libomp for ppc32 #30

barracuda156 · 2022-03-22T01:57:10Z

barracuda156
Mar 22, 2022

I was trying to build libomp for ppc, taking as an example implementation of mips from here: https://github.com/OMPI-X/libomp
However the build fails at the linking stage with:

Undefined symbols:
  "___sync_fetch_and_or_8", referenced from:
      ___kmp_suspend_64 in z_Linux_util.c.o
      ___kmp_suspend_64 in z_Linux_util.c.o
      ___kmp_suspend_oncore in z_Linux_util.c.o
      ___kmp_suspend_oncore in z_Linux_util.c.o
      __ZL33__kmp_hierarchical_barrier_gather12barrier_typeP8kmp_infoiiPFvPvS2_ES2_.isra.0 in kmp_barrier.cpp.o
      ___kmp_release_oncore in kmp_wait_release.cpp.o
  "___kmp_unnamed_critical_addr", referenced from:
      ___kmp_unnamed_critical_addr$non_lazy_ptr in kmp_gsupport.c.o
     (maybe you meant: ___kmp_unnamed_critical_addr$non_lazy_ptr)
  "___sync_fetch_and_and_8", referenced from:
      ___kmp_suspend_64 in z_Linux_util.c.o
      ___kmp_suspend_oncore in z_Linux_util.c.o
      ___kmp_resume_64 in z_Linux_util.c.o
      ___kmp_resume_oncore in z_Linux_util.c.o
      __ZL33__kmp_hierarchical_barrier_gather12barrier_typeP8kmp_infoiiPFvPvS2_ES2_.isra.0 in kmp_barrier.cpp.o
  "___sync_bool_compare_and_swap_8", referenced from:
      ___kmpc_atomic_float8_add in kmp_atomic.c.o
      ___kmpc_atomic_float8_sub in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_andb in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_div in kmp_atomic.c.o
      ___kmpc_atomic_fixed8u_div in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_mul in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_orb in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_shl in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_shr in kmp_atomic.c.o
      ___kmpc_atomic_fixed8u_shr in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_xor in kmp_atomic.c.o
      ___kmpc_atomic_float8_div in kmp_atomic.c.o
      ___kmpc_atomic_float8_mul in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_andl in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_andl in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_orl in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_max in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_min in kmp_atomic.c.o
      ___kmpc_atomic_float8_max in kmp_atomic.c.o
      ___kmpc_atomic_float8_min in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_neqv in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_eqv in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_add in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_sub in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_mul in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_div in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_mul_float8 in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_div_float8 in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_add_cmplx8 in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_sub_cmplx8 in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_mul_cmplx8 in kmp_atomic.c.o
      ___kmpc_atomic_cmplx4_div_cmplx8 in kmp_atomic.c.o
      ___kmpc_atomic_8 in kmp_atomic.c.o
      ___kmpc_dispatch_next_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8u in kmp_dispatch.cpp.o
      ___kmp_release_queuing_lock in kmp_lock.cpp.o
      ___kmp_test_drdpa_lock in kmp_lock.cpp.o
      __ZL45__kmp_acquire_nested_queuing_lock_with_checksP16kmp_queuing_locki in kmp_lock.cpp.o
      __ZL33__kmp_test_drdpa_lock_with_checksP14kmp_drdpa_locki in kmp_lock.cpp.o
      ___kmp_test_nested_drdpa_lock in kmp_lock.cpp.o
      __ZL40__kmp_test_nested_drdpa_lock_with_checksP14kmp_drdpa_locki in kmp_lock.cpp.o
      ___kmp_acquire_queuing_lock in kmp_lock.cpp.o
      __ZL38__kmp_acquire_queuing_lock_with_checksP16kmp_queuing_locki in kmp_lock.cpp.o
      ___kmp_acquire_nested_queuing_lock in kmp_lock.cpp.o
  "___sync_fetch_and_add_8", referenced from:
      ___kmpc_atomic_fixed8_add in kmp_atomic.c.o
      ___kmpc_atomic_fixed8_sub in kmp_atomic.c.o
      __ZL34__kmp_hierarchical_barrier_release12barrier_typeP8kmp_infoiiiPv in kmp_barrier.cpp.o
      __ZL34__kmp_hierarchical_barrier_release12barrier_typeP8kmp_infoiiiPv in kmp_barrier.cpp.o
      __ZL26__kmp_hyper_barrier_gather12barrier_typeP8kmp_infoiiPFvPvS2_ES2_.constprop.0 in kmp_barrier.cpp.o
      __ZL27__kmp_linear_barrier_gather12barrier_typeP8kmp_infoiiPFvPvS2_ES2_.constprop.0 in kmp_barrier.cpp.o
      __ZL25__kmp_tree_barrier_gather12barrier_typeP8kmp_infoiiPFvPvS2_ES2_.constprop.0 in kmp_barrier.cpp.o
      __ZL33__kmp_hierarchical_barrier_gather12barrier_typeP8kmp_infoiiPFvPvS2_ES2_.isra.0 in kmp_barrier.cpp.o
      __ZL28__kmp_linear_barrier_release12barrier_typeP8kmp_infoiiiPv in kmp_barrier.cpp.o
      __ZL26__kmp_tree_barrier_release12barrier_typeP8kmp_infoiiiPv in kmp_barrier.cpp.o
      __ZL27__kmp_hyper_barrier_release12barrier_typeP8kmp_infoiiiPv in kmp_barrier.cpp.o
      ___kmp_release_64 in kmp_wait_release.cpp.o
      ___kmpc_dispatch_next_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8u in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8u in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8u in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8u in kmp_dispatch.cpp.o
      ___kmpc_dispatch_next_8u in kmp_dispatch.cpp.o
      ___kmpc_dispatch_fini_8 in kmp_dispatch.cpp.o
      ___kmpc_dispatch_fini_8u in kmp_dispatch.cpp.o
      ___kmp_aux_dispatch_fini_chunk_8 in kmp_dispatch.cpp.o
      ___kmp_aux_dispatch_fini_chunk_8u in kmp_dispatch.cpp.o
      __ZL43__kmp_acquire_nested_drdpa_lock_with_checksP14kmp_drdpa_locki in kmp_lock.cpp.o
      ___kmp_acquire_drdpa_lock in kmp_lock.cpp.o
      __ZL36__kmp_acquire_drdpa_lock_with_checksP14kmp_drdpa_locki in kmp_lock.cpp.o
      ___kmp_acquire_nested_drdpa_lock in kmp_lock.cpp.o
  "___kmp_aux_dispatch_init_8u", referenced from:
      _GOMP_loop_ull_static_start in kmp_gsupport.c.o
      _GOMP_loop_ull_dynamic_start in kmp_gsupport.c.o
      _GOMP_loop_ull_guided_start in kmp_gsupport.c.o
      _GOMP_loop_ull_runtime_start in kmp_gsupport.c.o
      _GOMP_loop_ull_ordered_static_start in kmp_gsupport.c.o
      _GOMP_loop_ull_ordered_dynamic_start in kmp_gsupport.c.o
      _GOMP_loop_ull_ordered_guided_start in kmp_gsupport.c.o
      _GOMP_loop_ull_ordered_runtime_start in kmp_gsupport.c.o
ld: symbol(s) not found
collect2: error: ld returned 1 exit status

