undo reinterpret_cast

nod-ai · Oct 17, 2024 · 626057f · 626057f
1 parent 00e1b36
commit 626057f
Show file tree

Hide file tree

Showing 10 changed files with 138 additions and 103 deletions.
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/allocator.cc b/runtime/src/iree-amd-aie/driver/xrt-lite/allocator.cc
@@ -76,7 +76,9 @@ static iree_status_t iree_hal_xrt_lite_allocator_allocate_buffer(
   IREE_TRACE_ZONE_BEGIN(z0);
 
   iree_hal_xrt_lite_allocator* allocator =
-      reinterpret_cast<iree_hal_xrt_lite_allocator*>(base_allocator);
+      IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(base_allocator,
+                                            iree_hal_xrt_lite_allocator_vtable,
+                                            iree_hal_xrt_lite_allocator);
   iree_hal_buffer_params_t compat_params = *params;
   iree_hal_buffer_compatibility_t compatibility =
       iree_hal_xrt_lite_allocator_query_buffer_compatibility(
@@ -118,7 +120,9 @@ static void iree_hal_xrt_lite_allocator_deallocate_buffer(
   IREE_TRACE_ZONE_BEGIN(z0);
 
   iree_hal_xrt_lite_allocator* allocator =
-      reinterpret_cast<iree_hal_xrt_lite_allocator*>(base_allocator);
+      IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(base_allocator,
+                                            iree_hal_xrt_lite_allocator_vtable,
+                                            iree_hal_xrt_lite_allocator);
   bool was_imported = false;
   if (!was_imported) {
     IREE_STATISTICS(iree_hal_allocator_statistics_record_free(
@@ -157,10 +161,12 @@ iree_status_t iree_hal_xrt_lite_allocator_create(
 static void iree_hal_xrt_lite_allocator_destroy(
     iree_hal_allocator_t* base_allocator) {
   IREE_ASSERT_ARGUMENT(base_allocator);
-  iree_hal_xrt_lite_allocator* allocator =
-      reinterpret_cast<iree_hal_xrt_lite_allocator*>(base_allocator);
   IREE_TRACE_ZONE_BEGIN(z0);
 
+  iree_hal_xrt_lite_allocator* allocator =
+      IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(base_allocator,
+                                            iree_hal_xrt_lite_allocator_vtable,
+                                            iree_hal_xrt_lite_allocator);
   iree_hal_resource_release(&allocator->resource);
   iree_allocator_free(allocator->host_allocator, allocator);
 
@@ -172,7 +178,9 @@ static iree_allocator_t iree_hal_xrt_lite_allocator_host_allocator(
   IREE_TRACE_ZONE_BEGIN(z0);
 
   const iree_hal_xrt_lite_allocator* allocator =
-      reinterpret_cast<const iree_hal_xrt_lite_allocator*>(base_allocator);
+      IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(base_allocator,
+                                            iree_hal_xrt_lite_allocator_vtable,
+                                            const iree_hal_xrt_lite_allocator);
 
   IREE_TRACE_ZONE_END(z0);
   return allocator->host_allocator;

diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/buffer.cc b/runtime/src/iree-amd-aie/driver/xrt-lite/buffer.cc
@@ -24,8 +24,8 @@ static iree_status_t iree_hal_xrt_lite_buffer_invalidate_range(
     iree_device_size_t local_byte_length) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_buffer* buffer =
-      reinterpret_cast<iree_hal_xrt_lite_buffer*>(base_buffer);
+  iree_hal_xrt_lite_buffer* buffer = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_buffer, iree_hal_xrt_lite_buffer_vtable, iree_hal_xrt_lite_buffer);
   if (IREE_UNLIKELY(!buffer->bo)) {
     IREE_TRACE_ZONE_END(z0);
     return iree_make_status(
@@ -46,8 +46,8 @@ static iree_status_t iree_hal_xrt_lite_buffer_map_range(
     iree_hal_buffer_mapping_t* mapping) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_buffer* buffer =
-      reinterpret_cast<iree_hal_xrt_lite_buffer*>(base_buffer);
+  iree_hal_xrt_lite_buffer* buffer = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_buffer, iree_hal_xrt_lite_buffer_vtable, iree_hal_xrt_lite_buffer);
   IREE_RETURN_AND_END_ZONE_IF_ERROR(
       z0, iree_hal_buffer_validate_memory_type(
               iree_hal_buffer_memory_type(
@@ -87,8 +87,8 @@ static iree_status_t iree_hal_xrt_lite_buffer_flush_range(
     iree_device_size_t local_byte_length) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_buffer* buffer =
-      reinterpret_cast<iree_hal_xrt_lite_buffer*>(base_buffer);
+  iree_hal_xrt_lite_buffer* buffer = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_buffer, iree_hal_xrt_lite_buffer_vtable, iree_hal_xrt_lite_buffer);
   if (IREE_UNLIKELY(!buffer->bo)) {
     IREE_TRACE_ZONE_END(z0);
     return iree_make_status(
@@ -123,8 +123,8 @@ iree_status_t iree_hal_xrt_lite_buffer_wrap(
 
