[CUDA] Generates a call to libc abort in device code, which fails to link

[zwliew]

Reproducer:
https://godbolt.org/z/9qrdGE55d

When trying to call __builtin_abort() in __device__ or __global__ code, clang doesn't throw a compilation error, instead generating the call.uni abort instruction. ptxas then fails at link-time because abort does not exist.

By comparison, nvcc throws the error calling a __host__ function("__builtin_abort") from a __global__ function("f") is not allowed.

Perhaps clang should throw an error earlier too?

[llvmbot]

@llvm/issue-subscribers-libc

Author: Zhao Wei Liew (zwliew)

Reproducer: https://godbolt.org/z/9qrdGE55d

When trying to call __builtin_abort() in device or global code, clang doesn't throw a compilation error, instead generating the call.uni abort instruction. This then fails at link-time because abort does not exist.