CUDA: Add Device LTO support for nvcc

Fixes #22200

Modules/CheckIPOSupported.cmake

@@ -76,6 +76,23 @@ endmacro()
 
 # Run IPO/LTO test
 macro(_ipo_run_language_check language)
+  set(_C_ext "c")
+  set(_CXX_ext "cpp")
+  set(_Fortran_ext "f")
+  string(COMPARE EQUAL "${language}" "CUDA" is_cuda)
+
+  set(ext ${_${language}_ext})
+  if(NOT "${ext}" STREQUAL "")
+    set(copy_sources foo.${ext} main.${ext})
+  elseif(is_cuda)
+    if(_CMAKE_CUDA_IPO_SUPPORTED_BY_CMAKE)
+      set("${X_RESULT}" YES PARENT_SCOPE)
+    endif()
+    return()
+  else()
+    message(FATAL_ERROR "Language not supported")
+  endif()
+
   set(testdir "${CMAKE_CURRENT_BINARY_DIR}/CMakeFiles/_CMakeLTOTest-${language}")
 
   file(REMOVE_RECURSE "${testdir}")
@@ -91,17 +108,6 @@ macro(_ipo_run_language_check language)
 
   set(try_compile_src "${CMAKE_ROOT}/Modules/CheckIPOSupported")
 
-  set(_C_ext "c")
-  set(_CXX_ext "cpp")
-  set(_Fortran_ext "f")
-
-  set(ext ${_${language}_ext})
-  if(NOT "${ext}" STREQUAL "")
-    set(copy_sources foo.${ext} main.${ext})
-  else()
-    message(FATAL_ERROR "Language not supported")
-  endif()
-
   # Use:
   # * TRY_COMPILE_PROJECT_NAME
   # * CMAKE_VERSION
@@ -211,6 +217,11 @@ function(check_ipo_supported)
       list(APPEND languages "C")
     endif()
 
+    list(FIND enabled_languages "CUDA" result)
+    if(NOT result EQUAL -1)
+      list(APPEND languages "CUDA")
+    endif()
+
     list(FIND enabled_languages "Fortran" result)
     if(NOT result EQUAL -1)
       list(APPEND languages "Fortran")
@@ -219,7 +230,7 @@ function(check_ipo_supported)
     string(COMPARE EQUAL "${languages}" "" no_languages)
     if(no_languages)
       _ipo_not_supported(
-          "no C/CXX/Fortran languages found in ENABLED_LANGUAGES global property"
+          "no C/CXX/CUDA/Fortran languages found in ENABLED_LANGUAGES global property"
       )
       return()
     endif()
@@ -227,7 +238,7 @@ function(check_ipo_supported)
     set(languages "${X_LANGUAGES}")
 
     set(unsupported_languages "${languages}")
-    list(REMOVE_ITEM unsupported_languages "C" "CXX" "Fortran")
+    list(REMOVE_ITEM unsupported_languages "C" "CXX" "CUDA" "Fortran")
     string(COMPARE NOTEQUAL "${unsupported_languages}" "" has_unsupported)
     if(has_unsupported)
       _ipo_not_supported(

Modules/Compiler/Clang-CUDA.cmake

@@ -35,6 +35,10 @@ set(CMAKE_CUDA_RUNTIME_LIBRARY_LINK_OPTIONS_STATIC "cudadevrt;cudart_static")
 set(CMAKE_CUDA_RUNTIME_LIBRARY_LINK_OPTIONS_SHARED "cudadevrt;cudart")
 set(CMAKE_CUDA_RUNTIME_LIBRARY_LINK_OPTIONS_NONE   "")
 
+# Clang doesn't support CUDA device LTO
+set(_CMAKE_CUDA_IPO_SUPPORTED_BY_CMAKE NO)
+set(_CMAKE_CUDA_IPO_MAY_BE_SUPPORTED_BY_COMPILER NO)
+
 if(UNIX)
   list(APPEND CMAKE_CUDA_RUNTIME_LIBRARY_LINK_OPTIONS_STATIC "rt" "pthread" "dl")
 endif()

Modules/Compiler/NVIDIA-CUDA.cmake

@@ -48,6 +48,13 @@ if((NOT DEFINED CMAKE_DEPENDS_USE_COMPILER OR CMAKE_DEPENDS_USE_COMPILER)
   set(CMAKE_CUDA_DEPENDS_USE_COMPILER TRUE)
 endif()
 
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11.2)
+  set(_CMAKE_CUDA_IPO_SUPPORTED_BY_CMAKE YES)
+  set(_CMAKE_CUDA_IPO_MAY_BE_SUPPORTED_BY_COMPILER YES)
+
+  set(CMAKE_CUDA_DEVICE_LINK_OPTIONS_IPO " -dlto")
+endif()
+
 if(NOT "x${CMAKE_CUDA_SIMULATE_ID}" STREQUAL "xMSVC")
   set(CMAKE_CUDA_COMPILE_OPTIONS_PIE -Xcompiler=-fPIE)
   set(CMAKE_CUDA_COMPILE_OPTIONS_PIC -Xcompiler=-fPIC)
@@ -61,6 +68,7 @@ if(NOT "x${CMAKE_CUDA_SIMULATE_ID}" STREQUAL "xMSVC")
   string(APPEND CMAKE_CUDA_FLAGS_MINSIZEREL_INIT " -O1 -DNDEBUG")
   string(APPEND CMAKE_CUDA_FLAGS_RELWITHDEBINFO_INIT " -O2 -g -DNDEBUG")
 endif()
+
 set(CMAKE_SHARED_LIBRARY_CREATE_CUDA_FLAGS -shared)
 set(CMAKE_INCLUDE_SYSTEM_FLAG_CUDA -isystem=)