merge moar

7 files changed, 294 insertions, 163 deletions

diff --git a/.gitignore b/.gitignore
index bd1d485..dd81614 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,3 +1,4 @@
+build
 bin
 obj
 *.depend
diff --git a/src/ARM.h b/src/ARM.h
index 6b6e330..a8a7bb3 100644
--- a/src/ARM.h
+++ b/src/ARM.h
@@ -212,7 +212,8 @@ public:
     void UpdateDTCMSetting();
     void UpdateITCMSetting();
 
-    void UpdatePURegions();
+    void UpdatePURegion(u32 n);
+    void UpdatePURegions(bool update_all);
 
     u32 RandomLineIndex();
 
diff --git a/src/CP15.cpp b/src/CP15.cpp
index 4654a2a..d412db6 100644
--- a/src/CP15.cpp
+++ b/src/CP15.cpp
@@ -60,7 +60,7 @@ void ARMv5::CP15Reset()
     PU_DataRW = 0;
 
     memset(PU_Region, 0, 8*sizeof(u32));
-    UpdatePURegions();
+    UpdatePURegions(true);
 
     CurICacheLine = NULL;
 }
@@ -90,7 +90,7 @@ void ARMv5::CP15DoSavestate(Savestate* file)
     {
         UpdateDTCMSetting();
         UpdateITCMSetting();
-        UpdatePURegions();
+        UpdatePURegions(true);
     }
 }
 
@@ -126,27 +126,12 @@ void ARMv5::UpdateITCMSetting()
 }
 
 
-void ARMv5::UpdatePURegions()
+// covers updates to a specific PU region's cache/etc settings
+// (not to the region range/enabled status)
+void ARMv5::UpdatePURegion(u32 n)
 {
-    if (!(CP15Control & (1<<0)))
-    {
-        // PU disabled
-
-        u8 mask = 0x07;
-        if (CP15Control & (1<<2))  mask |= 0x30;
-        if (CP15Control & (1<<12)) mask |= 0x40;
-
-        memset(PU_UserMap, mask, 0x100000);
-        memset(PU_PrivMap, mask, 0x100000);
-
-        return;
-    }
-
-    memset(PU_UserMap, 0, 0x100000);
-    memset(PU_PrivMap, 0, 0x100000);
-
-    u32 coderw = PU_CodeRW;
-    u32 datarw = PU_DataRW;
+    u32 coderw = (PU_CodeRW >> (4*n)) & 0xF;
+    u32 datarw = (PU_DataRW >> (4*n)) & 0xF;
 
     u32 codecache, datacache, datawrite;
 
@@ -157,14 +142,14 @@ void ARMv5::UpdatePURegions()
     // 1/1: goes to cache
 
     if (CP15Control & (1<<12))
-        codecache = PU_CodeCacheable;
+        codecache = (PU_CodeCacheable >> n) & 0x1;
     else
         codecache = 0;
 
     if (CP15Control & (1<<2))
     {
-        datacache = PU_DataCacheable;
-        datawrite = PU_DataCacheWrite;
+        datacache = (PU_DataCacheable >> n) & 0x1;
+        datawrite = (PU_DataCacheWrite >> n) & 0x1;
     }
     else
     {
@@ -172,88 +157,102 @@ void ARMv5::UpdatePURegions()
         datawrite = 0;
     }
 
-    for (int n = 0; n < 8; n++)
+    u32 rgn = PU_Region[n];
+    if (!(rgn & (1<<0)))
     {
-        u32 rgn = PU_Region[n];
-        if (!(rgn & (1<<0)))
-        {
-            coderw >>= 4;
-            datarw >>= 4;
-            codecache >>= 1;
-            datacache >>= 1;
-            datawrite >>= 1;
-            continue;
-        }
+        return;
+    }
 
-        u32 start = rgn >> 12;
-        u32 sz = 2 << ((rgn >> 1) & 0x1F);
-        u32 end = start + (sz >> 12);
-        // TODO: check alignment of start
+    u32 start = rgn >> 12;
+    u32 sz = 2 << ((rgn >> 1) & 0x1F);
+    u32 end = start + (sz >> 12);
+    // TODO: check alignment of start
 
-        u8 usermask = 0;
-        u8 privmask = 0;
+    u8 usermask = 0;
+    u8 privmask = 0;
 
-        switch (datarw & 0xF)
-        {
-        case 0: break;
-        case 1: privmask |= 0x03; break;
-        case 2: privmask |= 0x03; usermask |= 0x01; break;
-        case 3: privmask |= 0x03; usermask |= 0x03; break;
-        case 5: privmask |= 0x01; break;
-        case 6: privmask |= 0x01; usermask |= 0x01; break;
-        default: printf("!! BAD DATARW VALUE %d\n", datarw&0xF);
-        }
+    switch (datarw)
+    {
+    case 0: break;
+    case 1: privmask |= 0x03; break;
+    case 2: privmask |= 0x03; usermask |= 0x01; break;
+    case 3: privmask |= 0x03; usermask |= 0x03; break;
+    case 5: privmask |= 0x01; break;
+    case 6: privmask |= 0x01; usermask |= 0x01; break;
+    default: printf("!! BAD DATARW VALUE %d\n", datarw&0xF);
+    }
 
