[WIP] Changes required for embedded GPU LLVM IR Flang runtime

clang/include/clang/Driver/Options.td

@@ -6141,6 +6141,7 @@ def : Flag<["-"], "nogpulib">,
 def : Flag<["-"], "nocudalib">, Alias<no_offloadlib>;
 def gpulibc : Flag<["-"], "gpulibc">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>,
   HelpText<"Link the LLVM C Library for GPUs">;
+def nogpuflangrt : Flag<["-"], "nogpuflangrt">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>;
 def nogpulibc : Flag<["-"], "nogpulibc">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>;
 def nodefaultlibs : Flag<["-"], "nodefaultlibs">,
   Visibility<[ClangOption, FlangOption]>;

clang/lib/Driver/ToolChains/Clang.cpp

@@ -9422,7 +9422,7 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
             Args.MakeArgString("--device-linker=" + TC.getTripleString() + "=" +
                                "-lclang_rt.builtins"));
       bool HasFlangRT = HasCompilerRT && C.getDriver().IsFlangMode();
-      if (HasFlangRT)
+      if (HasFlangRT && !Args.hasArg(options::OPT_nogpuflangrt))
         CmdArgs.push_back(
             Args.MakeArgString("--device-linker=" + TC.getTripleString() + "=" +
                                "-lflang_rt.runtime"));

flang-rt/CMakeLists.txt

@@ -152,6 +152,11 @@ 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 (FLANG_RT_EMBED_GPU_LLVM_IR)
+  add_compile_definitions(EMBED_FLANG_RT_GPU_LLVM_IR)
+endif ()
+
 if (WIN32)
   # Windows DLL currently not implemented.
   set(FLANG_RT_ENABLE_SHARED OFF)

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"
@@ -314,3 +315,4 @@ 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

flang-rt/lib/runtime/io-api-minimal.cpp

@@ -23,6 +23,7 @@ namespace Fortran::runtime::io {
 RT_EXT_API_GROUP_BEGIN
 #endif
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 Cookie IODEF(BeginExternalListOutput)(
     ExternalUnit unitNumber, const char *sourceFile, int sourceLine) {
   return BeginExternalListIO<Direction::Output, ExternalListIoStatementState>(
@@ -33,6 +34,7 @@ enum Iostat IODEF(EndIoStatement)(Cookie cookie) {
   IoStatementState &io{*cookie};
   return static_cast<enum Iostat>(io.EndIoStatement());
 }
+#endif
 
 template <int KIND, typename INT = CppTypeFor<TypeCategory::Integer, KIND>>
 inline RT_API_ATTRS bool FormattedScalarIntegerOutput(
@@ -45,6 +47,7 @@ inline RT_API_ATTRS bool FormattedScalarIntegerOutput(
   }
 }
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 bool IODEF(OutputInteger8)(Cookie cookie, std::int8_t n) {
   return FormattedScalarIntegerOutput<1>(*cookie, n, "OutputInteger8");
 }
@@ -60,6 +63,7 @@ bool IODEF(OutputInteger32)(Cookie cookie, std::int32_t n) {
 bool IODEF(OutputInteger64)(Cookie cookie, std::int64_t n) {
   return FormattedScalarIntegerOutput<8>(*cookie, n, "OutputInteger64");
 }
+#endif
 
 #ifdef __SIZEOF_INT128__
 bool IODEF(OutputInteger128)(Cookie cookie, common::int128_t n) {
@@ -79,13 +83,15 @@ inline RT_API_ATTRS bool FormattedScalarRealOutput(
   }
 }
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 bool IODEF(OutputReal32)(Cookie cookie, float x) {
   return FormattedScalarRealOutput<4>(*cookie, x, "OutputReal32");
 }
 
 bool IODEF(OutputReal64)(Cookie cookie, double x) {
   return FormattedScalarRealOutput<8>(*cookie, x, "OutputReal64");
 }
+#endif
 
 template <int KIND,
     typename REAL = typename RealOutputEditing<KIND>::BinaryFloatingPoint>
@@ -110,6 +116,7 @@ inline RT_API_ATTRS bool FormattedScalarComplexOutput(
   return false;
 }
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 bool IODEF(OutputComplex32)(Cookie cookie, float re, float im) {
   return FormattedScalarComplexOutput<4>(*cookie, re, im, "OutputComplex32");
 }
@@ -145,6 +152,7 @@ bool IODEF(OutputLogical)(Cookie cookie, bool truth) {
     return false;
   }
 }
+#endif
 
 } // namespace Fortran::runtime::io
 

flang-rt/lib/runtime/io-api.cpp

@@ -199,12 +199,14 @@ RT_API_ATTRS Cookie BeginExternalFormattedIO(const char *format,
   }
 }
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 Cookie IODEF(BeginExternalFormattedOutput)(const char *format,
     std::size_t formatLength, const Descriptor *formatDescriptor,
     ExternalUnit unitNumber, const char *sourceFile, int sourceLine) {
   return BeginExternalFormattedIO<Direction::Output>(format, formatLength,
       formatDescriptor, unitNumber, sourceFile, sourceLine);
 }
+#endif
 
 Cookie IODEF(BeginExternalFormattedInput)(const char *format,
     std::size_t formatLength, const Descriptor *formatDescriptor,

flang-rt/lib/runtime/stop.cpp

@@ -13,11 +13,15 @@
 #include "flang-rt/runtime/file.h"
 #include "flang-rt/runtime/io-error.h"
 #include "flang-rt/runtime/terminator.h"
+#if not defined(__AMDGPU__) && not defined(__NVPTX__)
 #include "flang/Runtime/main.h"
+#endif
 #include <cfenv>
 #include <cstdio>
 #include <cstdlib>
+#if not defined(__AMDGPU__) && not defined(__NVPTX__)
 #include <thread>
+#endif
 
