wrapper for s_pCurBufferPointer

Source/Core/VideoCommon/Src/DataReader.h

@@ -20,6 +20,8 @@
 #ifndef _DATAREADER_H
 #define _DATAREADER_H
 
+#include "VertexManagerBase.h"
+
 extern u8* g_pVideoData;
 
 #if _M_SSE >= 0x301 && !(defined __GNUC__ && !defined __SSSE3__)
@@ -145,4 +147,11 @@ __forceinline u8* DataGetPosition()
 	return g_pVideoData;
 }
 
+template <typename T>
+__forceinline void DataWrite(T data)
+{
+	*(T*)VertexManager::s_pCurBufferPointer = data;
+	VertexManager::s_pCurBufferPointer += sizeof(T);
+}
+
 #endif

Source/Core/VideoCommon/Src/VertexLoader_Color.cpp

@@ -34,8 +34,7 @@ extern int colElements[2];
 
 __forceinline void _SetCol(u32 val)
 {
-	*(u32*)VertexManager::s_pCurBufferPointer = val;
+	DataWrite(val);
-	VertexManager::s_pCurBufferPointer += 4;
 	colIndex++;
 }
 

Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp

@@ -59,16 +59,13 @@ template <typename T, int N>
 inline void ReadIndirect(const T* data)
 {
 	static_assert(3 == N || 9 == N, "N is only sane as 3 or 9!");
-	
-	auto const dest = reinterpret_cast<float*>(VertexManager::s_pCurBufferPointer);
 
 	for (int i = 0; i != N; ++i)
 	{
-		dest[i] = FracAdjust(Common::FromBigEndian(data[i]));
+		DataWrite(FracAdjust(Common::FromBigEndian(data[i])));
-		LOG_NORM();
 	}
 	
-	VertexManager::s_pCurBufferPointer += sizeof(float) * N;
+	LOG_NORM();
 }
 
 template <typename T, int N>

Source/Core/VideoCommon/Src/VertexLoader_Position.cpp

@@ -88,15 +88,10 @@ void LOADERDECL Pos_ReadDirect()
 {
 	static_assert(N <= 3, "N > 3 is not sane!");
 	
-	auto const dest = reinterpret_cast<float*>(VertexManager::s_pCurBufferPointer);
+	for (int i = 0; i < 3; ++i)
-	for (int i = 0; i != N; ++i)
+		DataWrite(i<N ? PosScale(DataRead<T>()) : 0.f);
-		dest[i] = PosScale(DataRead<T>());
-	
-	for (int i = N; i != 3; ++i)
-		dest[i] = 0.f;
 	
 	LOG_VTX();
-	VertexManager::s_pCurBufferPointer += sizeof(float) * 3;
 }
 
 template <typename I, typename T, int N>
@@ -109,16 +104,11 @@ void LOADERDECL Pos_ReadIndex()
 	if (index < std::numeric_limits<I>::max())
 	{
 		auto const data = reinterpret_cast<const T*>(cached_arraybases[ARRAY_POSITION] + (index * arraystrides[ARRAY_POSITION]));
-		auto const dest = reinterpret_cast<float*>(VertexManager::s_pCurBufferPointer);
 		
-		for (int i = 0; i != N; ++i)
+		for (int i = 0; i < 3; ++i)
-			dest[i] = PosScale(Common::FromBigEndian(data[i]));
+			DataWrite(i<N ? PosScale(Common::FromBigEndian(data[i])) : 0.f);
-		
-		for (int i = N; i != 3; ++i)
-			dest[i] = 0.f
 		
 		LOG_VTX();
-		VertexManager::s_pCurBufferPointer += sizeof(float) * 3;
 	}
 }
 
@@ -136,8 +126,8 @@ void LOADERDECL Pos_ReadIndex_Float_SSSE3()
 		GC_ALIGNED128(const __m128i a = _mm_loadu_si128((__m128i*)pData));
 		GC_ALIGNED128(__m128i b = _mm_shuffle_epi8(a, three ? kMaskSwap32_3 : kMaskSwap32_2));
 		_mm_storeu_si128((__m128i*)VertexManager::s_pCurBufferPointer, b);
-		LOG_VTX();
 		VertexManager::s_pCurBufferPointer += sizeof(float) * 3;
+		LOG_VTX();
 	}
 }
 #endif

Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp

@@ -35,14 +35,14 @@ template <>
 __forceinline void LOG_TEX<1>()
 {
 	// warning: mapping buffer should be disabled to use this
-	// PRIM_LOG("tex: %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0]);
+	// PRIM_LOG("tex: %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-1]);
 }
 
 template <>
 __forceinline void LOG_TEX<2>()
 {
 	// warning: mapping buffer should be disabled to use this
-	// PRIM_LOG("tex: %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0], ((float*)VertexManager::s_pCurBufferPointer)[1]);
+	// PRIM_LOG("tex: %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
 }
 
 extern int tcIndex;
@@ -66,14 +66,11 @@ float TCScale(float val)
 template <typename T, int N>
 void LOADERDECL TexCoord_ReadDirect()
 {
-	auto const dest = reinterpret_cast<float*>(VertexManager::s_pCurBufferPointer);
-	
 	for (int i = 0; i != N; ++i)
-		dest[i] = TCScale(DataRead<T>());
+		DataWrite(TCScale(DataRead<T>()));
 
 	LOG_TEX<N>();
 	
-	VertexManager::s_pCurBufferPointer += sizeof(float) * N;
 	++tcIndex;
 }
 
@@ -86,14 +83,10 @@ void LOADERDECL TexCoord_ReadIndex()
 	auto const data = reinterpret_cast<const T*>(cached_arraybases[ARRAY_TEXCOORD0 + tcIndex]
 		+ (index * arraystrides[ARRAY_TEXCOORD0 + tcIndex]));
 	
-	auto const dest = reinterpret_cast<float*>(VertexManager::s_pCurBufferPointer);
-	
 	for (int i = 0; i != N; ++i)
-		dest[i] = TCScale(Common::FromBigEndian(data[i]));
+		DataWrite(TCScale(Common::FromBigEndian(data[i])));
 	
 	LOG_TEX<N>();
-	
-	VertexManager::s_pCurBufferPointer += sizeof(float) * N;
 	++tcIndex;
 }
 
@@ -112,8 +105,8 @@ void LOADERDECL TexCoord_ReadIndex16_Short2_SSE4()
 	const __m128 e = _mm_load1_ps(&tcScale[tcIndex]);
 	const __m128 f = _mm_mul_ps(d, e);
 	_mm_storeu_ps((float*)VertexManager::s_pCurBufferPointer, f);
-	LOG_TEX<2>();
 	VertexManager::s_pCurBufferPointer += 8;
+	LOG_TEX<2>();
 	tcIndex++;
 }
 #endif
@@ -128,8 +121,8 @@ void LOADERDECL TexCoord_ReadIndex16_Float2_SSSE3()
 	GC_ALIGNED128(const __m128i a = _mm_loadl_epi64((__m128i*)pData));
 	GC_ALIGNED128(const __m128i b = _mm_shuffle_epi8(a, kMaskSwap32));
 	_mm_storel_epi64((__m128i*)VertexManager::s_pCurBufferPointer, b);
-	LOG_TEX<2>();
 	VertexManager::s_pCurBufferPointer += 8;
+	LOG_TEX<2>();
 	tcIndex++;
 }
 #endif

Source/Core/VideoCommon/Src/VideoCommon.h

@@ -91,7 +91,7 @@ struct TargetRectangle : public MathUtil::Rectangle<int>
 #endif
 
 // warning: mapping buffer should be disabled to use this
-// #define LOG_VTX() DEBUG_LOG(VIDEO, "vtx: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0], ((float*)VertexManager::s_pCurBufferPointer)[1], ((float*)VertexManager::s_pCurBufferPointer)[2]);
+// #define LOG_VTX() DEBUG_LOG(VIDEO, "vtx: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
 
 #define LOG_VTX()