[libc] Rework the GPU build to be a regular target (#81921)

Summary:
This is a massive patch because it reworks the entire build and
everything that depends on it. This is not split up because various bots
would fail otherwise. I will attempt to describe the necessary changes
here.

This patch completely reworks how the GPU build is built and targeted.
Previously, we used a standard runtimes build and handled both NVPTX and
AMDGPU in a single build via multi-targeting. This added a lot of
divergence in the build system and prevented us from doing various
things like building for the CPU / GPU at the same time, or exporting
the startup libraries or running tests without a full rebuild.

The new appraoch is to handle the GPU builds as strict cross-compiling
runtimes. The first step required
#81557 to allow the `LIBC`
target to build for the GPU without touching the other targets. This
means that the GPU uses all the same handling as the other builds in
`libc`.

The new expected way to build the GPU libc is with
`LLVM_LIBC_RUNTIME_TARGETS=amdgcn-amd-amdhsa;nvptx64-nvidia-cuda`.

The second step was reworking how we generated the embedded GPU library
by moving it into the library install step. Where we previously had one
`libcgpu.a` we now have `libcgpu-amdgpu.a` and `libcgpu-nvptx.a`. This
patch includes the necessary clang / OpenMP changes to make that not
break the bots when this lands.

We unfortunately still require that the NVPTX target has an `internal`
target for tests. This is because the NVPTX target needs to do LTO for
the provided version (The offloading toolchain can handle it) but cannot
use it for the native toolchain which is used for making tests.

This approach is vastly superior in every way, allowing us to treat the
GPU as a standard cross-compiling target. We can now install the GPU
utilities to do things like use the offload tests and other fun things.

Some certain utilities need to be built with 
`--target=${LLVM_HOST_TRIPLE}` as well. I think this is a fine
workaround as we
will always assume that the GPU `libc` is a cross-build with a
functioning host.

Depends on #81557

clang/lib/Driver/ToolChains/CommonArgs.cpp

@@ -1087,10 +1087,41 @@ static void addOpenMPDeviceLibC(const ToolChain &TC, const ArgList &Args,
                           "llvm-libc-decls");
   bool HasLibC = llvm::sys::fs::exists(LibCDecls) &&
                  llvm::sys::fs::is_directory(LibCDecls);
-  if (Args.hasFlag(options::OPT_gpulibc, options::OPT_nogpulibc, HasLibC)) {
-    CmdArgs.push_back("-lcgpu");
-    CmdArgs.push_back("-lmgpu");
+  if (!Args.hasFlag(options::OPT_gpulibc, options::OPT_nogpulibc, HasLibC))
+    return;
+
+  // We don't have access to the offloading toolchains here, so determine from
+  // the arguments if we have any active NVPTX or AMDGPU toolchains.
+  llvm::DenseSet<const char *> Libraries;
+  if (const Arg *Targets = Args.getLastArg(options::OPT_fopenmp_targets_EQ)) {
+    if (llvm::any_of(Targets->getValues(),
+                     [](auto S) { return llvm::Triple(S).isAMDGPU(); })) {
+      Libraries.insert("-lcgpu-amdgpu");
+      Libraries.insert("-lmgpu-amdgpu");
+    }
+    if (llvm::any_of(Targets->getValues(),
+                     [](auto S) { return llvm::Triple(S).isNVPTX(); })) {
+      Libraries.insert("-lcgpu-nvptx");
+      Libraries.insert("-lmgpu-nvptx");
+    }
   }
+
+  for (StringRef Arch : Args.getAllArgValues(options::OPT_offload_arch_EQ)) {
+    if (llvm::any_of(llvm::split(Arch, ","), [](StringRef Str) {
+          return IsAMDGpuArch(StringToCudaArch(Str));
+        })) {
+      Libraries.insert("-lcgpu-amdgpu");
+      Libraries.insert("-lmgpu-amdgpu");
+    }
+    if (llvm::any_of(llvm::split(Arch, ","), [](StringRef Str) {
+          return IsNVIDIAGpuArch(StringToCudaArch(Str));
+        })) {
+      Libraries.insert("-lcgpu-nvptx");
+      Libraries.insert("-lmgpu-nvptx");
+    }
+  }
+
+  llvm::append_range(CmdArgs, Libraries);
 }
 
