clang 16/17/18 fail to compiler ppc64le with -mfloat128

[munroesj52]

Clang fails the pveclib configure test for PVECLIB_FLOAT128_CFLAGS with -mfloat128. and no -mcpu. This should default to power8 and the libgcc runtime.

Clang uses /usr/include/powerpc64le-linux-gnu/bits/floatn.h from the GCC installation which defines:

/* The following builtins (suffixed with 'q') are available in GCC >= 6.2,
   which is the minimum version required for float128 support on powerpc64le.
   Since GCC 7.0 the builtins suffixed with f128 are also available, then
   there is no need to redefined them.  */
#  if !__GNUC_PREREQ (7, 0)
#   define __builtin_copysignf128 __builtin_copysignq
#   define __builtin_fabsf128 __builtin_fabsq
#   define __builtin_inff128 __builtin_infq
#   define __builtin_nanf128 __builtin_nanq
#   define __builtin_nansf128 __builtin_nansq
#  endif

But the configure test fails with:

conftest.c:78:13: error: use of unknown builtin '__builtin_copysignq' [-Wimplicit-function-declaration]
   78 |     return (__builtin_copysignf128(valx, valy));
      |             ^
/usr/include/powerpc64le-linux-gnu/bits/floatn.h:101:35: note: expanded from macro '__builtin_copysignf128'
  101 | #   define __builtin_copysignf128 __builtin_copysignq
      |                                   ^
1 error generated.

GCC still defines both __builtin_copysignf128 and __builtin_copysignq

But is appears that clang only defines __builtin_copysignf128 which leaves GCCs floatn.h in a confused state.

PVECLIB is @ https://github.com/open-power-sdk/pveclib
The pveclib configue test source is ./testprograms/test_ppc_F128.c

[EugeneZelenko]

Could you please try 20 or main branch? https://godbolt.org should be helpful.

[munroesj52]

No can do. Ubuntu 24.04 LTS will only install up to clang-19.

But you can download https://github.com/open-power-sdk/pveclib/blob/master/testprograms/test_ppc_F128math.c and compile it yourself.