 #ifdef HAVE_BACKTRACE
 #include BACKTRACE_HEADER
@@ -26,6 +30,7 @@
 extern "C" {
 
 [[maybe_unused]] static void DescribeIEEESignaledExceptions() {
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 #ifdef fetestexcept // a macro in some environments; omit std::
   auto excepts{fetestexcept(FE_ALL_EXCEPT)};
 #else
@@ -60,13 +65,15 @@ extern "C" {
 #endif
     std::fputc('\n', stderr);
   }
+#endif
 }
 
 static void CloseAllExternalUnits(const char *why) {
   Fortran::runtime::io::IoErrorHandler handler{why};
   Fortran::runtime::io::ExternalFileUnit::CloseAll(handler);
 }
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 [[noreturn]] RT_API_ATTRS void RTNAME(StopStatement)(
     int code, bool isErrorStop, bool quiet) {
 #if defined(RT_DEVICE_COMPILATION)
@@ -103,7 +110,9 @@ static void CloseAllExternalUnits(const char *why) {
   std::exit(code);
 #endif
 }
+#endif
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 [[noreturn]] RT_API_ATTRS void RTNAME(StopStatementText)(
     const char *code, std::size_t length, bool isErrorStop, bool quiet) {
 #if defined(RT_DEVICE_COMPILATION)
@@ -136,6 +145,7 @@ static void CloseAllExternalUnits(const char *why) {
   }
 #endif
 }
+#endif
 
 static bool StartPause() {
   if (Fortran::runtime::io::IsATerminal(0)) {
@@ -218,13 +228,14 @@ static RT_NOINLINE_ATTR void PrintBacktrace() {
 
 #endif
 }
-
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 [[noreturn]] RT_OPTNONE_ATTR void RTNAME(Abort)() {
 #ifdef HAVE_BACKTRACE
   PrintBacktrace();
 #endif
   std::abort();
 }
+#endif
 
 RT_OPTNONE_ATTR void FORTRAN_PROCEDURE_NAME(backtrace)() { PrintBacktrace(); }
 

flang/include/flang/Runtime/extensions.h

@@ -18,10 +18,13 @@
 
 #define FORTRAN_PROCEDURE_NAME(name) name##_
 
-#ifdef _WIN32
+#if defined (_WIN32)
 // UID and GID don't exist on Windows, these exist to avoid errors.
 typedef std::uint32_t uid_t;
 typedef std::uint32_t gid_t;
+#elif (defined(__AMDGPU__) || defined(__NVPTX__)) && defined (EMBED_FLANG_RT_GPU_LLVM_IR)
+typedef std::uint32_t uid_t;
+typedef std::uint32_t gid_t;
 #else
 #include "sys/types.h" //pid_t
 #endif

flang/include/flang/Runtime/main.h

@@ -11,11 +11,15 @@
 
 #include "flang/Runtime/c-or-cpp.h"
 #include "flang/Runtime/entry-names.h"
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 #include <thread>
+#endif
 
 struct EnvironmentDefaultList;
 
+#if (not defined(__AMDGPU__) && not defined(__NVPTX__)) || not defined(EMBED_FLANG_RT_GPU_LLVM_IR)
 std::thread::id RTNAME(GetMainThreadId)();
+#endif
 
 FORTRAN_EXTERN_C_BEGIN
 void RTNAME(ProgramStart)(

libc/CMakeLists.txt

@@ -67,6 +67,7 @@ if(NOT LIBC_NAMESPACE MATCHES "^__llvm_libc")
   message(FATAL_ERROR "Invalid LIBC_NAMESPACE. Must start with '__llvm_libc' was '${LIBC_NAMESPACE}'")
 endif()
 
+string(REPLACE "." "_" LIBC_NAMESPACE "${LIBC_NAMESPACE}")
 message(STATUS "Setting LIBC_NAMESPACE namespace to '${LIBC_NAMESPACE}'")
 add_compile_definitions(LIBC_NAMESPACE=${LIBC_NAMESPACE})
 

libcxx/src/string.cpp

@@ -360,9 +360,11 @@ wstring to_wstring(unsigned long val) { return i_to_string<wstring>(val); }
 wstring to_wstring(unsigned long long val) { return i_to_string<wstring>(val); }
 #endif
 
+#if not defined(__AMDGPU__) && not defined(__NVPTX__)
 string to_string(float val) { return as_string(snprintf, initial_string< string>()(), "%f", val); }
 string to_string(double val) { return as_string(snprintf, initial_string< string>()(), "%f", val); }
 string to_string(long double val) { return as_string(snprintf, initial_string< string>()(), "%Lf", val); }
+#endif
 
 #if _LIBCPP_HAS_WIDE_CHARACTERS
 wstring to_wstring(float val) { return as_string(get_swprintf(), initial_string<wstring>()(), L"%f", val); }

offload/plugins-nextgen/common/CMakeLists.txt

@@ -48,7 +48,6 @@ if(OFFLOAD_ENABLE_EMISSARY_APIS AND LIBOMPTARGET_BUILD_DEVICE_FORTRT)
   target_link_libraries(PluginCommon PRIVATE flang_rt.runtime
   -L${CMAKE_BINARY_DIR}/../../lib  -L${CMAKE_INSTALL_PREFIX}/lib)
 endif()
-
 if (OMPT_TARGET_DEFAULT AND LIBOMPTARGET_OMPT_SUPPORT)
   add_library(PluginOmpt STATIC OMPT/OmptTracing.cpp OMPT/OmptProfiler.cpp)
   target_include_directories(PluginOmpt PUBLIC