 void tools::addOpenMPRuntimeLibraryPath(const ToolChain &TC,

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

@@ -393,14 +393,28 @@
 //
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp \
 // RUN:      --libomptarget-nvptx-bc-path=%S/Inputs/libomptarget/libomptarget-nvptx-test.bc \
+// RUN:      --libomptarget-amdgpu-bc-path=%S/Inputs/hip_dev_lib/libomptarget-amdgpu-gfx803.bc \
 // RUN:      --cuda-path=%S/Inputs/CUDA_102/usr/local/cuda \
-// RUN:      --offload-arch=sm_52 -gpulibc -nogpuinc %s 2>&1 \
+// RUN:      --rocm-path=%S/Inputs/rocm \
+// RUN:      --offload-arch=sm_52,gfx803 -gpulibc -nogpuinc %s 2>&1 \
 // RUN:   | FileCheck --check-prefix=LIBC-GPU %s
-// LIBC-GPU: "-lcgpu"{{.*}}"-lmgpu"
+// RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp \
+// RUN:      --libomptarget-nvptx-bc-path=%S/Inputs/libomptarget/libomptarget-nvptx-test.bc \
+// RUN:      --libomptarget-amdgpu-bc-path=%S/Inputs/hip_dev_lib/libomptarget-amdgpu-gfx803.bc \
+// RUN:      --cuda-path=%S/Inputs/CUDA_102/usr/local/cuda \
+// RUN:      --rocm-path=%S/Inputs/rocm \
+// RUN:      -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_52 \
+// RUN:      -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx803 \
+// RUN:      -fopenmp-targets=nvptx64-nvidia-cuda,amdgcn-amd-amdhsa -gpulibc -nogpuinc %s 2>&1 \
+// RUN:   | FileCheck --check-prefix=LIBC-GPU %s
+// LIBC-GPU-DAG: "-lcgpu-amdgpu"
+// LIBC-GPU-DAG: "-lmgpu-amdgpu"
+// LIBC-GPU-DAG: "-lcgpu-nvptx"
+// LIBC-GPU-DAG: "-lmgpu-nvptx"
 
 // RUN:   %clang -### --target=x86_64-unknown-linux-gnu -fopenmp=libomp \
 // RUN:      --libomptarget-nvptx-bc-path=%S/Inputs/libomptarget/libomptarget-nvptx-test.bc \
 // RUN:      --cuda-path=%S/Inputs/CUDA_102/usr/local/cuda \
 // RUN:      --offload-arch=sm_52 -nogpulibc -nogpuinc %s 2>&1 \
 // RUN:   | FileCheck --check-prefix=NO-LIBC-GPU %s
-// NO-LIBC-GPU-NOT: "-lcgpu"{{.*}}"-lmgpu"
+// NO-LIBC-GPU-NOT: -lmgpu{{.*}}-lcgpu

libc/CMakeLists.txt

@@ -43,7 +43,7 @@ set(LIBC_NAMESPACE "__llvm_libc_${LLVM_VERSION_MAJOR}_${LLVM_VERSION_MINOR}_${LL
   CACHE STRING "The namespace to use to enclose internal implementations. Must start with '__llvm_libc'."
 )
 
-if(LLVM_LIBC_FULL_BUILD OR LIBC_GPU_BUILD OR LIBC_GPU_ARCHITECTURES)
+if(LLVM_LIBC_FULL_BUILD OR LLVM_LIBC_GPU_BUILD)
   if(NOT LIBC_HDRGEN_EXE)
     # We need to set up hdrgen first since other targets depend on it.
     add_subdirectory(utils/LibcTableGenUtil)
@@ -77,7 +77,7 @@ if(LIBC_HDRGEN_ONLY OR NEED_LIBC_HDRGEN)
   # to build libc-hdrgen and return.
 
   # Always make the RPC server availible to other projects for GPU mode.
-  if(LIBC_GPU_BUILD OR LIBC_GPU_ARCHITECTURES)
+  if(LLVM_LIBC_GPU_BUILD)
     add_subdirectory(utils/gpu/server)
   endif()
   return()
@@ -118,7 +118,7 @@ if(COMMAND_RETURN_CODE EQUAL 0)
   message(STATUS "Set COMPILER_RESOURCE_DIR to "
                  "${COMPILER_RESOURCE_DIR} using --print-resource-dir")
 else()
-  if (LIBC_TARGET_ARCHITECTURE_IS_GPU)
+  if (LIBC_TARGET_OS_IS_GPU)
     message(FATAL_ERROR "COMPILER_RESOURCE_DIR must be set for GPU builds")
   else()
     set(COMPILER_RESOURCE_DIR OFF)
@@ -216,11 +216,7 @@ foreach(config_path IN LISTS LIBC_CONFIG_JSON_FILE_LIST)
   load_libc_config(${config_path}/config.json ${cmd_line_conf})
 endforeach()
 
