Merge pull request #2480 from phire/GamesDoWeirdShit

Fix invalid pointer errors in Burnout 2.

Source/Core/VideoBackends/Software/CPMemLoader.cpp

@@ -23,11 +23,13 @@ void SWLoadCPReg(u32 sub_cmd, u32 value)
 	case 0x50:
 		g_main_cp_state.vtx_desc.Hex &= ~0x1FFFF;  // keep the Upper bits
 		g_main_cp_state.vtx_desc.Hex |= value;
+		g_main_cp_state.bases_dirty = true;
 		break;
 
 	case 0x60:
 		g_main_cp_state.vtx_desc.Hex &= 0x1FFFF;  // keep the lower 17Bits
 		g_main_cp_state.vtx_desc.Hex |= (u64)value << 17;
+		g_main_cp_state.bases_dirty = true;
 		break;
 
 	case 0x70:
@@ -48,7 +50,7 @@ void SWLoadCPReg(u32 sub_cmd, u32 value)
 	// Pointers to vertex arrays in GC RAM
 	case 0xA0:
 		g_main_cp_state.array_bases[sub_cmd & 0xF] = value;
-		cached_arraybases[sub_cmd & 0xF] = Memory::GetPointer(value);
+		g_main_cp_state.bases_dirty = true;
 		break;
 
 	case 0xB0:

Source/Core/VideoBackends/Software/SWVertexLoader.cpp

@@ -16,8 +16,10 @@
 #include "VideoBackends/Software/XFMemLoader.h"
 
 #include "VideoCommon/VertexLoaderBase.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexLoaderUtils.h"
 
+
 SWVertexLoader::SWVertexLoader() :
 	m_VertexSize(0)
 {
@@ -176,6 +178,8 @@ void SWVertexLoader::LoadVertex()
 	// reserve memory for the destination of the vertex loader
 	m_LoadedVertices.resize(vdec.stride + 4);
 
+	VertexLoaderManager::UpdateVertexArrayPointers();
+
 	// convert the vertex from the gc format to the videocommon (hardware optimized) format
 	u8* old = g_video_buffer_read_ptr;
 	int converted_vertices = m_CurrentLoader->RunVertices(

Source/Core/VideoCommon/CPMemory.cpp

@@ -7,8 +7,6 @@
 #include "VideoCommon/CPMemory.h"
 
 // CP state
-u8 *cached_arraybases[16];
-
 CPState g_main_cp_state;
 CPState g_preprocess_cp_state;
 
@@ -24,7 +22,10 @@ void DoCPState(PointerWrap& p)
 	p.DoArray(g_main_cp_state.vtx_attr, 8);
 	p.DoMarker("CP Memory");
 	if (p.mode == PointerWrap::MODE_READ)
+	{
 		CopyPreprocessCPStateFromMain();
+		g_main_cp_state.bases_dirty = true;
+	}
 }
 
 void CopyPreprocessCPStateFromMain()

Source/Core/VideoCommon/CPMemory.h

@@ -88,6 +88,12 @@ union TVtxDesc
 	{
 		u32 Hex0, Hex1;
 	};
+
+	// Easily index into the Position..Tex7Coord fields.
+	u32 GetVertexArrayStatus(int idx)
+	{
+		return (Hex >> (9 + idx * 2)) & 0x3;
+	}
 };
 
 union UVAT_group0
@@ -249,6 +255,7 @@ struct CPState final
 
 	// Attributes that actually belong to VertexLoaderManager:
 	BitSet32 attr_dirty;
+	bool bases_dirty;
 	VertexLoaderBase* vertex_loaders[8];
 };
 
@@ -261,7 +268,6 @@ extern void CopyPreprocessCPStateFromMain();
 extern CPState g_main_cp_state;
 extern CPState g_preprocess_cp_state;
 
-extern u8 *cached_arraybases[16];
 
 // Might move this into its own file later.
 void LoadCPReg(u32 SubCmd, u32 Value, bool is_preprocess = false);

Source/Core/VideoCommon/MainBase.cpp

@@ -212,7 +212,6 @@ void VideoBackendHardware::DoState(PointerWrap& p)
 	if (p.GetMode() == PointerWrap::MODE_READ)
 	{
 		m_invalid = true;
-		RecomputeCachedArraybases();
 
 		// Clear all caches that touch RAM
 		// (? these don't appear to touch any emulation state that gets saved. moved to on load only.)

