https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122281
--- Comment #18 from Benjamin Schulz <schulz.benjamin at googlemail dot com> --- What one sees with clang are indeed problems when initializing the runtime... ==11239== 104 bytes in 1 blocks are possibly lost in loss record 2,701 of 2,872 ==11239== at 0x4A668D8: malloc (vg_replace_malloc.c:447) ==11239== by 0x54F5C6B: ___kmp_allocate_align(unsigned long, unsigned long, char const*, int) (in /usr/lib64/libomp.so) ==11239== by 0x54FA630: ___kmp_allocate (in /usr/lib64/libomp.so) ==11239== by 0x558EBF9: KMPAffinity::pick_api() (in /usr/lib64/libomp.so) ==11239== by 0x554F31F: __kmp_env_initialize(char const*) (in /usr/lib64/libomp.so) ==11239== by 0x552E836: __kmp_do_serial_initialize() (in /usr/lib64/libomp.so) ==11239== by 0x552EDE6: __kmp_serial_initialize (in /usr/lib64/libomp.so) ==11239== by 0x55FBD06: ompt_libomp_connect (in /usr/lib64/libomp.so) ==11239== by 0x56A8CFB: llvm::omp::target::ompt::connectLibrary() (in /usr/lib64/libomptarget.so.21.1) ==11239== by 0x5696B20: initRuntime() (in /usr/lib64/libomptarget.so.21.1) ==11239== by 0x568884F: __tgt_register_lib (in /usr/lib64/libomptarget.so.21.1) ==11239== by 0x400439C: .omp_offloading.descriptor_reg (in /home/benni/projects/arraylibrary/openmp/sparsetests) problem with the clang code appears in connect library, init runtime and so on.. With OpenMP I am not the one who is manually connecting the cuda runtime, so that seems not to be my problem...