-if(LIBC_TARGET_ARCHITECTURE_IS_GPU)
-  set(LIBC_INCLUDE_DIR ${CMAKE_CURRENT_BINARY_DIR}/include)
-  set(LIBC_INSTALL_INCLUDE_DIR ${CMAKE_INSTALL_INCLUDEDIR}/gpu-none-llvm)
-  set(LIBC_LIBRARY_DIR ${LLVM_LIBRARY_OUTPUT_INTDIR})
-elseif(LLVM_ENABLE_PER_TARGET_RUNTIME_DIR AND LIBC_ENABLE_USE_BY_CLANG)
+if(LLVM_ENABLE_PER_TARGET_RUNTIME_DIR AND LIBC_ENABLE_USE_BY_CLANG)
   set(LIBC_INCLUDE_DIR ${LLVM_BINARY_DIR}/include/${LLVM_DEFAULT_TARGET_TRIPLE})
   set(LIBC_INSTALL_INCLUDE_DIR ${CMAKE_INSTALL_INCLUDEDIR}/${LLVM_DEFAULT_TARGET_TRIPLE})
   set(LIBC_LIBRARY_DIR ${LLVM_LIBRARY_OUTPUT_INTDIR}/${LLVM_DEFAULT_TARGET_TRIPLE})
@@ -235,7 +231,11 @@ else()
     set(LIBC_INCLUDE_DIR ${CMAKE_BINARY_DIR}/include)
     set(LIBC_LIBRARY_DIR ${CMAKE_BINARY_DIR}/lib${LLVM_LIBDIR_SUFFIX})
   endif()
-  set(LIBC_INSTALL_INCLUDE_DIR ${CMAKE_INSTALL_INCLUDEDIR})
+  if(LIBC_TARGET_OS_IS_GPU)
+    set(LIBC_INSTALL_INCLUDE_DIR ${CMAKE_INSTALL_INCLUDEDIR}/${LLVM_DEFAULT_TARGET_TRIPLE})
+  else()
+    set(LIBC_INSTALL_INCLUDE_DIR ${CMAKE_INSTALL_INCLUDEDIR})
+  endif()
 endif()
 
 if(LIBC_TARGET_TRIPLE)
@@ -247,7 +247,7 @@ else()
   set(LIBC_INSTALL_LIBRARY_DIR lib${LLVM_LIBDIR_SUFFIX})
 endif()
 
-if(LIBC_TARGET_ARCHITECTURE_IS_GPU)
+if(LIBC_TARGET_OS_IS_GPU)
   include(prepare_libc_gpu_build)
   set(LIBC_ENABLE_UNITTESTS OFF)
 endif()

libc/cmake/modules/LLVMLibCArchitectures.cmake

@@ -6,18 +6,6 @@
 # platform.
 # ------------------------------------------------------------------------------
 
-if(LIBC_GPU_BUILD OR LIBC_GPU_ARCHITECTURES)
-  # We set the generic target and OS to "gpu" here. More specific defintions
-  # for the exact target GPU are set up in prepare_libc_gpu_build.cmake.
-  set(LIBC_TARGET_OS "gpu")
-  set(LIBC_TARGET_ARCHITECTURE_IS_GPU TRUE)
-  set(LIBC_TARGET_ARCHITECTURE "gpu")
-  if(LIBC_TARGET_TRIPLE)
-    message(WARNING "LIBC_TARGET_TRIPLE is ignored as LIBC_GPU_BUILD is on. ")
-  endif()
-  return()
-endif()
-
 if(MSVC)
   # If the compiler is visual c++ or equivalent, we will assume a host build.
   set(LIBC_TARGET_OS ${CMAKE_HOST_SYSTEM_NAME})
@@ -59,6 +47,10 @@ function(get_arch_and_system_from_triple triple arch_var sys_var)
     set(target_arch "riscv32")
   elseif(target_arch MATCHES "^riscv64")
     set(target_arch "riscv64")
+  elseif(target_arch MATCHES "^amdgcn")
+    set(target_arch "amdgpu")
+  elseif(target_arch MATCHES "^nvptx64")
+    set(target_arch "nvptx")
   else()
     return()
   endif()
@@ -75,6 +67,12 @@ function(get_arch_and_system_from_triple triple arch_var sys_var)
     set(target_sys "darwin")
   endif()
 
