Merge topic 'add_visbility_flags_to_device_link_line'

0c56bdf91e CUDA: device linking obeys CMAKE_CUDA_VISIBILITY_PRESET setting

Acked-by: Kitware Robot <[email protected]>
Merge-request: !8065

Source/cmLocalGenerator.cxx

@@ -1408,7 +1408,7 @@ void cmLocalGenerator::GetDeviceLinkFlags(
                               linkPath);
   }
 
-  // iterate link deps and see if any of them need IPO
+  this->AddVisibilityPresetFlags(linkFlags, target, "CUDA");
 
   std::vector<std::string> linkOpts;
   target->GetLinkOptions(linkOpts, config, "CUDA");

Tests/CudaOnly/SeparateCompilation/CMakeLists.txt

@@ -15,6 +15,9 @@ get_property(sep_comp TARGET CUDASeparateLibA PROPERTY CUDA_SEPARABLE_COMPILATIO
 if(NOT sep_comp)
   message(FATAL_ERROR "CUDA_SEPARABLE_COMPILATION not initialized")
 endif()
+set_target_properties(CUDASeparateLibA
+                      PROPERTIES
+                      POSITION_INDEPENDENT_CODE ON)
 unset(CMAKE_CUDA_SEPARABLE_COMPILATION)
 
 if(CMAKE_CUDA_SIMULATE_ID STREQUAL "MSVC")
@@ -26,17 +29,24 @@ if(CMAKE_CUDA_SIMULATE_ID STREQUAL "MSVC")
   target_compile_options(CUDASeparateLibA PRIVATE -Xcompiler=-bigobj)
 endif()
 
-#Having file4/file5 in a shared library causes serious problems
-#with the nvcc linker and it will generate bad entries that will
-#cause a segv when trying to run the executable
+#Have file4 and file5 in different shared libraries so that we
+#verify that hidden visibility is passed to the device linker.
+#Otherwise we will get a segv when trying to run the executable
 #
-add_library(CUDASeparateLibB STATIC file4.cu file5.cu)
+add_library(CUDASeparateLibB SHARED file4.cu)
 target_compile_features(CUDASeparateLibB PRIVATE cuda_std_11)
 target_link_libraries(CUDASeparateLibB PRIVATE CUDASeparateLibA)
 
-set_target_properties(CUDASeparateLibA
-                      CUDASeparateLibB
-                      PROPERTIES CUDA_SEPARABLE_COMPILATION ON
-                      POSITION_INDEPENDENT_CODE ON)
+add_library(CUDASeparateLibC SHARED file5.cu)
+target_compile_features(CUDASeparateLibC PRIVATE cuda_std_11)
+target_link_libraries(CUDASeparateLibC PRIVATE CUDASeparateLibA)
+
+set_target_properties(CUDASeparateLibB
+                      CUDASeparateLibC
+                      PROPERTIES
+                      CUDA_SEPARABLE_COMPILATION ON
+                      POSITION_INDEPENDENT_CODE ON
+                      CUDA_VISIBILITY_PRESET hidden
+                      RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/main")
 
 add_subdirectory(main)

Tests/CudaOnly/SeparateCompilation/file1.h

@@ -1,5 +1,14 @@
 
 #pragma once
+
+#ifdef _WIN32
+#  define EXPORT __declspec(dllexport)
+#  define IMPORT __declspec(dllimport)
+#else
+#  define EXPORT __attribute__((__visibility__("default")))
+#  define IMPORT
+#endif
+
 struct result_type
 {
   int input;

Tests/CudaOnly/SeparateCompilation/file4.cu

@@ -15,7 +15,7 @@ static __global__ void file4_kernel(result_type& r, int x)
   result_type_dynamic rd = file2_func(x);
 }
 
-int file4_launch_kernel(int x)
+EXPORT int file4_launch_kernel(int x)
 {
   result_type r;
   file4_kernel<<<1, 1>>>(r, x);

Tests/CudaOnly/SeparateCompilation/file5.cu

@@ -15,7 +15,7 @@ static __global__ void file5_kernel(result_type& r, int x)
   result_type_dynamic rd = file2_func(x);
 }
 
-int file5_launch_kernel(int x)
+EXPORT int file5_launch_kernel(int x)
 {
   result_type r;
   file5_kernel<<<1, 1>>>(r, x);

Tests/CudaOnly/SeparateCompilation/main/CMakeLists.txt

@@ -1,5 +1,5 @@
 add_executable(CudaOnlySeparateCompilation main.cu)
-target_link_libraries(CudaOnlySeparateCompilation PRIVATE CUDASeparateLibB)
+target_link_libraries(CudaOnlySeparateCompilation PRIVATE CUDASeparateLibB CUDASeparateLibC)
 set_target_properties(CudaOnlySeparateCompilation PROPERTIES
   CUDA_STANDARD 11
   CUDA_STANDARD_REQUIRED TRUE

Tests/CudaOnly/SeparateCompilation/main/main.cu

@@ -4,8 +4,8 @@
 #include "../file1.h"
 #include "../file2.h"
 
-int file4_launch_kernel(int x);
-int file5_launch_kernel(int x);
+IMPORT int file4_launch_kernel(int x);
+IMPORT int file5_launch_kernel(int x);
 
 int choose_cuda_device()
 {