[Hostjit] Move Clang/LLD hostjit into a separate library

c/parallel.v2/CMakeLists.txt

@@ -65,6 +65,13 @@ if (CCCL_C_PARALLEL_V2_LIBRARY_OUTPUT_DIRECTORY)
   )
 endif()
 
+if (UNIX AND NOT APPLE)
+  set_target_properties(
+    cccl.c.parallel.v2
+    PROPERTIES BUILD_RPATH_USE_ORIGIN ON INSTALL_RPATH "$ORIGIN"
+  )
+endif()
+
 cccl_get_cub()
 cccl_get_cudatoolkit()
 cccl_get_thrust()

c/parallel.v2/src/hostjit/CMakeLists.txt

@@ -1,77 +1,49 @@
 cmake_minimum_required(VERSION 3.30)
 
-# --------------------------------------------------------------------------
-# LLVM/Clang/LLD — fetched via CPM as static libraries
-# --------------------------------------------------------------------------
-# CPM.cmake is at the cccl repo root: cccl/cmake/CPM.cmake
-# From c/parallel.v2/src/hostjit/ that's ../../../../cmake/CPM.cmake
-set(_cccl_cmake_dir "${CMAKE_CURRENT_SOURCE_DIR}/../../../../cmake")
-if (EXISTS "${_cccl_cmake_dir}/CPM.cmake")
-  include("${_cccl_cmake_dir}/CPM.cmake")
-else()
-  message(FATAL_ERROR "CPM.cmake not found at ${_cccl_cmake_dir}/CPM.cmake")
+# Find CUDA toolkit (may already be found by parent)
+if (NOT CUDAToolkit_FOUND)
+  find_package(CUDAToolkit)
 endif()
 
-if (MSVC AND CMAKE_BUILD_TYPE STREQUAL "Debug")
-  message(
-    FATAL_ERROR
-    "hostjit does not support Debug builds on Windows. "
-    "The statically-linked LLVM Debug build is too large and causes stack "
-    "overflows at runtime. Use MinSizeRel, Release, or RelWithDebInfo instead."
+# CCCL_SOURCE_DIR points to the cccl repo root
+# From c/parallel.v2/src/hostjit -> c/parallel.v2/src -> c/parallel.v2 -> c -> cccl
+cmake_path(GET CMAKE_CURRENT_SOURCE_DIR PARENT_PATH _src_dir) # c/parallel.v2/src
+cmake_path(GET _src_dir PARENT_PATH _c_parallel_dir) # c/parallel.v2
+cmake_path(GET _c_parallel_dir PARENT_PATH _c_dir) # c
+cmake_path(GET _c_dir PARENT_PATH _cccl_root) # cccl
+set(_hostjit_include_dir "${CMAKE_CURRENT_SOURCE_DIR}/include")
+
+set(LIBNVCC_CPM_CMAKE_PATH "${_cccl_root}/cmake/CPM.cmake")
+set(LIBNVCC_HEADER_INSTALL_DESTINATION "cuda/cccl/headers/libnvcc")
+set(LIBNVCC_CLANG_HEADER_INSTALL_DESTINATION "cuda/cccl/headers/clang")
+if (CCCL_C_PARALLEL_V2_LIBRARY_OUTPUT_DIRECTORY)
+  set(
+    LIBNVCC_LIBRARY_OUTPUT_DIRECTORY
+    "${CCCL_C_PARALLEL_V2_LIBRARY_OUTPUT_DIRECTORY}"
   )
 endif()
 
-set(HOSTJIT_LLVM_VERSION "llvmorg-22.1.1" CACHE STRING "LLVM git tag to fetch")
-
-# List options must be set before CPMAddPackage
-set(LLVM_ENABLE_PROJECTS "clang;lld" CACHE STRING "" FORCE)
-set(LLVM_TARGETS_TO_BUILD "X86;NVPTX" CACHE STRING "" FORCE)
+add_subdirectory(libnvcc)
 