Source/Core/VideoCommon/VertexLoaderARM64.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "VideoCommon/VertexLoaderARM64.h"
+#include "VideoCommon/VertexLoaderManager.h"
 
 using namespace Arm64Gen;
 
@@ -331,7 +332,7 @@ void VertexLoaderARM64::GenerateVertexLoader()
 	MOV(saved_count, count_reg);
 
 	MOVI2R(stride_reg, (u64)&g_main_cp_state.array_strides);
-	MOVI2R(arraybase_reg, (u64)&cached_arraybases);
+	MOVI2R(arraybase_reg, (u64)&VertexLoaderManager::cached_arraybases);
 	MOVI2R(scale_reg, (u64)&scale_factors);
 
 	const u8* loop_start = GetCodePtr();

Source/Core/VideoCommon/VertexLoaderManager.cpp

@@ -35,6 +35,8 @@ static std::mutex s_vertex_loader_map_lock;
 static VertexLoaderMap s_vertex_loader_map;
 // TODO - change into array of pointers. Keep a map of all seen so far.
 
+u8 *cached_arraybases[12];
+
 void Init()
 {
 	MarkAllDirty();
@@ -42,7 +44,6 @@ void Init()
 		map_entry = nullptr;
 	for (auto& map_entry : g_preprocess_cp_state.vertex_loaders)
 		map_entry = nullptr;
-	RecomputeCachedArraybases();
 	SETSTAT(stats.numVertexLoaders, 0);
 }
 
@@ -53,6 +54,27 @@ void Shutdown()
 	s_native_vertex_map.clear();
 }
 
+void UpdateVertexArrayPointers()
+{
+	// Anything to update?
+	if (!g_main_cp_state.bases_dirty)
+		return;
+
+	// Some games such as Burnout 2 can put invalid addresses into
+	// the array base registers. (see issue 8591)
+	// But the vertex arrays with invalid addresses aren't actually enabled.
+	// Note: Only array bases 0 through 11 are used by the Vertex loaders.
+	//       12 through 15 are used for loading data into xfmem.
+	for (int i = 0; i < 12; i++)
+	{
+		// Only update the array base if the vertex description states we are going to use it.
+		if (g_main_cp_state.vtx_desc.GetVertexArrayStatus(i) >= 0x2)
+			cached_arraybases[i] = Memory::GetPointer(g_main_cp_state.array_bases[i]);
+	}
+
+	g_main_cp_state.bases_dirty = false;
+}
+
 namespace
 {
 struct entry
@@ -136,6 +158,11 @@ static VertexLoaderBase* RefreshLoader(int vtx_attr_group, bool preprocess = fal
 	} else {
 		loader = state->vertex_loaders[vtx_attr_group];
 	}
+
+	// Lookup pointers for any vertex arrays.
+	if (!preprocess)
+		UpdateVertexArrayPointers();
+
 	return loader;
 }
 
@@ -203,12 +230,14 @@ void LoadCPReg(u32 sub_cmd, u32 value, bool is_preprocess)
 		state->vtx_desc.Hex &= ~0x1FFFF;  // keep the Upper bits
 		state->vtx_desc.Hex |= value;
 		state->attr_dirty = BitSet32::AllTrue(8);
+		state->bases_dirty = true;
 		break;
 
 	case 0x60:
 		state->vtx_desc.Hex &= 0x1FFFF;  // keep the lower 17Bits
 		state->vtx_desc.Hex |= (u64)value << 17;
 		state->attr_dirty = BitSet32::AllTrue(8);
+		state->bases_dirty = true;
 		break;
 
 	case 0x70:
@@ -232,8 +261,7 @@ void LoadCPReg(u32 sub_cmd, u32 value, bool is_preprocess)
 	// Pointers to vertex arrays in GC RAM
 	case 0xA0:
 		state->array_bases[sub_cmd & 0xF] = value;
-		if (update_global_state)
+		state->bases_dirty = true;
-			cached_arraybases[sub_cmd & 0xF] = Memory::GetPointer(value);
 		break;
 
 	case 0xB0:
@@ -262,11 +290,3 @@ void FillCPMemoryArray(u32 *memory)
 		memory[0xB0 + i] = g_main_cp_state.array_strides[i];
 	}
 }
-
-void RecomputeCachedArraybases()
-{
-	for (int i = 0; i < 16; i++)
-	{
-		cached_arraybases[i] = Memory::GetPointer(g_main_cp_state.array_bases[i]);
-	}
-}

Source/Core/VideoCommon/VertexLoaderManager.h

@@ -24,6 +24,9 @@ namespace VertexLoaderManager
 	void AppendListToString(std::string *dest);
 
 	NativeVertexFormat* GetCurrentVertexFormat();
+
+	// Resolved pointers to array bases. Used by vertex loaders.
+	extern u8 *cached_arraybases[12];
+	void UpdateVertexArrayPointers();
 }
 
