diff --git a/compiler/Makefile b/compiler/Makefile
index d033e7367..2357560c1 100644
--- a/compiler/Makefile
+++ b/compiler/Makefile
@@ -36,7 +36,7 @@ $(BUILD_DIR)/configured.stamp:
 	$(CMAKE_CCACHE_OPTIONS) \
 	$(CC_COMPILER_OPTION) \
 	$(CXX_COMPILER_OPTION) \
-	-DLLVM_ENABLE_PROJECTS=mlir \
+	-DLLVM_ENABLE_PROJECTS="mlir;clang;openmp" \
 	-DLLVM_BUILD_EXAMPLES=OFF \
 	-DLLVM_TARGETS_TO_BUILD="host" \
 	-DCMAKE_BUILD_TYPE=Release \
diff --git a/compiler/lib/Runtime/CMakeLists.txt b/compiler/lib/Runtime/CMakeLists.txt
index f33007b82..1fb9dda44 100644
--- a/compiler/lib/Runtime/CMakeLists.txt
+++ b/compiler/lib/Runtime/CMakeLists.txt
@@ -9,7 +9,7 @@ add_library(ConcretelangRuntime SHARED
 
 if(CONCRETELANG_PARALLEL_EXECUTION_ENABLED)
   add_library(DFRuntime SHARED DFRuntime.cpp)
-  target_link_libraries(DFRuntime PUBLIC pthread m dl HPX::hpx HPX::iostreams_component -rdynamic)
+  target_link_libraries(DFRuntime PUBLIC pthread m dl HPX::hpx HPX::iostreams_component -rdynamic -Wl,-wrap,main)
 
   install(TARGETS DFRuntime EXPORT DFRuntime)
   install(EXPORT DFRuntime DESTINATION "./")
diff --git a/compiler/lib/Runtime/DFRuntime.cpp b/compiler/lib/Runtime/DFRuntime.cpp
index a79abe09f..851077173 100644
--- a/compiler/lib/Runtime/DFRuntime.cpp
+++ b/compiler/lib/Runtime/DFRuntime.cpp
@@ -201,7 +201,9 @@ void _dfr_create_async_task(wfnptr wfn, size_t num_params, size_t num_outputs,
   }
 }
 
+/********************************/
 /* Distributed key management.  */
+/********************************/
 void _dfr_register_key(void *key, size_t key_id, size_t size) {
   node_level_key_manager->register_key(key, key_id, size);
 }
@@ -212,9 +214,13 @@ void *_dfr_get_key(size_t key_id) {
   return *node_level_key_manager->get_key(key_id).key.get();
 }
 
+/************************************/
+/*  Initialization & Finalization.  */
+/************************************/
 /* Runtime initialization and finalization.  */
 static inline void _dfr_stop_impl() {
-  hpx::apply([]() { hpx::finalize(); });
+  if (_dfr_is_root_node())
+    hpx::apply([]() { hpx::finalize(); });
   hpx::stop();
 }
 
@@ -228,35 +234,29 @@ static inline void _dfr_start_impl(int argc, char *argv[]) {
     hpx::start(nullptr, argc, argv);
   }
 
+  // Instantiate on each node
   new PbsKeyManager();
   new WorkFunctionRegistry();
 
-  if (!_dfr_is_root_node()) {
-    _dfr_stop_impl();
+  if (_dfr_is_root_node()) {
+    // Create compute server components on each node - from the root
+    // node only - and the corresponding compute client on the root
+    // node.
+    auto num_nodes = hpx::get_num_localities().get();
+    gcc = hpx::new_<GenericComputeClient[]>(
+              hpx::default_layout(hpx::find_all_localities()), num_nodes)
+              .get();
+  } else {
+    hpx::stop();
     exit(EXIT_SUCCESS);
   }
-
-  // Create compute server components on each node and the
-  // corresponding compute client.
-  auto num_nodes = hpx::get_num_localities().get();
-  gcc = hpx::new_<GenericComputeClient[]>(
-            hpx::default_layout(hpx::find_all_localities()), num_nodes)
-            .get();
 }
 
-// TODO: we need a better way to wrap main. For now loader --wrap and
-// main's constructor/destructor are not functional, but that should
-// replace the current, inefficient calls to _dfr_start/stop generated
-// in each compiled function.
-void _dfr_start() {
-  static int first_p = 0;
-  if (!first_p) {
-    _dfr_start_impl(0, nullptr);
-    first_p = 1;
-  } else {
-    hpx::resume();
-  }
-}
+/*  Start/stop functions to be called from within user code (or during
+    JIT invocation).  These serve to pause/resume the runtime
+    scheduler and to clean up used resources.  */
+void _dfr_start() { hpx::resume(); }
+
 void _dfr_stop() {
   hpx::suspend();
 
