Improve ppu_thread::stack_push

Save actual stack memory allocation size locally. (not in PS3 memory)

rpcs3/Emu/Cell/PPUThread.cpp

@@ -1292,26 +1292,31 @@ void ppu_thread::fast_call(u32 addr, u32 rtoc)
 	at_ret();
 }
 
-u32 ppu_thread::stack_push(u32 size, u32 align_v)
+std::pair<vm::addr_t, u32> ppu_thread::stack_push(u32 size, u32 align_v)
 {
 	if (auto cpu = get_current_cpu_thread<ppu_thread>())
 	{
 		ppu_thread& context = static_cast<ppu_thread&>(*cpu);
 
 		const u32 old_pos = vm::cast(context.gpr[1]);
-		context.gpr[1] -= utils::align(size + 4, 8); // room minimal possible size
+		context.gpr[1] -= size; // room minimal possible size
 		context.gpr[1] &= ~(u64{align_v} - 1); // fix stack alignment
 
-		if (old_pos >= context.stack_addr && old_pos < context.stack_addr + context.stack_size && context.gpr[1] < context.stack_addr)
+		auto is_stack = [&](u64 addr)
+		{
+			return addr >= context.stack_addr && addr < context.stack_addr + context.stack_size;
+		};
+
+		// TODO: This check does not care about custom stack memory
+		if (is_stack(old_pos) != is_stack(context.gpr[1]))
 		{
 			fmt::throw_exception("Stack overflow (size=0x%x, align=0x%x, SP=0x%llx, stack=*0x%x)", size, align_v, old_pos, context.stack_addr);
 		}
 		else
 		{
 			const u32 addr = static_cast<u32>(context.gpr[1]);
-			vm::_ref<nse_t<u32>>(addr + size) = old_pos;
 			std::memset(vm::base(addr), 0, size);
-			return addr;
+			return {vm::cast(addr), old_pos - addr};
 		}
 	}
 
@@ -1330,7 +1335,7 @@ void ppu_thread::stack_pop_verbose(u32 addr, u32 size) noexcept
 			return;
 		}
 
-		context.gpr[1] = vm::_ref<nse_t<u32>>(static_cast<u32>(context.gpr[1]) + size);
+		context.gpr[1] += size;
 		return;
 	}
 

rpcs3/Emu/Cell/PPUThread.h

@@ -308,7 +308,7 @@ public:
 	void exec_task();
 	void fast_call(u32 addr, u32 rtoc);
 
-	static u32 stack_push(u32 size, u32 align_v);
+	static std::pair<vm::addr_t, u32> stack_push(u32 size, u32 align_v);
 	static void stack_pop_verbose(u32 addr, u32 size) noexcept;
 };
 

rpcs3/Emu/Memory/vm_var.h

@@ -9,23 +9,23 @@ namespace vm
 	template <memory_location_t Location = vm::main>
 	struct page_allocator
 	{
-		static inline vm::addr_t alloc(u32 size, u32 align)
+		static inline std::pair<vm::addr_t, u32> alloc(u32 size, u32 align)
 		{
-			return vm::cast(vm::alloc(size, Location, std::max<u32>(align, 0x10000)));
+			return {vm::cast(vm::alloc(size, Location, std::max<u32>(align, 0x10000))), size};
 		}
 
-		static inline void dealloc(u32 addr, u32 /*size*/ = 0) noexcept
+		static inline void dealloc(u32 addr, u32 size) noexcept
 		{
-			ensure(vm::dealloc(addr, Location));
+			ensure(vm::dealloc(addr, Location) >= size);
 		}
 	};
 
 	template <typename T>
 	struct stack_allocator
 	{
-		static inline vm::addr_t alloc(u32 size, u32 align)
+		static inline std::pair<vm::addr_t, u32> alloc(u32 size, u32 align)
 		{
-			return vm::cast(T::stack_push(size, align));
+			return T::stack_push(size, align);
 		}
 
 		static inline void dealloc(u32 addr, u32 size) noexcept
@@ -40,6 +40,14 @@ namespace vm
 	{
 		using pointer = _ptr_base<T, const u32>;
 
+		const u32 m_mem_size;
+
+		_var_base(std::pair<vm::addr_t, u32> alloc_info)
+			: pointer(alloc_info.first)
+			, m_mem_size(alloc_info.second)
+		{
+		}
+
 	public:
 		// Unmoveable object
 		_var_base(const _var_base&) = delete;
@@ -49,7 +57,7 @@ namespace vm
 		using enable_bitcopy = std::false_type; // Disable bitcopy inheritence
 
 		_var_base()
-		    : pointer(A::alloc(sizeof(T), alignof(T)))
+		    : _var_base(A::alloc(sizeof(T), alignof(T)))
 		{
 		}
 
@@ -63,7 +71,7 @@ namespace vm
 		{
 			if (pointer::addr())
 			{
-				A::dealloc(pointer::addr(), sizeof(T));
+				A::dealloc(pointer::addr(), m_mem_size);
 			}
 		}
 	};
@@ -74,7 +82,15 @@ namespace vm
 	{
 		using pointer = _ptr_base<T, const u32>;
 
-		u32 m_size;
+		const u32 m_mem_size;
+		const u32 m_size;
+
+		_var_base(u32 count, std::pair<vm::addr_t, u32> alloc_info)
+			: pointer(alloc_info.first)
+			, m_mem_size(alloc_info.second)
+		    , m_size(u32{sizeof(T)} * count)
+		{
+		}
 
 	public:
 		_var_base(const _var_base&) = delete;
@@ -84,8 +100,7 @@ namespace vm
 		using enable_bitcopy = std::false_type; // Disable bitcopy inheritence
 
 		_var_base(u32 count)
-		    : pointer(A::alloc(u32{sizeof(T)} * count, alignof(T)))
+		    : _var_base(count, A::alloc(u32{sizeof(T)} * count, alignof(T)))
-		    , m_size(u32{sizeof(T)} * count)
 		{
 		}
 
@@ -101,7 +116,7 @@ namespace vm
 		{
 			if (pointer::addr())
 			{
-				A::dealloc(pointer::addr(), m_size);
+				A::dealloc(pointer::addr(), m_mem_size);
 			}
 		}