Fix opaque stencil rendering

drivers/gles3/rasterizer_scene_gles3.cpp

@@ -2529,7 +2529,6 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_
 		scene_state.set_gl_depth_func(GL_GEQUAL);
 		scene_state.enable_gl_scissor_test(false);
 		scene_state.enable_gl_stencil_test(false);
-		scene_state.set_gl_stencil_write_mask(255);
 
 		glColorMask(0, 0, 0, 0);
 		RasterizerGLES3::clear_depth(0.0);
@@ -3084,7 +3083,7 @@ void RasterizerSceneGLES3::_render_list_template(RenderListParameters *p_params,
 			}
 
 			// Stencil.
-			if (shader->stencil_enabled) {
+			if (p_pass_mode != PASS_MODE_DEPTH && shader->stencil_enabled) {
 				static const GLenum stencil_compare_table[GLES3::SceneShaderData::STENCIL_COMPARE_MAX] = {
 					GL_LESS,
 					GL_EQUAL,
@@ -3121,7 +3120,6 @@ void RasterizerSceneGLES3::_render_list_template(RenderListParameters *p_params,
 				scene_state.set_gl_stencil_op(GL_KEEP, stencil_op_dpfail, stencil_op_dppass);
 			} else {
 				scene_state.enable_gl_stencil_test(false);
-				scene_state.set_gl_stencil_write_mask(255);
 			}
 
 			if constexpr (p_pass_mode == PASS_MODE_COLOR || p_pass_mode == PASS_MODE_COLOR_TRANSPARENT) {

servers/rendering/renderer_rd/forward_clustered/scene_shader_forward_clustered.cpp

@@ -359,8 +359,9 @@ void SceneShaderForwardClustered::ShaderData::_create_pipeline(PipelineKey p_pip
 		}
 	}
 
-	depth_stencil_state.enable_stencil = stencil_enabled;
-	if (stencil_enabled) {
+	bool use_stencil = stencil_enabled && p_pipeline_key.version == PIPELINE_VERSION_COLOR_PASS;
+	depth_stencil_state.enable_stencil = use_stencil;
+	if (use_stencil) {
 		static const RD::CompareOperator stencil_compare_rd_table[STENCIL_COMPARE_MAX] = {
 			RD::COMPARE_OP_LESS,
 			RD::COMPARE_OP_EQUAL,

servers/rendering/renderer_rd/forward_clustered/scene_shader_forward_clustered.h

@@ -293,7 +293,7 @@ public:
 
 		_FORCE_INLINE_ bool uses_shared_shadow_material() const {
 			bool backface_culling = cull_mode == RS::CULL_MODE_BACK;
-			return !uses_particle_trails && !writes_modelview_or_projection && !uses_vertex && !uses_position && !uses_discard && !uses_depth_prepass_alpha && !uses_alpha_clip && !uses_alpha_antialiasing && backface_culling && !uses_point_size && !uses_world_coordinates && !wireframe && !uses_z_clip_scale;
+			return !uses_particle_trails && !writes_modelview_or_projection && !uses_vertex && !uses_position && !uses_discard && !uses_depth_prepass_alpha && !uses_alpha_clip && !uses_alpha_antialiasing && backface_culling && !uses_point_size && !uses_world_coordinates && !wireframe && !uses_z_clip_scale && !stencil_enabled;
 		}
 
 		virtual void set_code(const String &p_Code);

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp

@@ -889,7 +889,11 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 
 	// fill our render lists early so we can find out if we use various features
 	_fill_render_list(RENDER_LIST_OPAQUE, p_render_data, PASS_MODE_COLOR);
-	render_list[RENDER_LIST_OPAQUE].sort_by_key();
+	if (scene_state.used_opaque_stencil) {
+		render_list[RENDER_LIST_OPAQUE].sort_by_key_and_stencil();
+	} else {
+		render_list[RENDER_LIST_OPAQUE].sort_by_key();
+	}
 	render_list[RENDER_LIST_ALPHA].sort_by_reverse_depth_and_priority();
 	_fill_instance_data(RENDER_LIST_OPAQUE);
 	_fill_instance_data(RENDER_LIST_ALPHA);
@@ -2149,6 +2153,9 @@ void RenderForwardMobile::_fill_render_list(RenderListType p_render_list, const
 				if (surf->flags & GeometryInstanceSurfaceDataCache::FLAG_USES_DEPTH_TEXTURE) {
 					scene_state.used_depth_texture = true;
 				}
+				if ((surf->flags & GeometryInstanceSurfaceDataCache::FLAG_USES_STENCIL) && !force_alpha && (surf->flags & (GeometryInstanceSurfaceDataCache::FLAG_PASS_DEPTH | GeometryInstanceSurfaceDataCache::FLAG_PASS_OPAQUE))) {
+					scene_state.used_opaque_stencil = true;
+				}
 
