CUDA: Clang separable compilation

For NVCC the compiler takes care of device linking when passed the "-dlink"
flag.
Clang doesn't support such magic and requires the buildsystem to do the work
that NVCC does behind the scenes.

The implementation is based on Bazel's device linking documentation:
https://github.com/tensorflow/tensorflow/blob/7cabcdf073abad8c46e9dda62bb8fa4682d2061e/third_party/nccl/build_defs.bzl.tpl#L259

Closes: #20726

1 files changed, 6 insertions, 1 deletions

diff --git a/Source/cmMakefileTargetGenerator.h b/Source/cmMakefileTargetGenerator.h
index 1740d54..cb804e0 100644
--- a/Source/cmMakefileTargetGenerator.h
+++ b/Source/cmMakefileTargetGenerator.h
@@ -104,6 +104,10 @@ protected:
   void WriteObjectDependRules(cmSourceFile const& source,
                               std::vector<std::string>& depends);
 
+  // CUDA device linking.
+  void WriteDeviceLinkRule(std::vector<std::string>& commands,
+                           const std::string& output);
+
   // write the build rule for a custom command
   void GenerateCustomRuleFile(cmCustomCommandGenerator const& ccg);
 
@@ -127,7 +131,8 @@ protected:
   void DriveCustomCommands(std::vector<std::string>& depends);
 
   // append intertarget dependencies
-  void AppendTargetDepends(std::vector<std::string>& depends);
+  void AppendTargetDepends(std::vector<std::string>& depends,
+                           bool ignoreType = false);
 
   // Append object file dependencies.
   void AppendObjectDepends(std::vector<std::string>& depends);