ARROW-7725: [C++] Add infrastructure for unity builds and precompiled…

… headers

Unity builds and precompiled headers can be enabled "easily" in CMake 3.16:
https://cmake.org/cmake/help/v3.16/prop_tgt/UNITY_BUILD.html
https://cmake.org/cmake/help/v3.16/command/target_precompile_headers.html

They can make builds faster in some conditions, especially on CI with little parallelism and caching.

Some care is required to make them work:

- unity builds concatenate a number of C/C++ source files together,
  so there should be no symbol clashes, C++ namespace resolution issues
  or conflicting preprocessor definitions

- precompiled headers require that the same compile options are passed
  for all files using the header

Closes #6324 from pitrou/ARROW-7725-pch-unity-builds and squashes the following commits:

2811feb <Antoine Pitrou> ARROW-7725:  Add infrastructure for unity builds and precompiled headers

Authored-by: Antoine Pitrou <antoine@python.org>
Signed-off-by: Sutou Kouhei <kou@clear-code.com>

.github/workflows/cpp.yml

@@ -163,6 +163,7 @@ jobs:
       CMAKE_GENERATOR: ${{ matrix.generator }}
       CMAKE_INSTALL_LIBDIR: bin
       CMAKE_INSTALL_PREFIX: /usr
+      CMAKE_UNITY_BUILD: ON
       ARROW_HOME: /usr
       ARROW_DATASET: ON
       ARROW_JEMALLOC: OFF

appveyor.yml

@@ -69,12 +69,12 @@ environment:
       ARROW_S3: "ON"
       ARROW_BUILD_FLIGHT: "ON"
       ARROW_BUILD_GANDIVA: "ON"
-    # NOTE: Since ARROW-5403 we have disabled the static CRT build
-    # - JOB: "Static_Crt_Build"
-    #   GENERATOR: Ninja
     - JOB: "Build_Debug"
       GENERATOR: Ninja
       CONFIGURATION: "Debug"
+    # NOTE: Since ARROW-5403 we have disabled the static CRT build
+    # - JOB: "Static_Crt_Build"
+    #   GENERATOR: Ninja
     - JOB: "MinGW32"
       MINGW_ARCH: i686
       MINGW_PACKAGE_PREFIX: mingw-w64-i686

ci/appveyor-cpp-build.bat

@@ -85,11 +85,13 @@ if "%JOB%" == "Build_Debug" (
   pushd cpp\build-debug
 
   cmake -G "%GENERATOR%" ^
+        -DARROW_USE_PRECOMPILED_HEADERS=OFF ^
         -DARROW_VERBOSE_THIRDPARTY_BUILD=OFF ^
         -DARROW_BOOST_USE_SHARED=OFF ^
         -DARROW_BUILD_TESTS=ON ^
         -DARROW_BUILD_EXAMPLES=ON ^
         -DCMAKE_BUILD_TYPE=%CONFIGURATION% ^
+        -DCMAKE_UNITY_BUILD=ON ^
         -DARROW_BUILD_STATIC=OFF ^
         -DARROW_CXXFLAGS="/MP" ^
         -DARROW_ENABLE_TIMING_TESTS=OFF ^

ci/cpp-msvc-build-main.bat

@@ -62,6 +62,7 @@ cmake -G "%GENERATOR%" %CMAKE_ARGS% ^
       -DCMAKE_BUILD_TYPE=%CONFIGURATION% ^
       -DCMAKE_INSTALL_PREFIX=%CONDA_PREFIX%\Library ^
       -DCMAKE_VERBOSE_MAKEFILE=OFF ^
+      -DCMAKE_UNITY_BUILD=ON ^
       -DARROW_VERBOSE_THIRDPARTY_BUILD=OFF ^
       -DARROW_BOOST_USE_SHARED=OFF ^
       -DARROW_BUILD_STATIC=OFF ^

