From d146cbf3bcf5cf057f7049aff0182df7a6484130 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Alexander=20H=C3=BCck?= <hueck.alexander@gmail.com>
Date: Sun, 20 Apr 2025 18:01:25 +0200
Subject: [PATCH 1/3] MPI requirement optional

---
 CMakeLists.txt                                |  10 +-
 cmake/cusanToolchain.cmake                    |   2 +-
 lib/runtime/CMakeLists.txt                    | 126 +++++++++---------
 lib/runtime/TSanInterface.h                   |   6 +-
 scripts/CMakeLists.txt                        |   1 -
 test/CMakeLists.txt                           |  23 +++-
 test/kernel_analysis/03_struct_write.c        |   2 +
 test/kernel_analysis/08_big_struct_write.c    |   2 +
 test/lit.cfg                                  |   2 +
 test/lit.site.cfg.in                          |   3 +-
 test/pass/03_cuda_to_mpi.c                    |   4 +-
 test/pass/04_mpi_to_cuda.c                    |   2 +
 test/pass/05_cuda_to_mpi_stream.c             |   2 +
 test/pass/06_cuda_to_mpi_event.c              |   2 +
 test/pass/07_cuda_to_mpi_read.c               |   2 +
 test/pass/08_cudamemcpy_to_mpi.c              |   2 +
 test/pass/11_cuda_to_mpi_struct_of_buff.c     |   2 +
 .../18_cuda_to_mpi_event_query_busy_loop.c    |   1 +
 ...pi_send_cudaMemcpyAsyncH2H_implicit_sync.c |   1 +
 test/pass/20_cuda_to_mpi_send_ds_sync_w_r.c   |   2 +
 test/pass/21_chunked_streams_example.c        |   2 +
 test/pass/22_cuda_to_mpi_partial_buff_write.c |   2 +
 test/pass/26_malloc_pitch.c                   |   2 +
 test/pass/29_tsan_cuda_to_mpi.c               |   1 +
 test/pass/30_tsan_annotate_cuda_to_mpi.c      |   2 +
 test/pass/31_tsan_cuda_event.c                |   3 +-
 test/pass/32_tsan_async_copy.c                |   2 +-
 test/runtime/03_cuda_to_mpi.c                 |   2 +-
 test/runtime/04_mpi_to_cuda.c                 |   2 +-
 test/runtime/05_cuda_to_mpi_stream.c          |   2 +-
 test/runtime/06_cuda_to_mpi_event.c           |   2 +-
 test/runtime/07_cuda_to_mpi_read.c            |   2 +-
 test/runtime/08_cudamemcpy_to_mpi.c           |   2 +-
 test/runtime/11_cuda_to_mpi_struct_of_buff.c  |   2 +-
 test/runtime/12_struct_ptr.c                  |   2 +-
 .../18_cuda_to_mpi_event_query_busy_loop.c    |   2 +-
 ...pi_send_cudaMemcpyAsyncH2H_implicit_sync.c |   2 +-
 .../runtime/20_cuda_to_mpi_send_ds_sync_w_r.c |   2 +-
 test/runtime/21_chunked_streams_example.c     |   2 +-
 .../22_cuda_to_mpi_partial_buff_write.c       |   2 +-
 test/runtime/26_malloc_pitch.c                |   2 +-
 test/runtime/29_tsan_cuda_to_mpi.c            |   2 +-
 test/runtime/30_tsan_annotate_cuda_to_mpi.c   |   2 +-
 43 files changed, 148 insertions(+), 95 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 58f070b..69051f5 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -30,11 +30,6 @@ add_subdirectory(externals)
 add_subdirectory(lib)
 add_subdirectory(scripts)
 
-if(PROJECT_IS_TOP_LEVEL)
-  enable_testing()
-  add_subdirectory(test)
-endif()
-
 write_basic_package_version_file(
   ${CMAKE_CURRENT_BINARY_DIR}/cusanConfigVersion.cmake
   VERSION ${PROJECT_VERSION}
@@ -68,3 +63,8 @@ if(PROJECT_IS_TOP_LEVEL)
       QUIET_ON_EMPTY
     )
 endif()
+
+if(PROJECT_IS_TOP_LEVEL)
+  enable_testing()
+  add_subdirectory(test)
+endif()
diff --git a/cmake/cusanToolchain.cmake b/cmake/cusanToolchain.cmake
index 449a0ff..cf2a55a 100644
--- a/cmake/cusanToolchain.cmake
+++ b/cmake/cusanToolchain.cmake
@@ -26,7 +26,7 @@ string(COMPARE EQUAL "${CMAKE_SOURCE_DIR}" "${PROJECT_SOURCE_DIR}"
 )
 
