start raster scfg bit; add stencil buffer to reset

also fix a small issue that prevented compiling

src/DSi.cpp

@@ -3173,4 +3173,9 @@ void DSi::ARM7IOWrite32(u32 addr, u32 val)
     return NDS::ARM7IOWrite32(addr, val);
 }
 
+bool DSi::GetSCFGRasterBit()
+{
+    return SCFG_EXT[0] & (1<<2);
+}
+
 }

src/DSi.h

@@ -129,6 +129,8 @@ public:
     void ARM7IOWrite16(u32 addr, u16 val) override;
     void ARM7IOWrite32(u32 addr, u32 val) override;
 
+    bool GetSCFGRasterBit() override;
+
 public:
     DSi(DSiArgs&& args) noexcept;
     ~DSi() noexcept override;

src/GPU3D.cpp

@@ -2470,7 +2470,8 @@ void GPU3D::VBlank() noexcept
                     && RenderFogOffset == FogOffset * 0x200
                     && memcmp(RenderEdgeTable, EdgeTable, 8*2) == 0
                     && memcmp(RenderFogDensityTable + 1, FogDensityTable, 32) == 0
-                    && memcmp(RenderToonTable, ToonTable, 32*2) == 0;
+                    && memcmp(RenderToonTable, ToonTable, 32*2) == 0
+                    && RenderRasterRev == NDS.GetSCFGRasterBit();
             }
 
             RenderDispCnt = DispCnt;
@@ -2488,6 +2489,8 @@ void GPU3D::VBlank() noexcept
 
             RenderClearAttr1 = ClearAttr1;
             RenderClearAttr2 = ClearAttr2;
+            // CHECKME: is this actually latched?
+            RenderRasterRev = NDS.GetSCFGRasterBit();
         }
 
         if (FlushRequest)

src/GPU3D.h

@@ -271,6 +271,8 @@ public:
     u32 RenderClearAttr1 = 0;
     u32 RenderClearAttr2 = 0;
 
+    bool RenderRasterRev = false;
+
     bool RenderFrameIdentical = false; // not part of the hardware state, don't serialize
 
     bool AbortFrame = false;

src/GPU3D_Soft.cpp

@@ -121,6 +121,7 @@ void SoftRenderer::Reset(GPU& gpu)
     memset(ColorBuffer, 0, BufferSize * 2 * 4);
     memset(DepthBuffer, 0, BufferSize * 2 * 4);
     memset(AttrBuffer, 0, BufferSize * 2 * 4);
+    memset(StencilBuffer, 0, 256*2); // CHECKME?
 
     PrevIsShadowMask = false;
 
@@ -713,7 +714,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
     else
         fnDepthTest = DepthTest_LessThan;
 
-    if (!PrevIsShadowMask)
+    if (!PrevIsShadowMask && !gpu3d.RenderRasterRev) // the "Revised" Rasterizer Circuit bugs out stencil buffer clearing
         memset(&StencilBuffer[256 * (y&0x1)], 0, 256);
 
     PrevIsShadowMask = true;
@@ -771,8 +772,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
         std::swap(wl, wr);
         std::swap(zl, zr);
 
-        // CHECKME: edge fill rules for swapped opaque shadow mask polygons
-        if ((gpu3d.RenderDispCnt & ((1<<4)|(1<<5))) || wireframe)
+        if ((gpu3d.RenderDispCnt & ((1<<4)|(1<<5))) || !(polygon->Attr & (0x1F << 16)))
         {
             l_filledge = true;
             r_filledge = true;
@@ -799,8 +799,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
         rp->SlopeL.EdgeParams<false>(&l_edgelen, &l_edgecov);
         rp->SlopeR.EdgeParams<false>(&r_edgelen, &r_edgecov);
 
-        // CHECKME: edge fill rules for unswapped opaque shadow mask polygons
-        if ((gpu3d.RenderDispCnt & ((1<<4)|(1<<5))) || wireframe)
+        if ((gpu3d.RenderDispCnt & ((1<<4)|(1<<5))) || !(polygon->Attr & (0x1F << 16)))
         {
             l_filledge = true;
             r_filledge = true;

src/NDS.cpp

@@ -4334,4 +4334,9 @@ void NDS::ARM7IOWrite32(u32 addr, u32 val)
     Log(LogLevel::Debug, "unknown ARM7 IO write32 %08X %08X %08X\n", addr, val, ARM7.R[15]);
 }
 
+bool NDS::GetSCFGRasterBit()
+{
+    return false;
+}
+
 }

src/NDS.h

@@ -471,6 +471,8 @@ public: // TODO: Encapsulate the rest of these members
     void SetJITArgs(std::optional<JITArgs> args) noexcept {}
 #endif
 
+    virtual bool GetSCFGRasterBit();
+   
 private:
     void InitTimings();
     u32 SchedListMask;