https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126015
--- Comment #1 from Benjamin Schulz <schulz.benjamin at googlemail dot com> --- A harsher comment of this by googles ai was the following: Standard Compliance Failure: The OpenMP specification allows array sections (reduction(+:arr[0:2])) to be reduced natively. The compiler's job is to map this to the target hardware. Broken NVPTX Code Lowering Pass: GCC is lazily falling back on an x86-style __atomic_compare_exchange_16 function call instead of decomposing the 16-byte object into two individual 8-byte GPU math registers. Because the GPU architecture cannot resolve that symbol, GCC is essentially generating invalid device assembly that can never be linked on an NVIDIA system
