[OpenMP] Port the OpenMP device runtime to direct C++ compilation (#123673)

Summary:
This removes the use of OpenMP offloading to build the device runtime.
The main benefit here is that we no longer need to rely on offloading
semantics to build a device only runtime. Things like variants are now
no longer needed and can just be simple if-defs. In the future, I will
remove most of the special handling here and fold it into calls to the
`<gpuintrin.h>` functions instead. Additionally I will rework the
compilation to make this a separate runtime.

The current plan is to have this, but make including OpenMP and
offloading either automatically add it, or print a warning if it's
missing. This will allow us to use a normal CMake workflow and delete
all the weird 'lets pull the clang binary out of the build' business.
```
-DRUNTIMES_amdgcn-amd-amdhsa_LLVM_ENABLE_RUNTIMES=offload
-DLLVM_RUNTIME_TARGETS=amdgcn-amd-amdhsa
```

After that, linking the OpenMP device runtime will be `-Xoffload-linker
-lomp`. I.e. no more fat binary business.

Only look at the most recent commit since this includes the two
dependencies
(fix to AMDGPUEmitPrintfBinding and the PointerToMember bug).

Author: jhuber6

offload/DeviceRTL/CMakeLists.txt

@@ -95,11 +95,10 @@ 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 -nostdlibinc
-             -fopenmp -fopenmp-cuda-mode
-             -Wno-unknown-cuda-version -Wno-openmp-target
+set(bc_flags -c -flto -std=c++17 -fvisibility=hidden
+             ${clang_opt_flags} -nogpulib -nostdlibinc
+             -fno-rtti -fno-exceptions -fconvergent-functions
+             -Wno-unknown-cuda-version
              -DOMPTARGET_DEVICE_RUNTIME
              -I${include_directory}
              -I${devicertl_base_directory}/../include
@@ -123,8 +122,7 @@ function(compileDeviceRTLLibrary target_name target_triple)
     add_custom_command(OUTPUT ${outfile}
       COMMAND ${CLANG_TOOL}
       ${bc_flags}
-      -fopenmp-targets=${target_triple}
-      -Xopenmp-target=${target_triple} -march=
+      --target=${target_triple}
       ${target_bc_flags}
       -MD -MF ${depfile}
       ${infile} -o ${outfile}
@@ -242,10 +240,8 @@ function(compileDeviceRTLLibrary target_name target_triple)
     set(ide_target_name omptarget-ide-${target_name})
     add_library(${ide_target_name} STATIC EXCLUDE_FROM_ALL ${src_files})
     target_compile_options(${ide_target_name} PRIVATE
-      -fopenmp-targets=${target_triple} -Xopenmp-target=${target_triple} -march=
-      -fopenmp -fopenmp-cuda-mode -mllvm -openmp-opt-disable
-      -foffload-lto -fvisibility=hidden --offload-device-only
-      -nocudalib -nogpulib -nogpuinc -nostdlibinc -Wno-unknown-cuda-version
+      -fvisibility=hidden --target=${target_triple}
+      -nogpulib -nostdlibinc -Wno-unknown-cuda-version
     )
     target_compile_definitions(${ide_target_name} PRIVATE SHARED_SCRATCHPAD_SIZE=512)
     target_include_directories(${ide_target_name} PRIVATE

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/DeviceTypes.h

@@ -99,14 +99,7 @@ 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 };
@@ -163,8 +156,7 @@ 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

@@ -22,8 +22,6 @@
 // Forward declaration.
 struct KernelEnvironmentTy;
 
-#pragma omp begin declare target device_type(nohost)
-
 namespace ompx {
 
 namespace memory {
@@ -88,8 +86,7 @@ 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 {
 
@@ -115,8 +112,7 @@ 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,
@@ -378,6 +374,4 @@ 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 {
 
@@ -220,6 +218,4 @@ 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

offload/DeviceRTL/src/Allocator.cpp

@@ -19,8 +19,6 @@
 
 using namespace ompx;
 
-#pragma omp begin declare target device_type(nohost)
-
 [[gnu::used, gnu::retain, gnu::weak,
   gnu::visibility(
       "protected")]] DeviceMemoryPoolTy __omp_rtl_device_memory_pool;
@@ -77,5 +75,3 @@ void *allocator::alloc(uint64_t Size) { return BumpAllocator.alloc(Size); }
 void allocator::free(void *Ptr) { BumpAllocator.free(Ptr); }
 
 ///}
-
-#pragma omp end declare target

offload/DeviceRTL/src/Configuration.cpp

@@ -17,8 +17,6 @@
 
 using namespace ompx;
 
-#pragma omp begin declare target device_type(nohost)
-
 // Weak definitions will be overridden by CGOpenmpRuntimeGPU if enabled.
 [[gnu::weak]] extern const uint32_t __omp_rtl_debug_kind = 0;
 [[gnu::weak]] extern const uint32_t __omp_rtl_assume_no_thread_state = 0;
@@ -85,5 +83,3 @@ bool config::mayUseNestedParallelism() {
     return false;
   return state::getKernelEnvironment().Configuration.MayUseNestedParallelism;
 }
-
-#pragma omp end declare target