[WIP] Try to support header only mode for core library.

.github/workflows/build_test.yml

@@ -252,6 +252,14 @@ jobs:
             cxx_flags: -ftrapv
             cxx_standard: 23
 
+          - name: Header only (GCC-14 C++23)
+            extra_deps: g++-14
+            c_compiler: gcc-14
+            cxx_compiler: g++-14
+            cxx_flags: -ftrapv
+            cxx_standard: 23
+            extra_cmake_flags: -DHWY_CMAKE_HEADER_ONLY=ON
+
     steps:
       - name: Harden Runner
         uses: step-security/harden-runner@20cf305ff2072d973412fa9b1e3a4f227bda3c76 # v2.14.0

CMakeLists.txt

@@ -285,6 +285,7 @@ set(HWY_SOURCES
     hwy/x86_cpuid.h
 )
 
+
 if (NOT HWY_CMAKE_HEADER_ONLY)
   list(APPEND HWY_SOURCES
     hwy/abort.cc
@@ -500,10 +501,6 @@ if (HWY_DISABLE_FUTEX)
   list(APPEND HWY_FLAGS -DHWY_DISABLE_FUTEX)
 endif()
 
-if (HWY_CMAKE_HEADER_ONLY)
-  list(APPEND HWY_FLAGS -DHWY_HEADER_ONLY)
-endif()
-
 include(CheckIncludeFile)
 check_include_file(sys/auxv.h  HAVE_SYS_AUXV_H)
 check_include_file(asm/hwcap.h HAVE_ASM_HWCAP_H)
@@ -531,33 +528,78 @@ else()
   set(DLLEXPORT_TO_DEFINE "HWY_STATIC_DEFINE")
 endif()
 