+  # Setting OS name for GPU architectures.
+  list(GET triple_comps -1 gpu_target_sys)
+  if(gpu_target_sys MATCHES "^amdhsa" OR gpu_target_sys MATCHES "^cuda")
+    set(target_sys "gpu")
+  endif()
+
   set(${sys_var} ${target_sys} PARENT_SCOPE)
 endfunction(get_arch_and_system_from_triple)
 
@@ -156,6 +154,10 @@ elseif(LIBC_TARGET_ARCHITECTURE STREQUAL "riscv64")
 elseif(LIBC_TARGET_ARCHITECTURE STREQUAL "riscv32")
   set(LIBC_TARGET_ARCHITECTURE_IS_RISCV32 TRUE)
   set(LIBC_TARGET_ARCHITECTURE "riscv")
+elseif(LIBC_TARGET_ARCHITECTURE STREQUAL "amdgpu")
+  set(LIBC_TARGET_ARCHITECTURE_IS_AMDGPU TRUE)
+elseif(LIBC_TARGET_ARCHITECTURE STREQUAL "nvptx")
+  set(LIBC_TARGET_ARCHITECTURE_IS_NVPTX TRUE)
 else()
   message(FATAL_ERROR
           "Unsupported libc target architecture ${LIBC_TARGET_ARCHITECTURE}")
@@ -178,6 +180,8 @@ elseif(LIBC_TARGET_OS STREQUAL "darwin")
   set(LIBC_TARGET_OS_IS_DARWIN TRUE)
 elseif(LIBC_TARGET_OS STREQUAL "windows")
   set(LIBC_TARGET_OS_IS_WINDOWS TRUE)
+elseif(LIBC_TARGET_OS STREQUAL "gpu")
+  set(LIBC_TARGET_OS_IS_GPU TRUE)
 else()
   message(FATAL_ERROR
           "Unsupported libc target operating system ${LIBC_TARGET_OS}")

libc/cmake/modules/LLVMLibCCheckMPFR.cmake

@@ -2,7 +2,7 @@ set(LLVM_LIBC_MPFR_INSTALL_PATH "" CACHE PATH "Path to where MPFR is installed (
 
 if(LLVM_LIBC_MPFR_INSTALL_PATH)
   set(LIBC_TESTS_CAN_USE_MPFR TRUE)
-elseif(LIBC_TARGET_ARCHITECTURE_IS_GPU)
+elseif(LIBC_TARGET_OS_IS_GPU)
   set(LIBC_TESTS_CAN_USE_MPFR FALSE)
 else()
   try_compile(

libc/cmake/modules/LLVMLibCCompileOptionRules.cmake

@@ -82,10 +82,22 @@ function(_get_common_compile_options output_var flags)
     list(APPEND compile_options "/EHs-c-")
     list(APPEND compile_options "/GR-")
   endif()
-  if (LIBC_TARGET_ARCHITECTURE_IS_GPU)
+  if (LIBC_TARGET_OS_IS_GPU)
     list(APPEND compile_options "-nogpulib")
     list(APPEND compile_options "-fvisibility=hidden")
     list(APPEND compile_options "-fconvergent-functions")
+    list(APPEND compile_options "-flto")
+
+    if(LIBC_TARGET_ARCHITECTURE_IS_NVPTX)
+      list(APPEND compile_options "-Wno-unknown-cuda-version")
+      list(APPEND compile_options "SHELL:-mllvm -nvptx-emit-init-fini-kernel=false")
+      list(APPEND compile_options "--cuda-feature=+ptx63")
+      if(LIBC_CUDA_ROOT)
+        list(APPEND compile_options "--cuda-path=${LIBC_CUDA_ROOT}")
+      endif()
+    elseif(LIBC_TARGET_ARCHITECTURE_IS_AMDGPU)
+      list(APPEND compile_options "SHELL:-Xclang -mcode-object-version=none")
+    endif()
 
     # Manually disable all standard include paths and include the resource
     # directory to prevent system headers from being included.
@@ -138,73 +150,21 @@ function(_get_common_test_compile_options output_var flags)
   set(${output_var} ${compile_options} PARENT_SCOPE)
 endfunction()
 
