[compiler-rt] Support unit tests for the GPU build (llvm#187895)

Summary:
This PR enables the basic unit tests for builtins to be run on the GPU
architectures. Other targets like profiling are supported, but the
host-device natures will make it more difficult to adequately unit
test. It may be be possible to do basic tests there, to simply verify
that
counters are present and in the proper format for when they are copied
to the host.

Author: jhuber6

compiler-rt/cmake/Modules/CompilerRTUtils.cmake

@@ -424,6 +424,18 @@ macro(construct_compiler_rt_default_triple)
     endif()
   endif()
 
+  # Try to locate the GPU loader utility for GPU unit tests.
+  if(COMPILER_RT_GPU_BUILD AND NOT COMPILER_RT_EMULATOR)
+    get_filename_component(_compiler_path "${CMAKE_C_COMPILER}" DIRECTORY)
+    find_program(COMPILER_RT_EMULATOR
+                 NAMES llvm-gpu-loader NO_DEFAULT_PATH
+                 PATHS ${LLVM_BINARY_DIR}/bin ${_compiler_path})
+    if(COMPILER_RT_EMULATOR)
+      message(STATUS "Found GPU loader for testing: ${COMPILER_RT_EMULATOR}")
+    endif()
+    unset(_compiler_path)
+  endif()
+
   # Determine if test target triple is specified explicitly, and doesn't match the
   # default.
   if(NOT COMPILER_RT_DEFAULT_TARGET_TRIPLE STREQUAL LLVM_TARGET_TRIPLE)

compiler-rt/cmake/caches/GPU.cmake

@@ -1,7 +1,7 @@
 # This file sets up a CMakeCache for GPU builds of compiler-rt. This supports
 # amdgcn and nvptx builds targeting the builtins and profile libraries.
 
-set(COMPILER_RT_INCLUDE_TESTS OFF CACHE BOOL "")
+set(COMPILER_RT_INCLUDE_TESTS ON CACHE BOOL "")
 set(COMPILER_RT_HAS_SAFESTACK OFF CACHE BOOL "")
 set(COMPILER_RT_DEFAULT_TARGET_ONLY ON CACHE BOOL "")
 

compiler-rt/test/CMakeLists.txt

@@ -98,7 +98,8 @@ if(COMPILER_RT_CAN_EXECUTE_TESTS)
       compiler_rt_test_runtime(${sanitizer})
     endforeach()
   endif()
-  if(COMPILER_RT_BUILD_PROFILE AND COMPILER_RT_HAS_PROFILE)
+  if(COMPILER_RT_BUILD_PROFILE AND COMPILER_RT_HAS_PROFILE
+     AND NOT COMPILER_RT_GPU_BUILD)
     compiler_rt_test_runtime(profile)
   endif()
   if(COMPILER_RT_BUILD_CTX_PROFILE)

compiler-rt/test/builtins/CMakeLists.txt

@@ -49,6 +49,19 @@ foreach(arch ${BUILTIN_TEST_ARCH})
   set(BUILTINS_TEST_TARGET_ARCH ${arch})
   string(TOLOWER "-${arch}-${OS_NAME}" BUILTINS_TEST_CONFIG_SUFFIX)
   get_test_cc_for_arch(${arch} BUILTINS_TEST_TARGET_CC BUILTINS_TEST_TARGET_CFLAGS)
+
+  if(COMPILER_RT_GPU_BUILD)
+    if("${arch}" MATCHES "amdgcn")
+      list(APPEND BUILTINS_TEST_TARGET_CFLAGS
+           -Wno-multi-gpu -flto -mcpu=native -nogpulib -startfiles -stdlib)
+    elseif("${arch}" MATCHES "nvptx")
+      list(APPEND BUILTINS_TEST_TARGET_CFLAGS
+           -Wno-multi-gpu -flto -march=native -nogpulib -startfiles -stdlib)
+    endif()
+    list(APPEND BUILTINS_TEST_TARGET_CFLAGS --target=${COMPILER_RT_DEFAULT_TARGET_TRIPLE})
+    string(REPLACE ";" " " BUILTINS_TEST_TARGET_CFLAGS "${BUILTINS_TEST_TARGET_CFLAGS}")
+  endif()
+
   if (${arch} STREQUAL "armhf")
     list(APPEND BUILTINS_TEST_TARGET_CFLAGS -fomit-frame-pointer -DCOMPILER_RT_ARMHF_TARGET)
     string(REPLACE ";" " " BUILTINS_TEST_TARGET_CFLAGS "${BUILTINS_TEST_TARGET_CFLAGS}")

compiler-rt/test/builtins/Unit/lit.cfg.py

@@ -158,7 +158,7 @@ def get_libgcc_file_name():
 
 # FIXME: Right now we don't compile the C99 complex builtins when using
 # clang-cl. Fix that.
-if not is_msvc:
+if not is_msvc and config.target_arch not in ("amdgcn", "nvptx64"):
     config.available_features.add("c99-complex")
 
 builtins_is_msvc = get_required_attr(config, "builtins_is_msvc")

compiler-rt/test/lit.common.cfg.py

@@ -987,6 +987,8 @@ def is_windows_lto_supported():
 
 
 def target_page_size():
+    if config.target_arch in ("amdgcn", "nvptx64"):
+        return 4096
     try:
         proc = subprocess.Popen(
             f"{emulator or ''} python3",