   iree_hal_xrt_lite_buffer* buffer = nullptr;
   IREE_RETURN_AND_END_ZONE_IF_ERROR(
-      z0,
-      iree_allocator_malloc(host_allocator, sizeof(*buffer), (void**)&buffer));
+      z0, iree_allocator_malloc(host_allocator, sizeof(*buffer),
+                                reinterpret_cast<void**>(&buffer)));
   iree_hal_buffer_initialize(host_allocator, allocator, &buffer->base,
                              allocation_size, byte_offset, byte_length,
                              memory_type, allowed_access, allowed_usage,
@@ -138,11 +138,11 @@ iree_status_t iree_hal_xrt_lite_buffer_wrap(
 }
 
 static void iree_hal_xrt_lite_buffer_destroy(iree_hal_buffer_t* base_buffer) {
-  iree_hal_xrt_lite_buffer* buffer =
-      reinterpret_cast<iree_hal_xrt_lite_buffer*>(base_buffer);
-  iree_allocator_t host_allocator = base_buffer->host_allocator;
+  iree_hal_xrt_lite_buffer* buffer = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_buffer, iree_hal_xrt_lite_buffer_vtable, iree_hal_xrt_lite_buffer);
   IREE_TRACE_ZONE_BEGIN(z0);
 
+  iree_allocator_t host_allocator = base_buffer->host_allocator;
   if (buffer->release_callback.fn) {
     buffer->release_callback.fn(buffer->release_callback.user_data,
                                 base_buffer);
@@ -157,8 +157,8 @@ static void iree_hal_xrt_lite_buffer_destroy(iree_hal_buffer_t* base_buffer) {
 shim_xdna::bo* iree_hal_xrt_lite_buffer_handle(iree_hal_buffer_t* base_buffer) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_buffer* buffer =
-      reinterpret_cast<iree_hal_xrt_lite_buffer*>(base_buffer);
+  iree_hal_xrt_lite_buffer* buffer = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_buffer, iree_hal_xrt_lite_buffer_vtable, iree_hal_xrt_lite_buffer);
 
   IREE_TRACE_ZONE_END(z0);
   return buffer->bo;

diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/device.cc b/runtime/src/iree-amd-aie/driver/xrt-lite/device.cc
@@ -23,7 +23,6 @@ extern const iree_hal_device_vtable_t iree_hal_xrt_lite_device_vtable;
 
 struct iree_hal_xrt_lite_device {
   iree_hal_resource_t resource;
-  iree_string_view_t identifier;
   iree_allocator_t host_allocator;
   // TODO(max): not used because "device allocations" are performed through
   // device
@@ -32,6 +31,9 @@ struct iree_hal_xrt_lite_device {
   // since command buffers can contain inlined data
   iree_arena_block_pool_t block_pool;
   shim_xdna::device* shim_device;
+  // should come last; see the definition of total_size below in
+  // iree_hal_xrt_lite_device_create
+  iree_string_view_t identifier;
 
   iree_hal_xrt_lite_device(const iree_hal_xrt_lite_device_options* options,
                            iree_allocator_t host_allocator) {
@@ -53,12 +55,12 @@ struct iree_hal_xrt_lite_device {
 };
 
 static iree_status_t iree_hal_xrt_lite_device_create_executable_cache(
-    iree_hal_device_t* base_value, iree_string_view_t identifier,
+    iree_hal_device_t* base_device, iree_string_view_t identifier,
     iree_loop_t loop, iree_hal_executable_cache_t** out_executable_cache) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   IREE_TRACE_ZONE_END(z0);
   return iree_hal_xrt_lite_nop_executable_cache_create(
@@ -67,7 +69,7 @@ static iree_status_t iree_hal_xrt_lite_device_create_executable_cache(
 }
 
 static iree_status_t iree_hal_xrt_lite_device_create_command_buffer(
-    iree_hal_device_t* base_value, iree_hal_command_buffer_mode_t mode,
+    iree_hal_device_t* base_device, iree_hal_command_buffer_mode_t mode,
     iree_hal_command_category_t command_categories,
     iree_hal_queue_affinity_t queue_affinity, iree_host_size_t binding_capacity,
     iree_hal_command_buffer_t** out_command_buffer) {
@@ -79,8 +81,8 @@ static iree_status_t iree_hal_xrt_lite_device_create_command_buffer(
                             "unimplmented multi-shot command buffer");
   }
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   IREE_TRACE_ZONE_END(z0);
   return iree_hal_deferred_command_buffer_create(
@@ -89,29 +91,29 @@ static iree_status_t iree_hal_xrt_lite_device_create_command_buffer(
 }
 
 static iree_status_t iree_hal_xrt_lite_device_create_semaphore(
-    iree_hal_device_t* base_value, uint64_t initial_value,
+    iree_hal_device_t* base_device, uint64_t initial_value,
     iree_hal_semaphore_flags_t flags, iree_hal_semaphore_t** out_semaphore) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   IREE_TRACE_ZONE_END(z0);
   return iree_hal_xrt_lite_semaphore_create(device->host_allocator,
                                             initial_value, out_semaphore);
 }
 
