[CUDA] Temporarily undefine __noinline__ when including bits/shared_ptr_base.h

[tra]

Summary

This avoid CUDA compilation errors caused by CUDA headers defining noinline
and conflicting with noinline use in libstdc++.

[tra]

Removed warning.

[tra]

Fun tidbit: https://github.com/NVIDIA/thrust/issues/1703#issuecomment-1526604000

Indeed, I believe the nvcc frontend has special handling for that attribute expansion. clang would need to emulate that "special" handling

Right. The attribute((attribute((noinline)))) void foo(); gets magically transformed into __attribute((noinline)) void foo() by the time it makes it to the final host compilation. 😭

And the magic seems to work only for attribute((attribute((noinline)))). Any other variants I tried error out.

So, it's been a known issue in the CUDA headers, deliberately worked around in NVCC. And now the bug lives on and keeps giving...

[jlebar]

wow.

[yaxunl]

LGTM.

This issue does not affect HIP as HIP headers have removed __noinline__ as a macro since https://reviews.llvm.org/D124866 and __attribute__((__noinline__)) is equivalent to __attribute__((noinline)) for HIP.

clang/lib/Headers/CMakeLists.txt

set(files
   ${utility_files}
   )
 
 set(cuda_wrapper_files
   cuda_wrappers/algorithm
   cuda_wrappers/cmath
   cuda_wrappers/complex
   cuda_wrappers/new
+  cuda_wrappers/bits/shared_ptr_base.h
 )
 
 set(ppc_wrapper_files
   ppc_wrappers/mmintrin.h
   ppc_wrappers/xmmintrin.h
   ppc_wrappers/mm_malloc.h
   ppc_wrappers/emmintrin.h
   ppc_wrappers/pmmintrin.h

clang/lib/Headers/cuda_wrappers/bits/shared_ptr_base.h

+// CUDA headers define __noinline__ which interferes with libstdc++'s use of
+// `__attribute((__noinline__))`. In order to avoid compilation error,
+// temporarily unset __noinline__ when we include affected libstdc++ header.
+
+#pragma push_macro("__noinline__")
+#undef __noinline__
+#include_next "bits/shared_ptr_base.h"
+
+#pragma pop_macro("__noinline__")