[OpenMP] Remove 'libomptarget.devicertl.a' fatbinary and use static library (llvm#126143)

Summary:
Currently, we build a single `libomptarget.devicertl.a` which is a
fatbinary. It is a host object file that contains the embedded archive
files for both the NVIDIA and AMDGPU targets. This was done primarily as
a convenience due to naming conflicts. Now that the clang driver for the
GPU targets can appropriate link via the per-target runtime-dir, we can
just make two separate static libraries and remove the indirection.

This patch creates two new static libraries that get installed into
```
/lib/amdgcn-amd-amdhsa/libomp.a
/lib/nvptx64-nvidia-cuda/libomp.a
```
for AMDGPU and NVPTX respectively. The link job created by the linker
wrapper now simply needs to do `-lomp` and it will search those
directories and link those static libraries. This requires far less
special handling.

This patch is a precursor to changing the build system entirely to be a
runtimes based one. Soon this target will be a standard `add_library`
and done through the GPU runtime targets.

NOTE that this actually does remove an additional optimization step.
Previously we merged all of the files into a single bitcode object and
forcibly internalized some definitions. This, instead, just treats them
like a normal static library. This may possibly affect performance for
some files, but I think it's better overall to use static library
semantics because it allows us to have an 'include-what-you-use'
relationship with the library.

Performance testing will be required. If we really need the merged blob
then we can simply pack that into a new static library.

Author: jhuber6

clang/lib/Driver/ToolChains/Clang.cpp

@@ -9267,6 +9267,11 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
           A->render(Args, LinkerArgs);
       }
 
+      // If this is OpenMP the device linker will need `-lompdevice`.
+      if (Kind == Action::OFK_OpenMP && !Args.hasArg(OPT_no_offloadlib) &&
+          (TC->getTriple().isAMDGPU() || TC->getTriple().isNVPTX()))
+        LinkerArgs.emplace_back("-lompdevice");
+
       // Forward all of these to the appropriate toolchain.
       for (StringRef Arg : CompilerArgs)
         CmdArgs.push_back(Args.MakeArgString(

clang/lib/Driver/ToolChains/CommonArgs.cpp

@@ -1309,10 +1309,6 @@ bool tools::addOpenMPRuntime(const Compilation &C, ArgStringList &CmdArgs,
   if (IsOffloadingHost)
     CmdArgs.push_back("-lomptarget");
 
-  if (IsOffloadingHost &&
-      Args.hasFlag(options::OPT_offloadlib, options::OPT_no_offloadlib, true))
-    CmdArgs.push_back("-lomptarget.devicertl");
-
   addArchSpecificRPath(TC, Args, CmdArgs);
 
   addOpenMPRuntimeLibraryPath(TC, Args, CmdArgs);

clang/test/Driver/openmp-offload-gpu.c

@@ -324,18 +324,18 @@
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp --offload-arch=sm_52 --cuda-path=%S/Inputs/CUDA_102/usr/local/cuda \
 // RUN:     -foffload-lto %s 2>&1 | FileCheck --check-prefix=CHECK-LTO-LIBRARY %s
 
-// CHECK-LTO-LIBRARY: {{.*}}-lomptarget{{.*}}-lomptarget.devicertl
+// CHECK-LTO-LIBRARY: --device-linker={{.*}}=-lomp{{.*}}-lomptarget{{.*}}
 
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp --offload-arch=sm_52 --cuda-path=%S/Inputs/CUDA_102/usr/local/cuda \
 // RUN:     --libomptarget-nvptx-bc-path=%S/Inputs/libomptarget/libomptarget-nvptx-test.bc %s 2>&1 \
 // RUN:   | FileCheck --check-prefix=CHECK-NO-LTO-LIBRARY %s
 
-// CHECK-NO-LTO-LIBRARY: {{.*}}-lomptarget{{.*}}-lomptarget.devicertl
+// CHECK-NO-LTO-LIBRARY: --device-linker={{.*}}=-lomp{{.*}}-lomptarget{{.*}}
 
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp --offload-arch=sm_52 -nogpulib \
 // RUN:     -foffload-lto %s 2>&1 | FileCheck --check-prefix=CHECK-NO-LIBRARY %s
 
-// CHECK-NO-LIBRARY-NOT: {{.*}}-lomptarget{{.*}}-lomptarget.devicertl
+// CHECK-NO-LIBRARY-NOT: --device-linker={{.*}}=-lomp{{.*}}-lomptarget{{.*}}
 
