[Flang-RT] Changes required for embedded GPU LLVM IR Flang runtime

Author: DominikAdamski

flang-rt/CMakeLists.txt

@@ -152,6 +152,8 @@ if (NOT "${FLANG_RT_LIBCXX_PROVIDER}" IN_LIST FLANG_RT_SUPPORTED_PROVIDERS)
 endif ()
 
 option(FLANG_RT_ENABLE_STATIC "Build Flang-RT as a static library." ON)
+option(FLANG_RT_EMBED_GPU_LLVM_IR "Build Flang-RT as GPU LLVM IR library" ON)
+
 if (WIN32)
   # Windows DLL currently not implemented.
   set(FLANG_RT_ENABLE_SHARED OFF)
@@ -327,7 +329,9 @@ endif ()
 
 if (FLANG_RT_INCLUDE_TESTS)
   add_subdirectory(test)
-  add_subdirectory(unittests)
+  if (NOT "${LLVM_RUNTIMES_TARGET}" MATCHES "^amdgcn|^nvptx")
+    add_subdirectory(unittests)
+  endif()
 else ()
   add_custom_target(check-flang-rt)
 endif()

flang-rt/cmake/modules/AddFlangRT.cmake

@@ -219,6 +219,12 @@ function (add_flangrt_library name)
     # Minimum required C++ version for Flang-RT, even if CMAKE_CXX_STANDARD is defined to something else.
     target_compile_features(${tgtname} PRIVATE cxx_std_17)
 
+    # Determine which version of GPU Flang-RT we want to build:
+    # flang_rt.hostdevice or the implicitly linked device flang_rt.runtime.
+    if (FLANG_RT_EMBED_GPU_LLVM_IR)
+      target_compile_definitions(${tgtname} PRIVATE EMBED_FLANG_RT_GPU_LLVM_IR)
+    endif ()
+
     # When building the flang runtime if LTO is enabled the archive file
     # contains LLVM IR rather than object code. Currently flang is not
     # LTO aware so cannot link this file to compiled Fortran code.

flang-rt/include/flang-rt/runtime/lock.h

@@ -23,7 +23,9 @@
 #endif
 
 #if USE_PTHREADS
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 #include <pthread.h>
+#endif
 #elif defined(_WIN32)
 #include "flang/Common/windows-include.h"
 #else
@@ -45,6 +47,7 @@ class Lock {
   RT_API_ATTRS void Drop() {}
   RT_API_ATTRS bool TakeIfNoDeadlock() { return true; }
 #elif USE_PTHREADS
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
   Lock() { pthread_mutex_init(&mutex_, nullptr); }
   ~Lock() { pthread_mutex_destroy(&mutex_); }
   void Take() {
@@ -68,6 +71,14 @@ class Lock {
     isBusy_ = false;
     pthread_mutex_unlock(&mutex_);
   }
+#else
+  RT_API_ATTRS void Take(){}
+  RT_API_ATTRS bool TakeIfNoDeadlock() {return true;}
+  RT_API_ATTRS bool Try() {return true;}
+  RT_API_ATTRS void Drop() {}
+  Lock() {}
+  ~Lock() {}
+#endif
 #elif defined(_WIN32)
   Lock() { InitializeCriticalSection(&cs_); }
   ~Lock() { DeleteCriticalSection(&cs_); }
@@ -91,9 +102,11 @@ class Lock {
 #if RT_USE_PSEUDO_FILE_UNIT
   // No state.
 #elif USE_PTHREADS
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
   pthread_mutex_t mutex_{};
   volatile bool isBusy_{false};
   volatile pthread_t holder_;
+#endif
 #elif defined(_WIN32)
   CRITICAL_SECTION cs_;
 #else

flang-rt/include/flang-rt/runtime/tools.h

@@ -42,6 +42,13 @@
 #define RT_USE_PSEUDO_FILE_UNIT 1
 #endif
 
+#if (defined(__AMDGPU__) ||  defined(__NVPTX__)) && defined(EMBED_FLANG_RT_GPU_LLVM_IR)
+// Use the pseudo lock and pseudo file unit implementations
+// for the device.
+#define RT_USE_PSEUDO_LOCK 1
+#define RT_USE_PSEUDO_FILE_UNIT 1
+#endif
+
 namespace Fortran::runtime {
 
