Help compiler generate some better Unpack code.

jpd002 · Jul 16, 2024 · 3303a20 · 3303a20
1 parent f4b3d75
commit 3303a20
Show file tree

Hide file tree

Showing 2 changed files with 40 additions and 34 deletions.
diff --git a/Source/ee/Vif.cpp b/Source/ee/Vif.cpp
@@ -711,12 +711,10 @@ void CVif::Cmd_UNPACK(StreamType& stream, CODE nCommand, uint32 nDstAddr)
 	((*this).*(unpackFct))(stream, nCommand, nDstAddr);
 }
 
-uint32 CVif::GetMaskOp(unsigned int row, unsigned int col) const
+uint32 CVif::GetColMaskOp(unsigned int col) const
 {
-	if(col > 3) col = 3;
-	assert(row < 4);
-	unsigned int index = (col * 4) + row;
-	return (m_MASK >> (index * 2)) & 0x03;
+	assert(col < 4);
+	return (m_MASK >> (col * 8)) & 0xFF;
 }
 
 void CVif::PrepareMicroProgram()

diff --git a/Source/ee/Vif.h b/Source/ee/Vif.h
@@ -317,7 +317,7 @@ class CVif
 	void Cmd_STCOL(StreamType&, CODE);
 	void Cmd_STMASK(StreamType&, CODE);
 
-	inline uint32 GetMaskOp(unsigned int, unsigned int) const;
+	inline uint32 GetColMaskOp(unsigned int) const;
 
 	inline bool Unpack_S32(StreamType& stream, uint128& result)
 	{
@@ -577,41 +577,49 @@ class CVif
 			if(mustWrite)
 			{
 				auto dst = reinterpret_cast<uint128*>(vuMem + nDstAddr);
+				uint32 col = (m_writeTick > 3) ? 3 : m_writeTick;
+				uint32 colMask = useMask ? GetColMaskOp(col) : 0;
 
-				for(unsigned int i = 0; i < 4; i++)
+				if((colMask == 0) && (mode == 0))
 				{
-					uint32 maskOp = useMask ? GetMaskOp(i, m_writeTick) : MASK_DATA;
-
-					if(maskOp == MASK_DATA)
+					(*dst) = writeValue;
+				}
+				else
+				{
+					for(unsigned int i = 0; i < 4; i++)
 					{
-						if(mode == MODE_OFFSET)
+						uint32 maskOp = useMask ? (colMask >> (i * 2)) & 0x3 : MASK_DATA;
+
+						if(maskOp == MASK_DATA)
 						{
-							writeValue.nV[i] += m_R[i];
+							if(mode == MODE_OFFSET)
+							{
+								writeValue.nV[i] += m_R[i];
+							}
+							else if(mode == MODE_DIFFERENCE)
+							{
+								writeValue.nV[i] += m_R[i];
+								m_R[i] = writeValue.nV[i];
+							}
+
+							dst->nV[i] = writeValue.nV[i];
 						}
-						else if(mode == MODE_DIFFERENCE)
+						else if(maskOp == MASK_ROW)
 						{
-							writeValue.nV[i] += m_R[i];
-							m_R[i] = writeValue.nV[i];
+							dst->nV[i] = m_R[i];
+						}
+						else if(maskOp == MASK_COL)
+						{
+							dst->nV[i] = m_C[col];
+						}
+						else if(maskOp == MASK_MASK)
+						{
+							//Don't write anything
+						}
+						else
+						{
+							assert(0);
 						}
-
-						dst->nV[i] = writeValue.nV[i];
-					}
-					else if(maskOp == MASK_ROW)
-					{
-						dst->nV[i] = m_R[i];
-					}
-					else if(maskOp == MASK_COL)
-					{
-						int index = (m_writeTick > 3) ? 3 : m_writeTick;
-						dst->nV[i] = m_C[index];
-					}
-					else if(maskOp == MASK_MASK)
-					{
-						//Don't write anything
-					}
-					else
-					{
-						assert(0);
 					}
 				}