 find_package(CUDAToolkit REQUIRED)
-find_package(MPI REQUIRED)
+find_package(MPI)
 
 option(CUSAN_TEST_CONFIGURE_IDE "Add targets for tests to help the IDE with completion etc." ON)
 mark_as_advanced(CUSAN_TEST_CONFIGURE_IDE)
diff --git a/lib/runtime/CMakeLists.txt b/lib/runtime/CMakeLists.txt
index 741bbc2..5397584 100644
--- a/lib/runtime/CMakeLists.txt
+++ b/lib/runtime/CMakeLists.txt
@@ -1,67 +1,69 @@
-find_package(MPI)
-add_library(cusan_mpi_interceptor SHARED MPIInterception.cpp)
-target_link_libraries(cusan_mpi_interceptor
-  PRIVATE MPI::MPI_CXX
-)
-set_target_properties(
-  cusan_mpi_interceptor
-  PROPERTIES
-  OUTPUT_NAME "CusanMPIInterceptor"
-  EXPORT_NAME "MPIInterceptor"
-)
-add_library(cusan::MPI_Interceptor ALIAS cusan_mpi_interceptor)
-target_compile_features(cusan_mpi_interceptor PUBLIC cxx_std_17)
-target_compile_definitions(
-  cusan_mpi_interceptor
-  PRIVATE CUSAN_LOG_LEVEL=${CUSAN_LOG_LEVEL_RT}
-  LLVM_VERSION_MAJOR=${LLVM_VERSION_MAJOR}
-   $<$<BOOL:${CUSAN_FIBERPOOL}>:CUSAN_FIBERPOOL=1>
-   $<$<BOOL:${CUSAN_TYPEART}>:CUSAN_TYPEART=1>
-   $<$<BOOL:${CUSAN_SOFTCOUNTER}>:CUSAN_SOFTCOUNTER>
-)
-
-target_include_directories(cusan_mpi_interceptor ${warning_guard}
-  PUBLIC $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
-  PRIVATE $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/lib>
-)
-
-target_include_directories(cusan_mpi_interceptor
-  SYSTEM
-  PRIVATE
-  ${LLVM_INCLUDE_DIRS}
-)
-
-target_link_libraries(cusan_mpi_interceptor PRIVATE LLVMSupport)
-
-if(CUSAN_FIBERPOOL)
-  target_link_libraries(cusan_mpi_interceptor PUBLIC cusan::fiberpool)
-  set_target_properties(cusan_mpi_interceptor PROPERTIES
-        BUILD_WITH_INSTALL_RPATH FALSE
-        LINK_FLAGS "-Wl,-rpath,${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
+if (MPI_FOUND)
+  add_library(cusan_mpi_interceptor SHARED MPIInterception.cpp)
+  target_link_libraries(cusan_mpi_interceptor
+    PRIVATE MPI::MPI_CXX
+  )
+  set_target_properties(
+    cusan_mpi_interceptor
+    PROPERTIES
+    OUTPUT_NAME "CusanMPIInterceptor"
+    EXPORT_NAME "MPIInterceptor"
+  )
+  add_library(cusan::MPI_Interceptor ALIAS cusan_mpi_interceptor)
+  
+  target_compile_features(cusan_mpi_interceptor PUBLIC cxx_std_17)
+  target_compile_definitions(
+    cusan_mpi_interceptor
+    PRIVATE CUSAN_LOG_LEVEL=${CUSAN_LOG_LEVEL_RT}
+    LLVM_VERSION_MAJOR=${LLVM_VERSION_MAJOR}
+    $<$<BOOL:${CUSAN_FIBERPOOL}>:CUSAN_FIBERPOOL=1>
+    $<$<BOOL:${CUSAN_TYPEART}>:CUSAN_TYPEART=1>
+    $<$<BOOL:${CUSAN_SOFTCOUNTER}>:CUSAN_SOFTCOUNTER>
+  )
+
+  target_include_directories(cusan_mpi_interceptor ${warning_guard}
+    PUBLIC $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
+    PRIVATE $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/lib>
+  )
+
+  target_include_directories(cusan_mpi_interceptor
+    SYSTEM
+    PRIVATE
+    ${LLVM_INCLUDE_DIRS}
+  )
+
+  target_link_libraries(cusan_mpi_interceptor PRIVATE LLVMSupport)
+
+  if(CUSAN_FIBERPOOL)
+    target_link_libraries(cusan_mpi_interceptor PUBLIC cusan::fiberpool)
+    set_target_properties(cusan_mpi_interceptor PROPERTIES
+          BUILD_WITH_INSTALL_RPATH FALSE
+          LINK_FLAGS "-Wl,-rpath,${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
+  endif()
+
+  set(CONFIG_NAME cusanMPIInterceptor)
+  set(TARGETS_EXPORT_NAME ${CONFIG_NAME}Targets)
+
+  install(
+    TARGETS cusan_mpi_interceptor
+    EXPORT ${TARGETS_EXPORT_NAME}
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  )
+
+  install(
+    EXPORT ${TARGETS_EXPORT_NAME}
+    NAMESPACE cusan::
+    DESTINATION ${CUSAN_INSTALL_CONFIGDIR}
+  )
+
+  export(
+    EXPORT ${TARGETS_EXPORT_NAME}
+    FILE ${CMAKE_BINARY_DIR}/${TARGETS_EXPORT_NAME}.cmake
+    NAMESPACE cusan::
+  )
 endif()
 
