Merge pull request #115757 from blueskythlikesclouds/fix-accidental-write-combined-read

Fix accidental write-combined memory reads in canvas renderer.

servers/rendering/renderer_rd/renderer_canvas_render_rd.cpp

@@ -3231,7 +3231,8 @@ RendererCanvasRenderRD::InstanceData *RendererCanvasRenderRD::new_instance_data(
 		// instance_count must be > 0 to indicate the batch has been used when calling _new_batch, so we set a flag.
 		p_current_batch.instance_count = PUSH_DATA_INSTANCE_COUNT;
 	} else {
-		instance_data = &state.instance_data[state.instance_data_index];
+		// Return the intermediary instance data to prevent the caller from accidentally reading write-combined memory pages, which has huge performance implications.
+		instance_data = &state.intermediary_instance_data;
 	}
 
 	memcpy(instance_data, &template_instance, sizeof(InstanceData));
@@ -3287,6 +3288,7 @@ void RendererCanvasRenderRD::_add_to_batch(bool &r_batch_broken, Batch *&r_curre
 			r_current_batch->command_type == Item::Command::TYPE_NINEPATCH ||
 			r_current_batch->command_type == Item::Command::TYPE_PRIMITIVE);
 	r_current_batch->instance_count++;
+	memcpy(&state.instance_data[state.instance_data_index], &state.intermediary_instance_data, sizeof(InstanceData));
 	state.instance_data_index++;
 	if (state.instance_data_index >= state.max_instances_per_buffer) {
 		RD::get_singleton()->buffer_flush(r_current_batch->instance_buffer);

servers/rendering/renderer_rd/renderer_canvas_render_rd.h

@@ -615,6 +615,8 @@ class RendererCanvasRenderRD : public RendererCanvasRender {
 		InstanceData *prev_instance_data = nullptr;
 		uint32_t prev_instance_data_index = 0;
 
+		InstanceData intermediary_instance_data;
+
 		uint32_t max_instances_per_buffer = 16384;
 		uint32_t max_instance_buffer_size = 16384 * sizeof(InstanceData);