preland [OpenMP] Port the OpenMP device runtime to direct C++ compilation llvm#123673

Author: ronlieb

clang/lib/Headers/__clang_hip_math.h

@@ -21,8 +21,11 @@
 #include <limits.h>
 #include <stdint.h>
 #ifdef __OPENMP_AMDGCN__
+// FIXME: A hack for the OpenMP DeviceRTL's `LibM.h` that should be removed.
+#ifndef __OPENMP_SKIP_INCLUDE__
 #include <omp.h>
 #endif
+#endif
 #endif // !defined(__HIPCC_RTC__)
 
 // __DEVICE__ is a helper macro with common set of attributes for the wrappers

offload/DeviceRTL/CMakeLists.txt

@@ -90,7 +90,7 @@ set(src_files
 # propagation. That said, we will run the vectorizer again after the runtime
 # has been linked into the user program.
 set(clang_opt_flags -O3 -mllvm -openmp-opt-disable -DSHARED_SCRATCHPAD_SIZE=512 -mllvm -vectorize-slp=false )
-set(link_opt_flags  -O3        -openmp-opt-disable -attributor-enable=module -vectorize-slp=false )
+set(link_opt_flags  -O3 -openmp-opt-disable -attributor-enable=module -vectorize-slp=false )
 set(link_export_flag -passes=internalize -internalize-public-api-file=${source_directory}/exports)
 
 # If the user built with the GPU C library enabled we will use that instead.
@@ -103,11 +103,10 @@ set (LIBOMPTARGET_LLVM_INCLUDE_DIRS_DEVICERTL "${LIBOMPTARGET_LLVM_INCLUDE_DIRS}
 list(TRANSFORM LIBOMPTARGET_LLVM_INCLUDE_DIRS_DEVICERTL PREPEND "-I")
 
 # Set flags for LLVM Bitcode compilation.
-set(bc_flags -c -foffload-lto -std=c++17 -fvisibility=hidden
-              ${clang_opt_flags} --offload-device-only
-             -nocudalib -nogpulib -nogpuinc
-             -fopenmp -fopenmp-cuda-mode
-             -Wno-unknown-cuda-version -Wno-openmp-target
+set(bc_flags -c -flto -std=c++17 -fvisibility=hidden
+             ${clang_opt_flags} -nocudalib -nogpulib
+             -nogpuinc -nostdlibinc -Wno-unknown-cuda-version
+             -fno-rtti -fno-exceptions -fconvergent-functions
              -I${CMAKE_BINARY_DIR}/openmp/runtime/src # Need omp.h for LibM.
              -I${CMAKE_BINARY_DIR}/projects/openmp/runtime/src # Need omp.h for LibM.
              -I${CMAKE_BINARY_DIR}/runtime/src
@@ -139,9 +138,8 @@ function(compileDeviceRTLLibrary target_name target_triple)
     add_custom_target(${outfile}
       COMMAND ${CLANG_TOOL}
       ${bc_flags}
-      -fopenmp-targets=${target_triple}
-      -Xopenmp-target=${target_triple} -march=
       ${target_bc_flags}
+      --target=${target_triple}
       -MD -MF ${depfile}
       ${infile} -o ${outfile}
       DEPENDS ${infile} ${include_files}
@@ -310,10 +308,7 @@ function(compileDeviceRTLLibrary target_name target_triple)
     set(ide_target_name omptarget-ide-${target_name})
     add_library(${ide_target_name} STATIC EXCLUDE_FROM_ALL ${src_files})
     target_compile_options(${ide_target_name} PRIVATE
-      -fopenmp -fopenmp-cuda-mode
-      -fopenmp-targets=${target_triple} -Xopenmp-target=${target_triple} -march=
-      -mllvm -openmp-opt-disable
-      -foffload-lto -fvisibility=hidden --offload-device-only
+      --target=${target_triple} -flto -fvisibility=hidden
       -nocudalib -nogpulib -nogpuinc -nostdlibinc -Wno-unknown-cuda-version
     )
     target_compile_definitions(${ide_target_name} PRIVATE SHARED_SCRATCHPAD_SIZE=512)

offload/DeviceRTL/include/Allocator.h

@@ -17,8 +17,6 @@
 // Forward declaration.
 struct KernelEnvironmentTy;
 
-#pragma omp begin declare target device_type(nohost)
-
 namespace ompx {
 
 namespace allocator {
@@ -44,6 +42,4 @@ extern "C" {
 [[gnu::weak]] void free(void *Ptr);
 }
 
-#pragma omp end declare target
-
 #endif

offload/DeviceRTL/include/DevRTLExtras.h

@@ -13,13 +13,7 @@
 #ifndef OPENMP_LIBOMPTARGET_DEVICERTL_INCLUDE_DEVRTLEXTRAS_H
 #define OPENMP_LIBOMPTARGET_DEVICERTL_INCLUDE_DEVRTLEXTRAS_H
 
