vm: add map_self() method to utils::shm

Add complementary unmap_self() method.
Move VirtualMemory to util/vm.hpp
Minor associated include cleanup.
Move asm.h to util/asm.hpp

Utilities/JIT.cpp

@@ -5,7 +5,7 @@
 #include "util/logs.hpp"
 #include "mutex.h"
 #include "sysinfo.h"
-#include "VirtualMemory.h"
+#include "util/vm.hpp"
 #include <immintrin.h>
 #include <zlib.h>
 

Utilities/Thread.cpp

@@ -73,6 +73,7 @@
 #endif
 
 #include "sync.h"
+#include "util/vm.hpp"
 #include "util/logs.hpp"
 #include "Emu/Memory/vm_locking.h"
 

Utilities/typemap.h

@@ -4,7 +4,6 @@
 #include "mutex.h"
 #include "util/atomic.hpp"
 #include "util/typeindices.hpp"
-#include "VirtualMemory.h"
 #include <memory>
 
 namespace utils

rpcs3/Emu/CMakeLists.txt

@@ -39,6 +39,7 @@ target_sources(rpcs3_emu PRIVATE
 	../util/logs.cpp
 	../util/yaml.cpp
 	../util/cereal.cpp
+	../util/vm_native.cpp
 	../../Utilities/bin_patch.cpp
 	../../Utilities/cheat_info.cpp
 	../../Utilities/cond.cpp
@@ -54,7 +55,6 @@ target_sources(rpcs3_emu PRIVATE
 	../../Utilities/sysinfo.cpp
 	../../Utilities/Thread.cpp
 	../../Utilities/version.cpp
-	../../Utilities/VirtualMemory.cpp
 )
 
 target_include_directories(rpcs3_emu PUBLIC "${CMAKE_SOURCE_DIR}")

rpcs3/Emu/Cell/PPUAnalyser.cpp

@@ -6,7 +6,7 @@
 
 #include <unordered_set>
 #include "util/yaml.hpp"
-#include "Utilities/asm.h"
+#include "util/asm.hpp"
 
 LOG_CHANNEL(ppu_validator);
 

rpcs3/Emu/Cell/PPUInterpreter.cpp

@@ -4,7 +4,6 @@
 #include "Emu/Memory/vm_reservation.h"
 #include "Emu/system_config.h"
 #include "PPUThread.h"
-#include "Utilities/asm.h"
 #include "Utilities/sysinfo.h"
 #include "Emu/Cell/Common.h"
 
@@ -12,6 +11,8 @@
 #include <bit>
 #include <cmath>
 
+#include "util/asm.hpp"
+
 #if !defined(_MSC_VER) && defined(__clang__)
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wold-style-cast"
@@ -119,7 +120,7 @@ FORCE_INLINE auto ppu_feed_data(ppu_thread& ppu, u64 addr)
 		const std::byte* src;
 		u32 size;
 		u32 offs = 0;
-		
+
 		if (raddr / 128 == addr / 128)
 			src = &ppu.rdata[addr & 127], size = std::min<u32>(128 - (addr % 128), sizeof(T));
 		else

rpcs3/Emu/Cell/PPUModule.cpp

@@ -1,7 +1,6 @@
 #include "stdafx.h"
 #include "Emu/Cell/PPUModule.h"
 
-#include "Utilities/VirtualMemory.h"
 #include "Utilities/bin_patch.h"
 #include "Utilities/StrUtil.h"
 #include "Crypto/sha1.h"
@@ -747,7 +746,7 @@ static void ppu_check_patch_spu_images(const ppu_segment& seg)
 				if (!name.empty())
 				{
 					fmt::append(dump, "\n\tSPUNAME: '%s'", name);
-				}	
+				}
 			}
 		}
 

rpcs3/Emu/Cell/PPUThread.cpp

@@ -1,5 +1,4 @@
 #include "stdafx.h"
-#include "Utilities/VirtualMemory.h"
 #include "Utilities/sysinfo.h"
 #include "Utilities/JIT.h"
 #include "Crypto/sha1.h"
@@ -64,7 +63,7 @@
 #include <thread>
 #include <cfenv>
 #include <cctype>
-#include <string>
+#include "util/vm.hpp"
 
 const bool s_use_ssse3 = utils::has_ssse3();
 

rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp

@@ -8,10 +8,11 @@
 #include "SPUThread.h"
 #include "SPUInterpreter.h"
 #include "Utilities/sysinfo.h"
-#include "Utilities/asm.h"
 #include "PPUAnalyser.h"
 #include "Crypto/sha1.h"
 