@iains I thought it may be due to the fact that z_Linux_asm.s does’t have definitions for ppc32. (Though there is nearly nothing for mips either.)
I have looked into your repo, and libomp appears to lack ppc32, unless I miss something: https://github.com/iains/LLVM-7-branch/tree/7.1.1-Darwin-WIP/openmp (judging from errors I was getting, it won’t even configure for powerpc-apple-darwin).
Is it because it won’t work regardless (i.e. inherently impossible due to OS lacking required functionality) or just because you did not fix it yet?

As always, I am not expecting some ready-made solution “for me”. Rather I have some interest to work on this and will appreciate advices from people who understand much more than me.

P. S. @catap @kencu @fangism Tagging you in a case you may be interested.

iains · 2022-03-22T09:05:01Z

iains
Mar 22, 2022
Maintainer

libomp is not even on my horizon yet :-)
first, we need a complete implementation of the missing pieces for powerpc-darwin

current status:

MC layer in llvm-7.1.1 is OK (and works fine to build an assembler for both ppc and ppc64)
I have draft versions of clang ABI and back end lower that are reasonable for 32b powerpc, but still some work needed for ppc64.

It is possible that it might be better to switch to a later llvm/clang (specifically >= 13) since there is now some support for the AIX structure layout rules, powerpc darwin has a similar case (and having a basis would probably be bette than back porting).

but .. there are many things at higher priority at present, so not likely that I will be doing anything soon.

