[CUDA][HIP] add test for uses of std::array on device code

[jmmartinez]

First attempt at testing llvm/llvm-project#136133 .

Draft for the moment, I have to manually test this further.

[yxsamliu]

can std::array be parameter type of kernel or device functions?

can it be used as non-constexpr stack variable in kernel or device functions?

if so, can we have testcases for those?

[yxsamliu]

I am wondering whether we have a systematic way to use CUDA tests for HIP. Using macro to map CUDA entries to HIP is not only tedious but may impose restrictions on how those CUDA tests are written.

I am wondering whether we could hipify the .cu file instead. Currently /opt/rocm/bin/hipify-perl is available on the HIP buildbot. For each .cu file to be hipified, you can use add_custom_command to hipify it to a .hip file then test it.

[kzhuravl]

Can we try using hipify-clang instead? hipify-perl is going to be deprecated soon more than likely

@emankov

[jmmartinez]

I've just built the upstream buildbot image from https://github.com/ROCm/aomp/tree/aomp-dev/upstream-buildbots/Ubu22 and there is no hipify in Rocm's installation (it's quite a minimal install).

I'll check if I install them if the tests work. I agree that it would be much better.

[jmmartinez]

I've tried with hipify-perl and it works.

hipify-clang did not, it yield errors for every test.

One common error case was with multiple main definitions (even if they are on different ifdef branches):

#include <stdio.h>
#if __cplusplus < 201103L
int main() { puts("foo\n"); return 0; }
#else
int main() { puts("bar\n"); return 0; }
#endif

/tmp/foo.cu-5f2589.hip:5:5: error: redefinition of 'main'
    5 | int main() { puts("bar\n"); return 0; }
      |     ^
/tmp/foo.cu-5f2589.hip:3:5: note: previous definition is here
    3 | int main() { puts("foo\n"); return 0; }

I'll adapt these tests to run with hipify and remove the hacky header.