diff --git a/Source/cmMakefileTargetGenerator.cxx b/Source/cmMakefileTargetGenerator.cxx
index 6324b2ef83..98c61fe8be 100644
--- a/Source/cmMakefileTargetGenerator.cxx
+++ b/Source/cmMakefileTargetGenerator.cxx
@@ -1519,6 +1519,20 @@ void cmMakefileTargetGenerator::WriteDeviceLinkRule(
 
   // Link device code for each architecture.
   for (const std::string& architectureKind : architectures) {
+    std::string registerFileCmd;
+
+    // The generated register file contains macros that when expanded
+    // register the device routines. Because the routines are the same for
+    // all architectures the register file will be the same too. Thus
+    // generate it only on the first invocation to reduce overhead.
+    if (fatbinaryDepends.empty()) {
+      std::string const registerFileRel =
+        cmStrCat(relPath, relObjectDir, "cmake_cuda_register.h");
+      registerFileCmd =
+        cmStrCat(" --register-link-binaries=", registerFileRel);
+      cleanFiles.push_back(registerFileRel);
+    }
+
     // Clang always generates real code, so strip the specifier.
     const std::string architecture =
       architectureKind.substr(0, architectureKind.find('-'));
@@ -1528,20 +1542,6 @@ void cmMakefileTargetGenerator::WriteDeviceLinkRule(
     profiles += cmStrCat(" -im=profile=sm_", architecture, ",file=", cubin);
     fatbinaryDepends.emplace_back(cubin);
 
-    std::string registerFileCmd;
-
-    // The generated register file contains macros that when expanded
-    // register the device routines. Because the routines are the same for
-    // all architectures the register file will be the same too. Thus
-    // generate it only on the first invocation to reduce overhead.
-    if (fatbinaryDepends.size() == 1) {
-      std::string const registerFileRel =
-        cmStrCat(relPath, relObjectDir, "cmake_cuda_register.h");
-      registerFileCmd =
-        cmStrCat(" --register-link-binaries=", registerFileRel);
-      cleanFiles.push_back(registerFileRel);
-    }
-
     std::string command = cmStrCat(
       this->Makefile->GetRequiredDefinition("CMAKE_CUDA_DEVICE_LINKER"),
       " -arch=sm_", architecture, registerFileCmd, " -o=$@ ",
diff --git a/Source/cmNinjaNormalTargetGenerator.cxx b/Source/cmNinjaNormalTargetGenerator.cxx
index 5a4c6521d8..493bd4ac85 100644
--- a/Source/cmNinjaNormalTargetGenerator.cxx
+++ b/Source/cmNinjaNormalTargetGenerator.cxx
@@ -753,10 +753,6 @@ void cmNinjaNormalTargetGenerator::WriteDeviceLinkStatements(
     const std::string cubin =
       cmStrCat(ninjaOutputDir, "/sm_", architecture, ".cubin");
 
-    fatbinary.Variables["PROFILES"] +=
-      cmStrCat(" -im=profile=sm_", architecture, ",file=", cubin);
-    fatbinary.ExplicitDeps.emplace_back(cubin);
-
     cmNinjaBuild dlink(this->LanguageLinkerCudaDeviceRule(config));
     dlink.ExplicitDeps = explicitDeps;
     dlink.Outputs = { cubin };
@@ -766,11 +762,15 @@ void cmNinjaNormalTargetGenerator::WriteDeviceLinkStatements(
     // the device routines. Because the routines are the same for all
     // architectures the register file will be the same too. Thus generate it
     // only on the first invocation to reduce overhead.
-    if (fatbinary.ExplicitDeps.size() == 1) {
+    if (fatbinary.ExplicitDeps.empty()) {
       dlink.Variables["REGISTER"] = cmStrCat(
         "--register-link-binaries=", ninjaOutputDir, "/cmake_cuda_register.h");
     }
 
+    fatbinary.Variables["PROFILES"] +=
+      cmStrCat(" -im=profile=sm_", architecture, ",file=", cubin);
+    fatbinary.ExplicitDeps.emplace_back(cubin);
+
     this->GetGlobalGenerator()->WriteBuild(this->GetCommonFileStream(), dlink);
   }