Hide symbols by default for mac/linux (#3057)

CMakeLists.txt

@@ -323,9 +323,6 @@ target_include_directories(
   mlx PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
              $<INSTALL_INTERFACE:include>)
 
-# Do not add mlx_EXPORTS define for shared library.
-set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
-
 if(USE_SYSTEM_FMT)
   find_package(fmt REQUIRED)
 else()

mlx/CMakeLists.txt

@@ -25,6 +25,21 @@ target_compile_definitions(mlx_version PRIVATE MLX_VERSION="${MLX_VERSION}")
 target_include_directories(mlx_version PRIVATE ${PROJECT_SOURCE_DIR})
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:mlx_version>)
 
+# Do not export symbols by default.
+set_target_properties(
+  mlx mlx_version
+  PROPERTIES VISIBILITY_INLINES_HIDDEN ON
+             CXX_VISIBILITY_PRESET hidden
+             CUDA_VISIBILITY_PRESET hidden)
+
+# Define MLX_EXPORT for shared libraries.
+set_target_properties(mlx mlx_version PROPERTIES DEFINE_SYMBOL MLX_EXPORT)
+# Define MLX_STATIC for static libraries.
+if(NOT BUILD_SHARED_LIBS)
+  target_compile_definitions(mlx PUBLIC MLX_STATIC)
+  target_compile_definitions(mlx_version PUBLIC MLX_STATIC)
+endif()
+
 if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
   # Supress warnings: note: parameter passing for argument of type
   # 'std::pair<float, float>' when C++17 is enabled changed to match C++14 in
@@ -33,16 +48,6 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
 endif()
 
 if(MSVC)
-  # Windows DLLs have a 65535 symbol export limit. We use explicit exports via
-  # MLX_API macro (__declspec(dllexport)) on public API functions only. This
-  # avoids exporting internal template instantiations.
-  if(BUILD_SHARED_LIBS)
-    target_compile_definitions(mlx PRIVATE MLX_EXPORT)
-    target_compile_definitions(mlx_version PRIVATE MLX_EXPORT)
-  else()
-    target_compile_definitions(mlx PUBLIC MLX_STATIC)
-    target_compile_definitions(mlx_version PRIVATE MLX_STATIC)
-  endif()
   # Some of CUDA's headers include windows.h, which defines min/max macros.
   target_compile_definitions(mlx PRIVATE NOMINMAX)
   # Disable some MSVC warnings to speed up compilation.

mlx/api.h

@@ -2,30 +2,28 @@
 
 #pragma once
 
-// MLX_API macro for controlling symbol visibility
-//
-// On Windows, DLL symbols must be explicitly exported. This header provides
-// the MLX_API macro which:
-// - Expands to __declspec(dllexport) when building the MLX DLL (MLX_EXPORT
-// defined)
-// - Expands to __declspec(dllimport) when using the MLX DLL
-// - Expands to nothing for static library builds or non-Windows platforms
+// MLX_API macro for controlling symbol visibility, must add for public APIs.
 //
 // Usage:
-//   MLX_API array arange(double start, double stop, ...);
+//   MLX_API void some_function(...);
 //   class MLX_API SomeClass { ... };
 
-#if defined(_WIN32) || defined(_WIN64)
 #if defined(MLX_STATIC)
+
 // Static library build - no import/export decorations needed
 #define MLX_API
-#elif defined(MLX_EXPORT)
+
+#else
+
+// Shared library build.
+#if defined(_WIN32)
+#if defined(MLX_EXPORT)
 #define MLX_API __declspec(dllexport)
 #else
 #define MLX_API __declspec(dllimport)
-#endif
+#endif // defined(MLX_EXPORT)
 #else
-// On non-Windows platforms, symbols are visible by default
-// Could use __attribute__((visibility("default"))) if needed
-#define MLX_API
-#endif
+#define MLX_API __attribute__((visibility("default")))
+#endif // defined(_WIN32)
+
+#endif // defined(MLX_STATIC)

mlx/backend/cpu/encoder.h

@@ -12,7 +12,7 @@ namespace mlx::core::cpu {
 // Number of dispatches per scheduler task
 constexpr int DISPATCHES_PER_TASK = 10;
 