-set(CONFIG_NAME cusanMPIInterceptor)
-set(TARGETS_EXPORT_NAME ${CONFIG_NAME}Targets)
-
-install(
-  TARGETS cusan_mpi_interceptor
-  EXPORT ${TARGETS_EXPORT_NAME}
-  LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
-  ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
-)
-
-install(
-  EXPORT ${TARGETS_EXPORT_NAME}
-  NAMESPACE cusan::
-  DESTINATION ${CUSAN_INSTALL_CONFIGDIR}
-)
-
-export(
-  EXPORT ${TARGETS_EXPORT_NAME}
-  FILE ${CMAKE_BINARY_DIR}/${TARGETS_EXPORT_NAME}.cmake
-  NAMESPACE cusan::
-)
-
 add_library(cusan_Runtime SHARED CusanRuntime.cpp CusanRuntime_cudaSpecific.cpp)
 set_target_properties(
   cusan_Runtime
diff --git a/lib/runtime/TSanInterface.h b/lib/runtime/TSanInterface.h
index 324b688..3bfe367 100644
--- a/lib/runtime/TSanInterface.h
+++ b/lib/runtime/TSanInterface.h
@@ -1,9 +1,11 @@
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wformat"
-#pragma GCC diagnostic ignored "-Wmacro-redefined"
 #pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wformat-pedantic"
 #pragma GCC diagnostic ignored "-Wcast-qual"
+#if defined(__clang__)
+#pragma clang diagnostic ignored "-Wmacro-redefined"
+#pragma clang diagnostic ignored "-Wformat-pedantic"
+#endif
 
 #include "TSan_External.h"
 
diff --git a/scripts/CMakeLists.txt b/scripts/CMakeLists.txt
index 74d956c..b27f7fb 100644
--- a/scripts/CMakeLists.txt
+++ b/scripts/CMakeLists.txt
@@ -175,7 +175,6 @@ if(NOT CUSAN_TYPEART)
   )
 endif()
 
-find_package(MPI)
 if(MPI_C_FOUND)
   configure_cusan_script(
     ${CUSAN_WRAPPER} cusan-mpicc${CMAKE_DEBUG_POSTFIX}
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 64266d6..d13fe37 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -36,6 +36,7 @@ function(configure_cusan_lit_site_cfg input output)
 
   pythonize_bool(${CUSAN_TEST_WORKAROUNDS} CUSAN_TEST_WORKAROUND)
   pythonize_bool(${CUSAN_TYPEART} CUSAN_HAS_TYPEART)
+  pythonize_bool(${MPI_FOUND} CUSAN_HAS_MPI)
 
   cusan_target_generate_file(${input} ${output})
 endfunction()
@@ -90,16 +91,32 @@ function(cusan_add_lit_target)
   endforeach()
 endfunction()
 