0 replies

barracuda156 · 2022-03-22T22:45:23Z

barracuda156
Mar 22, 2022
Author

On a closer look, these are but few symbols:

___sync_fetch_and_or_8
___kmp_unnamed_critical_addr
___sync_bool_compare_and_swap_8
___kmp_aux_dispatch_init_8u

___sync_bool_compare_and_swap_8 and ___sync_fetch_and_or_8 are in kmp_os.h (notice __ instead of ___):

#elif (KMP_ASM_INTRINS && KMP_OS_UNIX) || !(KMP_ARCH_X86 || KMP_ARCH_X86_64)
# define KMP_TEST_THEN_ADD8(p, v)               __sync_fetch_and_add( (kmp_int8 *)(p), (v) )

/* cast p to correct type so that proper intrinsic will be used */
# define KMP_TEST_THEN_INC32(p)                 __sync_fetch_and_add( (kmp_int32 *)(p), 1 )
# define KMP_TEST_THEN_OR8(p, v)                __sync_fetch_and_or( (kmp_int8 *)(p), (v) )
# define KMP_TEST_THEN_AND8(p, v)               __sync_fetch_and_and( (kmp_int8 *)(p), (v) )
# define KMP_TEST_THEN_INC_ACQ32(p)             __sync_fetch_and_add( (kmp_int32 *)(p), 1 )
# define KMP_TEST_THEN_INC64(p)                 __sync_fetch_and_add( (kmp_int64 *)(p), 1LL )
# define KMP_TEST_THEN_INC_ACQ64(p)             __sync_fetch_and_add( (kmp_int64 *)(p), 1LL )
# define KMP_TEST_THEN_ADD4_32(p)               __sync_fetch_and_add( (kmp_int32 *)(p), 4 )
# define KMP_TEST_THEN_ADD4_ACQ32(p)            __sync_fetch_and_add( (kmp_int32 *)(p), 4 )
# define KMP_TEST_THEN_ADD4_64(p)               __sync_fetch_and_add( (kmp_int64 *)(p), 4LL )
# define KMP_TEST_THEN_ADD4_ACQ64(p)            __sync_fetch_and_add( (kmp_int64 *)(p), 4LL )
# define KMP_TEST_THEN_DEC32(p)                 __sync_fetch_and_sub( (kmp_int32 *)(p), 1 )
# define KMP_TEST_THEN_DEC_ACQ32(p)             __sync_fetch_and_sub( (kmp_int32 *)(p), 1 )
# define KMP_TEST_THEN_DEC64(p)                 __sync_fetch_and_sub( (kmp_int64 *)(p), 1LL )
# define KMP_TEST_THEN_DEC_ACQ64(p)             __sync_fetch_and_sub( (kmp_int64 *)(p), 1LL )
# define KMP_TEST_THEN_ADD32(p, v)              __sync_fetch_and_add( (kmp_int32 *)(p), (v) )
# define KMP_TEST_THEN_ADD64(p, v)              __sync_fetch_and_add( (kmp_int64 *)(p), (v) )

# define KMP_TEST_THEN_OR32(p, v)               __sync_fetch_and_or( (kmp_int32 *)(p), (v) )
# define KMP_TEST_THEN_AND32(p, v)              __sync_fetch_and_and( (kmp_int32 *)(p), (v) )
# define KMP_TEST_THEN_OR64(p, v)               __sync_fetch_and_or( (kmp_int64 *)(p), (v) )
# define KMP_TEST_THEN_AND64(p, v)              __sync_fetch_and_and( (kmp_int64 *)(p), (v) )

# define KMP_COMPARE_AND_STORE_ACQ8(p, cv, sv)  __sync_bool_compare_and_swap( (volatile kmp_uint8 *)(p),(kmp_uint8)(cv),(kmp_uint8)(sv) )
# define KMP_COMPARE_AND_STORE_REL8(p, cv, sv)  __sync_bool_compare_and_swap( (volatile kmp_uint8 *)(p),(kmp_uint8)(cv),(kmp_uint8)(sv) )
# define KMP_COMPARE_AND_STORE_ACQ16(p, cv, sv) __sync_bool_compare_and_swap( (volatile kmp_uint16 *)(p),(kmp_uint16)(cv),(kmp_uint16)(sv) )
# define KMP_COMPARE_AND_STORE_REL16(p, cv, sv) __sync_bool_compare_and_swap( (volatile kmp_uint16 *)(p),(kmp_uint16)(cv),(kmp_uint16)(sv) )
# define KMP_COMPARE_AND_STORE_ACQ32(p, cv, sv) __sync_bool_compare_and_swap( (volatile kmp_uint32 *)(p),(kmp_uint32)(cv),(kmp_uint32)(sv) )
# define KMP_COMPARE_AND_STORE_REL32(p, cv, sv) __sync_bool_compare_and_swap( (volatile kmp_uint32 *)(p),(kmp_uint32)(cv),(kmp_uint32)(sv) )
# define KMP_COMPARE_AND_STORE_ACQ64(p, cv, sv) __sync_bool_compare_and_swap( (volatile kmp_uint64 *)(p),(kmp_uint64)(cv),(kmp_uint64)(sv) )
# define KMP_COMPARE_AND_STORE_REL64(p, cv, sv) __sync_bool_compare_and_swap( (volatile kmp_uint64 *)(p),(kmp_uint64)(cv),(kmp_uint64)(sv) )
# define KMP_COMPARE_AND_STORE_PTR(p, cv, sv)   __sync_bool_compare_and_swap( (volatile void **)(p),(void *)(cv),(void *)(sv) )

