[DeviceRTL] Allow IsSPMDMode to be optimized out in LTO mode

A previous patch merged the static and bitcode versions of the
deviceRTL. We previously used the static library's separate compilation
to set a special flag that prevented `IsSPMDMode` from being put in the
used list and preventing it from being optimized out. When they were
merged we could no longer do this separate compilation that allowed
users of LTO to get more optimal code.

This patch rearranges the code. The `IsSPMDMode` global is now
transitively used by its inclusion in the changed `__keep_alive`
function. This allows us to then manually delete the `__keep_alive`
function from the module when building the static library via
`llvm-extract`. The result is that the bitcode library correctly will
maintain the needed shared state, while the static library will be able
to internalize it and optimize it out.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D135280

openmp/libomptarget/DeviceRTL/CMakeLists.txt

@@ -31,7 +31,8 @@ if (LLVM_DIR)
   find_program(PACKAGER_TOOL clang-offload-packager PATHS ${LLVM_TOOLS_BINARY_DIR} NO_DEFAULT_PATH)
   find_program(LINK_TOOL llvm-link PATHS ${LLVM_TOOLS_BINARY_DIR} NO_DEFAULT_PATH)
   find_program(OPT_TOOL opt PATHS ${LLVM_TOOLS_BINARY_DIR} NO_DEFAULT_PATH)
-  if ((NOT CLANG_TOOL) OR (NOT LINK_TOOL) OR (NOT OPT_TOOL))
+  find_program(EXTRACT_TOOL llvm-extract PATHS ${LLVM_TOOLS_BINARY_DIR} NO_DEFAULT_PATH)
+  if ((NOT CLANG_TOOL) OR (NOT LINK_TOOL) OR (NOT OPT_TOOL) OR (NOT EXTRACT_TOOL))
     libomptarget_say("Not building DeviceRTL. Missing clang: ${CLANG_TOOL}, llvm-link: ${LINK_TOOL} or opt: ${OPT_TOOL}")
     return()
   else()
@@ -44,6 +45,7 @@ elseif (LLVM_TOOL_CLANG_BUILD AND NOT CMAKE_CROSSCOMPILING AND NOT OPENMP_STANDA
   set(PACKAGER_TOOL $<TARGET_FILE:clang-offload-packager>)
   set(LINK_TOOL $<TARGET_FILE:llvm-link>)
   set(OPT_TOOL $<TARGET_FILE:opt>)
+  set(EXTRACT_TOOL $<TARGET_FILE:llvm-extract>)
   libomptarget_say("Building DeviceRTL. Using clang from in-tree build")
 else()
   libomptarget_say("Not building DeviceRTL. No appropriate clang found")
@@ -118,6 +120,7 @@ set(src_files
 set(clang_opt_flags -O3 -mllvm -openmp-opt-disable -DSHARED_SCRATCHPAD_SIZE=512)
 set(link_opt_flags  -O3        -openmp-opt-disable)
 set(link_export_flag -passes=internalize -internalize-public-api-file=${source_directory}/exports)
+set(link_extract_flag --func='__keep_alive' --delete)
 
 # Prepend -I to each list element
 set (LIBOMPTARGET_LLVM_INCLUDE_DIRS_DEVICERTL "${LIBOMPTARGET_LLVM_INCLUDE_DIRS}")
@@ -199,11 +202,18 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
       COMMENT "Optimizing LLVM bitcode ${bclib_name}"
   )
 
+  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/extracted_${bclib_name}
+      COMMAND ${EXTRACT_TOOL} ${link_extract_flag} ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
+                      -o ${CMAKE_CURRENT_BINARY_DIR}/extracted_${bclib_name}
+      DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
+      COMMENT "Extracting LLVM bitcode ${bclib_name}"
+  )
+
   # Package the bitcode in the bitcode and embed it in an ELF for the static library
   add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
       COMMAND ${PACKAGER_TOOL} -o ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
-        "--image=file=${CMAKE_CURRENT_BINARY_DIR}/${bclib_name},triple=${target_triple},arch=${target_cpu},kind=openmp"
-      DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
+        "--image=file=${CMAKE_CURRENT_BINARY_DIR}/extracted_${bclib_name},triple=${target_triple},arch=${target_cpu},kind=openmp"
+      DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/extracted_${bclib_name}
       COMMENT "Packaging LLVM offloading binary ${bclib_name}.out"
   )
 
@@ -232,6 +242,11 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
       DEPENDS opt
       APPEND)
   endif()
+  if("${EXTRACT_TOOL}" STREQUAL "$<TARGET_FILE:llvm-extract>")
+    add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/extracted_${bclib_name}
+      DEPENDS opt
+      APPEND)
+  endif()
   if("${PACKAGER_TOOL}" STREQUAL "$<TARGET_FILE:clang-offload-packager>")
     add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
       DEPENDS clang-offload-packager

openmp/libomptarget/DeviceRTL/include/Types.h

@@ -213,13 +213,6 @@ enum OMPTgtExecModeFlags : int8_t {
 #define CONSTANT(NAME)                                                         \
   [[clang::address_space(4)]] NAME [[clang::loader_uninitialized]]
 
-// Attribute to keep alive certain definition for the bitcode library.
-#ifdef LIBOMPTARGET_BC_TARGET
-#define KEEP_ALIVE __attribute__((used, retain))
-#else
-#define KEEP_ALIVE
-#endif
-
 ///}
 
 #endif

openmp/libomptarget/DeviceRTL/src/Mapping.cpp

@@ -276,7 +276,7 @@ uint32_t mapping::getNumberOfProcessorElements() {
 
 // TODO: This is a workaround for initialization coming from kernels outside of
 //       the TU. We will need to solve this more correctly in the future.
-int __attribute__((weak)) KEEP_ALIVE SHARED(IsSPMDMode);
+int __attribute__((weak)) SHARED(IsSPMDMode);
 
 void mapping::init(bool IsSPMD) {
   if (mapping::isInitialThreadInLevel0(IsSPMD))

openmp/libomptarget/DeviceRTL/src/Utils.cpp

@@ -19,16 +19,17 @@
 
 using namespace _OMP;
 
-namespace _OMP {
+extern "C" __attribute__((weak)) int IsSPMDMode;
+
 /// Helper to keep code alive without introducing a performance penalty.
-__attribute__((weak, optnone, cold)) KEEP_ALIVE void keepAlive() {
+extern "C" __attribute__((weak, optnone, cold, used, retain)) void
+__keep_alive() {
   __kmpc_get_hardware_thread_id_in_block();
   __kmpc_get_hardware_num_threads_in_block();
   __kmpc_get_warp_size();
-  __kmpc_barrier_simple_spmd(nullptr, 0);
-  __kmpc_barrier_simple_generic(nullptr, 0);
+  __kmpc_barrier_simple_spmd(nullptr, IsSPMDMode);
+  __kmpc_barrier_simple_generic(nullptr, IsSPMDMode);
 }
-} // namespace _OMP
 
 namespace impl {
 

openmp/libomptarget/DeviceRTL/src/exports

@@ -2,6 +2,9 @@ omp_*
 *llvm_*
 __kmpc_*
 
+__keep_alive
+IsSPMDMode
+
 memcmp
 printf
 __assert_fail