With gcc-12's libstdc++, clang-15 rc2 still cannot compile CUDA/HIP code because of __noinline__

[littlewu2508]

Although https://reviews.llvm.org/D124866 aims to fix bugs such as NVIDIA/thrust#1703 or https://bugs.gentoo.org/857126

When I tried clang-15 rc2, the issue still exists.

Reproduce method:

pure C++ test.C

#define __noinline__ __attribute__((noinline))
#include<future>

int main()
{
        return 0;
}

Run clang++ test.C

HIP code example test.hip

#include <hip/hip_runtime.h>
#include<future>

int main()
{
        return 0;
}

Run clang++ -nogpulib -nogpuinc -x hip test.hip

Both gives error: g++-v12/bits/shared_ptr_base.h:196:22: error: use of undeclared identifier 'noinline'; did you mean 'inline'?

[littlewu2508]

I looked at the commit blurb between rc3 and rc2, and I guess there is no fix focusing on this, so I decided to report. If it's fixed in rc3, please ignore me

[littlewu2508]

@yxsamliu CC'ing the author of https://reviews.llvm.org/D124866

[Artem-B]

It appears that something have changed in libstdc++ headers between gcc-11.3 and 12.1:
https://cuda.godbolt.org/z/EvzG5TPzr

[littlewu2508]

It appears that something have changed in libstdc++ headers between gcc-11.3 and 12.1: https://cuda.godbolt.org/z/EvzG5TPzr

Right, that's why many error emerges when gcc12 released: https://bugs.gentoo.org/857126

I thought https://reviews.llvm.org/D124866 is just for it but it does not.

[yxsamliu]

with gcc12.1

/opt/compiler-explorer/gcc-12.1.0/lib/gcc/x86_64-linux-gnu/12.1.0/../../../../include/c++/12.1.0/bits/shared_ptr_base.h:196:22: error: use of undeclared identifier 'noinline'; did you mean 'inline'?
      __attribute__((__noinline__))
                     ^
/opt/compiler-explorer/cuda/11.7.0//include/crt/host_defines.h:83:24: note: expanded from macro '__noinline__'
        __attribute__((noinline))

CUDA header defines __noinline__ as __attribute__((noinline)). gcc12 uses __attribute__((__noinline__)). This ends up as __attribute__((__attribute__((noinline)))). clang emits error for that. One workaround is to intercept host_defines.h and undef __noinline__ in clang default include header for CUDA. HIP is not subject to this issue.

[littlewu2508]

CUDA header defines __noinline__ as __attribute__((noinline)). gcc12 uses __attribute__((__noinline__)). This ends up as __attribute__((__attribute__((noinline)))). clang emits error for that. One workaround is to intercept host_defines.h and undef __noinline__ in clang default include header for CUDA.

Yes, that's what I'm currently doing, but it's a bit ugly and requires a big patch for large projects, which means maintaining difficulties.

HIP is not subject to this issue.

HIP also suffers: https://github.com/ROCm-Developer-Tools/hipamd/blob/bf49b11d07064ec56ea130c3a73beb216d81f582/include/hip/amd_detail/host_defines.h#L51

And that's why bugs like https://bugs.gentoo.org/857126 is present.

[yxsamliu]

CUDA header defines __noinline__ as __attribute__((noinline)). gcc12 uses __attribute__((__noinline__)). This ends up as __attribute__((__attribute__((noinline)))). clang emits error for that. One workaround is to intercept host_defines.h and undef __noinline__ in clang default include header for CUDA.

Yes, that's what I'm currently doing, but it's a bit ugly and requires a big patch for large projects, which means maintaining difficulties.

HIP is not subject to this issue.

HIP also suffers: https://github.com/ROCm-Developer-Tools/hipamd/blob/bf49b11d07064ec56ea130c3a73beb216d81f582/include/hip/amd_detail/host_defines.h#L51

And that's why bugs like https://bugs.gentoo.org/857126 is present.

There is a condition for that macro !__has_feature(cuda_noinline_keyword). My patch https://reviews.llvm.org/D124866 enable that feature, therefore HIP no longer defines that macro. In that sense, I said 'HIP is not subject to this issue'.

[littlewu2508]

@yxsamliu I carefully read your patch, and figured out I misunderstood it previously. With your patch, CUDA/HIP can compile without define #define __noinline__ __attribute__((noinline)) in their api headers (I also made a MWE which confirms that). So with your patch, and removing #define __noinline__ __attribute__((noinline)) as well, problem with gcc-12 solved.

Thanks for your extra explanation!