VS: Fix CUDA CodeGeneration field for CMAKE_CUDA_ARCHITECTURES=OFF

If we add no CUDA architecture flags, write an empty `CodeGeneration`
field to tell the CUDA Toolkit's VS integration not to add its default
flags.  We already use this approach for `CMAKE_CUDA_ARCHITECTURES=all`
so that the `-arch=all` flag can be added via `AdditionalOptions`.

Fixes: #23490

1 files changed, 5 insertions, 4 deletions

diff --git a/Source/cmVisualStudioGeneratorOptions.cxx b/Source/cmVisualStudioGeneratorOptions.cxx
index 9045a4d..0b27dfb 100644
--- a/Source/cmVisualStudioGeneratorOptions.cxx
+++ b/Source/cmVisualStudioGeneratorOptions.cxx
@@ -171,14 +171,15 @@ void cmVisualStudioGeneratorOptions::FixCudaCodeGeneration()
     code.clear();
   }
 
-  if (arch.empty() && gencode.empty()) {
-    return;
-  }
-
   // Create a CodeGeneration field with [arch],[code] syntax in each entry.
   // CUDA will convert it to `-gencode=arch=[arch],code="[code],[arch]"`.
   FlagValue& result = this->FlagMap["CodeGeneration"];
 
+  // If there are no flags, leave the CodeGeneration field empty.
+  if (arch.empty() && gencode.empty()) {
+    return;
+  }
+
   // First entries for the -arch=<arch> [-code=<code>,...] pair.
   if (!arch.empty()) {
     std::string arch_name = arch[0];