-        switch (coderw & 0xF)
+    switch (coderw)
+    {
+    case 0: break;
+    case 1: privmask |= 0x04; break;
+    case 2: privmask |= 0x04; usermask |= 0x04; break;
+    case 3: privmask |= 0x04; usermask |= 0x04; break;
+    case 5: privmask |= 0x04; break;
+    case 6: privmask |= 0x04; usermask |= 0x04; break;
+    default: printf("!! BAD CODERW VALUE %d\n", datarw&0xF);
+    }
+
+    if (datacache & 0x1)
+    {
+        privmask |= 0x10;
+        usermask |= 0x10;
+
+        if (datawrite & 0x1)
         {
-        case 0: break;
-        case 1: privmask |= 0x04; break;
-        case 2: privmask |= 0x04; usermask |= 0x04; break;
-        case 3: privmask |= 0x04; usermask |= 0x04; break;
-        case 5: privmask |= 0x04; break;
-        case 6: privmask |= 0x04; usermask |= 0x04; break;
-        default: printf("!! BAD CODERW VALUE %d\n", datarw&0xF);
+            privmask |= 0x20;
+            usermask |= 0x20;
         }
+    }
 
-        if (datacache & 0x1)
-        {
-            privmask |= 0x10;
-            usermask |= 0x10;
+    if (codecache & 0x1)
+    {
+        privmask |= 0x40;
+        usermask |= 0x40;
+    }
 
-            if (datawrite & 0x1)
-            {
-                privmask |= 0x20;
-                usermask |= 0x20;
-            }
-        }
+    //printf("PU region %d: %08X-%08X, user=%02X priv=%02X\n", n, start<<12, end<<12, usermask, privmask);
 
-        if (codecache & 0x1)
-        {
-            privmask |= 0x40;
-            usermask |= 0x40;
-        }
+    for (u32 i = start; i < end; i++)
+    {
+        PU_UserMap[i] = usermask;
+        PU_PrivMap[i] = privmask;
+    }
 
-        printf("PU region %d: %08X-%08X, user=%02X priv=%02X\n", n, start<<12, end<<12, usermask, privmask);
+    UpdateRegionTimings(start<<12, end<<12);
+}
 
-        for (u32 i = start; i < end; i++)
-        {
-            PU_UserMap[i] = usermask;
-            PU_PrivMap[i] = privmask;
-        }
+void ARMv5::UpdatePURegions(bool update_all)
+{
+    if (!(CP15Control & (1<<0)))
+    {
+        // PU disabled
 
-        coderw >>= 4;
-        datarw >>= 4;
-        codecache >>= 1;
-        datacache >>= 1;
-        datawrite >>= 1;
+        u8 mask = 0x07;
+        if (CP15Control & (1<<2))  mask |= 0x30;
+        if (CP15Control & (1<<12)) mask |= 0x40;
+
+        memset(PU_UserMap, mask, 0x100000);
+        memset(PU_PrivMap, mask, 0x100000);
+
+        UpdateRegionTimings(0x00000000, 0xFFFFFFFF);
+        return;
+    }
+
+    if (update_all)
+    {
+        memset(PU_UserMap, 0, 0x100000);
+        memset(PU_PrivMap, 0, 0x100000);
+    }
 
-        // TODO: this will not be enough if they change their PU regions after the intial setup
-        //UpdateRegionTimings(start<<12, end<<12);
+    for (int n = 0; n < 8; n++)
+    {
+        UpdatePURegion(n);
     }
 
     // TODO: this is way unoptimized
     // should be okay unless the game keeps changing shit, tho
-    UpdateRegionTimings(0x00000000, 0xFFFFFFFF);
+    if (update_all) UpdateRegionTimings(0x00000000, 0xFFFFFFFF);
 }
 
 void ARMv5::UpdateRegionTimings(u32 addrstart, u32 addrend)
@@ -303,7 +302,6 @@ u32 ARMv5::RandomLineIndex()
     return (RNGSeed >> 17) & 0x3;
 }
 