ci/scripts/cpp_build.sh

@@ -88,6 +88,7 @@ cmake -G "${CMAKE_GENERATOR:-Ninja}" \
       -DARROW_USE_CCACHE=${ARROW_USE_CCACHE:-ON} \
       -DARROW_USE_GLOG=${ARROW_USE_GLOG:-ON} \
       -DARROW_USE_LD_GOLD=${ARROW_USE_LD_GOLD:-OFF} \
+      -DARROW_USE_PRECOMPILED_HEADERS=${ARROW_USE_PRECOMPILED_HEADERS:-OFF} \
       -DARROW_USE_STATIC_CRT=${ARROW_USE_STATIC_CRT:-OFF} \
       -DARROW_USE_UBSAN=${ARROW_USE_UBSAN:-OFF} \
       -DARROW_VERBOSE_THIRDPARTY_BUILD=${ARROW_VERBOSE_THIRDPARTY_BUILD:-OFF} \
@@ -107,6 +108,7 @@ cmake -G "${CMAKE_GENERATOR:-Ninja}" \
       -DCMAKE_CXX_FLAGS=${CXXFLAGS:-} \
       -DCMAKE_INSTALL_LIBDIR=${CMAKE_INSTALL_LIBDIR:-lib} \
       -DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX:-${ARROW_HOME}} \
+      -DCMAKE_UNITY_BUILD=${CMAKE_UNITY_BUILD:-OFF} \
       -Dgflags_SOURCE=${gflags_SOURCE:-AUTO} \
       -DgRPC_SOURCE=${gRPC_SOURCE:-AUTO} \
       -DGTest_SOURCE=${GTest_SOURCE:-AUTO} \
@@ -123,7 +125,7 @@ cmake -G "${CMAKE_GENERATOR:-Ninja}" \
       ${CMAKE_ARGS} \
       ${source_dir}
 
-cmake --build . --target install
+time cmake --build . --target install
 
 popd
 

cpp/CMakeLists.txt

@@ -151,6 +151,11 @@ if(ARROW_USE_CCACHE)
   endif(CCACHE_FOUND)
 endif()
 
+if(ARROW_USE_PRECOMPILED_HEADERS AND ${CMAKE_VERSION} VERSION_LESS "3.16.0")
+  message(WARNING "Precompiled headers need CMake 3.16.0 or later, disabling")
+  set(ARROW_USE_PRECOMPILED_HEADERS OFF)
+endif()
+
 if(ARROW_OPTIONAL_INSTALL)
   # Don't make the "install" target depend on the "all" target
   set(CMAKE_SKIP_INSTALL_ALL_DEPENDENCY true)

cpp/cmake_modules/BuildUtils.cmake

@@ -122,12 +122,20 @@ function(ADD_THIRDPARTY_LIB LIB_NAME)
   endif()
 endfunction()
 
+function(REUSE_PRECOMPILED_HEADER_LIB TARGET_NAME LIB_NAME)
+  if(ARROW_USE_PRECOMPILED_HEADERS)
+    target_precompile_headers(${TARGET_NAME} REUSE_FROM ${LIB_NAME})
+  endif()
+endfunction()
+
 # \arg OUTPUTS list to append built targets to
 function(ADD_ARROW_LIB LIB_NAME)
   set(options BUILD_SHARED BUILD_STATIC)
-  set(one_value_args CMAKE_PACKAGE_NAME PKG_CONFIG_NAME SHARED_LINK_FLAGS)
+  set(one_value_args CMAKE_PACKAGE_NAME PKG_CONFIG_NAME SHARED_LINK_FLAGS
+                     PRECOMPILED_HEADER_LIB)
   set(multi_value_args
       SOURCES
+      PRECOMPILED_HEADERS
       OUTPUTS
       STATIC_LINK_LIBS
       SHARED_LINK_LIBS
@@ -173,21 +181,26 @@ function(ADD_ARROW_LIB LIB_NAME)
     # because of dllexport declarations on Windows.
     # The Xcode generator doesn't reliably work with Xcode as target names are not
     # guessed correctly.
-    set(LIB_DEPS ${ARG_SOURCES})
-    set(EXTRA_DEPS ${ARG_DEPENDENCIES})
-
-    if(ARG_EXTRA_INCLUDES)
-      set(LIB_INCLUDES ${ARG_EXTRA_INCLUDES})
-    endif()
+    set(USE_OBJLIB OFF)
   else()
-    # Otherwise, generate a single "objlib" from all C++ modules and link
+    set(USE_OBJLIB ON)
+  endif()
+
+  if(USE_OBJLIB)
+    # Generate a single "objlib" from all C++ modules and link
     # that "objlib" into each library kind, to avoid compiling twice
     add_library(${LIB_NAME}_objlib OBJECT ${ARG_SOURCES})
     # Necessary to make static linking into other shared libraries work properly
     set_property(TARGET ${LIB_NAME}_objlib PROPERTY POSITION_INDEPENDENT_CODE 1)
     if(ARG_DEPENDENCIES)
       add_dependencies(${LIB_NAME}_objlib ${ARG_DEPENDENCIES})
     endif()
+    if(ARG_PRECOMPILED_HEADER_LIB)
+      reuse_precompiled_header_lib(${LIB_NAME}_objlib ${ARG_PRECOMPILED_HEADER_LIB})
+    endif()
+    if(ARG_PRECOMPILED_HEADERS AND ARROW_USE_PRECOMPILED_HEADERS)
+      target_precompile_headers(${LIB_NAME}_objlib PRIVATE ${ARG_PRECOMPILED_HEADERS})
+    endif()
     set(LIB_DEPS $<TARGET_OBJECTS:${LIB_NAME}_objlib>)
     set(LIB_INCLUDES)
     set(EXTRA_DEPS)
@@ -202,6 +215,15 @@ function(ADD_ARROW_LIB LIB_NAME)
     if(ARG_PRIVATE_INCLUDES)
       target_include_directories(${LIB_NAME}_objlib PRIVATE ${ARG_PRIVATE_INCLUDES})
     endif()
+  else()
+    # Prepare arguments for separate compilation of static and shared libs below
+    # TODO: add PCH directives
+    set(LIB_DEPS ${ARG_SOURCES})
+    set(EXTRA_DEPS ${ARG_DEPENDENCIES})
+
+    if(ARG_EXTRA_INCLUDES)
+      set(LIB_INCLUDES ${ARG_EXTRA_INCLUDES})
+    endif()
   endif()
 