-/// Base type declarations for freestanding mode
-///
-///{
-using uint64_t = unsigned long;
-// TODO: Properly implement this
-using uintptr_t = uint64_t;
-///}
+#include <stdint.h>
 
 /// Macros for allocating variables in different address spaces.
 ///{

offload/DeviceRTL/include/DeviceTypes.h

@@ -100,14 +100,7 @@ struct TaskDescriptorTy {
   TaskFnTy TaskFn;
 };
 
-#pragma omp begin declare variant match(device = {arch(amdgcn)})
 using LaneMaskTy = uint64_t;
-#pragma omp end declare variant
-
-#pragma omp begin declare variant match(                                       \
-        device = {arch(amdgcn)}, implementation = {extension(match_none)})
-using LaneMaskTy = uint64_t;
-#pragma omp end declare variant
 
 namespace lanes {
 enum : LaneMaskTy { All = ~(LaneMaskTy)0 };
@@ -164,8 +157,7 @@ typedef enum omp_allocator_handle_t {
 #define OMP_PRAGMA(STR) __PRAGMA(omp STR)
 
 #define SHARED(NAME)                                                           \
-  NAME [[clang::loader_uninitialized]];                                        \
-  OMP_PRAGMA(allocate(NAME) allocator(omp_pteam_mem_alloc))
+  [[clang::address_space(3)]] NAME [[clang::loader_uninitialized]];
 
 // TODO: clang should use address space 5 for omp_thread_mem_alloc, but right
 //       now that's not the case.

offload/DeviceRTL/include/DeviceUtils.h

@@ -15,8 +15,6 @@
 #include "DeviceTypes.h"
 #include "Shared/Utils.h"
 
-#pragma omp begin declare target device_type(nohost)
-
 namespace utils {
 
 template <typename T> struct type_identity {
@@ -95,6 +93,4 @@ bool isThreadLocalMemPtr(void *Ptr);
 
 } // namespace utils
 
-#pragma omp end declare target
-
 #endif

offload/DeviceRTL/include/Mapping.h

@@ -24,12 +24,8 @@ enum {
   DIM_Z = 2,
 };
 
-#pragma omp begin declare target device_type(nohost)
-
 inline constexpr uint32_t MaxThreadsPerTeam = 1024;
 
-#pragma omp end declare target
-
 /// Initialize the mapping machinery.
 void init(bool IsSPMD);
 

offload/DeviceRTL/include/State.h

@@ -23,8 +23,6 @@
 // Forward declaration.
 struct KernelEnvironmentTy;
 
-#pragma omp begin declare target device_type(nohost)
-
 namespace ompx {
 
 namespace memory {
@@ -89,8 +87,7 @@ struct TeamStateTy {
   ParallelRegionFnTy ParallelRegionFnVar;
 };
 
-extern TeamStateTy TeamState;
-#pragma omp allocate(TeamState) allocator(omp_pteam_mem_alloc)
+extern TeamStateTy [[clang::address_space(3)]] TeamState;
 
 struct ThreadStateTy {
 
@@ -116,8 +113,7 @@ struct ThreadStateTy {
   }
 };
 
-extern ThreadStateTy **ThreadStates;
-#pragma omp allocate(ThreadStates) allocator(omp_pteam_mem_alloc)
+extern ThreadStateTy **[[clang::address_space(3)]] ThreadStates;
 
 /// Initialize the state machinery. Must be called by all threads.
 void init(bool IsSPMD, KernelEnvironmentTy &KernelEnvironment,
@@ -379,6 +375,4 @@ inline state::Value<uint32_t, state::VK_RunSched> RunSched;
 
 } // namespace ompx
 
-#pragma omp end declare target
-
 #endif

offload/DeviceRTL/include/Synchronization.h

@@ -15,8 +15,6 @@
 #include "DeviceTypes.h"
 #include "DeviceUtils.h"
 
-#pragma omp begin declare target device_type(nohost)
-
 namespace ompx {
 namespace atomic {
 
@@ -221,6 +219,4 @@ void system(atomic::OrderingTy Ordering);
 
 } // namespace ompx
 
-#pragma omp end declare target
-
 #endif

offload/DeviceRTL/include/Workshare.h

@@ -12,8 +12,6 @@
 #ifndef OMPTARGET_WORKSHARE_H
 #define OMPTARGET_WORKSHARE_H
 
-#pragma omp begin declare target device_type(nohost)
-
 namespace ompx {
 
 namespace workshare {
@@ -25,6 +23,4 @@ void init(bool IsSPMD);
 
 } // namespace ompx
 
-#pragma omp end declare target
-
 #endif