-struct CommandEncoder {
+struct MLX_API CommandEncoder {
   CommandEncoder(Stream stream) : stream_(stream) {}
 
   CommandEncoder(const CommandEncoder&) = delete;
@@ -62,6 +62,6 @@ struct CommandEncoder {
   int num_ops_{0};
 };
 
-CommandEncoder& get_command_encoder(Stream stream);
+MLX_API CommandEncoder& get_command_encoder(Stream stream);
 
 } // namespace mlx::core::cpu

mlx/backend/metal/device.h

@@ -20,7 +20,7 @@ using MTLFCList =
 
 struct DeviceStream;
 
-struct CommandEncoder {
+struct MLX_API CommandEncoder {
   explicit CommandEncoder(DeviceStream& stream);
   CommandEncoder(const CommandEncoder&) = delete;
   CommandEncoder& operator=(const CommandEncoder&) = delete;
@@ -146,7 +146,7 @@ struct DeviceStream {
   std::vector<array> temporaries;
 };
 
-class Device {
+class MLX_API Device {
  public:
   Device();
   Device(const Device&) = delete;
@@ -261,7 +261,7 @@ class Device {
   int max_mb_per_buffer_;
 };
 
-Device& device(mlx::core::Device);
+MLX_API Device& device(mlx::core::Device);
 
 std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool();
 

mlx/backend/metal/utils.h

@@ -10,8 +10,8 @@
 
 namespace mlx::core {
 
-std::string type_to_name(const Dtype& t);
-std::string type_to_name(const array& a);
+MLX_API std::string type_to_name(const Dtype& t);
+MLX_API std::string type_to_name(const array& a);
 
 // Compute the grid and block dimensions, check backend/common/utils.h for docs.
 MTL::Size get_block_dims(int dim0, int dim1, int dim2, int pow2 = 10);

mlx/dtype.cpp

@@ -92,6 +92,21 @@ Dtype::Kind kindof(const Dtype& t) {
   return type_kinds[static_cast<int>(t.val())];
 }
 
+template class MLX_API TypeToDtype<bool>;
+template class MLX_API TypeToDtype<uint8_t>;
+template class MLX_API TypeToDtype<uint16_t>;
+template class MLX_API TypeToDtype<uint32_t>;
+template class MLX_API TypeToDtype<uint64_t>;
+template class MLX_API TypeToDtype<int8_t>;
+template class MLX_API TypeToDtype<int16_t>;
+template class MLX_API TypeToDtype<int32_t>;
+template class MLX_API TypeToDtype<int64_t>;
+template class MLX_API TypeToDtype<float16_t>;
+template class MLX_API TypeToDtype<float>;
+template class MLX_API TypeToDtype<double>;
+template class MLX_API TypeToDtype<bfloat16_t>;
+template class MLX_API TypeToDtype<complex64_t>;
+
 template <>
 TypeToDtype<bool>::operator Dtype() {
   return bool_;

mlx/primitives.h

@@ -173,7 +173,7 @@ class Abs : public UnaryPrimitive {
   DEFINE_INPUT_OUTPUT_SHAPE()
 };
 
-class Add : public UnaryPrimitive {
+class MLX_API Add : public UnaryPrimitive {
  public:
   explicit Add(Stream stream) : UnaryPrimitive(stream) {}
 
@@ -624,7 +624,7 @@ class Ceil : public UnaryPrimitive {
   DEFINE_INPUT_OUTPUT_SHAPE()
 };
 
-class Compiled : public Primitive {
+class MLX_API Compiled : public Primitive {
  public:
   /*
    * The inputs, outputs and tape are either tracers or constants.
@@ -1018,7 +1018,7 @@ class ErfInv : public UnaryPrimitive {
   DEFINE_INPUT_OUTPUT_SHAPE()
 };
 
-class Exp : public UnaryPrimitive {
+class MLX_API Exp : public UnaryPrimitive {
  public:
   explicit Exp(Stream stream) : UnaryPrimitive(stream) {}
 
@@ -1768,7 +1768,7 @@ class Reshape : public UnaryPrimitive {
   Shape shape_;
 };
 
-class Reduce : public UnaryPrimitive {
+class MLX_API Reduce : public UnaryPrimitive {
  public:
   enum ReduceType { And, Or, Sum, Prod, Min, Max };