   set(RUNTIME_INSTALL_DIR bin)
@@ -212,6 +234,10 @@ function(ADD_ARROW_LIB LIB_NAME)
       add_dependencies(${LIB_NAME}_shared ${EXTRA_DEPS})
     endif()
 
+    if(ARG_PRECOMPILED_HEADER_LIB)
+      reuse_precompiled_header_lib(${LIB_NAME}_shared ${ARG_PRECOMPILED_HEADER_LIB})
+    endif()
+
     if(ARG_OUTPUTS)
       list(APPEND ${ARG_OUTPUTS} ${LIB_NAME}_shared)
     endif()
@@ -293,6 +319,10 @@ function(ADD_ARROW_LIB LIB_NAME)
       add_dependencies(${LIB_NAME}_static ${EXTRA_DEPS})
     endif()
 
+    if(ARG_PRECOMPILED_HEADER_LIB)
+      reuse_precompiled_header_lib(${LIB_NAME}_static ${ARG_PRECOMPILED_HEADER_LIB})
+    endif()
+
     if(ARG_OUTPUTS)
       list(APPEND ${ARG_OUTPUTS} ${LIB_NAME}_static)
     endif()
@@ -516,9 +546,10 @@ endfunction()
 # names must exist
 function(ADD_TEST_CASE REL_TEST_NAME)
   set(options NO_VALGRIND ENABLED)
-  set(one_value_args)
+  set(one_value_args PRECOMPILED_HEADER_LIB)
   set(multi_value_args
       SOURCES
+      PRECOMPILED_HEADERS
       STATIC_LINK_LIBS
       EXTRA_LINK_LIBS
       EXTRA_INCLUDES
@@ -577,6 +608,14 @@ function(ADD_TEST_CASE REL_TEST_NAME)
     target_link_libraries(${TEST_NAME} PRIVATE ${ARROW_TEST_LINK_LIBS})
   endif()
 
+  if(ARG_PRECOMPILED_HEADER_LIB)
+    reuse_precompiled_header_lib(${TEST_NAME} ${ARG_PRECOMPILED_HEADER_LIB})
+  endif()
+
+  if(ARG_PRECOMPILED_HEADERS AND ARROW_USE_PRECOMPILED_HEADERS)
+    target_precompile_headers(${TEST_NAME} PRIVATE ${ARG_PRECOMPILED_HEADERS})
+  endif()
+
   if(ARG_EXTRA_LINK_LIBS)
     target_link_libraries(${TEST_NAME} PRIVATE ${ARG_EXTRA_LINK_LIBS})
   endif()

cpp/cmake_modules/DefineOptions.cmake

@@ -92,6 +92,9 @@ if("${CMAKE_SOURCE_DIR}" STREQUAL "${CMAKE_CURRENT_SOURCE_DIR}")
 
   define_option(ARROW_USE_LD_GOLD "Use ld.gold for linking on Linux (if available)" OFF)
 
+  define_option(ARROW_USE_PRECOMPILED_HEADERS "Use precompiled headers when compiling"
+                OFF)
+
   # Disable this option to exercise non-SIMD fallbacks
   define_option(ARROW_USE_SIMD "Build with SIMD optimizations" ON)
 

cpp/cmake_modules/SetupCxxFlags.cmake

@@ -227,6 +227,12 @@ elseif(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
     # Work around https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43407
     set(CXX_COMMON_FLAGS "${CXX_COMMON_FLAGS} -Wno-attributes")
   endif()
+
+  if(CMAKE_UNITY_BUILD)
+    # Work around issue similar to https://bugs.webkit.org/show_bug.cgi?id=176869
+    set(CXX_ONLY_FLAGS "${CXX_ONLY_FLAGS} -Wno-subobject-linkage")
+  endif()
+
 elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang"
        OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
   # Clang options for all builds