-void RecomputeCachedArraybases();

Source/Core/VideoCommon/VertexLoaderX64.cpp

@@ -7,6 +7,7 @@
 #include "Common/Intrinsics.h"
 #include "Common/JitRegister.h"
 #include "Common/x64ABI.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexLoaderX64.h"
 
 using namespace Gen;
@@ -58,7 +59,7 @@ OpArg VertexLoaderX64::GetVertexAddr(int array, u64 attribute)
 		}
 		// TODO: Move cached_arraybases into CPState and use MDisp() relative to a constant register loaded with &g_main_cp_state.
 		IMUL(32, scratch1, M(&g_main_cp_state.array_strides[array]));
-		MOV(64, R(scratch2), M(&cached_arraybases[array]));
+		MOV(64, R(scratch2), M(&VertexLoaderManager::cached_arraybases[array]));
 		return MRegSum(scratch1, scratch2);
 	}
 	else

Source/Core/VideoCommon/VertexLoader_Color.cpp

@@ -6,6 +6,7 @@
 
 #include "VideoCommon/VertexLoader.h"
 #include "VideoCommon/VertexLoader_Color.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexManagerBase.h"
 #include "VideoCommon/VideoCommon.h"
 
@@ -100,7 +101,7 @@ template <typename I>
 void Color_ReadIndex_16b_565(VertexLoader* loader)
 {
 	auto const Index = DataRead<I>();
-	u16 val = Common::swap16(*(const u16 *)(cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex])));
+	u16 val = Common::swap16(*(const u16 *)(VertexLoaderManager::cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex])));
 	_SetCol565(loader, val);
 }
 
@@ -108,7 +109,7 @@ template <typename I>
 void Color_ReadIndex_24b_888(VertexLoader* loader)
 {
 	auto const Index = DataRead<I>();
-	const u8 *iAddress = cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]);
+	const u8 *iAddress = VertexLoaderManager::cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]);
 	_SetCol(loader, _Read24(iAddress));
 }
 
@@ -116,7 +117,7 @@ template <typename I>
 void Color_ReadIndex_32b_888x(VertexLoader* loader)
 {
 	auto const Index = DataRead<I>();
-	const u8 *iAddress = cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]);
+	const u8 *iAddress = VertexLoaderManager::cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]);
 	_SetCol(loader, _Read24(iAddress));
 }
 
@@ -124,7 +125,7 @@ template <typename I>
 void Color_ReadIndex_16b_4444(VertexLoader* loader)
 {
 	auto const Index = DataRead<I>();
-	u16 val = *(const u16 *)(cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]));
+	u16 val = *(const u16 *)(VertexLoaderManager::cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]));
 	_SetCol4444(loader, val);
 }
 
@@ -132,7 +133,7 @@ template <typename I>
 void Color_ReadIndex_24b_6666(VertexLoader* loader)
 {
 	auto const Index = DataRead<I>();
-	const u8* pData = cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]) - 1;
+	const u8* pData = VertexLoaderManager::cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]) - 1;
 	u32 val = Common::swap32(pData);
 	_SetCol6666(loader, val);
 }
@@ -141,7 +142,7 @@ template <typename I>
 void Color_ReadIndex_32b_8888(VertexLoader* loader)
 {
 	auto const Index = DataRead<I>();
-	const u8 *iAddress = cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]);
+	const u8 *iAddress = VertexLoaderManager::cached_arraybases[ARRAY_COLOR + loader->m_colIndex] + (Index * g_main_cp_state.array_strides[ARRAY_COLOR + loader->m_colIndex]);
 	_SetCol(loader, _Read32(iAddress));
 }
 

Source/Core/VideoCommon/VertexLoader_Normal.cpp

@@ -8,6 +8,7 @@
 #include "Common/CommonTypes.h"
 #include "VideoCommon/VertexLoader.h"
 #include "VideoCommon/VertexLoader_Normal.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexManagerBase.h"
 #include "VideoCommon/VideoCommon.h"
 
@@ -71,7 +72,7 @@ __forceinline void Normal_Index_Offset()
 	static_assert(std::is_unsigned<I>::value, "Only unsigned I is sane!");
 
 	auto const index = DataRead<I>();
-	auto const data = reinterpret_cast<const T*>(cached_arraybases[ARRAY_NORMAL]
+	auto const data = reinterpret_cast<const T*>(VertexLoaderManager::cached_arraybases[ARRAY_NORMAL]
 	                + (index * g_main_cp_state.array_strides[ARRAY_NORMAL]) + sizeof(T) * 3 * Offset);
 	ReadIndirect<T, N * 3>(data);
 }

