>>> See: http://www.ipv6proxy.net/go.php?u=http://ampere2.nyi.freebsd.org/data/main-arm64-default/pf44e1c1de734_s63ca9ea4f3/logs/errors/mongodb44-4.4.6.log
> >>>>> And look for:
> >>>>> ld.lld: error: undefined symbol: __aarch64_ldadd8_acq_rel
> >>>>> There are a bunch of similar symbols not found while linking.
> >>>>> This compiles fine using llvm11 or on amd64.
> >>>
> >>> It turns out clang 12 now enables -moutline-atomics by default, but we
> >>> don't yet expose the necessary functions from compiler-rt. As a
> >>> temporary workaround, compile with -mno-outline-atomics. Meanwhile,
> >>> adding these functions to the compiler-rt lib is on my TODO list.
> > ...
> >> Hi Dimitry,
> >>
> >> I upgraded my poudriere and tried building mongodb50 with -moutline-atomics enabled.
> >> [00:00:56] Host OSVERSION: 1400032
> >> [00:00:56] Jail OSVERSION: 1400032
> >> ...
> >>
> >> [09:33:20] ld.lld: error: undefined hidden symbol: __aarch64_have_lse_atomics
> >> [09:33:20] >>> referenced by outline_atomic_cas1_1.S:121
> >> [09:33:20] >>>               outline_atomic_cas1_1.o:(__aarch64_cas1_relax) in archive /usr/lib/libgcc.a
> >> [09:33:20] >>> referenced by outline_atomic_cas1_1.S:121
> >> [09:33:20] >>>               outline_atomic_cas1_1.o:(__aarch64_cas1_relax) in archive /usr/lib/libgcc.a
> >> [09:33:20] >>> referenced by outline_atomic_cas1_2.S:121
> >> [09:33:20] >>>               outline_atomic_cas1_2.o:(__aarch64_cas1_acq) in archive /usr/lib/libgcc.a
> >> [09:33:20] >>> referenced 57 more times
> >> [09:33:20] c++: error: linker command failed with exit code 1 (use -v to see invocation)
> >> [09:33:22] scons: *** [build/opt/mongo/shell/mongo] Error 1
> >> [09:33:58] ld.lld: error: undefined hidden symbol: __aarch64_have_lse_atomics
> >> [09:33:58] >>> referenced by outline_atomic_cas1_1.S:121
> >> [09:33:58] >>>               outline_atomic_cas1_1.o:(__aarch64_cas1_relax) in archive /usr/lib/libgcc.a
> >> [09:33:58] >>> referenced by outline_atomic_cas1_1.S:121
> >> [09:33:58] >>>               outline_atomic_cas1_1.o:(__aarch64_cas1_relax) in archive /usr/lib/libgcc.a
> >> [09:33:58] >>> referenced by outline_atomic_cas1_2.S:121
> >> [09:33:58] >>>               outline_atomic_cas1_2.o:(__aarch64_cas1_acq) in archive /usr/lib/libgcc.a
> >> [09:33:58] >>> referenced 57 more times
> >> [09:33:59] c++: error: linker command failed with exit code 1 (use -v to see invocation)
> >> [09:33:59] scons: *** [build/opt/mongo/db/mongod] Error 1
> >
> > It turns out that this function is defined in
> > contrib/llvm-project/compiler-rt/lib/builtins/cpu_model.c, which we have
> > compiled only for x86 until now. But upstream added the
> > __aarch64_have_lse_atomics symbol and an initialization function for it.
> >
> > I tried simply enabling this file for the libcompiler-rt build on
> > aarch64, but it turns out that the implementation is rather Linuxy, and
> > does not compile on FreeBSD:
> >
> > contrib/llvm-project/compiler-rt/lib/builtins/cpu_model.c:778:25: error: implicit declaration of function 'getauxval' is invalid in C99 [-Werror,-Wimplicit-function-declaration]
> >    unsigned long hwcap = getauxval(AT_HWCAP);
> >                          ^
> > 1 error generated.
> >
> > The implementation looks like:
> >
> > // LSE support detection for out-of-line atomics
> > // using HWCAP and Auxiliary vector
> > _Bool __aarch64_have_lse_atomics
> >      __attribute__((visibility("hidden"), nocommon));
> > #if defined(__has_include)
> > #if __has_include(<sys/auxv.h>)
> > #include <sys/auxv.h>
> > #ifndef AT_HWCAP
> > #define AT_HWCAP 16
> > #endif
> > #ifndef HWCAP_ATOMICS
> > #define HWCAP_ATOMICS (1 << 8)
> > #endif
> > static void CONSTRUCTOR_ATTRIBUTE init_have_lse_atomics(void) {
> >    unsigned long hwcap = getauxval(AT_HWCAP);
> >    __aarch64_have_lse_atomics = (hwcap & HWCAP_ATOMICS) != 0;
> > }
> >
> > but in FreeBSD we don't have getauxval(), and I'm unsure if AT_HWCAP
> > works on aarch64.
> >
> > As this is all outside my domain of knowledge I'd appreciate if somebody
> > knows how to retrieve this information on our AArch64 implementation. :)
> >
> > -Dimitry
> >
> 
> We already have elf_aux_info(3) as counterpart to getauxval() and AT_HWCAP and AT_HWCAP2 are supported on arm64 as well, so the fix is trivial.
> I'll try  do it over the weekend, but I'm having some weird HW issues on my build box so, I'm not sure if it will be possible...
> 
> Michal
> 
> 
> 


Hi,

Dusted off my rusty C skills and came up with a patch.
I hope the attachment comes through.

It is currently building mongodb again to see if poudriere becomes happy. Will report back in a couple of hours.

Regards,
Ronald.
 
Received on Fri Sep 03 2021 - 11:29:01 UTC

Original text of this message