a few attempts at optimization

melonDS.cbp

@@ -51,12 +51,12 @@
 				<Option type="1" />
 				<Option compiler="gcc" />
 				<Compiler>
-					<Add option="-O2" />
+					<Add option="-O3" />
 					<Add option="-m64" />
-					<Add option="-D_FILE_OFFSET_BITS=64" />
-					<Add option="-D__WXMSW__" />
 					<Add option="-I$(TARGET_COMPILER_DIR)/lib/wx/include/msw-unicode-static-3.0" />
 					<Add option="-I$(TARGET_COMPILER_DIR)/include/wx-3.0" />
+					<Add option="-D_FILE_OFFSET_BITS=64" />
+					<Add option="-D__WXMSW__" />
 				</Compiler>
 				<Linker>
 					<Add option="-s" />
@@ -116,6 +116,45 @@
 					<Add option="-m64" />
 				</Linker>
 			</Target>
+			<Target title="Profile Windows">
+				<Option platforms="Windows;" />
+				<Option output="bin/Release/melonDS" prefix_auto="1" extension_auto="1" />
+				<Option object_output="obj/Release/" />
+				<Option type="1" />
+				<Option compiler="gcc" />
+				<Compiler>
+					<Add option="-O2" />
+					<Add option="-m64" />
+					<Add option="-g" />
+					<Add option="-I$(TARGET_COMPILER_DIR)/lib/wx/include/msw-unicode-static-3.0" />
+					<Add option="-I$(TARGET_COMPILER_DIR)/include/wx-3.0" />
+					<Add option="-D_FILE_OFFSET_BITS=64" />
+					<Add option="-D__WXMSW__" />
+				</Compiler>
+				<Linker>
+					<Add option="-m64" />
+					<Add library=":libwx_mswu_core-3.0.a" />
+					<Add library=":libwx_baseu-3.0.a" />
+					<Add library="SDL2" />
+					<Add library=":libpng.a" />
+					<Add library=":libjpeg.a" />
+					<Add library=":libtiff.a" />
+					<Add library=":libz.a" />
+					<Add library="rpcrt4" />
+					<Add library="oleaut32" />
+					<Add library="ole32" />
+					<Add library="uuid" />
+					<Add library="winspool" />
+					<Add library="winmm" />
+					<Add library="shell32" />
+					<Add library="comctl32" />
+					<Add library="comdlg32" />
+					<Add library="advapi32" />
+					<Add library="wsock32" />
+					<Add library="oleacc" />
+					<Add library="gdi32" />
+				</Linker>
+			</Target>
 		</Build>
 		<Compiler>
 			<Add option="-Wall" />
@@ -126,6 +165,7 @@
 			<Option compilerVar="WINDRES" />
 			<Option target="Debug Windows" />
 			<Option target="Release Windows" />
+			<Option target="Profile Windows" />
 		</Unit>
 		<Unit filename="src/ARM.cpp" />
 		<Unit filename="src/ARM.h" />
@@ -176,6 +216,7 @@
 		<Unit filename="xp.manifest">
 			<Option target="Debug Windows" />
 			<Option target="Release Windows" />
+			<Option target="Profile Windows" />
 		</Unit>
 		<Extensions>
 			<code_completion />

src/ARM.cpp

@@ -332,7 +332,7 @@ s32 ARM::Execute()
         else if (NDS::HaltInterrupted(Num))
         {
             Halted = 0;
-            if (NDS::IME[Num]&1)
+            if (NDS::IME[Num] & 0x1)
                 TriggerIRQ();
         }
         else
@@ -403,9 +403,9 @@ s32 ARM::Execute()
                 Cycles = CyclesToRun;
             break;
         }
-        if (NDS::HaltInterrupted(Num))
+        if (NDS::IF[Num] & NDS::IE[Num])
         {
-            if (NDS::IME[Num]&1)
+            if (NDS::IME[Num] & 0x1)
                 TriggerIRQ();
         }
     }