-CPMAddPackage(
-  NAME llvm_project
-  GIT_REPOSITORY https://github.com/llvm/llvm-project.git
-  GIT_TAG ${HOSTJIT_LLVM_VERSION}
-  GIT_SHALLOW ON
-  SOURCE_SUBDIR llvm
-  EXCLUDE_FROM_ALL YES
-  OPTIONS
-    "LLVM_BUILD_LLVM_C_DYLIB OFF"
-    "LLVM_BUILD_TOOLS OFF"
-    "LLVM_BUILD_UTILS OFF"
-    "LLVM_BUILD_RUNTIME OFF"
-    "LLVM_BUILD_RUNTIMES OFF"
-    "LLVM_INCLUDE_BENCHMARKS OFF"
-    "LLVM_INCLUDE_DOCS OFF"
-    "LLVM_INCLUDE_EXAMPLES OFF"
-    "LLVM_INCLUDE_RUNTIMES OFF"
-    "LLVM_INCLUDE_TESTS OFF"
-    "LLVM_INCLUDE_TOOLS ON"
-    "LLVM_INCLUDE_UTILS OFF"
-    "LLVM_ENABLE_ZLIB OFF"
-    "LLVM_ENABLE_ZSTD OFF"
-    "LLVM_ENABLE_TERMINFO OFF"
-    "LLVM_ENABLE_BINDINGS OFF"
-    "CLANG_BUILD_TOOLS OFF"
-    "CLANG_ENABLE_ARCMT OFF"
-    "CLANG_ENABLE_STATIC_ANALYZER OFF"
+# Windows does not have an RPATH equivalent. Keep libnvcc.dll in the same
+# runtime directory as CCCL's tests/executables so the loader can find it at
+# process startup when no explicit library output directory was requested.
+if (
+  WIN32
+  AND NOT CCCL_C_PARALLEL_V2_LIBRARY_OUTPUT_DIRECTORY
+  AND DEFINED CCCL_EXECUTABLE_OUTPUT_DIR
 )
-
-# Ensure the clang resource directory exists
-file(
-  MAKE_DIRECTORY "${llvm_project_BINARY_DIR}/lib/clang/${LLVM_VERSION_MAJOR}"
-)
-
-# Find CUDA toolkit (may already be found by parent)
-if (NOT CUDAToolkit_FOUND)
-  find_package(CUDAToolkit)
+  set_target_properties(
+    libnvcc
+    PROPERTIES RUNTIME_OUTPUT_DIRECTORY "${CCCL_EXECUTABLE_OUTPUT_DIR}"
+  )
 endif()
 
 # --------------------------------------------------------------------------
 # hostjit library
 # --------------------------------------------------------------------------
 add_library(
   cccl.c.parallel.v2.hostjit_lib
-  compiler.cpp
   config.cpp
   loader.cpp
   jit_compiler.cpp
@@ -82,129 +54,42 @@ add_library(
   codegen/cub_call.cpp
 )
 
-# CCCL_SOURCE_DIR points to the cccl repo root
-# From c/parallel.v2/src/hostjit -> c/parallel.v2/src -> c/parallel.v2 -> c -> cccl
-cmake_path(GET CMAKE_CURRENT_SOURCE_DIR PARENT_PATH _src_dir) # c/parallel.v2/src
-cmake_path(GET _src_dir PARENT_PATH _c_parallel_dir) # c/parallel.v2
-cmake_path(GET _c_parallel_dir PARENT_PATH _c_dir) # c
-cmake_path(GET _c_dir PARENT_PATH _cccl_root) # cccl
-
 target_include_directories(
   cccl.c.parallel.v2.hostjit_lib
-  PUBLIC
-    ${CMAKE_CURRENT_SOURCE_DIR}/include
-    ${_c_parallel_dir}/include
-    ${llvm_project_SOURCE_DIR}/llvm/include
-    ${llvm_project_BINARY_DIR}/include
-    ${llvm_project_SOURCE_DIR}/clang/include
-    ${llvm_project_BINARY_DIR}/tools/clang/include
-    ${llvm_project_SOURCE_DIR}/lld/include
-    ${llvm_project_BINARY_DIR}/tools/lld/include
+  PUBLIC ${_hostjit_include_dir} ${_c_parallel_dir}/include
 )
 
 target_compile_definitions(
   cccl.c.parallel.v2.hostjit_lib
   PRIVATE
     CCCL_C_EXPERIMENTAL=1
     CCCL_SOURCE_DIR="${_cccl_root}"
-    CLANG_RESOURCE_DIR="${llvm_project_BINARY_DIR}/lib/clang/${LLVM_VERSION_MAJOR}"
-    CLANG_HEADERS_DIR="${llvm_project_SOURCE_DIR}/clang/lib/Headers"
-    HOSTJIT_INCLUDE_DIR="${CMAKE_CURRENT_SOURCE_DIR}/include"
+    HOSTJIT_INCLUDE_DIR="${_hostjit_include_dir}"
 )
 
