rsx: Fixups

rpcs3/Emu/RSX/Common/surface_store.h

@@ -196,7 +196,7 @@ namespace rsx
 				{
 					auto surface = Traits::get(e.second);
 
-					if (new_surface->last_use_tag > surface->last_use_tag ||
+					if (new_surface->last_use_tag >= surface->last_use_tag ||
 						new_surface == surface ||
 						address == e.first)
 					{
@@ -274,15 +274,6 @@ namespace rsx
 				for (const auto& e : list2) surface_info.push_back(e);
 			}
 
-			if (UNLIKELY(surface_info.size() > 1))
-			{
-				// Sort with oldest first for early exit
-				std::sort(surface_info.begin(), surface_info.end(), [](const auto& a, const auto& b)
-				{
-					return (a.second->last_use_tag < b.second->last_use_tag);
-				});
-			}
-
 			// TODO: Modify deferred_clip_region::direct_copy() to take a few more things into account!
 			const areau child_region = new_surface->get_normalized_memory_area();
 			const auto child_w = child_region.width();
@@ -343,12 +334,6 @@ namespace rsx
 					continue;
 				}
 
-				if (child_w == size.width && child_h == size.height && surface_info.size() > 1)
-				{
-					// If the write covers the whole area, discard anything older
-					new_surface->clear_rw_barrier();
-				}
-
 				// TODO: Eventually need to stack all the overlapping regions, but for now just do the latest rect in the space
 				deferred_clipped_region<surface_type> region;
 				region.src_x = src_offset.x;

rpcs3/Emu/RSX/Common/surface_utils.h

@@ -318,6 +318,37 @@ namespace rsx
 			old_contents.clear();
 		}
 
+		template <typename T>
+		u32 prepare_rw_barrier_for_transfer(T *target)
+		{
+			if (old_contents.size() <= 1)
+				return 0;
+
+			// Sort here before doing transfers since surfaces may have been updated in the meantime
+			std::sort(old_contents.begin(), old_contents.end(), [](auto& a, auto &b)
+			{
+				auto _a = static_cast<T*>(a.source);
+				auto _b = static_cast<T*>(b.source);
+				return (_a->last_use_tag < _b->last_use_tag);
+			});
+
+			// Try and optimize by omitting possible overlapped transfers
+			for (size_t i = old_contents.size() - 1; i > 0 /* Intentional */; i--)
+			{
+				old_contents[i].init_transfer(target);
+
+				const auto dst_area = old_contents[i].dst_rect();
+				if (unsigned(dst_area.x2) == target->width() && unsigned(dst_area.y2) == target->height() &&
+					!dst_area.x1 && !dst_area.y1)
+				{
+					// This transfer will overwrite everything older
+					return u32(i);
+				}
+			}
+
+			return 0;
+		}
+
 		template<typename T>
 		void set_old_contents(T* other)
 		{

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -623,11 +623,15 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 		return;
 	}
 
-	for (auto &section : old_contents)
+	const bool dst_is_depth = !!(aspect() & gl::image_aspect::depth);
+	const auto dst_bpp = get_bpp();
+	unsigned first = prepare_rw_barrier_for_transfer(this);
+
+	for (auto i = first; i < old_contents.size(); ++i)
 	{
+		auto &section = old_contents[i];
 		auto src_texture = gl::as_rtt(section.source);
 		const auto src_bpp = src_texture->get_bpp();
-		const auto dst_bpp = get_bpp();
 		rsx::typeless_xfer typeless_info{};
 
 		if (get_internal_format() == src_texture->get_internal_format())
@@ -649,7 +653,6 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 			}
 		}
 
-		const bool dst_is_depth = !!(aspect() & gl::image_aspect::depth);
 		section.init_transfer(this);
 
 		if (state_flags & rsx::surface_state_flags::erase_bkgnd)

rpcs3/Emu/RSX/VK/VKRenderTargets.h

@@ -315,14 +315,20 @@ namespace vk
 				return;
 			}
 
+			// Memory transfers
+			vk::image *target_image = (samples() > 1) ? get_resolve_target() : this;
+			vk::blitter hw_blitter;
 			bool optimize_copy = true;
-			for (auto &section : old_contents)
+			const auto dst_bpp = get_bpp();
+			unsigned first = prepare_rw_barrier_for_transfer(this);
+
+			for (auto i = first; i < old_contents.size(); ++i)
 			{
+				auto &section = old_contents[i];
 				auto src_texture = static_cast<vk::render_target*>(section.source);
 				src_texture->read_barrier(cmd);
 
 				const auto src_bpp = src_texture->get_bpp();
-				const auto dst_bpp = get_bpp();
 				rsx::typeless_xfer typeless_info{};
 
 				if (src_texture->info.format == info.format)
@@ -342,9 +348,7 @@ namespace vk
 					}
 				}
 
-				vk::blitter hw_blitter;
 				section.init_transfer(this);
-
 				auto src_area = section.src_rect();
 				auto dst_area = section.dst_rect();
 
@@ -354,7 +358,6 @@ namespace vk
 					this->transform_pixels_to_samples(dst_area);
 				}
 
-				vk::image *target_image = (samples() > 1) ? get_resolve_target() : this;
 				bool memory_load = true;
 				if (dst_area.x1 == 0 && dst_area.y1 == 0 &&
 					unsigned(dst_area.x2) == target_image->width() && unsigned(dst_area.y2) == target_image->height())