+#include "util/asm.hpp"
+
 #include <cmath>
 #include <mutex>
 #include <thread>

rpcs3/Emu/Cell/SPUInterpreter.cpp

@@ -3,10 +3,11 @@
 
 #include "Utilities/JIT.h"
 #include "Utilities/sysinfo.h"
-#include "Utilities/asm.h"
 #include "SPUThread.h"
 #include "Emu/Cell/Common.h"
 
+#include "util/asm.hpp"
+
 #include <cmath>
 #include <cfenv>
 

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1,6 +1,5 @@
 #include "stdafx.h"
 #include "Utilities/JIT.h"
-#include "Utilities/asm.h"
 #include "Utilities/date_time.h"
 #include "Utilities/sysinfo.h"
 #include "Emu/Memory/vm.h"
@@ -31,6 +30,8 @@
 #include <atomic>
 #include <thread>
 #include <shared_mutex>
+#include "util/vm.hpp"
+#include "util/asm.hpp"
 
 using spu_rdata_t = decltype(spu_thread::rdata);
 

rpcs3/Emu/Cell/lv2/sys_memory.cpp

@@ -1,12 +1,13 @@
 #include "stdafx.h"
 #include "sys_memory.h"
 
-#include "Utilities/VirtualMemory.h"
 #include "Emu/Memory/vm_locking.h"
 #include "Emu/CPU/CPUThread.h"
 #include "Emu/Cell/ErrorCodes.h"
 #include "Emu/IdManager.h"
 
+#include "util/vm.hpp"
+
 LOG_CHANNEL(sys_memory);
 
 //

rpcs3/Emu/Cell/lv2/sys_mmapper.cpp

@@ -5,11 +5,12 @@
 #include "sys_ppu_thread.h"
 #include "Emu/Cell/lv2/sys_event.h"
 #include "Emu/Memory/vm_var.h"
-#include "Utilities/VirtualMemory.h"
 #include "sys_memory.h"
 #include "sys_sync.h"
 #include "sys_process.h"
 
+#include "util/vm.hpp"
+
 LOG_CHANNEL(sys_mmapper);
 
 lv2_memory::lv2_memory(u32 size, u32 align, u64 flags, lv2_memory_container* ct)

rpcs3/Emu/Cell/lv2/sys_mmapper.h

@@ -8,6 +8,11 @@
 
 struct lv2_memory_container;
 
+namespace utils
+{
+	class shm;
+}
+
 struct lv2_memory : lv2_obj
 {
 	static const u32 id_base = 0x08000000;

rpcs3/Emu/Cell/lv2/sys_time.cpp

@@ -3,7 +3,8 @@
 
 #include "Emu/system_config.h"
 #include "Emu/Cell/ErrorCodes.h"
-#include "Utilities/asm.h"
+
+#include "util/asm.hpp"
 
 #ifdef _WIN32
 

rpcs3/Emu/Memory/vm.cpp

@@ -7,7 +7,6 @@
 
 #include "Utilities/mutex.h"
 #include "Utilities/Thread.h"
-#include "Utilities/VirtualMemory.h"
 #include "Utilities/address_range.h"
 #include "Emu/CPU/CPUThread.h"
 #include "Emu/Cell/lv2/sys_memory.h"
@@ -18,6 +17,9 @@
 #include <deque>
 #include <shared_mutex>
 
+#include "util/vm.hpp"
+#include "util/asm.hpp"
+
 LOG_CHANNEL(vm_log, "VM");
 
 namespace vm
@@ -301,9 +303,9 @@ namespace vm
 		// Block or signal new range locks
 		g_range_lock = addr | u64{size} << 32 | flags;
 
-		_mm_prefetch(g_range_lock_set + 0, _MM_HINT_T0);
+		utils::prefetch_read(g_range_lock_set + 0);
-		_mm_prefetch(g_range_lock_set + 2, _MM_HINT_T0);
+		utils::prefetch_read(g_range_lock_set + 2);
-		_mm_prefetch(g_range_lock_set + 4, _MM_HINT_T0);
+		utils::prefetch_read(g_range_lock_set + 4);
 
 		const auto range = utils::address_range::start_length(addr, size);
 
