diff --git a/.github/workflows/generate_cmake_matrix.py b/.github/workflows/generate_cmake_matrix.py
index 0fcbfd489..4a07b6a6d 100644
--- a/.github/workflows/generate_cmake_matrix.py
+++ b/.github/workflows/generate_cmake_matrix.py
@@ -11,10 +11,14 @@
         "toolchain": ["llvm", "gcc"],
         "arch_flags": ["-march=native", "-march=x86-64", "native"]
     },
-    "windows-2022": {
-        "toolchain": ["msvc", "llvm"],
+    "windows-2022-msvc": {
+        "toolchain": ["msvc"],
         "arch_flags": ["/arch:AVX2", "/arch:SSE2", "native"]
     },
+    "windows-2022-llvm": {
+        "toolchain": ["msvc"],
+        "arch_flags": ["-march=native", "-march=x86-64", "native"]
+    },
     "macos-13": {
         "toolchain": ["llvm", "gcc-14"],
         "arch_flags": ["-march=native", "-march=x86-64", "native"]
diff --git a/CMakeLists.txt b/CMakeLists.txt
index ee9e3ef27..97c83806d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -29,7 +29,7 @@ set(FINUFFT_FFTW_SUFFIX "OpenMP" CACHE STRING "Suffix for FFTW libraries (e.g. O
 # if FINUFFT_USE_CUDA is OFF, the following options are ignored
 set(FINUFFT_CUDA_ARCHITECTURES "native" CACHE STRING "CUDA architectures to build for (e.g. 60;70;75;)")
 # if FINUFFT_USE_CPU is OFF, the following options are ignored
-set(FINUFFT_ARCH_FLAGS native CACHE STRING "Compiler flags for specifying target architecture, defaults to -march=native")
+set(FINUFFT_ARCH_FLAGS "native" CACHE STRING "Compiler flags for specifying target architecture, defaults to -march=native")
 # sphinx tag (don't remove): @cmake_opts_end
 cmake_dependent_option(FINUFFT_ENABLE_INSTALL "Disable installation in the case of python builds" OFF "FINUFFT_BUILD_PYTHON" OFF)
 cmake_dependent_option(FINUFFT_STATIC_LINKING "Disable static libraries in the case of python builds" ON "NOT FINUFFT_BUILD_PYTHON" OFF)
@@ -77,10 +77,10 @@ message(
   STATUS "FINUFFT RelWithDebInfo flags: ${FINUFFT_CXX_FLAGS_RELWITHDEBINFO}")
 
 if(FINUFFT_ARCH_FLAGS STREQUAL "native")
-  set(FINUFFT_ARCH_FLAGS -march=native)
+  set(FINUFFT_ARCH_FLAGS CACHE STRING -march=native FORCE)
   filter_supported_compiler_flags(FINUFFT_ARCH_FLAGS FINUFFT_ARCH_FLAGS)
   if(NOT FINUFFT_ARCH_FLAGS)
-    set(FINUFFT_ARCH_FLAGS -mtune=native)
+    set(FINUFFT_ARCH_FLAGS CACHE STRING -mtune=native FORCE)
     filter_supported_compiler_flags(FINUFFT_ARCH_FLAGS FINUFFT_ARCH_FLAGS)
   endif()
   if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
@@ -172,15 +172,7 @@ function(finufft_link_test target)
   set_target_properties(
     ${target} PROPERTIES MSVC_RUNTIME_LIBRARY
                          "MultiThreaded$<$<CONFIG:Debug>:Debug>")
-  if(WIN32)
-    # copy the .dll file to the same folder as the executable on windows make we
-    # should provide a helper cmake function for the users to use
-    add_custom_command(
-      TARGET ${target}
-      POST_BUILD
-      COMMAND ${CMAKE_COMMAND} -E copy_directory $<TARGET_FILE_DIR:finufft>
-              $<TARGET_FILE_DIR:${target}>)
-  endif()
+  copy_dll(finufft ${target})
 endfunction()
 
 # Utility function to set finufft compilation options.
diff --git a/cmake/utils.cmake b/cmake/utils.cmake
index 909cbfaf7..4dbf1e03b 100644
--- a/cmake/utils.cmake
+++ b/cmake/utils.cmake
@@ -40,20 +40,47 @@ function(check_arch_support)
   if(RUN_OUTPUT MATCHES "AVX512")
     set(FINUFFT_ARCH_FLAGS
         "/arch:AVX512"
-        CACHE STRING "Compiler flags for specifying target architecture.")
+        CACHE STRING FORCE)
   elseif(RUN_OUTPUT MATCHES "AVX")
     set(FINUFFT_ARCH_FLAGS
         "/arch:AVX"
-        CACHE STRING "Compiler flags for specifying target architecture.")
+        CACHE STRING FORCE)
   elseif(RUN_OUTPUT MATCHES "SSE")
     set(FINUFFT_ARCH_FLAGS
         "/arch:SSE"
-        CACHE STRING "Compiler flags for specifying target architecture.")
+        CACHE STRING FORCE)
   else()
     set(FINUFFT_ARCH_FLAGS
         ""
-        CACHE STRING "Compiler flags for specifying target architecture.")
+        CACHE STRING FORCE)
   endif()
   message(STATUS "CPU supports: ${RUN_OUTPUT}")
   message(STATUS "Using MSVC flags: ${FINUFFT_ARCH_FLAGS}")
 endfunction()
+
+function(copy_dll source_target destination_target)
+  if(NOT WIN32)
+    return()
+  endif()
+  # Get the binary directory of the destination target
+  get_target_property(DESTINATION_DIR ${destination_target} BINARY_DIR)
+  set(DESTINATION_FILE ${DESTINATION_DIR}/$<TARGET_FILE_NAME:${source_target}>)
+  if(NOT EXISTS ${DESTINATION_FILE})
+    message(
+      STATUS
+        "Copying ${source_target} to ${DESTINATION_DIR} directory for ${destination_target}"
+    )
+    # Define the custom command to copy the source target to the destination
+    # directory
+    add_custom_command(
+      TARGET ${destination_target}
+      POST_BUILD
+      COMMAND ${CMAKE_COMMAND} -E copy $<TARGET_FILE:${source_target}>
+              ${DESTINATION_FILE}
+      COMMENT "Copying ${source_target} to ${destination_target} directory")
+  endif()
+  # Unset the variables to leave a clean state
+  unset(DESTINATION_DIR)
+  unset(SOURCE_FILE)
+  unset(DESTINATION_FILE)
+endfunction()