[CUDA] Relax restrictions on GPU-side variadic functions

[tra]

Summary

Allow parsing GPU-side variadic functions when we're compiling with CUDA-9 or
newer. We still do not allow accessing variadic arguments.

CUDA-9 was the version which introduced PTX-6.0 which allows implementing
variadic functions, so older versions can't have variadics in principle.

This is required for dealing with headers in recent CUDA versions that rely on
variadic function declarations in some of the templated code in libcu++.
E.g. https://github.com/llvm/llvm-project/issues/58410

[jlebar]

This seems a little dangerous -- we're saying the frontend will accept this but we can't generate code for it? What happens if we try to generate code? Do we get some sort of error, or do we silently fail?

[yaxunl]

could make the diagnostic about variadic functions a deferred diagnostic, then the diag will only be emitted if the function is really emitted.

[tra]

In D150718#4348737, @jlebar wrote:

This seems a little dangerous -- we're saying the frontend will accept this but we can't generate code for it? What happens if we try to generate code? Do we get some sort of error, or do we silently fail?

This will allow generation of variadic functions, as long as they do not use va_arg (in other words, as long as they do not access the variadic arguments).
It's safe to do now, because NVPTX backend already has variadics support

In case va_arg is encountered, clang will produce correct diagnostics and it is already covered by test/SemaCUDA/vararg.cu

In D150718#4349747, @yaxunl wrote:

could make the diagnostic about variadic functions a deferred diagnostic, then the diag will only be emitted if the function is really emitted.

The short-term need is only for being able to parse variadic function declarations. Being able to emit some of them is nice, but not useful in practice. No point blocking the ones we can emit, but there's not much benefit in extending that set with deferred diags either.

I do plan to enable full support for variadics soon, anyways.

[tra]

test.

[jlebar]

OK, sgtm.

clang/lib/Driver/ToolChains/Cuda.cpp

void CudaToolChain::addClangTargetOptions(
 
   if (DeviceOffloadingKind == Action::OFK_Cuda) {
     CC1Args.append(
         {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
 
     if (DriverArgs.hasFlag(options::OPT_fcuda_approx_transcendentals,
                            options::OPT_fno_cuda_approx_transcendentals, false))
       CC1Args.push_back("-fcuda-approx-transcendentals");
+
+    // Unsized function arguments used for variadics were introduced in CUDA-9.0
+    // We still do not support generating code that actually uses variadic
+    // arguments yet, but we do need to allow parsing them as recent CUDA
+    // headers rely on that. https://github.com/llvm/llvm-project/issues/58410
+    if (CudaInstallation.version() >= CudaVersion::CUDA_90)
+      CC1Args.push_back("-fcuda-allow-variadic-functions");
   }
 
   if (DriverArgs.hasArg(options::OPT_nogpulib))
     return;
 
   if (DeviceOffloadingKind == Action::OFK_OpenMP &&
       DriverArgs.hasArg(options::OPT_S))
     return;