Simple file fails to compile with -O2 or -O3

[Michael Poole]

I have a new laptop with Ubuntu 16.06 (xenial) installed onto a new partition, plus nvidia-cuda-toolkit 7.5.18-0ubuntu1 from multiverse.

$ echo '#include <stdio.h>' > dummy.cpp
$ g++ -c -O2 dummy.cpp
$ cp dummy.cpp dummy.cu
$ nvcc -c -O2 dummy.cu
/usr/include/string.h: In function ‘void* __mempcpy_inline(void*, const void*, size_t)’:
/usr/include/string.h:652:42: error: ‘memcpy’ was not declared in this scope
   return (char *) memcpy (__dest, __src, __n) + __n;
                                          ^

nvcc can compile it at -O0 and -O1, and it is a conforming translation unit (modulo the addition of at least one definition after the #include statement; including such a definition does not affect the error).

[Graham Inggs (ginggs)]

Please try the following workaround:

$ nvcc -c -O2 -D_FORCE_INLINES dummy.cu

[Graham Inggs (ginggs)]

Please also try editing /usr/include/common_functions.h and commenting out the line containing (as per the attached patch:

__cdecl memcpy(void*, const void*, size_t) __THROW;

[Michael Poole (mdpoole)]

The workarounds (either one separately, or both) resolve the compile issue for me. Thank you for the pointers!

[Graham Inggs (ginggs)]

@mdpoole: thanks for confirming!

This seems to be fixed in the CUDA 8.0 release candidate.
I can't see that Nvidia changed anything related to this in common_functions.h, so I suggest using -D_FORCE_INLINES for now.