# define KMP_COMPARE_AND_STORE_RET8(p, cv, sv)  __sync_val_compare_and_swap( (volatile kmp_uint8 *)(p),(kmp_uint8)(cv),(kmp_uint8)(sv) )
# define KMP_COMPARE_AND_STORE_RET16(p, cv, sv) __sync_val_compare_and_swap( (volatile kmp_uint16 *)(p),(kmp_uint16)(cv),(kmp_uint16)(sv) )
# define KMP_COMPARE_AND_STORE_RET32(p, cv, sv) __sync_val_compare_and_swap( (volatile kmp_uint32 *)(p),(kmp_uint32)(cv),(kmp_uint32)(sv) )
# define KMP_COMPARE_AND_STORE_RET64(p, cv, sv) __sync_val_compare_and_swap( (volatile kmp_uint64 *)(p),(kmp_uint64)(cv),(kmp_uint64)(sv) )

#define KMP_XCHG_FIXED8(p, v)                   __sync_lock_test_and_set( (volatile kmp_uint8 *)(p), (kmp_uint8)(v) )
#define KMP_XCHG_FIXED16(p, v)                  __sync_lock_test_and_set( (volatile kmp_uint16 *)(p), (kmp_uint16)(v) )
#define KMP_XCHG_FIXED32(p, v)                  __sync_lock_test_and_set( (volatile kmp_uint32 *)(p), (kmp_uint32)(v) )
#define KMP_XCHG_FIXED64(p, v)                  __sync_lock_test_and_set( (volatile kmp_uint64 *)(p), (kmp_uint64)(v) )

extern kmp_int8 __kmp_test_then_add8( volatile kmp_int8 *p, kmp_int8 v );
extern kmp_int8 __kmp_test_then_or8( volatile kmp_int8 *p, kmp_int8 v );
extern kmp_int8 __kmp_test_then_and8( volatile kmp_int8 *p, kmp_int8 v );
inline kmp_real32 KMP_XCHG_REAL32( volatile kmp_real32 *p, kmp_real32 v)
{
    kmp_int32 tmp = __sync_lock_test_and_set( (kmp_int32*)p, *(kmp_int32*)&v);
    return *(kmp_real32*)&tmp;
}

inline kmp_real64 KMP_XCHG_REAL64( volatile kmp_real64 *p, kmp_real64 v)
{
    kmp_int64 tmp = __sync_lock_test_and_set( (kmp_int64*)p, *(kmp_int64*)&v);
    return *(kmp_real64*)&tmp;
}

___kmp_unnamed_critical_addr is in z_Linux_asm.s:

# if KMP_ARCH_X86
#  if KMP_OS_DARWIN
        .data
        .comm .gomp_critical_user_,32
        .data
        .globl ___kmp_unnamed_critical_addr
___kmp_unnamed_critical_addr:
        .long .gomp_critical_user_
#  else /* Linux* OS */
        .data
        .comm .gomp_critical_user_,32,8
        .data
	ALIGN 4
        .global __kmp_unnamed_critical_addr
__kmp_unnamed_critical_addr:
        .4byte .gomp_critical_user_
        .type __kmp_unnamed_critical_addr,@object
        .size __kmp_unnamed_critical_addr,4
#  endif /* KMP_OS_DARWIN */
# endif /* KMP_ARCH_X86 */

# if KMP_ARCH_X86_64
#  if KMP_OS_DARWIN
        .data
        .comm .gomp_critical_user_,32
        .data
        .globl ___kmp_unnamed_critical_addr
___kmp_unnamed_critical_addr:
        .quad .gomp_critical_user_
#  else /* Linux* OS */
        .data
        .comm .gomp_critical_user_,32,8
        .data
	ALIGN 8
        .global __kmp_unnamed_critical_addr
__kmp_unnamed_critical_addr:
        .8byte .gomp_critical_user_
        .type __kmp_unnamed_critical_addr,@object
        .size __kmp_unnamed_critical_addr,8
#  endif /* KMP_OS_DARWIN */
# endif /* KMP_ARCH_X86_64 */