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp --offload-arch=sm_52 -nogpulib \
 // RUN:     -Xoffload-linker a -Xoffload-linker-nvptx64-nvidia-cuda b -Xoffload-linker-nvptx64 c \

offload/DeviceRTL/CMakeLists.txt

@@ -16,23 +16,10 @@ if (LLVM_DIR)
   # Builds that use pre-installed LLVM have LLVM_DIR set.
   # A standalone or LLVM_ENABLE_RUNTIMES=openmp build takes this route
   find_program(CLANG_TOOL clang PATHS ${LLVM_TOOLS_BINARY_DIR} NO_DEFAULT_PATH)
-  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) OR (NOT PACKAGER_TOOL))
-    message(STATUS "Not building DeviceRTL. Missing clang: ${CLANG_TOOL}, llvm-link: ${LINK_TOOL}, opt: ${OPT_TOOL}, or clang-offload-packager: ${PACKAGER_TOOL}")
-    return()
-  else()
-    message(STATUS "Building DeviceRTL. Using clang: ${CLANG_TOOL}, llvm-link: ${LINK_TOOL} and opt: ${OPT_TOOL}")
-  endif()
 elseif (LLVM_TOOL_CLANG_BUILD AND NOT CMAKE_CROSSCOMPILING AND NOT OPENMP_STANDALONE_BUILD)
   # LLVM in-tree builds may use CMake target names to discover the tools.
   # A LLVM_ENABLE_PROJECTS=openmp build takes this route
   set(CLANG_TOOL $<TARGET_FILE:clang>)
-  set(PACKAGER_TOOL $<TARGET_FILE:clang-offload-packager>)
-  set(LINK_TOOL $<TARGET_FILE:llvm-link>)
-  set(OPT_TOOL $<TARGET_FILE:opt>)
-  message(STATUS "Building DeviceRTL. Using clang from in-tree build")
 else()
   message(STATUS "Not building DeviceRTL. No appropriate clang found")
   return()
@@ -82,8 +69,6 @@ set(src_files
 # propagation. That said, we will run the vectorizer again after the runtime
 # has been linked into the user program.
 set(clang_opt_flags -O3 -mllvm -openmp-opt-disable -DSHARED_SCRATCHPAD_SIZE=512 -mllvm -vectorize-slp=false )
-set(link_opt_flags  -O3        -openmp-opt-disable -attributor-enable=module -vectorize-slp=false )
-set(link_export_flag -passes=internalize -internalize-public-api-file=${source_directory}/exports)
 
 # If the user built with the GPU C library enabled we will use that instead.
 if(${LIBOMPTARGET_GPU_LIBC_SUPPORT})
@@ -107,15 +92,13 @@ set(bc_flags -c -flto -std=c++17 -fvisibility=hidden
 )
 
 # first create an object target
-add_library(omptarget.devicertl.all_objs OBJECT IMPORTED)
 function(compileDeviceRTLLibrary target_name target_triple)
   set(target_bc_flags ${ARGN})
 
-  set(bc_files "")
   foreach(src ${src_files})
     get_filename_component(infile ${src} ABSOLUTE)
     get_filename_component(outfile ${src} NAME)
-    set(outfile "${outfile}-${target_name}.bc")
+    set(outfile "${outfile}-${target_name}.o")
     set(depfile "${outfile}.d")
 
     # Passing an empty CPU to -march= suppressed target specific metadata.
@@ -142,99 +125,40 @@ function(compileDeviceRTLLibrary target_name target_triple)
     endif()
     set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${outfile})
 