-add_library(hwy ${HWY_LIBRARY_TYPE} ${HWY_SOURCES})
-if(NOT HAVE_SYS_AUXV_H)
-  target_compile_definitions(hwy PUBLIC TOOLCHAIN_MISS_SYS_AUXV_H)
-endif()
-if(NOT HAVE_ASM_HWCAP_H)
-  target_compile_definitions(hwy PUBLIC TOOLCHAIN_MISS_ASM_HWCAP_H)
-endif()
-target_compile_definitions(hwy PUBLIC "${DLLEXPORT_TO_DEFINE}")
-target_compile_options(hwy PRIVATE ${HWY_FLAGS})
-set_property(TARGET hwy PROPERTY POSITION_INDEPENDENT_CODE ON)
-set_target_properties(hwy PROPERTIES VERSION ${LIBRARY_VERSION} SOVERSION ${LIBRARY_SOVERSION})
-target_include_directories(hwy PUBLIC
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
-    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
-target_compile_features(hwy PUBLIC cxx_std_11)
-if (NOT HWY_CXX_STD_TGT_COMPILE_FEATURE STREQUAL "cxx_std_11")
-  target_compile_features(hwy PRIVATE ${HWY_CXX_STD_TGT_COMPILE_FEATURE})
-endif()
-set_target_properties(hwy PROPERTIES
-  LINK_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/hwy/hwy.version)
-# For GCC __atomic_store_8, see #887
-target_link_libraries(hwy PRIVATE ${ATOMICS_LIBRARIES})
-# not supported by MSVC/Clang, safe to skip (we use DLLEXPORT annotations)
-if(UNIX AND NOT APPLE)
-  set_property(TARGET hwy APPEND_STRING PROPERTY
-    LINK_FLAGS " -Wl,--version-script=${CMAKE_CURRENT_SOURCE_DIR}/hwy/hwy.version")
-endif()
+if (NOT HWY_CMAKE_HEADER_ONLY)
+  add_library(hwy ${HWY_LIBRARY_TYPE} ${HWY_SOURCES})
+  if (NOT HAVE_SYS_AUXV_H)
+    target_compile_definitions(hwy PUBLIC TOOLCHAIN_MISS_SYS_AUXV_H)
+  endif ()
+  if (NOT HAVE_ASM_HWCAP_H)
+    target_compile_definitions(hwy PUBLIC TOOLCHAIN_MISS_ASM_HWCAP_H)
+  endif ()
+  target_compile_definitions(hwy PUBLIC "${DLLEXPORT_TO_DEFINE}")
+  target_compile_definitions(hwy PUBLIC -DHWY_HEADER_ONLY=0)
+  target_compile_options(hwy PRIVATE ${HWY_FLAGS})
+  set_property(TARGET hwy PROPERTY POSITION_INDEPENDENT_CODE ON)
+  set_target_properties(hwy PROPERTIES VERSION ${LIBRARY_VERSION} SOVERSION ${LIBRARY_SOVERSION})
+  target_include_directories(hwy PUBLIC
+          $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
+          $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
+  target_compile_features(hwy PUBLIC cxx_std_11)
+  if (NOT HWY_CXX_STD_TGT_COMPILE_FEATURE STREQUAL "cxx_std_11")
+    target_compile_features(hwy PRIVATE ${HWY_CXX_STD_TGT_COMPILE_FEATURE})
+  endif ()
+  set_target_properties(hwy PROPERTIES
+          LINK_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/hwy/hwy.version)
+  # For GCC __atomic_store_8, see #887
+  target_link_libraries(hwy PRIVATE ${ATOMICS_LIBRARIES})
+  # not supported by MSVC/Clang, safe to skip (we use DLLEXPORT annotations)
+  if (UNIX AND NOT APPLE)
+    set_property(TARGET hwy APPEND_STRING PROPERTY
+            LINK_FLAGS " -Wl,--version-script=${CMAKE_CURRENT_SOURCE_DIR}/hwy/hwy.version")
+  endif ()
+else()
+  # to uncomment this, finally
+#  add_library(hwy INTERFACE)
+#  target_compile_definitions(hwy INTERFACE -DHWY_HEADER_ONLY=1)
+#  if (HWY_WARNINGS_ARE_ERRORS)
+#    target_include_directories(hwy INTERFACE
+#            $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
+#            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
+#  else()
+#    target_include_directories(hwy SYSTEM INTERFACE
+#            $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
+#            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
+#  endif ()
+#  target_link_libraries(hwy INTERFACE ${ATOMICS_LIBRARIES})
+
+  add_library(hwy STATIC)
+  target_compile_features(hwy PUBLIC cxx_std_11)
+  if (NOT HWY_CXX_STD_TGT_COMPILE_FEATURE STREQUAL "cxx_std_11")
+    target_compile_features(hwy PRIVATE ${HWY_CXX_STD_TGT_COMPILE_FEATURE})
+  endif ()
+  target_compile_definitions(hwy PUBLIC -DHWY_HEADER_ONLY=1)
+  if (HWY_WARNINGS_ARE_ERRORS)
+    target_include_directories(hwy PUBLIC
+            $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
+  else ()
+    target_include_directories(hwy SYSTEM PUBLIC
+            $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
+            $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
+  endif ()
+  target_link_libraries(hwy PRIVATE ${ATOMICS_LIBRARIES})
+  target_sources(hwy PRIVATE
+#          hwy/abort.cc
+#          hwy/aligned_allocator.cc
+#          hwy/nanobenchmark.cc
+          hwy/per_target.cc
+#          hwy/perf_counters.cc
+#          hwy/print.cc
+#          hwy/profiler.cc
+#          hwy/targets.cc
+#          hwy/timer.cc
+  )
+endif ()
 
 if (HWY_ENABLE_CONTRIB)
 add_library(hwy_contrib ${HWY_LIBRARY_TYPE} ${HWY_CONTRIB_SOURCES})

hwy/abort.cc

@@ -20,19 +20,19 @@
 
 namespace hwy {
 
-namespace {
+namespace detail {
 
-std::atomic<WarnFunc>& AtomicWarnFunc() {
+HWY_HEADER_ONLY_FUNC std::atomic<WarnFunc>& AtomicWarnFunc() {
   static std::atomic<WarnFunc> func;
   return func;
 }
 
-std::atomic<AbortFunc>& AtomicAbortFunc() {
+HWY_HEADER_ONLY_FUNC std::atomic<AbortFunc>& AtomicAbortFunc() {
   static std::atomic<AbortFunc> func;
   return func;
 }
 
-std::string GetBaseName(std::string const& file_name) {
+HWY_HEADER_ONLY_FUNC std::string GetBaseName(std::string const& file_name) {
   auto last_slash = file_name.find_last_of("/\\");
   return file_name.substr(last_slash + 1);
 }
@@ -43,55 +43,57 @@ std::string GetBaseName(std::string const& file_name) {
 // is required to safely implement `SetWarnFunc`. As a workaround, we store a
 // copy here, update it when called, and return a reference to the copy. This
 // has the added benefit of protecting the actual pointer from modification.
-HWY_DLLEXPORT WarnFunc& GetWarnFunc() {
+HWY_HEADER_ONLY_FUNC HWY_DLLEXPORT WarnFunc& GetWarnFunc() {
   static WarnFunc func;
-  func = AtomicWarnFunc().load();
+  func = detail::AtomicWarnFunc().load();
   return func;
 }
 