#endif /* KMP_GOMP_COMPAT */

___kmp_aux_dispatch_init_8u defined in ktm.h:

#ifdef KMP_GOMP_COMPAT

extern void __kmp_aux_dispatch_init_4( ident_t *loc, kmp_int32 gtid,
    enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st,
    kmp_int32 chunk, int push_ws );
extern void __kmp_aux_dispatch_init_4u( ident_t *loc, kmp_int32 gtid,
    enum sched_type schedule, kmp_uint32 lb, kmp_uint32 ub, kmp_int32 st,
    kmp_int32 chunk, int push_ws );
extern void __kmp_aux_dispatch_init_8( ident_t *loc, kmp_int32 gtid,
    enum sched_type schedule, kmp_int64 lb, kmp_int64 ub, kmp_int64 st,
    kmp_int64 chunk, int push_ws );
extern void __kmp_aux_dispatch_init_8u( ident_t *loc, kmp_int32 gtid,
    enum sched_type schedule, kmp_uint64 lb, kmp_uint64 ub, kmp_int64 st,
    kmp_int64 chunk, int push_ws );
extern void __kmp_aux_dispatch_fini_chunk_4( ident_t *loc, kmp_int32 gtid );
extern void __kmp_aux_dispatch_fini_chunk_8( ident_t *loc, kmp_int32 gtid );
extern void __kmp_aux_dispatch_fini_chunk_4u( ident_t *loc, kmp_int32 gtid );
extern void __kmp_aux_dispatch_fini_chunk_8u( ident_t *loc, kmp_int32 gtid );

0 replies

barracuda156 · 2022-03-22T23:50:44Z

barracuda156
Mar 22, 2022
Author

UPD. In fact the latter is a result of my partial editing away of failing functions. Initially I was getting these errors:

/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp: In function 'kmp_int32 compare_and_swap(volatile T*, T, T)':
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp:257:62: warning: no return statement in function returning non-void [-Wreturn-type]
  257 | compare_and_swap( volatile T *p, T c, T s ) { KMP_ASSERT(0); };
      |                                                              ^
In file included from /opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp.h:97,
                 from /opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp:28:
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp: In instantiation of 'void __kmp_dispatch_init(ident_t*, int, sched_type, T, T, typename traits_t<T>::signed_t, typename traits_t<T>::signed_t, int) [with T = long long int; ident_t = ident; typename traits_t<T>::signed_t = long long int]':
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp:2580:47:   required from here
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_debug.h:43:58: error: narrowing conversion of '-1' from 'int' to 'long unsigned int' [-Wnarrowing]
   43 | #define __KMP_BUILD_ASSERT( expr, suffix )  typedef char __kmp_build_check_##suffix[ (expr) ? 1 : -1 ]
      |                                                          ^~~~~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_debug.h:44:45: note: in expansion of macro '__KMP_BUILD_ASSERT'
   44 | #define _KMP_BUILD_ASSERT( expr, suffix )   __KMP_BUILD_ASSERT( (expr), suffix )
      |                                             ^~~~~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_debug.h:46:49: note: in expansion of macro '_KMP_BUILD_ASSERT'
   46 |     #define KMP_BUILD_ASSERT( expr )            _KMP_BUILD_ASSERT( (expr), __LINE__ )
      |                                                 ^~~~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp:609:5: note: in expansion of macro 'KMP_BUILD_ASSERT'
  609 |     KMP_BUILD_ASSERT( sizeof( dispatch_shared_info_template< UT > ) == sizeof( dispatch_shared_info ) );
      |     ^~~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_debug.h:43:93: error: size '-1' of array is negative
   43 | #define __KMP_BUILD_ASSERT( expr, suffix )  typedef char __kmp_build_check_##suffix[ (expr) ? 1 : -1 ]
      |                                                                                      ~~~~~~~^~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_debug.h:44:45: note: in expansion of macro '__KMP_BUILD_ASSERT'
   44 | #define _KMP_BUILD_ASSERT( expr, suffix )   __KMP_BUILD_ASSERT( (expr), suffix )
      |                                             ^~~~~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_debug.h:46:49: note: in expansion of macro '_KMP_BUILD_ASSERT'
   46 |     #define KMP_BUILD_ASSERT( expr )            _KMP_BUILD_ASSERT( (expr), __LINE__ )
      |                                                 ^~~~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_PPCSnowLeopardPorts_lang_libomp/libomp/work/openmp-381/final/runtime/src/kmp_dispatch.cpp:609:5: note: in expansion of macro 'KMP_BUILD_ASSERT'
  609 |     KMP_BUILD_ASSERT( sizeof( dispatch_shared_info_template< UT > ) == sizeof( dispatch_shared_info ) );
      |     ^~~~~~~~~~~~~~~~
make[2]: *** [src/CMakeFiles/omp.dir/kmp_dispatch.cpp.o] Error 1
make[2]: *** Waiting for unfinished jobs....