+
+if(NOT MPI_FOUND)
+  # mock target to make test generation work
+  add_library(cusan_mpi_interceptor_mock SHARED IMPORTED)
+  set_target_properties(
+    cusan_mpi_interceptor_mock
+    PROPERTIES
+    OUTPUT_NAME "CusanMPIInterceptorPlaceholder"
+    IMPORTED_LOCATION "${PROJECT_SOURCE_DIR}/lib/runtime"
+  )
+  set(_cusan_interceptor_target cusan_mpi_interceptor_mock)
+  add_library(cusan::MPI_Interceptor ALIAS cusan_mpi_interceptor_mock)
+endif()
+
 set(CUSAN_TEST_DEPENDS
   cusan::TransformPass
-  cusan::MPI_Interceptor
   cusan::Runtime
   cusan::Analysis
-  
 )
 
+if(FOUND_MPI)
+  list(APPEND CUSAN_TEST_DEPENDS cusan::MPI_Interceptor)
+endif()
+
 if(CUSAN_TYPEART)
-list(APPEND CUSAN_TEST_DEPENDS typeart::Runtime typeart::TransformPass)
+  list(APPEND CUSAN_TEST_DEPENDS typeart::Runtime typeart::TransformPass)
 endif()
 
 set(CUSAN_SUITES
diff --git a/test/kernel_analysis/03_struct_write.c b/test/kernel_analysis/03_struct_write.c
index c001e03..a6edeef 100644
--- a/test/kernel_analysis/03_struct_write.c
+++ b/test/kernel_analysis/03_struct_write.c
@@ -22,6 +22,8 @@
 // CHECK-NEXT: subarg: {{.*}}ptr: 0, rw: ReadWrite
 // CHECK-NOT: Handling Arg:
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/kernel_analysis/08_big_struct_write.c b/test/kernel_analysis/08_big_struct_write.c
index 5d34cd5..3df5664 100644
--- a/test/kernel_analysis/08_big_struct_write.c
+++ b/test/kernel_analysis/08_big_struct_write.c
@@ -22,6 +22,8 @@
 // CHECK-NEXT: subarg: {{.*}}ptr: 0, rw: ReadWrite
 // CHECK-NOT: Handling Arg:
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/lit.cfg b/test/lit.cfg
index cf956ce..909e384 100644
--- a/test/lit.cfg
+++ b/test/lit.cfg
@@ -45,6 +45,8 @@ if config.cusan_typeart:
 
 if config.mpiexec:
   config.available_features.add('mpiexec')
+if config.cusan_mpi:
+  config.available_features.add('mpi')
 
 config.substitutions.append(("%clang-cpp", clang_cpp))
 config.substitutions.append(("%clang-cc", clang_cc))
diff --git a/test/lit.site.cfg.in b/test/lit.site.cfg.in
index 70030c3..8f655dd 100644
--- a/test/lit.site.cfg.in
+++ b/test/lit.site.cfg.in
@@ -28,9 +28,10 @@ config.cusan_pass_dir = "@CUSAN_PASS_DIR@"
 config.cusan_runtime_dir = "@CUSAN_RUNTIME_DIR@"
 config.cusan_script_dir = "@CUSAN_SCRIPT_DIR@"
 config.cusan_pass = "$<TARGET_FILE_NAME:cusan::TransformPass>"
-config.cusan_mpi_interceptor = "$<TARGET_FILE:cusan::MPI_Interceptor>"
+config.cusan_mpi_interceptor = "$<TARGET_FILE_NAME:cusan::MPI_Interceptor>"
 config.cusan_use_workarounds = @CUSAN_TEST_WORKAROUND@
 config.cusan_typeart = @CUSAN_HAS_TYPEART@
+config.cusan_mpi = @CUSAN_HAS_MPI@
 