@@ -274,7 +274,32 @@ void _dfr_stop() {
   }
 }
 
+/*******************/
+/*  Main wrapper.  */
+/*******************/
+extern "C" {
+extern int main(int argc, char *argv[]); // __attribute__((weak));
+extern int __real_main(int argc, char *argv[]) __attribute__((weak));
+int __wrap_main(int argc, char *argv[]) {
+  int r;
+  // Initialize and immediately suspend the HPX runtime.
+  _dfr_start_impl(0, nullptr);
+  hpx::suspend();
+  // Run the actual main function. Within there should be a call to
+  // _dfr_start to resume execution of the HPX scheduler if needed.
+  r = __real_main(argc, argv);
+  // By default all _dfr_start should be matched to a _dfr_stop, so we
+  // need to resume before being able to finalize.
+  hpx::resume();
+  _dfr_stop_impl();
+
+  return r;
+}
+}
+
+/**********************/
 /*  Debug interface.  */
+/**********************/
 size_t _dfr_debug_get_node_id() { return hpx::get_locality_id(); }
 
 size_t _dfr_debug_get_worker_id() { return hpx::get_worker_thread_num(); }
diff --git a/compiler/src/CMakeLists.txt b/compiler/src/CMakeLists.txt
index 0944d9899..168fdab9f 100644
--- a/compiler/src/CMakeLists.txt
+++ b/compiler/src/CMakeLists.txt
@@ -28,6 +28,7 @@ if(CONCRETELANG_PARALLEL_EXECUTION_ENABLED)
     -Wl,-rpath,${HPX_DIR}/../../
     -Wl,--no-as-needed
     DFRuntime
+    omp
     )
 else()
   target_link_libraries(concretecompiler
diff --git a/compiler/tests/unittest/CMakeLists.txt b/compiler/tests/unittest/CMakeLists.txt
index b592af8f1..716dac76e 100644
--- a/compiler/tests/unittest/CMakeLists.txt
+++ b/compiler/tests/unittest/CMakeLists.txt
@@ -3,14 +3,28 @@ enable_testing()
 include_directories(${PROJECT_SOURCE_DIR}/include)
 
 if(CONCRETELANG_PARALLEL_EXECUTION_ENABLED)
+  set (CMAKE_LIBRARY_OUTPUT_DIRECTORY, ${CMAKE_BINARY_DIR}/lib/)
+  set (CMAKE_SHARED_LINKER_FLAGS)
+  add_custom_command(
+    OUTPUT "${CMAKE_BINARY_DIR}/lib/gtest_main_shared_src/TestMain.cpp.o"
+    COMMAND mkdir -p ${CMAKE_BINARY_DIR}/lib/gtest_main_shared_src
+    COMMAND ${CMAKE_C_COMPILER_AR} -x ${CMAKE_BINARY_DIR}/lib/libgtest_main.a --output ${CMAKE_BINARY_DIR}/lib/gtest_main_shared_src
+    DEPENDS gtest_main
+  )
+  add_library(gtest_main_shared SHARED
+    "${CMAKE_BINARY_DIR}/lib/gtest_main_shared_src/TestMain.cpp.o"
+    )
+  set_target_properties(gtest_main_shared PROPERTIES LINKER_LANGUAGE C)
   add_compile_options(
     -DCONCRETELANG_PARALLEL_TESTING_ENABLED
     )
   link_libraries(
-    -Wl,-rpath,${CMAKE_BINARY_DIR}/lib/Runtime
+    -Wl,-rpath,${CMAKE_BINARY_DIR}/lib/
     -Wl,-rpath,${HPX_DIR}/../../
     -Wl,--no-as-needed
     DFRuntime
+    gtest_main_shared
+    omp
   )
 endif()
 
@@ -111,7 +125,7 @@ if(CONCRETELANG_PARALLEL_EXECUTION_ENABLED)
     end_to_end_jit_dfr
     gtest_main
     ConcretelangSupport
-    -Wl,-rpath,${CMAKE_BINARY_DIR}/lib/Runtime
+    -Wl,-rpath,${CMAKE_BINARY_DIR}/lib/
     -Wl,-rpath,${HPX_DIR}/../../
     -Wl,--no-as-needed
     DFRuntime
@@ -121,7 +135,7 @@ if(CONCRETELANG_PARALLEL_EXECUTION_ENABLED)
     end_to_end_jit_auto_parallelization
     gtest_main
     ConcretelangSupport
-    -Wl,-rpath,${CMAKE_BINARY_DIR}/lib/Runtime
+    -Wl,-rpath,${CMAKE_BINARY_DIR}/lib/
     -Wl,-rpath,${HPX_DIR}/../../
     -Wl,--no-as-needed
     DFRuntime