I found a document which is not immediately related but deals with atomic operations on Darwin PPC: https://svn.blender.org/svnroot/bf-blender/trunk/lib/darwin-9.x.universal/openimageio/build.patch?p=46445
It seem to follow that gnu implementation may just not work.

0 replies

barracuda156 · 2022-04-07T16:39:44Z

barracuda156
Apr 7, 2022
Author

@iains In fact it looks like it is lacking the code for Big Endian in general. See this: https://github.com/iains/LLVM-7-branch/blob/7.1.1-Darwin-WIP/openmp/runtime/src/kmp.h

/* ToDo: Fix word ordering for big-endian machines. */

0 replies

barracuda156 · 2022-09-12T16:43:56Z

barracuda156
Sep 12, 2022
Author

@iains I will give it another try, afresh.

Gonna commit attempts to implement support for Darwin PPC in libomp here: https://github.com/barracuda156/llvm-project/tree/libomp-14-ppc/openmp

Disclaimer: this is WIP, may contain errors and omissions, and eventually may not lead anywhere. But I will try :)

0 replies

barracuda156 · 2022-09-12T22:38:04Z

barracuda156
Sep 12, 2022
Author

@iains UPD: I think I got somewhat more correct settings now (and did not have to cut anything out), build still fails at linking, this is what needs to be fixed:

___sync_fetch_and_or_8
___sync_fetch_and_and_8
___sync_fetch_and_add_8
___sync_add_and_fetch_8
___sync_bool_compare_and_swap_8

Plus ___kmp_unnamed_critical_addr – this has to be added in z_Linux_asm.s.

P. S. @catap suggested to build compiler-rt libs and use it with GCC: catap/llvm-project@ac23587

14 replies

iains Oct 23, 2022
Maintainer

addition: the other significant chunk of work is amending the searches for SDKs to deal with needing 10.5 to build ppc/ppc64 and then other minimum versions to support other parts of compiler-rt. I think I had to declare some extra vars and then also make sure we do not try to build things like the sanitisers where they are unsupported (at least, for now).

iains Oct 23, 2022
Maintainer

BTW I do not think we really need libomp, compiler-rt or libcxx for GCC since we already have libgomp, libgcc and libstdc++. We can use an installed libc++ (when we have suitable headers installed). compiler-rt is tied to clang.
as for libomp .. I'd expect that the code-gen from clang is different from GCC .. so that it's not obvious it would work..

barracuda156 Oct 23, 2022
Author

addition: the other significant chunk of work is amending the searches for SDKs to deal with needing 10.5 to build ppc/ppc64 and then other minimum versions to support other parts of compiler-rt. I think I had to declare some extra vars and then also make sure we do not try to build things like the sanitisers where they are unsupported (at least, for now).

@iains This part I have sort of done – perhaps not in a generally appropriate way, but functional enough. Since I began doing this on Rosetta (for random reasons), I also had to ensure that ppc is picked when cpu is x86, and that is working now.

barracuda156 Oct 23, 2022
Author

unfortunately compiler-rt is quite a maze with all the different requirements, AFAIR first I had to add ppc and ppc64 and then to declare a clause that has the right C and assembler files in it ... (remember that we are the only 32b PPC platform around at the moment). You need to make sure that the save/rest regs things are included too.

@iains It seems that there is i386 and mips32 support, at least partial, or maybe just that code was not fully removed.

As noted, I am part way through this - so you are likely repeating some of my work ..

Yes, I guess so. Perhaps it’s a better idea to wait for your fixes first, since they gonna be a definite improvement over mine, and it makes little sense indeed to do the same work twice.
If you consider having a WIP branch for this, we – with @catap – could work together on fixing this. @catap already fixed building Clang-11 with GCC, we just need to add PPC parts.

barracuda156 Nov 26, 2022
Author

@iains UPD: I think I got somewhat more correct settings now (and did not have to cut anything out), build still fails at linking, this is what needs to be fixed:
___sync_fetch_and_or_8
___sync_fetch_and_and_8
___sync_fetch_and_add_8
___sync_add_and_fetch_8
___sync_bool_compare_and_swap_8
Plus ___kmp_unnamed_critical_addr – this has to be added in z_Linux_asm.s.