@@ -497,9 +499,9 @@ namespace vm
 
 			g_range_lock = addr | range_locked;
 
-			_mm_prefetch(g_range_lock_set + 0, _MM_HINT_T0);
+			utils::prefetch_read(g_range_lock_set + 0);
-			_mm_prefetch(g_range_lock_set + 2, _MM_HINT_T0);
+			utils::prefetch_read(g_range_lock_set + 2);
-			_mm_prefetch(g_range_lock_set + 4, _MM_HINT_T0);
+			utils::prefetch_read(g_range_lock_set + 4);
 
 			const auto range = utils::address_range::start_length(addr, 128);
 

rpcs3/Emu/Memory/vm.h

@@ -3,10 +3,14 @@
 #include <map>
 #include <memory>
 #include "Utilities/types.h"
-#include "Utilities/VirtualMemory.h"
 #include "Utilities/StrFmt.h"
 #include "Utilities/BEType.h"
 
+namespace utils
+{
+	class shm;
+}
+
 namespace vm
 {
 	extern u8* const g_base_addr;

rpcs3/Emu/RSX/RSXThread.cpp

@@ -14,11 +14,12 @@
 #include "Overlays/overlay_perf_metrics.h"
 #include "Utilities/date_time.h"
 #include "Utilities/span.h"
-#include "Utilities/asm.h"
 #include "Utilities/StrUtil.h"
 
 #include <cereal/archives/binary.hpp>
 
+#include "util/asm.hpp"
+
 #include <sstream>
 #include <thread>
 #include <unordered_set>

rpcs3/Emu/RSX/rsx_cache.h

@@ -1,5 +1,4 @@
 #pragma once
-#include "Utilities/VirtualMemory.h"
 #include "Utilities/hash.h"
 #include "Utilities/File.h"
 #include "Utilities/lockless.h"
@@ -15,6 +14,9 @@
 #include <thread>
 #include <chrono>
 
+
+#include "util/vm.hpp"
+
 namespace rsx
 {
 	enum protection_policy

rpcs3/emucore.vcxproj

@@ -174,7 +174,7 @@
     </ClCompile>
     <ClCompile Include="..\Utilities\Thread.cpp" />
     <ClCompile Include="..\Utilities\version.cpp" />
-    <ClCompile Include="..\Utilities\VirtualMemory.cpp" />
+    <ClCompile Include="util\vm_native.cpp" />
     <ClCompile Include="Emu\Cell\lv2\sys_config.cpp" />
     <ClCompile Include="Crypto\md5.cpp">
       <PrecompiledHeader Condition="'$(Configuration)|$(Platform)'=='Release - LLVM|x64'">NotUsing</PrecompiledHeader>
@@ -550,7 +550,8 @@
     <ClInclude Include="..\Utilities\Timer.h" />
     <ClInclude Include="..\Utilities\types.h" />
     <ClInclude Include="..\Utilities\version.h" />
-    <ClInclude Include="..\Utilities\VirtualMemory.h" />
+    <ClInclude Include="util\vm.hpp" />
+    <ClInclude Include="util\asm.hpp" />
     <ClInclude Include="Crypto\aes.h" />
     <ClInclude Include="Crypto\aesni.h" />
     <ClInclude Include="Crypto\ec.h" />

rpcs3/emucore.vcxproj.filters

@@ -188,7 +188,7 @@
     <ClCompile Include="Emu\RSX\Common\VertexProgramDecompiler.cpp">
       <Filter>Emu\GPU\RSX\Common</Filter>
     </ClCompile>
-    <ClCompile Include="..\Utilities\VirtualMemory.cpp">
+    <ClCompile Include="util\vm_native.cpp">
       <Filter>Utilities</Filter>
     </ClCompile>
     <ClCompile Include="Emu\Cell\SPURecompiler.cpp">
@@ -1165,7 +1165,10 @@
     <ClInclude Include="Emu\Cell\Common.h">
       <Filter>Emu\Cell</Filter>
     </ClInclude>
-    <ClInclude Include="..\Utilities\VirtualMemory.h">
+    <ClInclude Include="util\vm.hpp">
+      <Filter>Utilities</Filter>
+    </ClInclude>
+    <ClInclude Include="util\asm.hpp">
       <Filter>Utilities</Filter>
     </ClInclude>
     <ClInclude Include="Emu\Cell\SPUASMJITRecompiler.h">

rpcs3/util/asm.hpp

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "types.h"
+#include "Utilities/types.h"
 
 extern bool g_use_rtm;
 extern u64 g_rtm_tx_limit1;
@@ -67,10 +67,17 @@ namespace utils
 		}
 	};
 