-# Obtains NVPTX specific arguments for compilation.
-# The PTX feature is primarily based on the CUDA toolchain version. We want to
-# be able to target NVPTX without an existing CUDA installation, so we need to
-# set this manually. This simply sets the PTX feature to the minimum required
-# for the features we wish to use on that target. The minimum PTX features used
-# here roughly corresponds to the CUDA 9.0 release.
-# Adjust as needed for desired PTX features.
-function(get_nvptx_compile_options output_var gpu_arch)
-  set(nvptx_options "")
-  list(APPEND nvptx_options "-march=${gpu_arch}")
-  list(APPEND nvptx_options "-Wno-unknown-cuda-version")
-  list(APPEND nvptx_options "SHELL:-mllvm -nvptx-emit-init-fini-kernel=false")
-  if(${gpu_arch} STREQUAL "sm_35")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_37")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_50")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_52")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_53")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_60")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_61")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_62")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_70")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_72")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_75")
-    list(APPEND nvptx_options "--cuda-feature=+ptx63")
-  elseif(${gpu_arch} STREQUAL "sm_80")
-    list(APPEND nvptx_options "--cuda-feature=+ptx72")
-  elseif(${gpu_arch} STREQUAL "sm_86")
-    list(APPEND nvptx_options "--cuda-feature=+ptx72")
-  elseif(${gpu_arch} STREQUAL "sm_89")
-    list(APPEND nvptx_options "--cuda-feature=+ptx72")
-  elseif(${gpu_arch} STREQUAL "sm_90")
-    list(APPEND nvptx_options "--cuda-feature=+ptx72")
-  else()
-    message(FATAL_ERROR "Unknown Nvidia GPU architecture '${gpu_arch}'")
-  endif()
-
-  if(LIBC_CUDA_ROOT)
-    list(APPEND nvptx_options "--cuda-path=${LIBC_CUDA_ROOT}")
-  endif()
-  set(${output_var} ${nvptx_options} PARENT_SCOPE)
-endfunction()
-
 function(_get_hermetic_test_compile_options output_var flags)
   _get_compile_options_from_flags(compile_flags ${flags})
   list(APPEND compile_options ${LIBC_COMPILE_OPTIONS_DEFAULT} ${compile_flags}
        ${flags} -fpie -ffreestanding -fno-exceptions -fno-rtti)
 
   # The GPU build requires overriding the default CMake triple and architecture.
-  if(LIBC_GPU_TARGET_ARCHITECTURE_IS_AMDGPU)
+  if(LIBC_TARGET_ARCHITECTURE_IS_AMDGPU)
     list(APPEND compile_options
          -nogpulib -mcpu=${LIBC_GPU_TARGET_ARCHITECTURE} -flto
-         --target=${LIBC_GPU_TARGET_TRIPLE}
          -mcode-object-version=${LIBC_GPU_CODE_OBJECT_VERSION})
-  elseif(LIBC_GPU_TARGET_ARCHITECTURE_IS_NVPTX)
-    get_nvptx_compile_options(nvptx_options ${LIBC_GPU_TARGET_ARCHITECTURE})
+  elseif(LIBC_TARGET_ARCHITECTURE_IS_NVPTX)
     list(APPEND compile_options
-         -nogpulib ${nvptx_options} -fno-use-cxa-atexit --target=${LIBC_GPU_TARGET_TRIPLE})
+         "SHELL:-mllvm -nvptx-emit-init-fini-kernel=false"
+         --cuda-path=${LIBC_CUDA_ROOT}
+         -nogpulib -march=${LIBC_GPU_TARGET_ARCHITECTURE} -fno-use-cxa-atexit)
   endif()
   set(${output_var} ${compile_options} PARENT_SCOPE)
 endfunction()

libc/cmake/modules/LLVMLibCHeaderRules.cmake

@@ -139,7 +139,7 @@ function(add_gen_header target_name)
             ${hdrgen_deps}
   )
 
-  if(LIBC_TARGET_ARCHITECTURE_IS_GPU)
+  if(LIBC_TARGET_OS_IS_GPU)
     file(MAKE_DIRECTORY ${LIBC_INCLUDE_DIR}/llvm-libc-decls)
     set(decl_out_file ${LIBC_INCLUDE_DIR}/llvm-libc-decls/${relative_path})
     add_custom_command(