 static iree_status_t iree_hal_xrt_lite_device_queue_execute(
-    iree_hal_device_t* base_value, iree_hal_queue_affinity_t queue_affinity,
+    iree_hal_device_t* base_device, iree_hal_queue_affinity_t queue_affinity,
     const iree_hal_semaphore_list_t wait_semaphore_list,
     const iree_hal_semaphore_list_t signal_semaphore_list,
     iree_host_size_t command_buffer_count,
     iree_hal_command_buffer_t* const* command_buffers,
     iree_hal_buffer_binding_table_t const* binding_tables) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   for (iree_host_size_t i = 0; i < command_buffer_count; i++) {
     iree_hal_command_buffer_t* xrt_command_buffer = nullptr;
@@ -136,26 +138,26 @@ static iree_status_t iree_hal_xrt_lite_device_queue_execute(
 }
 
 static void iree_hal_xrt_lite_device_replace_device_allocator(
-    iree_hal_device_t* base_value, iree_hal_allocator_t* new_allocator) {
+    iree_hal_device_t* base_device, iree_hal_allocator_t* new_allocator) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
   iree_hal_allocator_retain(new_allocator);
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
   device->device_allocator = new_allocator;
   iree_hal_allocator_release(device->device_allocator);
 
   IREE_TRACE_ZONE_END(z0);
 }
 
 static iree_status_t iree_hal_xrt_lite_device_query_i64(
-    iree_hal_device_t* base_value, iree_string_view_t category,
+    iree_hal_device_t* base_device, iree_string_view_t category,
     iree_string_view_t key, int64_t* out_value) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
   *out_value = 0;
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
   if (iree_string_view_equal(category, IREE_SV("hal.device.id"))) {
     *out_value =
         iree_string_view_match_pattern(device->identifier, key) ? 1 : 0;
@@ -172,16 +174,16 @@ static iree_status_t iree_hal_xrt_lite_device_query_i64(
 }
 
 static iree_status_t iree_hal_xrt_lite_device_queue_alloca(
-    iree_hal_device_t* base_value, iree_hal_queue_affinity_t queue_affinity,
+    iree_hal_device_t* base_device, iree_hal_queue_affinity_t queue_affinity,
     const iree_hal_semaphore_list_t wait_semaphore_list,
     const iree_hal_semaphore_list_t signal_semaphore_list,
     iree_hal_allocator_pool_t pool, iree_hal_buffer_params_t params,
     iree_device_size_t allocation_size,
     iree_hal_buffer_t** IREE_RESTRICT out_buffer) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
   IREE_RETURN_AND_END_ZONE_IF_ERROR(
       z0, iree_hal_semaphore_list_wait(wait_semaphore_list,
                                        iree_infinite_timeout()));
@@ -196,21 +198,21 @@ static iree_status_t iree_hal_xrt_lite_device_queue_alloca(
 }
 
 static iree_string_view_t iree_hal_xrt_lite_device_id(
-    iree_hal_device_t* base_value) {
+    iree_hal_device_t* base_device) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   IREE_TRACE_ZONE_END(z0);
   return device->identifier;
 }
 
-static void iree_hal_xrt_lite_device_destroy(iree_hal_device_t* base_value) {
+static void iree_hal_xrt_lite_device_destroy(iree_hal_device_t* base_device) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   iree_hal_allocator_release(device->device_allocator);
   delete device->shim_device;
@@ -220,22 +222,22 @@ static void iree_hal_xrt_lite_device_destroy(iree_hal_device_t* base_value) {
 };
 
 static iree_allocator_t iree_hal_xrt_lite_device_host_allocator(
-    iree_hal_device_t* base_value) {
+    iree_hal_device_t* base_device) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   IREE_TRACE_ZONE_END(z0);
   return device->host_allocator;
 }
 
 static iree_hal_allocator_t* iree_hal_xrt_lite_device_device_allocator(
-    iree_hal_device_t* base_value) {
+    iree_hal_device_t* base_device) {
   IREE_TRACE_ZONE_BEGIN(z0);
 
-  iree_hal_xrt_lite_device* device =
-      reinterpret_cast<iree_hal_xrt_lite_device*>(base_value);
+  iree_hal_xrt_lite_device* device = IREE_HAL_XRT_LITE_CHECKED_VTABLE_CAST(
+      base_device, iree_hal_xrt_lite_device_vtable, iree_hal_xrt_lite_device);
 
   IREE_TRACE_ZONE_END(z0);
   return device->device_allocator;
@@ -262,8 +264,8 @@ iree_status_t iree_hal_xrt_lite_device_create(
 
   iree_hal_xrt_lite_device* device = nullptr;
   iree_host_size_t total_size = iree_sizeof_struct(*device) + identifier.size;
-  IREE_RETURN_IF_ERROR(
-      iree_allocator_malloc(host_allocator, total_size, (void**)&device));
+  IREE_RETURN_IF_ERROR(iree_allocator_malloc(
+      host_allocator, total_size, reinterpret_cast<void**>(&device)));
   device = new (device) iree_hal_xrt_lite_device(options, host_allocator);
   iree_string_view_append_to_buffer(
       identifier, &device->identifier,