P. S. @catap suggested to build compiler-rt libs and use it with GCC: catap/llvm-project@ac23587

@iains @catap @kencu Are ___sync* 64-bit atomics actually supported on ppc32? I am getting a similar failure with another port now, and using -std=gnu++11 to enable extensions does not help, as well as linking to libatomic. (Well, there are no __sync* symbols there anyway, if we are to believe nm.)

I have seen some mentions of these being broken, like: ofiwg/libfabric#3431
https://bz.apache.org/bugzilla/show_bug.cgi?id=63566
https://bugzilla.redhat.com/show_bug.cgi?id=699149#c1

barracuda156 · 2022-09-13T02:13:24Z

barracuda156
Sep 13, 2022
Author

UPD. About missing symbols, it may be that this has not been defined for Darwin PPC:

# if KMP_OS_DARWIN
#  define KMP_PREFIX_UNDERSCORE(x) _##x  // extra underscore for OS X* symbols
#  define KMP_LABEL(x) L_##x             // form the name of label
# else // !KMP_OS_DARWIN
#  define KMP_PREFIX_UNDERSCORE(x) x  // no extra underscore for Linux* OS symbols
#  define KMP_LABEL(x) .L_##x         // local label hidden from backtraces
# endif

Extra _ should have been added.

What may be a real problem is fixing PPC assembler for Darwin in z_Linux_asm.s. Though it may work without assembler, possibly. MIPS does not have any there.

2 replies

iains Sep 13, 2022
Maintainer

UPD. About missing symbols, it may be that this has not been defined for Darwin PPC:

# if KMP_OS_DARWIN
#  define KMP_PREFIX_UNDERSCORE(x) _##x  // extra underscore for OS X* symbols
#  define KMP_LABEL(x) L_##x             // form the name of label
# else // !KMP_OS_DARWIN
#  define KMP_PREFIX_UNDERSCORE(x) x  // no extra underscore for Linux* OS symbols
#  define KMP_LABEL(x) .L_##x         // local label hidden from backtraces
# endif

Well, I would hope that KMP_OS_DARWIN would be true for *-darwin* not just x86. The compiler's term for this is __USER_LABEL_PREFIX__ which is predefined to _ by both GCC and clang. There are a few platforms that use this.

So - check the configuration and make sure that KMP_OS_DARWIN is defined for all Darwin.

What may be a real problem is fixing PPC assembler for Darwin in z_Linux_asm.s. Though it may work without assembler, possibly. MIPS does not have any there.

Hard to say, not sure what you are trying to achieve - is there no z_Darwin_asm for X86? .. the underlying OS operates the same way for powerpc, x86 and arm64 Darwin, so even though the asm would be different - the required functionality could be similar (although I am guessing here because I don't actually know what the asm needs to do ;) )

barracuda156 Sep 13, 2022
Author

@iains The issue is that the definition adding _ was set within arch code blocks, so it is there for x86* and aarch64, while ppc64 block only has non-Darwin half. KMP_OS_DARWIN itself is arch-agnostic.
No idea why it was done that way, unless someone wanted to break Darwin ppc on purpose :)

I will look into assembler part today.

kencu · 2022-10-27T14:31:58Z

kencu
Oct 27, 2022

most of the symbols missing in your link are found in libatomic.

configure.ldflags-append -latomic

this is a frequently-required library to support atomics. It should probably be added automatically to the link lines by gcc on all builds, but for some reason gcc upstream specifically doesn’t want to do that, so you’re stuck adding it manually when this happens.

2 replies

iains Oct 27, 2022
Maintainer

Is there any reference to discussion with upstream on (not) adding -latomic by default?

FWIW: it would create an unnecessary dependency for exes/dylibs*** that have no need of external atomics support (i.e only using values that can be handled inline, or no atomics at all - which probably covers 99.99% if Intel cases).

It's just that powerpc needs it for 8byte quantities, which do occur often enough to make it an annoyance (but it might still be better to take those case-by-case rather than adding to powerpc-darwin always).

***I guess unless you default to dead stripping dylibs for ld64.

barracuda156 Oct 27, 2022
Author

most of the symbols missing in your link are found in libatomic.

configure.ldflags-append -latomic

this is a frequently-required library to support atomics. It should probably be added automatically to the link lines by gcc on all builds, but for some reason gcc upstream specifically doesn’t want to do that, so you’re stuck adding it manually when this happens.