 # Let the main config do the real work.
 config.loaded_site_config = True
diff --git a/test/pass/03_cuda_to_mpi.c b/test/pass/03_cuda_to_mpi.c
index 82d6a98..5186403 100644
--- a/test/pass/03_cuda_to_mpi.c
+++ b/test/pass/03_cuda_to_mpi.c
@@ -2,11 +2,11 @@
 // RUN: %rm-file %t.yaml 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
-
-
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaMemcpy({{i8\*|ptr}} {{.*}}[[target:%[0-9a-z]+]], {{i8\*|ptr}} {{.*}}[[from:%[0-9a-z]+]], 
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_memcpy({{i8\*|ptr}} {{.*}}[[target]], {{i8\*|ptr}} {{.*}}[[from]],
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/04_mpi_to_cuda.c b/test/pass/04_mpi_to_cuda.c
index e27f41a..6196dcc 100644
--- a/test/pass/04_mpi_to_cuda.c
+++ b/test/pass/04_mpi_to_cuda.c
@@ -8,6 +8,8 @@
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaMemcpy({{i8\*|ptr}} {{.*}}[[target:%[0-9a-z]+]], {{i8\*|ptr}} {{.*}}[[from:%[0-9a-z]+]],
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_memcpy({{i8\*|ptr}} {{.*}}[[target]], {{i8\*|ptr}} {{.*}}[[from]],
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/05_cuda_to_mpi_stream.c b/test/pass/05_cuda_to_mpi_stream.c
index bdb304f..6a22f3e 100644
--- a/test/pass/05_cuda_to_mpi_stream.c
+++ b/test/pass/05_cuda_to_mpi_stream.c
@@ -10,6 +10,8 @@
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaStreamSynchronize
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_sync_stream
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/06_cuda_to_mpi_event.c b/test/pass/06_cuda_to_mpi_event.c
index 2253965..50bdd34 100644
--- a/test/pass/06_cuda_to_mpi_event.c
+++ b/test/pass/06_cuda_to_mpi_event.c
@@ -10,6 +10,8 @@
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaMemcpy({{i8\*|ptr}} {{.*}}[[target:%[0-9a-z]+]], {{i8\*|ptr}} {{.*}}[[from:%[0-9a-z]+]], 
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_memcpy({{i8\*|ptr}} {{.*}}[[target]], {{i8\*|ptr}} {{.*}}[[from]],
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/07_cuda_to_mpi_read.c b/test/pass/07_cuda_to_mpi_read.c
index c1ce739..f379fb7 100644
--- a/test/pass/07_cuda_to_mpi_read.c
+++ b/test/pass/07_cuda_to_mpi_read.c
@@ -11,6 +11,8 @@
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaMemcpy({{i8\*|ptr}} {{.*}}[[target:%[0-9a-z]+]], {{i8\*|ptr}} {{.*}}[[from:%[0-9a-z]+]],
 // CHECK-LLVM-IR: {{(call|invoke)}} void @_cusan_memcpy({{i8\*|ptr}} {{.*}}[[target]], {{i8\*|ptr}} {{.*}}[[from]],
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/08_cudamemcpy_to_mpi.c b/test/pass/08_cudamemcpy_to_mpi.c
index 3d1139c..1d553a2 100644
--- a/test/pass/08_cudamemcpy_to_mpi.c
+++ b/test/pass/08_cudamemcpy_to_mpi.c
@@ -15,6 +15,8 @@
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaStreamSynchronize
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_sync_stream
 
+// REQUIRES: mpi
+
 // clang-format on
 
 // Tsan sometimes crashes with this test it seems
diff --git a/test/pass/11_cuda_to_mpi_struct_of_buff.c b/test/pass/11_cuda_to_mpi_struct_of_buff.c
index 2dcea96..3a2001d 100644
--- a/test/pass/11_cuda_to_mpi_struct_of_buff.c
+++ b/test/pass/11_cuda_to_mpi_struct_of_buff.c
@@ -3,6 +3,8 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
+
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamCreate
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_create_stream
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamCreate
diff --git a/test/pass/18_cuda_to_mpi_event_query_busy_loop.c b/test/pass/18_cuda_to_mpi_event_query_busy_loop.c
index cd00aae..8c590cc 100644
--- a/test/pass/18_cuda_to_mpi_event_query_busy_loop.c
+++ b/test/pass/18_cuda_to_mpi_event_query_busy_loop.c
@@ -3,6 +3,7 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
 
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaStreamCreate
 // CHECK-LLVM-IR: {{(call|invoke)}} void @_cusan_create_stream
diff --git a/test/pass/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c b/test/pass/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c
index 1c5d299..248c3a6 100644
--- a/test/pass/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c
+++ b/test/pass/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c
@@ -3,6 +3,7 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
 
