rsx: Fix and improve fp program data invalidation

rpcs3/Emu/RSX/Capture/rsx_capture.cpp

@@ -51,10 +51,7 @@ namespace rsx
 			std::unordered_set<u64>& mem_changes = frame_capture.replay_commands.back().memory_state;
 
 			// capture fragment shader mem
-			const u32 shader_program = method_registers.shader_program_address();
+			const auto [program_offset, program_location] = method_registers.shader_program_address();
-
-			const u32 program_location = (shader_program & 0x3) - 1;
-			const u32 program_offset   = (shader_program & ~0x3);
 
 			const u32 addr          = get_address(program_offset, program_location, HERE);
 			const auto program_info = program_hash_util::fragment_program_utils::analyse_fragment_program(vm::base(addr));

rpcs3/Emu/RSX/RSXFragmentProgram.h

@@ -230,6 +230,7 @@ struct RSXFragmentProgram
 	void *addr;
 	u32 offset;
 	u32 ucode_length;
+	u32 total_length;
 	u32 ctrl;
 	u16 unnormalized_coords;
 	u16 redirected_textures;

rpcs3/Emu/RSX/RSXThread.cpp

@@ -1585,10 +1585,7 @@ namespace rsx
 		m_graphics_state &= ~(rsx::pipeline_state::fragment_program_dirty);
 		auto &result = current_fragment_program = {};
 
-		const u32 shader_program = rsx::method_registers.shader_program_address();
+		const auto [program_offset, program_location] = method_registers.shader_program_address();
-
-		const u32 program_location = (shader_program & 0x3) - 1;
-		const u32 program_offset = (shader_program & ~0x3);
 
 		result.addr = vm::base(rsx::get_address(program_offset, program_location, HERE));
 		current_fp_metadata = program_hash_util::fragment_program_utils::analyse_fragment_program(result.addr);
@@ -1596,6 +1593,7 @@ namespace rsx
 		result.addr = (static_cast<u8*>(result.addr) + current_fp_metadata.program_start_offset);
 		result.offset = program_offset + current_fp_metadata.program_start_offset;
 		result.ucode_length = current_fp_metadata.program_ucode_length;
+		result.total_length = result.ucode_length + current_fp_metadata.program_start_offset;
 		result.valid = true;
 		result.ctrl = rsx::method_registers.shader_control() & (CELL_GCM_SHADER_CONTROL_32_BITS_EXPORTS | CELL_GCM_SHADER_CONTROL_DEPTH_EXPORT);
 		result.texcoord_control_mask = rsx::method_registers.texcoord_control_mask();
@@ -1737,6 +1735,22 @@ namespace rsx
 		}
 	}
 
+	bool thread::invalidate_fragment_program(u32 dst_dma, u32 dst_offset, u32 size)
+	{
+		const auto [shader_offset, shader_dma] = rsx::method_registers.shader_program_address();
+
+		if ((dst_dma & CELL_GCM_LOCATION_MAIN) == shader_dma &&
+		address_range::start_length(shader_offset, current_fragment_program.total_length).overlaps(
+			address_range::start_length(dst_offset, size))) [[unlikely]]
+		{
+			// Data overlaps
+			m_graphics_state |= rsx::pipeline_state::fragment_program_dirty;
+			return true;
+		}
+
+		return false;
+	}
+
 	void thread::reset()
 	{
 		rsx::method_registers.reset();

rpcs3/Emu/RSX/RSXThread.h

@@ -711,6 +711,8 @@ namespace rsx
 		 * returns whether surface is a render target and surface pitch in native format
 		 */
 		void get_current_fragment_program(const std::array<std::unique_ptr<rsx::sampled_image_descriptor_base>, rsx::limits::fragment_textures_count>& sampler_descriptors);
+	public:
+		bool invalidate_fragment_program(u32 dst_dma, u32 dst_offset, u32 size);
 
 	public:
 		u64 target_rsx_flip_time = 0;

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -876,16 +876,22 @@ namespace rsx
 					{
 						// Move last 32 bits
 						reinterpret_cast<u32*>(dst)[0] = reinterpret_cast<const u32*>(src)[count - 1];
-					}
+						rsx->invalidate_fragment_program(dst_dma, dst_offset, 4);
-					else if (dst_dma & CELL_GCM_LOCATION_MAIN)
-					{
-						// May overlap
-						std::memmove(dst, src, data_length);
 					}
 					else
 					{
-						// Never overlaps
+						if (dst_dma & CELL_GCM_LOCATION_MAIN)
-						std::memcpy(dst, src, data_length);
+						{
+							// May overlap
+							std::memmove(dst, src, data_length);
+						}
+						else
+						{
+							// Never overlaps
+							std::memcpy(dst, src, data_length);
+						}
+
+						rsx->invalidate_fragment_program(dst_dma, dst_offset, count * 4);
 					}
 
 					break;
@@ -912,6 +918,7 @@ namespace rsx
 					{
 						// Move last 16 bits
 						dst[0] = convert(src[count - 1]);
+						rsx->invalidate_fragment_program(dst_dma, dst_offset, 2);
 						break;
 					}
 
@@ -920,6 +927,7 @@ namespace rsx
 						dst[i] = convert(src[i]);
 					}
 
+					rsx->invalidate_fragment_program(dst_dma, dst_offset, count * 2);
 					break;
 				}
 				default:
@@ -930,12 +938,6 @@ namespace rsx
 
 				//res->release(0);
 
-				if (!(dst_dma & CELL_GCM_LOCATION_MAIN))
-				{
-					// Set this flag on LOCAL memory transfer
-					rsx->m_graphics_state |= rsx::pipeline_state::fragment_program_dirty;
-				}
-
 				// Skip "handled methods"
 				rsx->fifo_ctrl->skip_methods(count - 1);
 			}
@@ -1086,6 +1088,8 @@ namespace rsx
 				const u32 nb_lines = std::min(clip_h, in_h);
 				const u32 data_length = nb_lines * src_line_length;
 
+				rsx->invalidate_fragment_program(dst_dma, dst_offset, data_length);
+
 				if (const auto result = rsx->read_barrier(src_address, data_length, false);
 					result == rsx::result_zcull_intr)
 				{
@@ -1099,6 +1103,8 @@ namespace rsx
 			else
 			{
 				const u32 data_length = in_pitch * (in_h - 1) + src_line_length;
+
+				rsx->invalidate_fragment_program(dst_dma, dst_offset, data_length);
 				rsx->read_barrier(src_address, data_length, true);
 			}
 
@@ -1437,6 +1443,8 @@ namespace rsx
 			const auto write_address = get_address(dst_offset, dst_dma, HERE);
 			const auto data_length = in_pitch * (line_count - 1) + line_length;
 
+			rsx->invalidate_fragment_program(dst_dma, dst_offset, data_length);
+	
 			if (const auto result = rsx->read_barrier(read_address, data_length, !is_block_transfer);
 				result == rsx::result_zcull_intr)
 			{

rpcs3/Emu/RSX/rsx_methods.h

@@ -1331,9 +1331,10 @@ namespace rsx
 			return decode<NV4097_SET_VERTEX_DATA_BASE_INDEX>().vertex_data_base_index();
 		}
 
-		u32 shader_program_address() const
+		std::pair<u32, u32> shader_program_address() const
 		{
-			return decode<NV4097_SET_SHADER_PROGRAM>().shader_program_address();
+			const u32 shader_address = decode<NV4097_SET_SHADER_PROGRAM>().shader_program_address();
+			return { shader_address & ~3, (shader_address & 3) - 1 };
 		}
 
 		u32 transform_program_start() const