Hi, I’m in the process of migrating a C++ project over to CUDA and running into a few compiler issues along the way. One of them is related to using C++17 for CTAD with std::tuple in host code. A small reproducer is given below:
#include <tuple>
auto foo(int a, int b) { return std::tuple{a, b}; }
int main() { foo(3, 2); }
When compiling with the newest nvcc:
$ uname -r
5.8.0-50-generic (ubuntu 20.04)
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Sun_Mar_21_19:15:46_PDT_2021
Cuda compilation tools, release 11.3, V11.3.58
Build cuda_11.3.r11.3/compiler.29745058_0
$ nvcc -std=c++17 example1.cu
/usr/include/c++/9/tuple(580): internal error: assertion failed at: "il_to_str.c", line 1994 in form_type_specifier
1 catastrophic error detected in the compilation of "example1.cu".
Compilation aborted.
Aborted (core dumped)
An older version of nvcc (11.1) doesn’t seem to have this issue:
$ path/to/old/cuda/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Mon_Oct_12_20:09:46_PDT_2020
Cuda compilation tools, release 11.1, V11.1.105
Build cuda_11.1.TC455_06.29190527_0
$ path/to/old/cuda/bin/nvcc -std=c++17 example1.cu
$
Furthermore, manually specifying the template arguments for std::tuple seems to address the issue as well:
#include <tuple>
auto foo(int a, int b) { return std::tuple<int,int>{a, b}; }
int main() { foo(3, 2); }
$ nvcc example2.cu -std=c++17
$ path/to/old/cuda/bin/nvcc example2.cu -std=c++17
$ (nvcc v11.1 and v11.3 both work!)
The documentation says to expect nvcc >= v11.0 to support C++17 language features (with a few exceptions seemingly unrelated to this issue).
Is this a known limitation, or a regression in the 11.3 version of nvcc?