src/DMA.cpp

@@ -233,6 +233,23 @@ s32 DMA::Run(s32 cycles)
     }
     else
     {
+        // optimized path for typical GXFIFO DMA
+        if (CPU == 0 && (CurSrcAddr>>24) == 0x02 && CurDstAddr == 0x04000400 && DstAddrInc == 0)
+        {
+            while (IterCount > 0 && cycles > 0)
+            {
+                GPU3D::WriteToGXFIFO(*(u32*)&NDS::MainRAM[CurSrcAddr&0x3FFFFF]);
+
+                s32 c = (Waitstates[1][0x2] + Waitstates[1][0x4]);
+                cycles -= c;
+                NDS::RunTimingCriticalDevices(0, c);
+
+                CurSrcAddr += SrcAddrInc<<2;
+                IterCount--;
+                RemCount--;
+            }
+        }
+
         u32 (*readfn)(u32) = CPU ? NDS::ARM7Read32 : NDS::ARM9Read32;
         void (*writefn)(u32,u32) = CPU ? NDS::ARM7Write32 : NDS::ARM9Write32;
 

src/GPU3D.cpp

@@ -1807,6 +1807,45 @@ u32* GetLine(int line)
 }
 
 
+void WriteToGXFIFO(u32 val)
+{
+    if (NumCommands == 0)
+    {
+        NumCommands = 4;
+        CurCommand = val;
+        ParamCount = 0;
+        TotalParams = CmdNumParams[CurCommand & 0xFF];
+
+        if (TotalParams > 0) return;
+    }
+    else
+        ParamCount++;
+
+    for (;;)
+    {
+        if ((CurCommand & 0xFF) || (NumCommands == 4 && CurCommand == 0))
+        {
+            CmdFIFOEntry entry;
+            entry.Command = CurCommand & 0xFF;
+            entry.Param = val;
+            CmdFIFOWrite(entry);
+        }
+
+        if (ParamCount >= TotalParams)
+        {
+            CurCommand >>= 8;
+            NumCommands--;
+            if (NumCommands == 0) break;
+
+            ParamCount = 0;
+            TotalParams = CmdNumParams[CurCommand & 0xFF];
+        }
+        if (ParamCount < TotalParams)
+            break;
+    }
+}
+
+
 u8 Read8(u32 addr)
 {
     printf("unknown GPU3D read8 %08X\n", addr);
@@ -2012,41 +2051,7 @@ void Write32(u32 addr, u32 val)
 
     if (addr >= 0x04000400 && addr < 0x04000440)
     {
-        if (NumCommands == 0)
+        WriteToGXFIFO(val);
-        {
-            NumCommands = 4;
-            CurCommand = val;
-            ParamCount = 0;
-            TotalParams = CmdNumParams[CurCommand & 0xFF];
-
-            if (TotalParams > 0) return;
-        }
-        else
-            ParamCount++;
-
-        for (;;)
-        {
-            if ((CurCommand & 0xFF) || (NumCommands == 4 && CurCommand == 0))
-            {
-                CmdFIFOEntry entry;
-                entry.Command = CurCommand & 0xFF;
-                entry.Param = val;
-                CmdFIFOWrite(entry);
-            }
-
-            if (ParamCount >= TotalParams)
-            {
-                CurCommand >>= 8;
-                NumCommands--;
-                if (NumCommands == 0) break;
-
-                ParamCount = 0;
-                TotalParams = CmdNumParams[CurCommand & 0xFF];
-            }
-            if (ParamCount < TotalParams)
-                break;
-        }
-
         return;
     }
 

src/GPU3D.h

@@ -85,6 +85,8 @@ void VBlank();
 void VCount215();
 u32* GetLine(int line);
 
+void WriteToGXFIFO(u32 val);
+
 u8 Read8(u32 addr);
 u16 Read16(u32 addr);
 u32 Read32(u32 addr);

src/GPU3D_Soft.cpp

@@ -541,7 +541,6 @@ bool DepthTest(s32 oldz, s32 z)
 
 u32 RenderPixel(Polygon* polygon, u8 vr, u8 vg, u8 vb, s16 s, s16 t)
 {
-    u32 attr = polygon->Attr;
     u8 r, g, b, a;
 
     u32 blendmode = (polygon->Attr >> 4) & 0x3;
@@ -910,7 +909,10 @@ void RenderPolygon(Polygon* polygon)
 
             // wireframe polygons. really ugly, but works
             if (wireframe && edge==0)
+            {
+                x = r_edgestart + 1;
                 continue;
+            }
 
             u32 pixeladdr = (y*256) + x;
             u32 attr = polygon->Attr & 0x3F008000;