Source/Core/VideoCommon/VertexLoader_Position.cpp

@@ -7,6 +7,7 @@
 #include "Common/CommonTypes.h"
 #include "VideoCommon/VertexLoader.h"
 #include "VideoCommon/VertexLoader_Position.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexManagerBase.h"
 #include "VideoCommon/VideoCommon.h"
 
@@ -46,7 +47,7 @@ void LOADERDECL Pos_ReadIndex(VertexLoader* loader)
 
 	auto const index = DataRead<I>();
 	loader->m_vertexSkip = index == std::numeric_limits<I>::max();
-	auto const data = reinterpret_cast<const T*>(cached_arraybases[ARRAY_POSITION] + (index * g_main_cp_state.array_strides[ARRAY_POSITION]));
+	auto const data = reinterpret_cast<const T*>(VertexLoaderManager::cached_arraybases[ARRAY_POSITION] + (index * g_main_cp_state.array_strides[ARRAY_POSITION]));
 	auto const scale = loader->m_posScale;
 	DataReader dst(g_vertex_manager_write_ptr, nullptr);
 

Source/Core/VideoCommon/VertexLoader_TextCoord.cpp

@@ -7,6 +7,7 @@
 #include "Common/CommonTypes.h"
 #include "VideoCommon/VertexLoader.h"
 #include "VideoCommon/VertexLoader_TextCoord.h"
+#include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexManagerBase.h"
 #include "VideoCommon/VideoCommon.h"
 
@@ -67,7 +68,7 @@ void LOADERDECL TexCoord_ReadIndex(VertexLoader* loader)
 	static_assert(std::is_unsigned<I>::value, "Only unsigned I is sane!");
 
 	auto const index = DataRead<I>();
-	auto const data = reinterpret_cast<const T*>(cached_arraybases[ARRAY_TEXCOORD0 + loader->m_tcIndex]
+	auto const data = reinterpret_cast<const T*>(VertexLoaderManager::cached_arraybases[ARRAY_TEXCOORD0 + loader->m_tcIndex]
 	                + (index * g_main_cp_state.array_strides[ARRAY_TEXCOORD0 + loader->m_tcIndex]));
 	auto const scale = loader->m_tcScale[loader->m_tcIndex];
 	DataReader dst(g_vertex_manager_write_ptr, nullptr);

Source/UnitTests/VideoCommon/VertexLoaderTest.cpp

@@ -16,6 +16,7 @@
 #include "VideoCommon/DataReader.h"
 #include "VideoCommon/OpcodeDecoding.h"
 #include "VideoCommon/VertexLoaderBase.h"
+#include "VideoCommon/VertexLoaderManager.h"
 
 TEST(VertexLoaderUID, UniqueEnough)
 {
@@ -151,7 +152,7 @@ TEST_P(VertexLoaderParamTest, PositionAll)
 				Input<u8>(i);
 			else
 				Input<u16>(i);
-		cached_arraybases[ARRAY_POSITION] = m_src.GetPointer();
+		VertexLoaderManager::cached_arraybases[ARRAY_POSITION] = m_src.GetPointer();
 		g_main_cp_state.array_strides[ARRAY_POSITION] = elements * elem_size;
 	}
 	CreateAndCheckSizes(input_size, elements * sizeof(float));
@@ -192,7 +193,7 @@ TEST_F(VertexLoaderTest, PositionIndex16FloatXY)
 	m_vtx_attr.g0.PosFormat = FORMAT_FLOAT;
 	CreateAndCheckSizes(sizeof(u16), 2 * sizeof(float));
 	Input<u16>(1); Input<u16>(0);
-	cached_arraybases[ARRAY_POSITION] = m_src.GetPointer();
+	VertexLoaderManager::cached_arraybases[ARRAY_POSITION] = m_src.GetPointer();
 	g_main_cp_state.array_strides[ARRAY_POSITION] = sizeof(float); // ;)
 	Input(1.f); Input(2.f); Input(3.f);
 	RunVertices(2);
@@ -297,9 +298,9 @@ TEST_F(VertexLoaderTest, LargeFloatVertexSpeed)
 
 	CreateAndCheckSizes(33, 156);
 
-	for (int i = 0; i < 16; i++)
+	for (int i = 0; i < 12; i++)
 	{
-		cached_arraybases[i] = m_src.GetPointer();
+		VertexLoaderManager::cached_arraybases[i] = m_src.GetPointer();
 		g_main_cp_state.array_strides[i] = 129;
 	}