-HWY_DLLEXPORT AbortFunc& GetAbortFunc() {
+HWY_HEADER_ONLY_FUNC HWY_DLLEXPORT AbortFunc& GetAbortFunc() {
   static AbortFunc func;
-  func = AtomicAbortFunc().load();
+  func = detail::AtomicAbortFunc().load();
   return func;
 }
 
+HWY_HEADER_ONLY_FUNC
 HWY_DLLEXPORT WarnFunc SetWarnFunc(WarnFunc func) {
-  return AtomicWarnFunc().exchange(func);
+  return detail::AtomicWarnFunc().exchange(func);
 }
 
+HWY_HEADER_ONLY_FUNC
 HWY_DLLEXPORT AbortFunc SetAbortFunc(AbortFunc func) {
-  return AtomicAbortFunc().exchange(func);
+  return detail::AtomicAbortFunc().exchange(func);
 }
 
-HWY_DLLEXPORT void HWY_FORMAT(3, 4)
+HWY_HEADER_ONLY_FUNC HWY_DLLEXPORT void HWY_FORMAT(3, 4)
     Warn(const char* file, int line, const char* format, ...) {
   char buf[800];
   va_list args;
   va_start(args, format);
   vsnprintf(buf, sizeof(buf), format, args);
   va_end(args);
 
-  WarnFunc handler = AtomicWarnFunc().load();
+  WarnFunc handler = detail::AtomicWarnFunc().load();
   if (handler != nullptr) {
     handler(file, line, buf);
   } else {
-    fprintf(stderr, "Warn at %s:%d: %s\n", GetBaseName(file).data(), line, buf);
+    fprintf(stderr, "Warn at %s:%d: %s\n", detail::GetBaseName(file).data(), line, buf);
   }
 }
 
-HWY_DLLEXPORT HWY_NORETURN void HWY_FORMAT(3, 4)
+HWY_HEADER_ONLY_FUNC HWY_DLLEXPORT HWY_NORETURN void HWY_FORMAT(3, 4)
     Abort(const char* file, int line, const char* format, ...) {
   char buf[800];
   va_list args;
   va_start(args, format);
   vsnprintf(buf, sizeof(buf), format, args);
   va_end(args);
 
-  AbortFunc handler = AtomicAbortFunc().load();
+  AbortFunc handler = detail::AtomicAbortFunc().load();
   if (handler != nullptr) {
     handler(file, line, buf);
   } else {
-    fprintf(stderr, "Abort at %s:%d: %s\n", GetBaseName(file).data(), line,
+    fprintf(stderr, "Abort at %s:%d: %s\n", detail::GetBaseName(file).data(), line,
             buf);
   }
 

hwy/aligned_allocator.cc

@@ -25,7 +25,7 @@
 #include "hwy/base.h"
 
 namespace hwy {
-namespace {
+namespace detail {
 
 #if HWY_ARCH_RISCV && defined(__riscv_v_intrinsic) && \
     __riscv_v_intrinsic >= 11000
@@ -54,6 +54,7 @@ struct AllocationHeader {
 #pragma pack(pop)
 
 // Returns a 'random' (cyclical) offset for AllocateAlignedBytes.
+HWY_HEADER_ONLY_FUNC
 size_t NextAlignedOffset() {
   static std::atomic<size_t> next{0};
   static_assert(kAlias % kAlignment == 0, "kAlias must be a multiple");
@@ -66,8 +67,10 @@ size_t NextAlignedOffset() {
 
 }  // namespace
 
+HWY_HEADER_ONLY_FUNC
 HWY_DLLEXPORT void* AllocateAlignedBytes(const size_t payload_size,
                                          AllocPtr alloc_ptr, void* opaque_ptr) {
+  using namespace hwy::detail;
   HWY_ASSERT(payload_size != 0);  // likely a bug in caller
   if (payload_size >= std::numeric_limits<size_t>::max() / 2) {
     HWY_DASSERT(false && "payload_size too large");
@@ -114,8 +117,10 @@ HWY_DLLEXPORT void* AllocateAlignedBytes(const size_t payload_size,
   return HWY_ASSUME_ALIGNED(reinterpret_cast<void*>(payload), kAlignment);
 }
 
