diff --git a/clang/tools/amdgpu-arch/CMakeLists.txt b/clang/tools/amdgpu-arch/CMakeLists.txt
--- a/clang/tools/amdgpu-arch/CMakeLists.txt
+++ b/clang/tools/amdgpu-arch/CMakeLists.txt
@@ -10,9 +10,12 @@
 
 add_clang_tool(amdgpu-arch AMDGPUArch.cpp)
 
+option(CLANG_FORCE_DLOPEN_LIBHSA "Build with 'dlopened' libhsa" OFF)
+
 # If we find the HSA runtime we link with it directly.
 find_package(hsa-runtime64 QUIET 1.2.0 HINTS ${CMAKE_INSTALL_PREFIX} PATHS /opt/rocm)
-if(hsa-runtime64_FOUND AND NOT (LLVM_BUILD_32_BITS OR CMAKE_SIZEOF_VOID_P EQUAL 4))
+if(hsa-runtime64_FOUND AND NOT
+   (LLVM_BUILD_32_BITS OR CMAKE_SIZEOF_VOID_P EQUAL 4 OR CLANG_FORCE_DLOPEN_LIBHSA))
   set_target_properties(amdgpu-arch PROPERTIES INSTALL_RPATH_USE_LINK_PATH ON)
   target_link_libraries(amdgpu-arch PRIVATE hsa-runtime64::hsa-runtime64)
 else()
diff --git a/clang/tools/nvptx-arch/CMakeLists.txt b/clang/tools/nvptx-arch/CMakeLists.txt
--- a/clang/tools/nvptx-arch/CMakeLists.txt
+++ b/clang/tools/nvptx-arch/CMakeLists.txt
@@ -9,10 +9,12 @@
 set(LLVM_LINK_COMPONENTS Support)
 add_clang_tool(nvptx-arch NVPTXArch.cpp)
 
-find_package(CUDAToolkit QUIET)
+option(CLANG_FORCE_DLOPEN_LIBCUDA "Build with 'dlopened' libcuda" OFF)
 
 # If we found the CUDA library directly we just dynamically link against it.
-if(CUDAToolkit_FOUND AND NOT (LLVM_BUILD_32_BITS OR CMAKE_SIZEOF_VOID_P EQUAL 4))
+find_package(CUDAToolkit QUIET)
+if(CUDAToolkit_FOUND AND NOT
+   (LLVM_BUILD_32_BITS OR CMAKE_SIZEOF_VOID_P EQUAL 4 OR CLANG_FORCE_DLOPEN_LIBCUDA))
   target_link_libraries(nvptx-arch PRIVATE CUDA::cuda_driver)
 else()
   target_compile_definitions(nvptx-arch PRIVATE "DYNAMIC_CUDA")