-if (CUDAToolkit_FOUND)
-  target_include_directories(
+if (DEFINED CLANG_HEADERS_DIR)
+  target_compile_definitions(
     cccl.c.parallel.v2.hostjit_lib
-    PUBLIC ${CUDAToolkit_INCLUDE_DIRS}
+    PRIVATE CLANG_HEADERS_DIR="${CLANG_HEADERS_DIR}"
   )
+endif()
+
+if (CUDAToolkit_FOUND)
   cmake_path(GET CUDAToolkit_BIN_DIR PARENT_PATH CUDA_TOOLKIT_ROOT_FROM_CMAKE)
   target_compile_definitions(
     cccl.c.parallel.v2.hostjit_lib
-    PRIVATE
-      CUDA_TOOLKIT_PATH="${CUDA_TOOLKIT_ROOT_FROM_CMAKE}"
-      CUDA_SDK_VERSION="${CUDAToolkit_VERSION_MAJOR}.0"
+    PRIVATE CUDA_TOOLKIT_PATH="${CUDA_TOOLKIT_ROOT_FROM_CMAKE}"
   )
 endif()
 
-# Link against LLVM/Clang/LLD
-target_link_libraries(
-  cccl.c.parallel.v2.hostjit_lib
-  PUBLIC
-    # LLVM
-    LLVMCore
-    LLVMSupport
-    LLVMIRReader
-    LLVMMC
-    LLVMObject
-    LLVMX86CodeGen
-    LLVMX86AsmParser
-    LLVMX86Desc
-    LLVMX86Info
-    LLVMNVPTXCodeGen
-    LLVMNVPTXDesc
-    LLVMNVPTXInfo
-    LLVMLinker
-    LLVMPasses
-    # Clang
-    clangAST
-    clangBasic
-    clangCodeGen
-    clangDriver
-    clangFrontend
-    clangFrontendTool
-    clangLex
-    clangParse
-    clangSema
-    clangEdit
-    clangAnalysis
-    clangRewrite
-    clangSerialization
-    # LLD
-    $<IF:$<PLATFORM_ID:Windows>,lldCOFF,lldELF>
-    lldCommon
-)
-
 if (NOT WIN32)
   target_link_libraries(cccl.c.parallel.v2.hostjit_lib PUBLIC dl)
 endif()
 
+target_link_libraries(cccl.c.parallel.v2.hostjit_lib PUBLIC libnvcc)
+
 if (CUDAToolkit_FOUND)
-  target_link_libraries(
-    cccl.c.parallel.v2.hostjit_lib
-    PUBLIC CUDA::cuda_driver CUDA::cudart
-  )
-  if (WIN32)
-    # On Windows, static CUDA libs are built with /MT which conflicts with
-    # the project's dynamic CRT (/MD). Use dynamic variants instead.
-    target_link_libraries(
-      cccl.c.parallel.v2.hostjit_lib
-      PUBLIC CUDA::nvJitLink CUDA::nvfatbin
-    )
-  else()
-    # Prefer static CUDA libs on Linux for self-contained binaries. If the
-    # toolchain (e.g. lite/pip CUDA installs or some Docker images) only ships
-    # the dynamic variants, fall back to those rather than failing configure.
-    foreach (_cudalib nvJitLink nvptxcompiler nvfatbin)
-      if (TARGET "CUDA::${_cudalib}_static")
-        target_link_libraries(
-          cccl.c.parallel.v2.hostjit_lib
-          PUBLIC "CUDA::${_cudalib}_static"
-        )
-      elseif (TARGET "CUDA::${_cudalib}")
-        target_link_libraries(
-          cccl.c.parallel.v2.hostjit_lib
-          PUBLIC "CUDA::${_cudalib}"
-        )
-      else()
-        message(
-          FATAL_ERROR
-          "hostjit needs CUDA::${_cudalib}[_static] but neither variant was "
-          "found by FindCUDAToolkit. Install the full CUDA toolkit "
-          "(libnvjitlink-dev / libnvfatbin-dev or equivalent)."
-        )
-      endif()
-    endforeach()
-  endif()
+  target_link_libraries(cccl.c.parallel.v2.hostjit_lib PUBLIC CUDA::cudart)
 endif()
 
 if (NOT MSVC)
@@ -216,49 +101,8 @@ set_target_properties(
   PROPERTIES CXX_STANDARD 20 POSITION_INDEPENDENT_CODE ON
 )
 