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaStreamCreate
 // CHECK-LLVM-IR: {{(call|invoke)}} void @_cusan_create_stream
diff --git a/test/pass/20_cuda_to_mpi_send_ds_sync_w_r.c b/test/pass/20_cuda_to_mpi_send_ds_sync_w_r.c
index f86f3a0..871b3ab 100644
--- a/test/pass/20_cuda_to_mpi_send_ds_sync_w_r.c
+++ b/test/pass/20_cuda_to_mpi_send_ds_sync_w_r.c
@@ -3,6 +3,8 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
+
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamCreate
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_create_stream
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaMemset
diff --git a/test/pass/21_chunked_streams_example.c b/test/pass/21_chunked_streams_example.c
index 13562f7..03774ce 100644
--- a/test/pass/21_chunked_streams_example.c
+++ b/test/pass/21_chunked_streams_example.c
@@ -3,6 +3,8 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
+
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamCreate
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_create_stream
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamCreate
diff --git a/test/pass/22_cuda_to_mpi_partial_buff_write.c b/test/pass/22_cuda_to_mpi_partial_buff_write.c
index c037c8a..2f23ae8 100644
--- a/test/pass/22_cuda_to_mpi_partial_buff_write.c
+++ b/test/pass/22_cuda_to_mpi_partial_buff_write.c
@@ -3,6 +3,8 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
+
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamCreate
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_create_stream
 // CHECK-LLVM-IR: {{call|invoke}} i32 @cudaStreamSynchronize
diff --git a/test/pass/26_malloc_pitch.c b/test/pass/26_malloc_pitch.c
index b71af1a..7a0a97c 100644
--- a/test/pass/26_malloc_pitch.c
+++ b/test/pass/26_malloc_pitch.c
@@ -6,6 +6,8 @@
 // {{.*}}[[from:%[0-9a-z]+]], CHECK-LLVM-IR: {{call|invoke}} void @_cusan_memcpy({{i8\*|ptr}} {{.*}}[[target]],
 // {{i8\*|ptr}} {{.*}}[[from]],
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/29_tsan_cuda_to_mpi.c b/test/pass/29_tsan_cuda_to_mpi.c
index 7608db6..9ab6fdc 100644
--- a/test/pass/29_tsan_cuda_to_mpi.c
+++ b/test/pass/29_tsan_cuda_to_mpi.c
@@ -3,6 +3,7 @@
 
 // RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
 
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaMemcpy
 // CHECK-LLVM-IR: {{(call|invoke)}} void @_cusan_memcpy
diff --git a/test/pass/30_tsan_annotate_cuda_to_mpi.c b/test/pass/30_tsan_annotate_cuda_to_mpi.c
index 506eff1..335e372 100644
--- a/test/pass/30_tsan_annotate_cuda_to_mpi.c
+++ b/test/pass/30_tsan_annotate_cuda_to_mpi.c
@@ -8,6 +8,8 @@
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaFree
 // CHECK-LLVM-IR: {{(call|invoke)}} void @_cusan_device_free
 
+// REQUIRES: mpi
+
 // clang-format on
 
 #include "../support/gpu_mpi.h"
diff --git a/test/pass/31_tsan_cuda_event.c b/test/pass/31_tsan_cuda_event.c
index 918ba99..311dc93 100644
--- a/test/pass/31_tsan_cuda_event.c
+++ b/test/pass/31_tsan_cuda_event.c
@@ -1,8 +1,9 @@
 // clang-format off
 // RUN: %rm-file %t.yaml 
 
-// RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
+// RUN: %wrapper-cc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
+// REQUIRES: mpi
 
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaEventCreate
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_create_event
diff --git a/test/pass/32_tsan_async_copy.c b/test/pass/32_tsan_async_copy.c
index cc7c12f..d30220d 100644
--- a/test/pass/32_tsan_async_copy.c
+++ b/test/pass/32_tsan_async_copy.c
@@ -1,7 +1,7 @@
 // clang-format off
 // RUN: %rm-file %t.yaml 
 
-// RUN: %wrapper-mpicc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
+// RUN: %wrapper-cc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
 