+HWY_HEADER_ONLY_FUNC
 HWY_DLLEXPORT void FreeAlignedBytes(const void* aligned_pointer,
                                     FreePtr free_ptr, void* opaque_ptr) {
+  using namespace hwy::detail;
   if (aligned_pointer == nullptr) return;
 
   const uintptr_t payload = reinterpret_cast<uintptr_t>(aligned_pointer);
@@ -131,10 +136,12 @@ HWY_DLLEXPORT void FreeAlignedBytes(const void* aligned_pointer,
 }
 
 // static
+HWY_HEADER_ONLY_FUNC
 HWY_DLLEXPORT void AlignedDeleter::DeleteAlignedArray(void* aligned_pointer,
                                                       FreePtr free_ptr,
                                                       void* opaque_ptr,
                                                       ArrayDeleter deleter) {
+  using namespace hwy::detail;
   if (aligned_pointer == nullptr) return;
 
   const uintptr_t payload = reinterpret_cast<uintptr_t>(aligned_pointer);

hwy/aligned_allocator.h

@@ -424,4 +424,8 @@ class AlignedNDArray {
 };
 
 }  // namespace hwy
+
+#if HWY_HEADER_ONLY
+#include "hwy/aligned_allocator.cc"
+#endif  // HWY_HEADER_ONLY
 #endif  // HIGHWAY_HWY_ALIGNED_ALLOCATOR_H_

hwy/auto_tune.h

@@ -28,7 +28,7 @@
 
 // configuration to allow auto_tune to use std::sort instead of VQSort
 // (also enabled in header only mode).
-#if defined(HWY_HEADER_ONLY)
+#if HWY_HEADER_ONLY
 #define HWY_AUTOTUNE_STDSORT
 #endif
 

hwy/base.h

@@ -21,9 +21,13 @@
 // IWYU pragma: begin_exports
 #include <stddef.h>
 #include <stdint.h>
-#if defined(HWY_HEADER_ONLY)
+#if HWY_HEADER_ONLY
 #include <stdarg.h>
 #include <stdio.h>
+
+#define HWY_HEADER_ONLY_FUNC inline
+#else
+#define HWY_HEADER_ONLY_FUNC
 #endif
 
 #if !defined(HWY_NO_LIBCXX)
@@ -284,38 +288,6 @@ namespace hwy {
 //------------------------------------------------------------------------------
 // Abort / Warn
 
-#if defined(HWY_HEADER_ONLY)
-HWY_DLLEXPORT inline void HWY_FORMAT(3, 4)
-    Warn(const char* file, int line, const char* format, ...) {
-  char buf[800];
-  va_list args;
-  va_start(args, format);
-  vsnprintf(buf, sizeof(buf), format, args);
-  va_end(args);
-
-  fprintf(stderr, "Warn at %s:%d: %s\n", file, line, buf);
-}
-
-HWY_DLLEXPORT HWY_NORETURN inline void HWY_FORMAT(3, 4)
-    Abort(const char* file, int line, const char* format, ...) {
-  char buf[800];
-  va_list args;
-  va_start(args, format);
-  vsnprintf(buf, sizeof(buf), format, args);
-  va_end(args);
-
-  fprintf(stderr, "Abort at %s:%d: %s\n", file, line, buf);
-
-  fflush(stderr);
-
-// Now terminate the program:
-#if HWY_ARCH_RISCV
-  exit(1);  // trap/abort just freeze Spike.
-#else
-  abort();  // Compile error without this due to HWY_NORETURN.
-#endif
-}
-#else  // !HWY_HEADER_ONLY
 // Interfaces for custom Warn/Abort handlers.
 typedef void (*WarnFunc)(const char* file, int line, const char* message);
 
@@ -350,8 +322,6 @@ HWY_DLLEXPORT void HWY_FORMAT(3, 4)
 HWY_DLLEXPORT HWY_NORETURN void HWY_FORMAT(3, 4)
     Abort(const char* file, int line, const char* format, ...);
 
-#endif  // HWY_HEADER_ONLY
-
 #define HWY_WARN(format, ...) \
   ::hwy::Warn(__FILE__, __LINE__, format, ##__VA_ARGS__)
 
@@ -3290,4 +3260,7 @@ HWY_API void PreventElision(T&& output) {
 
 }  // namespace hwy
 
+#if HWY_HEADER_ONLY
+#include "hwy/abort.cc"
+#endif  // HWY_HEADER_ONLY
 #endif  // HIGHWAY_HWY_BASE_H_