Fmt: Compile error under visual c++ 2017 and CUDA 10.1

[hbsnmyj]

Including the fmt header using nvcc CUDA compiler with msvc 2017 triggers the following compiling errors:

...\fmt\core.h(662): error C2059: syntax error: ''
...\fmt\core.h(662): error C2062: type 'unknown-type' unexpected
...\fmt\core.h(662): error C2143: syntax error: missing ';' before '{'
...\fmt\core.h(662): error C2447: '{': missing function header (old-style formal list?)
...\fmt\core.h(666): error C2059: syntax error: ''
...\fmt\core.h(666): error C2062: type 'unknown-type' unexpected
...\fmt\core.h(666): error C2143: syntax error: missing ';' before '{'
...\fmt\core.h(666): error C2447: '{': missing function header (old-style formal list?)

The parse error can be resolved by rewriting FMT_MAKE_VALUE macro manually using SFINAE, but after that nvcc emits a new bunch of errors. Both gcc+nvcc and clang under Linux work fine.

I think this is similar with # 1080.

MSVC version: 2017 15.9.11
nvcc version: V10.1.105
Compile command used: nvcc -DFMT_HEADER_ONLY=1 -isystem=... -D_WINDOWS .\test-fmt.cu

Error code:

#include <fmt/core.h>

int main() {
    return 0;
}

[vitaut]

I'm not familiar with CUDA but PRs to fix/workaround this are welcome.

[luncliff]

Hello @hbsnmyj, I'm interested in this issue.

I created a new test for importing fmt in CUDA source code. This is the commit: https://github.com/luncliff/fmt/commit/ba82f542d61073d36ab17bc3a86f530d7e59d739

I can't provide my build/test log since Windows CI services don't support CUDA.
But you can try with my fork.

For now it seems like suppressing FMT_DEPRECATED makes the code compilable.

static_assert(__cplusplus >= 201402L, "expect C++ 2014 for nvcc");

#if defined(__CUDACC__)
#   define FMT_DEPRECATED
#endif
#include <fmt/core.h>

#include <cuda.h>
#include <iostream>

using namespace std;

extern auto make_message_cpp() -> std::string;
extern auto make_message_cuda() -> std::string;

int main(int, char*[]){
    cout << make_message_cuda() << endl;
    cout << make_message_cpp() << endl;
}

auto make_message_cuda() -> std::string{
    return fmt::format("nvcc         \t: __cplusplus == {}", __cplusplus);
}

I'd like to try your alternative (which used SFINAE), could you provide the code here if you still have it?

Thanks in advance.

[luncliff]

@vitaut, would it be critical if users suppress FMT_DEPRECATED like the comment above?

We can consider adding check for __CUDACC__ macro in the header file, but I think we had better avoid the solution since NVCC has a gap from major C++ compilers.

I believe most of the CUDA source codes are being maintained in a relatively small pieces.
And they can use some extern functions which implements string rendering instead of including fmt's headers in .cu source codes

[vitaut]

would it be critical if users suppress FMT_DEPRECATED like the comment above?

I think it's a reasonable workaround, so I'll be happy to accept a PR that does this (with a short explanatory comment). I'm about to add a similar workaround for the Intel compiler that doesn't understand [[deprecated]] in type aliases.

[vitaut]

@hbsnmyj, does the newly added cuda-test work on your system?

[vitaut]

Closing for now since the test shows that {fmt} can work with CUDA. If you still get an error feel free to reopen but please provide more repro details, particularly the {fmt} version/commit.