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaStreamCreate 
diff --git a/test/runtime/03_cuda_to_mpi.c b/test/runtime/03_cuda_to_mpi.c
index d7f11b6..e2dcd9d 100644
--- a/test/runtime/03_cuda_to_mpi.c
+++ b/test/runtime/03_cuda_to_mpi.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/04_mpi_to_cuda.c b/test/runtime/04_mpi_to_cuda.c
index 370f48f..77f8915 100644
--- a/test/runtime/04_mpi_to_cuda.c
+++ b/test/runtime/04_mpi_to_cuda.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/05_cuda_to_mpi_stream.c b/test/runtime/05_cuda_to_mpi_stream.c
index af827b9..4ff4400 100644
--- a/test/runtime/05_cuda_to_mpi_stream.c
+++ b/test/runtime/05_cuda_to_mpi_stream.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/06_cuda_to_mpi_event.c b/test/runtime/06_cuda_to_mpi_event.c
index 0585479..65ef15e 100644
--- a/test/runtime/06_cuda_to_mpi_event.c
+++ b/test/runtime/06_cuda_to_mpi_event.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/07_cuda_to_mpi_read.c b/test/runtime/07_cuda_to_mpi_read.c
index e367e2a..d1ee1fc 100644
--- a/test/runtime/07_cuda_to_mpi_read.c
+++ b/test/runtime/07_cuda_to_mpi_read.c
@@ -2,7 +2,7 @@
 // RUN: %wrapper-mpicxx %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t.exe 2>&1 | %filecheck --allow-empty %s
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/08_cudamemcpy_to_mpi.c b/test/runtime/08_cudamemcpy_to_mpi.c
index b8fa5bf..4aedbcc 100644
--- a/test/runtime/08_cudamemcpy_to_mpi.c
+++ b/test/runtime/08_cudamemcpy_to_mpi.c
@@ -2,7 +2,7 @@
 // RUN: %wrapper-mpicxx %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t.exe
 // RUN: %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t.exe 2>&1 | %filecheck %s --allow-empty
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/11_cuda_to_mpi_struct_of_buff.c b/test/runtime/11_cuda_to_mpi_struct_of_buff.c
index 4a3141d..8455f74 100644
--- a/test/runtime/11_cuda_to_mpi_struct_of_buff.c
+++ b/test/runtime/11_cuda_to_mpi_struct_of_buff.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/12_struct_ptr.c b/test/runtime/12_struct_ptr.c
index 5a71312..41ab21d 100644
--- a/test/runtime/12_struct_ptr.c
+++ b/test/runtime/12_struct_ptr.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // CHECK-DAG: data race
 
diff --git a/test/runtime/18_cuda_to_mpi_event_query_busy_loop.c b/test/runtime/18_cuda_to_mpi_event_query_busy_loop.c
index 73674a7..0b82829 100644
--- a/test/runtime/18_cuda_to_mpi_event_query_busy_loop.c
+++ b/test/runtime/18_cuda_to_mpi_event_query_busy_loop.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c b/test/runtime/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c
index 61eee40..23d9a9c 100644
--- a/test/runtime/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c
+++ b/test/runtime/19_cuda_to_mpi_send_cudaMemcpyAsyncH2H_implicit_sync.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/20_cuda_to_mpi_send_ds_sync_w_r.c b/test/runtime/20_cuda_to_mpi_send_ds_sync_w_r.c
index 820912b..7f88a11 100644
--- a/test/runtime/20_cuda_to_mpi_send_ds_sync_w_r.c
+++ b/test/runtime/20_cuda_to_mpi_send_ds_sync_w_r.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/21_chunked_streams_example.c b/test/runtime/21_chunked_streams_example.c
index fa4f5d0..6b7dee5 100644
--- a/test/runtime/21_chunked_streams_example.c
+++ b/test/runtime/21_chunked_streams_example.c
@@ -2,7 +2,7 @@
 // RUN: %wrapper-mpicxx %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t.exe 2>&1 | %filecheck %s -DFILENAME=%s
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/22_cuda_to_mpi_partial_buff_write.c b/test/runtime/22_cuda_to_mpi_partial_buff_write.c
index 8a29dd9..2639ce8 100644
--- a/test/runtime/22_cuda_to_mpi_partial_buff_write.c
+++ b/test/runtime/22_cuda_to_mpi_partial_buff_write.c
@@ -2,7 +2,7 @@
 // RUN: %wrapper-mpicxx %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t.exe 2>&1 | %filecheck %s --allow-empty
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/26_malloc_pitch.c b/test/runtime/26_malloc_pitch.c
index 3d7524f..0dac955 100644
--- a/test/runtime/26_malloc_pitch.c
+++ b/test/runtime/26_malloc_pitch.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/29_tsan_cuda_to_mpi.c b/test/runtime/29_tsan_cuda_to_mpi.c
index 14ebfa1..2ac700f 100644
--- a/test/runtime/29_tsan_cuda_to_mpi.c
+++ b/test/runtime/29_tsan_cuda_to_mpi.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 
diff --git a/test/runtime/30_tsan_annotate_cuda_to_mpi.c b/test/runtime/30_tsan_annotate_cuda_to_mpi.c
index 24e59ce..19c6d97 100644
--- a/test/runtime/30_tsan_annotate_cuda_to_mpi.c
+++ b/test/runtime/30_tsan_annotate_cuda_to_mpi.c
@@ -5,7 +5,7 @@
 // RUN: %wrapper-mpicxx -DCUSAN_SYNC %clang_args -x cuda -g %s -gencode arch=compute_70,code=sm_70 -o %cusan_test_dir/%basename_t-sync.exe
 // RUN: %cusan_ldpreload %tsan-options %mpi-exec -n 2 %cusan_test_dir/%basename_t-sync.exe 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-SYNC
 
