Fix BE endianess arch support in semaphore_406e (RPCS3#6116)

Add raw() methods for endianness support types and make use of it.

Utilities/BEType.h

@@ -490,6 +490,11 @@ class se_t<T, true, Align>
 		return storage::from(m_data);
 	}
 
+	stype& raw()
+	{
+		return m_data;
+	}
+
 	se_t& operator=(const se_t&) = default;
 
 	se_t& operator=(type value)
@@ -533,6 +538,11 @@ class se_t<T, false, Align>
 		return std::bit_cast<type>(m_data);
 	}
 
+	stype& raw()
+	{
+		return m_data;
+	}
+
 	se_t& operator=(const se_t& value) = default;
 
 	se_t& operator=(type value)

rpcs3/Emu/RSX/Capture/rsx_capture.cpp

@@ -370,26 +370,24 @@ namespace rsx
 			frame_capture_data::tile_state tilestate;
 			for (u32 i = 0; i < limits::tiles_count; ++i)
 			{
-				// Avoid byteswapping
-				auto tile = rsx->tiles[i].pack();
+				const auto tile = rsx->tiles[i].pack();
 				auto& tstate = tilestate.tiles[i];
-				tstate.tile = std::bit_cast<u32>(tile.tile);
-				tstate.limit = std::bit_cast<u32>(tile.limit);
-				tstate.pitch = rsx->tiles[i].binded ? std::bit_cast<u32>(tile.pitch) : 0;
-				tstate.format = rsx->tiles[i].binded ? std::bit_cast<u32>(tile.format) : 0;
+				tstate.tile = tile.tile;
+				tstate.limit = tile.limit;
+				tstate.pitch = rsx->tiles[i].binded ? u32{tile.pitch} : 0;
+				tstate.format = rsx->tiles[i].binded ? u32{tile.format} : 0;
 			}
 
 			for (u32 i = 0; i < limits::zculls_count; ++i)
 			{
-				// Avoid byteswapping
-				auto zc = rsx->zculls[i].pack();
+				const auto zc = rsx->zculls[i].pack();
 				auto& zcstate = tilestate.zculls[i];
-				zcstate.region = std::bit_cast<u32>(zc.region);
-				zcstate.size = std::bit_cast<u32>(zc.size);
-				zcstate.start = std::bit_cast<u32>(zc.start);
-				zcstate.offset = std::bit_cast<u32>(zc.offset);
-				zcstate.status0 = rsx->zculls[i].binded ? std::bit_cast<u32>(zc.status0) : 0;
-				zcstate.status1 = rsx->zculls[i].binded ? std::bit_cast<u32>(zc.status1) : 0;
+				zcstate.region = zc.region;
+				zcstate.size = zc.size;
+				zcstate.start = zc.start;
+				zcstate.offset = zc.offset;
+				zcstate.status0 = rsx->zculls[i].binded ? u32{zc.status0} : 0;
+				zcstate.status1 = rsx->zculls[i].binded ? u32{zc.status1} : 0;
 			}
 
 			const u64 tsnum = XXH64(&tilestate, sizeof(frame_capture_data::tile_state), 0);

rpcs3/Emu/RSX/CgBinaryFragmentProgram.cpp

@@ -228,7 +228,8 @@ void CgBinaryDisasm::TaskFP()
 	verify(HERE), ((m_buffer_size - m_offset) % sizeof(u32) == 0);
 	for (u32 i = 0; i < (m_buffer_size - m_offset) / sizeof(u32); i++)
 	{
-		data[i] = se_storage<u32>::swap(data[i]); // WTF, cannot use be_t<> there?
+		// Get BE data
+		data[i] = be_t<u32>{data[i]}.raw();
 	}
 
 	enum

rpcs3/Emu/RSX/CgBinaryProgram.h

@@ -375,7 +375,7 @@ class CgBinaryDisasm
 			verify(HERE), (m_buffer_size - m_offset) % sizeof(u32) == 0;
 			for (u32 i = 0; i < (m_buffer_size - m_offset) / sizeof(u32); i++)
 			{
-				vdata[i] = se_storage<u32>::swap(vdata[i]); // WTF, cannot use be_t<> there?
+				vdata[i] = be_t<u32>{vdata[i]}.raw();
 			}
 
 			for (u32 i = 0; i < prog.ucodeSize / sizeof(u32); i++)

rpcs3/Emu/RSX/RSXThread.cpp

@@ -325,7 +325,7 @@ namespace rsx
 	{
 		//Endianness is swapped because common upload code expects input in BE
 		//TODO: Implement fast upload path for LE inputs and do away with this
-		element_push_buffer.push_back(se_storage<u32>::swap(index));
+		element_push_buffer.push_back(be_t<u32>{index}.raw());
 	}
 
 	u32 thread::get_push_buffer_index_count() const

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -64,12 +64,10 @@ namespace rsx
 			rsx->sync_point_request = true;
 			const u32 addr = get_address(method_registers.semaphore_offset_406e(), method_registers.semaphore_context_dma_406e());
 
-#ifdef IS_LE_MACHINE
-			arg = se_storage<u32>::swap(arg);
-			const auto& sema = vm::_ref<le_t<u32>>(addr);
-#else
-			const auto& sema = vm::_ref<u32>(addr);
-#endif
+			// Get raw BE value
+			arg = be_t<u32>{arg}.raw();
+			const auto& sema = vm::_ref<nse_t<u32>>(addr);
+
 			// TODO: Remove vblank semaphore hack
 			if (sema == arg || addr == rsx->ctxt_addr + 0x30) return;
 
@@ -244,8 +242,8 @@ namespace rsx
 			{
 			case rsx::vertex_base_type::ub:
 			case rsx::vertex_base_type::ub256:
-				// One-way byteswap
-				arg = se_storage<u32>::swap(arg);
+				// Get BE data
+				arg = be_t<u32>{arg}.raw();
 				break;
 			}