Fix buffer overflow on Mac

src/Renderer/QuadRasterizer.cpp

@@ -53,8 +53,8 @@ namespace sw
 
 		Do
 		{
-			Int yMin = *Pointer<Int>(primitive + OFFSET(Primitive,yMin));
-			Int yMax = *Pointer<Int>(primitive + OFFSET(Primitive,yMax));
+			yMin = *Pointer<Int>(primitive + OFFSET(Primitive,yMin));
+			yMax = *Pointer<Int>(primitive + OFFSET(Primitive,yMax));
 
 			Int cluster2 = cluster + cluster;
 			yMin += clusterCount * 2 - 2 - cluster2;
@@ -63,7 +63,7 @@ namespace sw
 
 			If(yMin < yMax)
 			{
-				rasterize(yMin, yMax);
+				rasterize();
 			}
 
 			primitive += sizeof(Primitive) * state.multiSample;
@@ -90,7 +90,7 @@ namespace sw
 		Return();
 	}
 
-	void QuadRasterizer::rasterize(Int &yMin, Int &yMax)
+	void QuadRasterizer::rasterize()
 	{
 		Pointer<Byte> cBuffer[RENDERTARGETS];
 		Pointer<Byte> zBuffer;
@@ -114,7 +114,7 @@ namespace sw
 			sBuffer = *Pointer<Pointer<Byte>>(data + OFFSET(DrawData,stencilBuffer)) + yMin * *Pointer<Int>(data + OFFSET(DrawData,stencilPitchB));
 		}
 
-		Int y = yMin;
+		y = yMin;
 
 		Do
 		{
@@ -287,7 +287,7 @@ namespace sw
 						cMask[q] = SignMask(PackSigned(mask, mask)) & 0x0000000F;
 					}
 
-					quad(cBuffer, zBuffer, sBuffer, cMask, x, y);
+					quad(cBuffer, zBuffer, sBuffer, cMask, x);
 				}
 			}
 

src/Renderer/QuadRasterizer.hpp

@@ -40,11 +40,15 @@ namespace sw
 
 		UInt occlusion;
 
+		Int y;
+		Int yMin;
+		Int yMax;
+
 #if PERF_PROFILE
 		Long cycles[PERF_TIMERS];
 #endif
 
-		virtual void quad(Pointer<Byte> cBuffer[4], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x, Int &y) = 0;
+		virtual void quad(Pointer<Byte> cBuffer[4], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x) = 0;
 
 		bool interpolateZ() const;
 		bool interpolateW() const;
@@ -54,7 +58,7 @@ namespace sw
 		const PixelShader *const shader;
 
 	private:
-		void rasterize(Int &yMin, Int &yMax);
+		void rasterize();
 	};
 }
 

src/Shader/PixelRoutine.cpp

@@ -48,7 +48,7 @@ namespace sw
 	{
 	}
 
-	void PixelRoutine::quad(Pointer<Byte> cBuffer[RENDERTARGETS], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x, Int &y)
+	void PixelRoutine::quad(Pointer<Byte> cBuffer[RENDERTARGETS], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x)
 	{
 		#if PERF_PROFILE
 			Long pipeTime = Ticks();
@@ -1684,10 +1684,17 @@ namespace sw
 					c23 |= masked;
 				}
 
-				c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD23Q) + xMask * 8);
-				value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD23Q) + xMask * 8);
-				c23 |= value;
-				*Pointer<Short4>(buffer) = c23;
+#ifdef __APPLE__
+				// On Mac we render directly to an IOSurface that isn't vertically padded. So we
+				// only render the bottom half of quads when it won't overflow the buffer.
+				If ((y + 1) < yMax)
+#endif
+				{
+					c23 &= *Pointer<Short4>(constants + OFFSET(Constants,maskD23Q) + xMask * 8);
+					value &= *Pointer<Short4>(constants + OFFSET(Constants,invMaskD23Q) + xMask * 8);
+					c23 |= value;
+					*Pointer<Short4>(buffer) = c23;
+				}
 			}
 			break;
 		case FORMAT_A8B8G8R8:

src/Shader/PixelRoutine.hpp

@@ -47,7 +47,7 @@ namespace sw
 		virtual Bool alphaTest(Int cMask[4]) = 0;
 		virtual void rasterOperation(Float4 &fog, Pointer<Byte> cBuffer[4], Int &x, Int sMask[4], Int zMask[4], Int cMask[4]) = 0;
 
-		virtual void quad(Pointer<Byte> cBuffer[4], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x, Int &y);
+		virtual void quad(Pointer<Byte> cBuffer[4], Pointer<Byte> &zBuffer, Pointer<Byte> &sBuffer, Int cMask[4], Int &x);
 
 		void alphaTest(Int &aMask, Short4 &alpha);
 		void alphaToCoverage(Int cMask[4], Float4 &alpha);