-
-// Rotate helpers
 #if defined(__GNUG__)
 
+	inline void prefetch_read(const void* ptr)
+	{
+#if __has_builtin(__builtin_prefetch)
+		return __builtin_prefetch(ptr);
+#else
+		__asm__ volatile ("prefetcht0 0(%[ptr])" : : [ptr] "r" (ptr));
+#endif
+	}
+
 	inline u8 rol8(u8 x, u8 n)
 	{
 #if __has_builtin(__builtin_rotateleft8)
@@ -200,6 +207,11 @@ namespace utils
 	}
 
 #elif defined(_MSC_VER)
+	inline void prefetch_read(const void* ptr)
+	{
+		return _mm_prefetch(reinterpret_cast<const char*>(ptr), _MM_HINT_T0);
+	}
+
 	inline u8 rol8(u8 x, u8 n)
 	{
 		return _rotl8(x, n);

rpcs3/util/vm.hpp

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "types.h"
+#include "Utilities/types.h"
 
 namespace utils
 {
@@ -49,6 +49,7 @@ namespace utils
 #endif
 		u32 m_size;
 		u32 m_flags;
+		void* m_ptr;
 
 	public:
 		explicit shm(u32 size, u32 flags = 0);
@@ -65,12 +66,24 @@ namespace utils
 		// Map shared memory over reserved memory region, which is unsafe (non-atomic) under Win32
 		u8* map_critical(void* ptr, protection prot = protection::rw);
 
+		// Map shared memory into its own storage (not mapped by default)
+		u8* map_self(protection prot = protection::rw);
+
 		// Unmap shared memory
 		void unmap(void* ptr) const;
 
 		// Unmap shared memory, undoing map_critical
 		void unmap_critical(void* ptr);
 
+		// Unmap shared memory, undoing map_self()
+		void unmap_self();
+
+		// Get memory mapped by map_self()
+		u8* get() const
+		{
+			return reinterpret_cast<u8*>(m_ptr);
+		}
+
 		u32 size() const
 		{
 			return m_size;

rpcs3/util/vm_native.cpp

@@ -1,6 +1,6 @@
 #include "stdafx.h"
 #include "util/logs.hpp"
-#include "VirtualMemory.h"
+#include "util/vm.hpp"
 #ifdef _WIN32
 #include <Windows.h>
 #else
@@ -183,6 +183,7 @@ namespace utils
 	shm::shm(u32 size, u32 flags)
 		: m_size(::align(size, 0x10000))
 		, m_flags(flags)
+		, m_ptr(0)
 	{
 #ifdef _WIN32
 		m_handle = ::CreateFileMappingW(INVALID_HANDLE_VALUE, NULL, PAGE_EXECUTE_READWRITE, 0, m_size, NULL);
@@ -211,6 +212,8 @@ namespace utils
 
 	shm::~shm()
 	{
+		this->unmap_self();
+
 #ifdef _WIN32
 		::CloseHandle(m_handle);
 #else
@@ -284,6 +287,16 @@ namespace utils
 		return this->map(target, prot);
 	}
 
+	u8* shm::map_self(protection prot)
+	{
+		if (!m_ptr)
+		{
+			m_ptr = this->map(nullptr, prot);
+		}
+
+		return static_cast<u8*>(m_ptr);
+	}
+
 	void shm::unmap(void* ptr) const
 	{
 #ifdef _WIN32
@@ -325,4 +338,13 @@ namespace utils
 		}
 #endif
 	}
+
+	void shm::unmap_self()
+	{
+		if (m_ptr)
+		{
+			this->unmap(m_ptr);
+			m_ptr = nullptr;
+		}
+	}
 }