-// REQUIRES: mpiexec
+// REQUIRES: mpi
 
 // clang-format on
 

From 850f74accbc2840557b9579899660d09735ff4dd Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Alexander=20H=C3=BCck?= <hueck.alexander@gmail.com>
Date: Sun, 20 Apr 2025 18:05:46 +0200
Subject: [PATCH 2/3] Revert target gen expression

---
 test/lit.site.cfg.in | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/lit.site.cfg.in b/test/lit.site.cfg.in
index 8f655dd..d6cde3f 100644
--- a/test/lit.site.cfg.in
+++ b/test/lit.site.cfg.in
@@ -28,7 +28,7 @@ config.cusan_pass_dir = "@CUSAN_PASS_DIR@"
 config.cusan_runtime_dir = "@CUSAN_RUNTIME_DIR@"
 config.cusan_script_dir = "@CUSAN_SCRIPT_DIR@"
 config.cusan_pass = "$<TARGET_FILE_NAME:cusan::TransformPass>"
-config.cusan_mpi_interceptor = "$<TARGET_FILE_NAME:cusan::MPI_Interceptor>"
+config.cusan_mpi_interceptor = "$<TARGET_FILE:cusan::MPI_Interceptor>"
 config.cusan_use_workarounds = @CUSAN_TEST_WORKAROUND@
 config.cusan_typeart = @CUSAN_HAS_TYPEART@
 config.cusan_mpi = @CUSAN_HAS_MPI@

From 74d4c7a8d0a8e7bc1359827e3fc4a477837e2e21 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Alexander=20H=C3=BCck?= <hueck.alexander@gmail.com>
Date: Sun, 20 Apr 2025 18:10:10 +0200
Subject: [PATCH 3/3] Fix test

---
 test/pass/31_tsan_cuda_event.c | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/test/pass/31_tsan_cuda_event.c b/test/pass/31_tsan_cuda_event.c
index 311dc93..eb819ea 100644
--- a/test/pass/31_tsan_cuda_event.c
+++ b/test/pass/31_tsan_cuda_event.c
@@ -3,8 +3,6 @@
 
 // RUN: %wrapper-cc %clang-pass-only-args --cusan-kernel-data=%t.yaml -x cuda --cuda-gpu-arch=sm_72 %s 2>&1 | %filecheck %s  -DFILENAME=%s --allow-empty --check-prefix CHECK-LLVM-IR
 
-// REQUIRES: mpi
-
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaEventCreate
 // CHECK-LLVM-IR: {{call|invoke}} void @_cusan_create_event
 // CHECK-LLVM-IR: {{(call|invoke)}} i32 @cudaStreamCreate
@@ -16,8 +14,6 @@
 
 // clang-format on
 
-#include "../support/gpu_mpi.h"
-
 #include <unistd.h>
 
 __global__ void kernel(int* arr, const int N) {  // CHECK-DAG: [[FILENAME]]:[[@LINE]]
@@ -28,10 +24,6 @@ __global__ void kernel(int* arr, const int N) {  // CHECK-DAG: [[FILENAME]]:[[@L
 }
 
 int main(int argc, char* argv[]) {
-  if (!has_gpu_aware_mpi()) {
-    printf("This example is designed for CUDA-aware MPI. Exiting.\n");
-    return 1;
-  }
   cudaEvent_t first_finished_event;
   cudaEventCreate(&first_finished_event);
   cudaStream_t stream1;