 class Terminator;

flang-rt/lib/runtime/CMakeLists.txt

@@ -178,7 +178,11 @@ else ()
 endif ()
 
 if ("${LLVM_RUNTIMES_TARGET}" MATCHES "^amdgcn|^nvptx")
-  set(sources ${gpu_sources})
+  if (FLANG_RT_EMBED_GPU_LLVM_IR)
+    set(sources ${supported_sources} ${gpu_sources})
+  else ()
+    set(sources ${gpu_sources})
+  endif ()
 elseif(FLANG_RT_EXPERIMENTAL_OFFLOAD_SUPPORT STREQUAL "CUDA")
   set(sources ${supported_sources})
 else ()

flang-rt/lib/runtime/assign.cpp

@@ -7,7 +7,9 @@
 //===----------------------------------------------------------------------===//
 
 #include "flang/Runtime/assign.h"
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 #include "flang/Runtime/stop.h"
+#endif
 #include "flang-rt/runtime/assign-impl.h"
 #include "flang-rt/runtime/derived.h"
 #include "flang-rt/runtime/descriptor.h"
@@ -862,17 +864,6 @@ void RTDEF(AssignPolymorphic)(Descriptor &to, const Descriptor &from,
           PolymorphicLHS);
 }
 
-#if defined(OMP_OFFLOAD_BUILD)
-// To support a recently added use of variant in the OpenMP offload build,
-// added an abort wrapper which calls the flang-rt FortranAAbort.
-// Avoids the following linker error:
-//   ld.lld: error: undefined symbol: abort
-//   >>> referenced by /tmp/device_aassign.amdgcn.gfx90a-34a7ed.img.lto.o:(std::__throw_bad_variant_access(char const*))
-extern "C" void abort(void) {
-  RTNAME(Abort)();
-}
-#endif
-
 RT_EXT_API_GROUP_END
 } // extern "C"
 } // namespace Fortran::runtime

flang-rt/lib/runtime/descriptor.cpp

@@ -8,7 +8,9 @@
 
 #include "flang-rt/runtime/descriptor.h"
 #include "ISO_Fortran_util.h"
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 #include "memory.h"
+#endif
 #include "flang-rt/runtime/allocator-registry.h"
 #include "flang-rt/runtime/derived.h"
 #include "flang-rt/runtime/stat.h"

flang-rt/lib/runtime/edit-input.cpp

@@ -569,7 +569,11 @@ static RT_API_ATTRS void RaiseFPExceptions(
 #ifdef feraisexcept // a macro in some environments; omit std::
 #define RAISE feraiseexcept
 #else
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined (EMBED_FLANG_RT_GPU_LLVM_IR)
 #define RAISE std::feraiseexcept
+#else
+#define RAISE
+#endif
 #endif
 #endif // !defined(RT_DEVICE_COMPILATION)
 

flang-rt/lib/runtime/environment.cpp

@@ -6,6 +6,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#if (not defined (__AMDGPU__) && not defined(__NVPTX__)) || not defined (EMBED_FLANG_RT_GPU_LLVM_IR)
 #include "flang-rt/runtime/environment.h"
 #include "environment-default-list.h"
 #include "memory.h"
@@ -334,3 +335,4 @@ bool RTNAME(RegisterConfigureEnv)(
 } // extern "C"
 
 } // namespace Fortran::runtime
+#endif

flang-rt/lib/runtime/file.cpp

@@ -6,6 +6,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 #include "flang-rt/runtime/file.h"
 #include "flang-rt/runtime/memory.h"
 #include "flang-rt/runtime/tools.h"
@@ -486,3 +487,4 @@ RT_API_ATTRS std::int64_t SizeInBytes(const char *path) {
 #endif // defined(RT_DEVICE_COMPILATION)
 
 } // namespace Fortran::runtime::io
+#endif