@kencu That was done, of course. It does not work in this case, for w/e reason.

barracuda156 · 2022-11-26T06:30:06Z

barracuda156
Nov 26, 2022
Author

(I am afraid the update gets lost above in sub-thread, so returning here.)

I have established that at least on PPC there seems to be no way to make 64-bit atomics of a form __sync* work. Nothing helps, neither -latomic, nor turning on extensions (via CMake or directly by -std=gnu* flag).
I have two cases in question here: libomp and rocksdb.

However, simple substitution with new __atomic* versions works fine. Like this, as an example:

template <typename T, typename U>
__attribute__((always_inline)) static inline T toku_sync_fetch_and_sub(T *addr,
                                                                       U diff) {
  paranoid_invariant(!crosses_boundary(addr, sizeof *addr));
#if defined(__APPLE__) && defined(__GNUC__)
  return __atomic_fetch_sub(addr, diff, __ATOMIC_SEQ_CST);
#else
  return __sync_fetch_and_sub(addr, diff);
#endif
}

I do not get why this should be the case, but result is consistent (at least libomp I tried to build on 10.6.8 and likely 10.5.8) for ppc32 and gcc11/gcc12 alike.

@iains Any idea why is that the case?

P. S. I saw some reports of GCC 64-bit atomics not working on ppc32 and mips elsewhere (FreeBSD, Linux), but not sure issues are identical. Here I have specifically __sync* ones not recognized.

2 replies

iains Nov 26, 2022
Maintainer

(I am afraid the update gets lost above in sub-thread, so returning here.)

I have established that at least on PPC there seems to be no way to make 64-bit atomics of a form __sync* work. Nothing helps, neither -latomic, nor turning on extensions (via CMake or directly by -std=gnu* flag). I have two cases in question here: libomp and rocksdb.

(1) There are no 64 bit (or larger) lock-free atomics on PPC 32 - this is a constraint of the ISA, not the software.

However, simple substitution with new __atomic* versions works fine. Like this, as an example:

This is the correct thing to use; the symbols exported by libatomic implement a locked set of load/store/etc. for 64bits and larger.

@iains Any idea why is that the case?

hopefully (1) and (2) explain this.

P. S. I saw some reports of GCC 64-bit atomics not working on ppc32 and mips elsewhere (FreeBSD, Linux), but not sure issues are identical. Here I have specifically __sync* ones not recognized.

Because of (1) above.

barracuda156 Nov 26, 2022
Author

@iains Thank you very much!

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

libomp for ppc32 #30

{{title}}

Replies: 9 comments 20 replies

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

Select a reply

libomp for ppc32 #30

barracuda156 Mar 22, 2022

Replies: 9 comments · 20 replies

iains Mar 22, 2022 Maintainer

barracuda156 Mar 22, 2022 Author

barracuda156 Mar 22, 2022 Author

barracuda156 Apr 7, 2022 Author

barracuda156 Sep 12, 2022 Author

barracuda156 Sep 12, 2022 Author

iains Oct 23, 2022 Maintainer

iains Oct 23, 2022 Maintainer

barracuda156 Oct 23, 2022 Author

barracuda156 Oct 23, 2022 Author

barracuda156 Nov 26, 2022 Author

barracuda156 Sep 13, 2022 Author

iains Sep 13, 2022 Maintainer

barracuda156 Sep 13, 2022 Author

kencu Oct 27, 2022

iains Oct 27, 2022 Maintainer

barracuda156 Oct 27, 2022 Author

barracuda156 Nov 26, 2022 Author

iains Nov 26, 2022 Maintainer

barracuda156 Nov 26, 2022 Author

barracuda156
Mar 22, 2022

Replies: 9 comments 20 replies

iains
Mar 22, 2022
Maintainer

barracuda156
Mar 22, 2022
Author

barracuda156
Mar 22, 2022
Author

barracuda156
Apr 7, 2022
Author

barracuda156
Sep 12, 2022
Author

barracuda156
Sep 12, 2022
Author

iains Oct 23, 2022
Maintainer

iains Oct 23, 2022
Maintainer

barracuda156 Oct 23, 2022
Author

barracuda156 Oct 23, 2022
Author

barracuda156 Nov 26, 2022
Author

barracuda156
Sep 13, 2022
Author

iains Sep 13, 2022
Maintainer

barracuda156 Sep 13, 2022
Author

kencu
Oct 27, 2022

iains Oct 27, 2022
Maintainer

barracuda156 Oct 27, 2022
Author

barracuda156
Nov 26, 2022
Author

iains Nov 26, 2022
Maintainer

barracuda156 Nov 26, 2022
Author