-    list(APPEND bc_files ${outfile})
+    list(APPEND obj_files ${CMAKE_CURRENT_BINARY_DIR}/${outfile})
   endforeach()
-
-  set(bclib_name "libomptarget-${target_name}.bc")
-
-  # Link to a bitcode library.
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/linked_${bclib_name}
-      COMMAND ${LINK_TOOL}
-        -o ${CMAKE_CURRENT_BINARY_DIR}/linked_${bclib_name} ${bc_files}
-      DEPENDS ${bc_files}
-      COMMENT "Linking LLVM bitcode ${bclib_name}"
-  )
-
-  if(TARGET llvm-link)
-    add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/linked_${bclib_name}
-      DEPENDS llvm-link
-      APPEND)
-  endif()
-
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/internalized_${bclib_name}
-      COMMAND ${OPT_TOOL} ${link_export_flag} ${CMAKE_CURRENT_BINARY_DIR}/linked_${bclib_name}
-                      -o ${CMAKE_CURRENT_BINARY_DIR}/internalized_${bclib_name}
-      DEPENDS ${source_directory}/exports ${CMAKE_CURRENT_BINARY_DIR}/linked_${bclib_name}
-      COMMENT "Internalizing LLVM bitcode ${bclib_name}"
-  )
-  if(TARGET opt)
-    add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/internalized_${bclib_name}
-      DEPENDS opt
-      APPEND)
-  endif()
-
-  add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
-      COMMAND ${OPT_TOOL} ${link_opt_flags} ${CMAKE_CURRENT_BINARY_DIR}/internalized_${bclib_name}
-                      -o ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
-      DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/internalized_${bclib_name}
-      COMMENT "Optimizing LLVM bitcode ${bclib_name}"
-  )
-  if(TARGET opt)
-    add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
-      DEPENDS opt
-      APPEND)
-  endif()
-
-  set(bclib_target_name "omptarget-${target_name}-bc")
-  add_custom_target(${bclib_target_name} DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name})
-
-  # Copy library to destination.
-  add_custom_command(TARGET ${bclib_target_name} POST_BUILD
-                    COMMAND ${CMAKE_COMMAND} -E copy ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
-                    ${LIBOMPTARGET_LIBRARY_DIR})
-  add_dependencies(omptarget.devicertl.${target_name} ${bclib_target_name})
-
-  set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${bclib_name} ${LIBOMPTARGET_LIBRARY_DIR}/${bclib_name})
-
-  # Install bitcode library under the lib destination folder.
-  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name} DESTINATION "${OFFLOAD_INSTALL_LIBDIR}")
-
-  set(target_feature "")
-  if("${target_triple}" STREQUAL "nvptx64-nvidia-cuda")
-    set(target_feature "feature=+ptx63")
-  endif()
-
-  # 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},${target_feature},triple=${target_triple},arch=generic,kind=openmp"
-      DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
-      COMMENT "Packaging LLVM offloading binary ${bclib_name}.out"
-  )
-  if(TARGET clang-offload-packager)
-    add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
-      DEPENDS clang-offload-packager
-      APPEND)
-  endif()
-
-  set(output_name "${CMAKE_CURRENT_BINARY_DIR}/devicertl-${target_name}.o")
-  add_custom_command(OUTPUT ${output_name}
-    COMMAND ${CLANG_TOOL} --std=c++17 -c -nostdlib
-            -Xclang -fembed-offload-object=${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
-            -o ${output_name}
-            ${source_directory}/Stub.cpp
-    DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name} ${source_directory}/Stub.cpp
-    COMMENT "Embedding LLVM offloading binary in devicertl-${target_name}.o"
-    VERBATIM
+  # Trick to combine these into a bitcode file via the linker's LTO pass. This
+  # is used to provide the legacy `libomptarget-<name>.bc` files. Hack this
+  # through as an executable to get it to use the relocatable link.
+  add_executable(libomptarget-${target_name} ${obj_files})
+  set_target_properties(libomptarget-${target_name} PROPERTIES
+    RUNTIME_OUTPUT_DIRECTORY ${LIBOMPTARGET_LLVM_LIBRARY_INTDIR}
+    LINKER_LANGUAGE CXX
+    BUILD_RPATH ""
+    INSTALL_RPATH ""
+    RUNTIME_OUTPUT_NAME libomptarget-${target_name}.bc)
+  target_compile_options(libomptarget-${target_name} PRIVATE "--target=${target_triple}")
+  target_link_options(libomptarget-${target_name} PRIVATE "--target=${target_triple}"
+                      "-r" "-nostdlib" "-flto" "-Wl,--lto-emit-llvm")
+  install(TARGETS libomptarget-${target_name}
+          PERMISSIONS OWNER_WRITE OWNER_READ GROUP_READ WORLD_READ
+          DESTINATION ${OFFLOAD_INSTALL_LIBDIR})
+
+  add_library(omptarget.${target_name}.all_objs OBJECT IMPORTED)
+  set_property(TARGET omptarget.${target_name}.all_objs APPEND PROPERTY IMPORTED_OBJECTS
+               ${LIBOMPTARGET_LLVM_LIBRARY_INTDIR}/libomptarget-${target_name}.bc)
+
+  # Archive all the object files generated above into a static library
+  add_library(omptarget.${target_name} STATIC)
+  set_target_properties(omptarget.${target_name} PROPERTIES
+    ARCHIVE_OUTPUT_DIRECTORY "${LIBOMPTARGET_LLVM_LIBRARY_INTDIR}/${target_triple}"
+    ARCHIVE_OUTPUT_NAME ompdevice
+    LINKER_LANGUAGE CXX
   )
-  if(TARGET clang)
-    add_custom_command(OUTPUT ${output_name}
-      DEPENDS clang
-      APPEND)
-  endif()
+  target_link_libraries(omptarget.${target_name} PRIVATE omptarget.${target_name}.all_objs)
 
