rsx: Fix segmented memory access for rsx::super_ptr

rpcs3/Emu/RSX/GL/GLTextureCache.h

@@ -559,6 +559,7 @@ namespace gl
 				}
 			}
 
+			flush_io();
 			glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 			glBindBuffer(GL_PIXEL_PACK_BUFFER, GL_NONE);
 

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -323,6 +323,7 @@ namespace vk
 				}
 			}
 
+			flush_io();
 			dma_buffer->unmap();
 			reset_write_statistics();
 

rpcs3/Emu/RSX/rsx_cache.h

@@ -43,6 +43,8 @@ namespace rsx
 
 				*first = cpu_address_base + confirmed_range.first;
 				*last = cpu_address_base + valid_limit - 4;
+
+				locked_memory_ptr.flush();
 			}
 		}
 
@@ -319,13 +321,13 @@ namespace rsx
 		}
 
 		template <typename T = void>
-		T* get_raw_ptr(u32 offset = 0) const
+		T* get_raw_ptr(u32 offset = 0)
 		{
 			verify(HERE), locked_memory_ptr;
 			return locked_memory_ptr.get<T>(offset);
 		}
 
-		bool test_memory_head() const
+		bool test_memory_head()
 		{
 			if (!locked_memory_ptr)
 			{
@@ -336,7 +338,7 @@ namespace rsx
 			return (*first == (cpu_address_base + confirmed_range.first));
 		}
 
-		bool test_memory_tail() const
+		bool test_memory_tail()
 		{
 			if (!locked_memory_ptr)
 			{
@@ -348,6 +350,11 @@ namespace rsx
 			return (*last == (cpu_address_base + valid_limit - 4));
 		}
 
+		void flush_io() const
+		{
+			locked_memory_ptr.flush();
+		}
+
 		std::pair<u32, u32> get_confirmed_range() const
 		{
 			if (confirmed_range.second == 0)

rpcs3/Emu/RSX/rsx_utils.cpp

@@ -109,14 +109,40 @@ namespace rsx
 			}
 		}
 
-		auto result = vm::get_super_ptr<u8>(addr, len - 1);
-		if (!result)
+		if (auto result = vm::get_super_ptr<u8>(addr, len - 1))
 		{
-			//Probably allocated as split blocks??
-			LOG_ERROR(RSX, "Could not get super_ptr for memory block 0x%x+0x%x", addr, len);
+			return { result };
 		}
 
-		return { result };
+		//Probably allocated as split blocks. Try to grab separate chunks
+		std::vector<weak_ptr::memory_block_t> blocks;
+		const u32 limit = addr + len;
+		u32 next = addr;
+		u32 remaining = len;
+
+		while (true)
+		{
+			auto region = vm::get(vm::any, next)->get(next, 1);
+			if (!region.second)
+			{
+				break;
+			}
+
+			const u32 block_offset = next - region.first;
+			const u32 block_length = std::min(remaining, region.second->size() - block_offset);
+			std::shared_ptr<u8> _ptr = { region.second, region.second->get(block_offset, block_length) };
+			blocks.push_back({_ptr, block_length});
+
+			remaining -= block_length;
+			next = region.first + region.second->size();
+			if (next >= limit)
+			{
+				return { blocks };
+			}
+		}
+
+		LOG_ERROR(RSX, "Could not get super_ptr for memory block 0x%x+0x%x", addr, len);
+		return {};
 	}
 
 	/* Fast image scaling routines

rpcs3/Emu/RSX/rsx_utils.h

@@ -35,21 +35,60 @@ namespace rsx
 	//Weak pointer without lock semantics
 	//Backed by a real shared_ptr for non-rsx memory
 	//Backed by a global shared pool for rsx memory
-	struct weak_ptr
+	class weak_ptr
 	{
-		void* _ptr;
-		std::shared_ptr<u8> _extern;
+	public:
+		using memory_block_t = std::pair<std::shared_ptr<u8>, u32>;
 
+	private:
+		void* _ptr = nullptr;
+		std::vector<memory_block_t> _blocks;
+		std::vector<u8> io_cache;
+		bool contiguous = true;
+		bool synchronized = true;
+
+	public:
 		weak_ptr(void* raw, bool is_rsx_mem = true)
 		{
 			_ptr = raw;
-			if (!is_rsx_mem) _extern.reset((u8*)raw);
+
+			if (!is_rsx_mem)
+			{
+				_blocks.push_back({});
+				_blocks.back().first.reset((u8*)raw);
+			}
 		}
 
 		weak_ptr(std::shared_ptr<u8>& block)
 		{
-			_extern = block;
-			_ptr = _extern.get();
+			_blocks.push_back({ block, 0 });
+			_ptr = block.get();
+		}
+
+		weak_ptr(std::vector<memory_block_t>& blocks)
+		{
+			verify(HERE), blocks.size() > 0;
+
+			_blocks = std::move(blocks);
+			_ptr = nullptr;
+
+			if (blocks.size() == 1)
+			{
+				_ptr = _blocks[0].first.get();
+				contiguous = true;
+			}
+			else
+			{
+				u32 block_length = 0;
+				for (const auto &block : _blocks)
+				{
+					block_length += block.second;
+				}
+
+				io_cache.resize(block_length);
+				contiguous = false;
+				synchronized = false;
+			}
 		}
 
 		weak_ptr()
@@ -58,14 +97,52 @@ namespace rsx
 		}
 
 		template <typename T = void>
-		T* get(u32 offset = 0) const
+		T* get(u32 offset = 0)
 		{
-			return (T*)((u8*)_ptr + offset);
+			if (contiguous)
+			{
+				return (T*)((u8*)_ptr + offset);
+			}
+			else
+			{
+				if (!synchronized)
+					sync();
+
+				return (T*)(io_cache.data() + offset);
+			}
+		}
+
+		void sync()
+		{
+			if (synchronized)
+				return;
+
+			u8* dst = (u8*)io_cache.data();
+			for (const auto &block : _blocks)
+			{
+				memcpy(dst, block.first.get(), block.second);
+				dst += block.second;
+			}
+
+			synchronized = true;
+		}
+
+		void flush() const
+		{
+			if (contiguous)
+				return;
+
+			u8* src = (u8*)io_cache.data();
+			for (const auto &block : _blocks)
+			{
+				memcpy(block.first.get(), src, block.second);
+				src += block.second;
+			}
 		}
 
 		operator bool() const
 		{
-			return (_ptr != nullptr);
+			return (_ptr != nullptr || _blocks.size() > 1);
 		}
 	};