 			} else if (p_pass_mode == PASS_MODE_SHADOW || p_pass_mode == PASS_MODE_SHADOW_DP) {
 				if (surf->flags & GeometryInstanceSurfaceDataCache::FLAG_PASS_SHADOW) {
@@ -2672,6 +2679,10 @@ void RenderForwardMobile::_geometry_instance_add_surface_with_material(GeometryI
 		flags |= GeometryInstanceSurfaceDataCache::FLAG_USES_DOUBLE_SIDED_SHADOWS;
 	}
 
+	if (p_material->shader_data->stencil_enabled) {
+		flags |= GeometryInstanceSurfaceDataCache::FLAG_USES_STENCIL;
+	}
+
 	if (p_material->shader_data->uses_alpha_pass()) {
 		flags |= GeometryInstanceSurfaceDataCache::FLAG_PASS_ALPHA;
 		if (p_material->shader_data->uses_depth_in_alpha_pass()) {

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h

@@ -289,6 +289,7 @@ private:
 		bool used_screen_texture = false;
 		bool used_depth_texture = false;
 		bool used_lightmap = false;
+		bool used_opaque_stencil = false;
 
 		struct ShadowPass {
 			uint32_t element_from;
@@ -338,6 +339,22 @@ private:
 			sorter.sort(elements.ptr() + p_from, p_size);
 		}
 
+		struct SortByKeyAndStencil {
+			_FORCE_INLINE_ bool operator()(const GeometryInstanceSurfaceDataCache *A, const GeometryInstanceSurfaceDataCache *B) const {
+				bool a_stencil = A->flags & GeometryInstanceSurfaceDataCache::FLAG_USES_STENCIL;
+				bool b_stencil = B->flags & GeometryInstanceSurfaceDataCache::FLAG_USES_STENCIL;
+				if (a_stencil != b_stencil) {
+					return a_stencil < b_stencil;
+				}
+				return (A->sort.sort_key2 == B->sort.sort_key2) ? (A->sort.sort_key1 < B->sort.sort_key1) : (A->sort.sort_key2 < B->sort.sort_key2);
+			}
+		};
+
+		void sort_by_key_and_stencil() {
+			SortArray<GeometryInstanceSurfaceDataCache *, SortByKeyAndStencil> sorter;
+			sorter.sort(elements.ptr(), elements.size());
+		}
+
 		struct SortByDepth {
 			_FORCE_INLINE_ bool operator()(const GeometryInstanceSurfaceDataCache *A, const GeometryInstanceSurfaceDataCache *B) const {
 				return (A->owner->depth < B->owner->depth);
@@ -448,6 +465,7 @@ protected:
 			FLAG_USES_NORMAL_TEXTURE = 16384,
 			FLAG_USES_DOUBLE_SIDED_SHADOWS = 32768,
 			FLAG_USES_PARTICLE_TRAILS = 65536,
+			FLAG_USES_STENCIL = 131072,
 		};
 
 		union {

servers/rendering/renderer_rd/forward_mobile/scene_shader_forward_mobile.h

@@ -290,7 +290,7 @@ public:
 		}
 
 		_FORCE_INLINE_ bool uses_shared_shadow_material() const {
-			return !uses_particle_trails && !writes_modelview_or_projection && !uses_vertex && !uses_discard && !uses_depth_prepass_alpha && !uses_alpha_clip && !uses_alpha_antialiasing && !uses_world_coordinates && !wireframe;
+			return !uses_particle_trails && !writes_modelview_or_projection && !uses_vertex && !uses_discard && !uses_depth_prepass_alpha && !uses_alpha_clip && !uses_alpha_antialiasing && !uses_world_coordinates && !wireframe && !stencil_enabled;
 		}
 
 		virtual void set_code(const String &p_Code);