-  set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${output_name})
-  set_property(TARGET omptarget.devicertl.all_objs APPEND PROPERTY IMPORTED_OBJECTS ${output_name})
+  install(TARGETS omptarget.${target_name}
+          ARCHIVE DESTINATION "lib${LLVM_LIBDIR_SUFFIX}/${target_triple}")
 
   if (CMAKE_EXPORT_COMPILE_COMMANDS)
     set(ide_target_name omptarget-ide-${target_name})
@@ -254,18 +178,10 @@ function(compileDeviceRTLLibrary target_name target_triple)
   endif()
 endfunction()
 
-add_custom_target(omptarget.devicertl.amdgpu)
-compileDeviceRTLLibrary(amdgpu amdgcn-amd-amdhsa -Xclang -mcode-object-version=none)
-
-add_custom_target(omptarget.devicertl.nvptx)
-compileDeviceRTLLibrary(nvptx nvptx64-nvidia-cuda --cuda-feature=+ptx63)
-
-# Archive all the object files generated above into a static library
-add_library(omptarget.devicertl STATIC)
-set_target_properties(omptarget.devicertl PROPERTIES
-  ARCHIVE_OUTPUT_DIRECTORY "${LIBOMPTARGET_LLVM_LIBRARY_INTDIR}"
-  LINKER_LANGUAGE CXX
-)
-target_link_libraries(omptarget.devicertl PRIVATE omptarget.devicertl.all_objs)
+if(NOT LLVM_TARGETS_TO_BUILD OR "AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)
+  compileDeviceRTLLibrary(amdgpu amdgcn-amd-amdhsa -Xclang -mcode-object-version=none)
+endif()
 
-install(TARGETS omptarget.devicertl ARCHIVE DESTINATION ${OFFLOAD_INSTALL_LIBDIR})
+if(NOT LLVM_TARGETS_TO_BUILD OR "NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
+  compileDeviceRTLLibrary(nvptx nvptx64-nvidia-cuda --cuda-feature=+ptx63)
+endif()

offload/DeviceRTL/src/exports

@@ -164,7 +164,7 @@ else: # Unices
     if config.cuda_libdir:
         config.test_flags += " -Wl,-rpath," + config.cuda_libdir
     if config.libomptarget_current_target.startswith('nvptx'):
-        config.test_flags_clang += " --libomptarget-nvptx-bc-path=" + config.library_dir + '/DeviceRTL'
+        config.test_flags_clang += " --libomptarget-nvptx-bc-path=" + config.llvm_library_intdir
     if config.libomptarget_current_target.endswith('-LTO'):
         config.test_flags += " -foffload-lto"
     if config.libomptarget_current_target.endswith('-JIT-LTO') and evaluate_bool_env(
@@ -173,22 +173,6 @@ else: # Unices
         config.test_flags += " -foffload-lto"
         config.test_flags += " -Wl,--embed-bitcode"
 
-def remove_suffix_if_present(name):
-    if name.endswith('-LTO'):
-        return name[:-4]
-    elif name.endswith('-JIT-LTO'):
-        return name[:-8]
-    else:
-        return name
-
-def add_libraries(source):
-    if config.libomptarget_has_libc:
-        return source + " -Xoffload-linker " + "-lc " + \
-               "-Xoffload-linker " + "-lm " + \
-               config.llvm_library_intdir + "/libomptarget.devicertl.a"
-    else:
-        return source + " " + config.llvm_library_intdir + "/libomptarget.devicertl.a"
-
 # Add platform targets
 host_targets = [
     "aarch64-unknown-linux-gnu",
@@ -207,6 +191,24 @@ if config.libomptarget_current_target.startswith('amdgcn'):
 if config.libomptarget_current_target in host_targets:
     config.available_features.add('host')
 
+def remove_suffix_if_present(name):
+    if name.endswith('-LTO'):
+        return name[:-4]
+    elif name.endswith('-JIT-LTO'):
+        return name[:-8]
+    else:
+        return name
+
+def add_libraries(source):
+    if "gpu" not in config.available_features:
+        return source
+    if config.libomptarget_has_libc:
+        return source + " -Xoffload-linker -lc " + \
+               "-Xoffload-linker -lm " + \
+               "-Xoffload-linker -lompdevice"
+    else:
+        return source + " " + "-Xoffload-lnker -lompdevice"
+
 # substitutions
 # - for targets that exist in the system create the actual command.
 # - for valid targets that do not exist in the system, return false, so that the