rsx: Implement workaround for pitch compatibility of overlapping sections with blit engine targets

- Avoid generating surfaces that introduce pitch conflict!

Author: kd-11

rpcs3/Emu/RSX/Common/surface_store.h

@@ -112,11 +112,21 @@ namespace rsx
 					}
 					else
 					{
+						const bool is_pitch_compatible = Traits::surface_is_pitch_compatible(found->second, prev_surface->get_rsx_pitch());
+						if (!is_pitch_compatible && found->second->last_use_tag >= prev_surface->last_use_tag) [[unlikely]]
+						{
+							// HACK: A newer surface with an incompatible pitch owns the memory, do not evict.
+							// TODO: Pitch conversion is required to resolve this properly.
+							rsx_log.warning("[SURFACE CACHE] Discarding block at 0x%x from surface at 0x%x (pitch=%u); block is owned by a newer surface with pitch=%u.",
+								new_address, prev_surface->base_addr, prev_surface->get_rsx_pitch(), found->second->get_rsx_pitch());
+							return;
+						}
+
 						invalidate(found->second);
 						data.erase(new_address);
 
 						auto &old = invalidated_resources.back();
-						if (Traits::surface_is_pitch_compatible(old, prev_surface->get_rsx_pitch()))
+						if (is_pitch_compatible)
 						{
 							if (old->last_use_tag >= prev_surface->last_use_tag) [[unlikely]]
 							{
@@ -1128,6 +1138,69 @@ namespace rsx
 			return nullptr;
 		}
 
+		// Workaround to handle overlapping surfaces with differing pitch
+		// For a surface region defined by range [address, length] and pitch, we return the max length we can write without clobbering a surface of different pitch.
+		// TODO: Re-evaluate usefulness once pitch-conversion work is completed.
+		u32 truncate_memory_range_by_pitch(u32 address, u32 pitch, u32 length, u64 reference_tag)
+		{
+			if (!length || !pitch)
+			{
+				return length;
+			}
+
+			const auto test_range = rsx::address_range32::start_length(address, length);
+			const auto test_height = static_cast<u16>(utils::aligned_div(length, pitch));
+			u32 limit = address + length;
+
+			auto process_list_function = [&](surface_ranged_map& data)
+			{
+				for (auto it = data.begin_range(test_range); it != data.end(); ++it)
+				{
+					// Only a surface that begins after us can shorten the length. An exact address match is (hackishly) handled in bind_surface_address instead.
+					const auto base_address = it->first;
+					if (base_address <= address || base_address >= limit)
+					{
+						continue;
+					}
+
+					const auto surface = Traits::get(it->second);
+					if (!surface->get_memory_range().overlaps(test_range))
+					{
+						continue;
+					}
+
+					if (rsx::pitch_compatible(surface, pitch, test_height))
+					{
+						// Normal inheritance resolves this overlap. Nothing to do.
+						continue;
+					}
+
+					if (surface->last_use_tag <= reference_tag)
+					{
+						// Stale data, safe to swallow.
+						continue;
+					}
+
+					// min() is already handled in the original address bounds check
+					limit = base_address;
+				}
+			};
+
+			if (m_render_targets_memory_range.valid() &&
+				test_range.overlaps(m_render_targets_memory_range))
+			{
+				process_list_function(m_render_targets_storage);
+			}
+
+			if (m_depth_stencil_memory_range.valid() &&
+				test_range.overlaps(m_depth_stencil_memory_range))
+			{
+				process_list_function(m_depth_stencil_storage);
+			}
+
+			return limit - address;
+		}
+
 		/**
 		 * Invalidates surface that exists at an address
 		 */

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -1702,6 +1702,24 @@ namespace rsx
 				{
 					block_end = write_end;
 				}
+				else if (!use_null_region && block_end > write_end)
+				{
+					// Before attempting to create any surface that would exceed our bounds, ensure we don't end up in a pitch conflict!
+					// TODO: Remove this when pitch conversion is implemented or at least revise usefulness. It may still be beneficial to some extent.
+					const auto prev_surface = m_rtts.get_surface_at(dst_base_address);
+					const u64 reference_tag = prev_surface ? prev_surface->last_use_tag : 0ull;
+
+					const u32 block_length = block_end - dst_base_address;
+					const u32 safe_length = m_rtts.truncate_memory_range_by_pitch(dst_base_address, dst.pitch, block_length, reference_tag);
+
+					if (safe_length < block_length) [[ unlikely ]]
+					{
+						// We have pitch conflicts. Play it safe.
+						const auto aligned_safe_length = (safe_length / dst.pitch) * dst.pitch; // <- Align safe_length down to a multiple of pitch.
+						const auto proposed_end = aligned_safe_length + dst_base_address;
+						block_end = std::max(proposed_end, write_end);
+					}
+				}
 
 				const u32 usable_section_length = std::max(write_end, block_end) - dst_base_address;
 				dst_dimensions.height = align2(usable_section_length, dst.pitch) / dst.pitch;