-# --------------------------------------------------------------------------
-# Install clang headers into wheel (for self-sufficient packaging)
-# --------------------------------------------------------------------------
-# Clang CUDA headers we still use from the LLVM source tree.
-# We DON'T install device_functions, math, or libdevice_declares — our local
-# copies in cuda_minimal/ replace them.
-set(
-  _clang_cuda_headers_needed
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__clang_cuda_math_forward_declares.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__clang_cuda_builtin_vars.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__clang_cuda_cmath.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__clang_cuda_intrinsics.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__clang_cuda_complex_builtins.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__clang_cuda_texture_intrinsics.h"
-)
-install(
-  FILES ${_clang_cuda_headers_needed}
-  DESTINATION "cuda/cccl/headers/clang"
-)
-
-# Clang builtin C headers needed by our stubs and CUDA toolkit headers.
-file(
-  GLOB _clang_stddef_headers
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__stddef_*.h"
-)
-set(
-  _clang_c_headers
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/limits.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/stddef.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/stdint.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__stddef_header_macro.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/float.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/__float_header_macro.h"
-  "${llvm_project_SOURCE_DIR}/clang/lib/Headers/inttypes.h"
-  ${_clang_stddef_headers}
-)
-install(FILES ${_clang_c_headers} DESTINATION "cuda/cccl/headers/clang")
-
 # Hostjit's minimal CUDA runtime headers (replacements for upstream clang headers)
-set(
-  _hostjit_cuda_minimal_dir
-  "${CMAKE_CURRENT_SOURCE_DIR}/include/hostjit/cuda_minimal"
-)
+set(_hostjit_cuda_minimal_dir "${_hostjit_include_dir}/hostjit/cuda_minimal")
 file(GLOB _hostjit_cuda_minimal_headers "${_hostjit_cuda_minimal_dir}/*.h")
 install(
   FILES ${_hostjit_cuda_minimal_headers}
@@ -274,7 +118,7 @@ install(
 )
 
 # On Windows with multi-config generators (Visual Studio), exclude hostjit
-# targets from Debug builds — the LLVM Debug build causes stack overflows.
+# targets that depend on libnvcc from Debug builds.
 if (MSVC)
   set_target_properties(
     cccl.c.parallel.v2.hostjit_lib

c/parallel.v2/src/hostjit/codegen/bitcode.cpp

@@ -102,16 +102,31 @@ bool BitcodeCollector::compile_and_add(const char* source, size_t source_size, c
     return true;
   }
 
-  hostjit::CUDACompiler compiler;
   std::string src(source, source_size);
-  auto result = compiler.compileToDeviceBitcode(src, config_);
-  if (!result.success)
+  auto path = make_temp_path("cccl_" + name + "_", unique_id_, ".bc");
+
+  std::vector<std::string> options;
+  config_.appendCommandLineArguments(options);
+  auto option_ptrs = hostjit::detail::make_libnvcc_option_ptrs(options);
+
+  hostjit::detail::LibnvccProgramGuard program;
+  auto create_result = libnvccCreateProgram(&program.program, src.c_str(), "input.cu");
+  if (create_result != LIBNVCC_SUCCESS)
   {
-    fprintf(stderr, "\nERROR compiling %s to bitcode: %s\n", name.c_str(), result.diagnostics.c_str());
+    fprintf(stderr, "\nERROR creating libnvcc program for %s: %s\n", name.c_str(), libnvccGetErrorString(create_result));
     return false;
   }
-  auto path = make_temp_path("cccl_" + name + "_", unique_id_, ".bc");
-  if (write_file(result.bitcode.data(), result.bitcode.size(), path))
+
+  auto result = libnvccCompileProgramToDeviceBitcode(
+    program.program, path.c_str(), static_cast<int>(option_ptrs.size()), option_ptrs.data());
+  if (result != LIBNVCC_SUCCESS)
+  {
+    auto log = hostjit::detail::get_libnvcc_program_log(program.program);
+    fprintf(stderr, "\nERROR compiling %s to bitcode: %s\n", name.c_str(), log.c_str());
+    return false;
+  }
+
+  if (std::filesystem::exists(path))
   {
     config_.device_bitcode_files.push_back(path);
     temp_paths_.push_back(path);