-int zog=1;
 void ARMv5::ICacheLookup(u32 addr)
 {
     u32 tag = addr & 0xFFFFF800;
@@ -312,25 +310,25 @@ void ARMv5::ICacheLookup(u32 addr)
     id <<= 2;
     if (ICacheTags[id+0] == tag)
     {
-        CodeCycles = 1;zog=1;
+        CodeCycles = 1;
         CurICacheLine = &ICache[(id+0) << 5];
         return;
     }
     if (ICacheTags[id+1] == tag)
     {
-        CodeCycles = 1;zog=2;
+        CodeCycles = 1;
         CurICacheLine = &ICache[(id+1) << 5];
         return;
     }
     if (ICacheTags[id+2] == tag)
     {
-        CodeCycles = 1;zog=3;
+        CodeCycles = 1;
         CurICacheLine = &ICache[(id+2) << 5];
         return;
     }
     if (ICacheTags[id+3] == tag)
     {
-        CodeCycles = 1;zog=4;
+        CodeCycles = 1;
         CurICacheLine = &ICache[(id+3) << 5];
         return;
     }
@@ -418,10 +416,13 @@ void ARMv5::CP15Write(u32 id, u32 val)
             val &= 0x000FF085;
             CP15Control &= ~0x000FF085;
             CP15Control |= val;
-            printf("CP15Control = %08X (%08X->%08X)\n", CP15Control, old, val);
+            //printf("CP15Control = %08X (%08X->%08X)\n", CP15Control, old, val);
             UpdateDTCMSetting();
             UpdateITCMSetting();
-            if ((old & 0x1005) != (val & 0x1005)) UpdatePURegions();
+            if ((old & 0x1005) != (val & 0x1005))
+            {
+                UpdatePURegions((old & 0x1) != (val & 0x1));
+            }
             if (val & (1<<7)) printf("!!!! ARM9 BIG ENDIAN MODE. VERY BAD. SHIT GONNA ASPLODE NOW\n");
             if (val & (1<<13)) ExceptionBase = 0xFFFF0000;
             else               ExceptionBase = 0x00000000;
@@ -430,63 +431,100 @@ void ARMv5::CP15Write(u32 id, u32 val)
 
 
     case 0x200: // data cacheable
-        PU_DataCacheable = val;
-        printf("PU: DataCacheable=%08X\n", val);
-        UpdatePURegions();
+        {
+            u32 diff = PU_DataCacheable ^ val;
+            PU_DataCacheable = val;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (1<<i)) UpdatePURegion(i);
+            }
+        }
         return;
 
     case 0x201: // code cacheable
-        PU_CodeCacheable = val;
-        printf("PU: CodeCacheable=%08X\n", val);
-        UpdatePURegions();
+        {
+            u32 diff = PU_CodeCacheable ^ val;
+            PU_CodeCacheable = val;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (1<<i)) UpdatePURegion(i);
+            }
+        }
         return;
 
 
     case 0x300: // data cache write-buffer
-        PU_DataCacheWrite = val;
-        printf("PU: DataCacheWrite=%08X\n", val);
-        UpdatePURegions();
+        {
+            u32 diff = PU_DataCacheWrite ^ val;
+            PU_DataCacheWrite = val;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (1<<i)) UpdatePURegion(i);
+            }
+        }
         return;
 
 
     case 0x500: // legacy data permissions
-        PU_DataRW = 0;
-        PU_DataRW |= (val & 0x0003);
-        PU_DataRW |= ((val & 0x000C) << 2);
-        PU_DataRW |= ((val & 0x0030) << 4);
-        PU_DataRW |= ((val & 0x00C0) << 6);
-        PU_DataRW |= ((val & 0x0300) << 8);
-        PU_DataRW |= ((val & 0x0C00) << 10);
-        PU_DataRW |= ((val & 0x3000) << 12);
-        PU_DataRW |= ((val & 0xC000) << 14);
-        printf("PU: DataRW=%08X (legacy %08X)\n", PU_DataRW, val);
-        UpdatePURegions();
+        {
+            u32 old = PU_DataRW;
+            PU_DataRW = 0;
+            PU_DataRW |= (val & 0x0003);
+            PU_DataRW |= ((val & 0x000C) << 2);
+            PU_DataRW |= ((val & 0x0030) << 4);
+            PU_DataRW |= ((val & 0x00C0) << 6);
+            PU_DataRW |= ((val & 0x0300) << 8);
+            PU_DataRW |= ((val & 0x0C00) << 10);
+            PU_DataRW |= ((val & 0x3000) << 12);
+            PU_DataRW |= ((val & 0xC000) << 14);
+            u32 diff = old ^ PU_DataRW;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (0xF<<(i*4))) UpdatePURegion(i);
+            }
+        }
         return;
 
     case 0x501: // legacy code permissions
-        PU_CodeRW = 0;
-        PU_CodeRW |= (val & 0x0003);
-        PU_CodeRW |= ((val & 0x000C) << 2);
-        PU_CodeRW |= ((val & 0x0030) << 4);
-        PU_CodeRW |= ((val & 0x00C0) << 6);
-        PU_CodeRW |= ((val & 0x0300) << 8);
-        PU_CodeRW |= ((val & 0x0C00) << 10);
-        PU_CodeRW |= ((val & 0x3000) << 12);
-        PU_CodeRW |= ((val & 0xC000) << 14);
-        printf("PU: CodeRW=%08X (legacy %08X)\n", PU_CodeRW, val);
-        UpdatePURegions();
+        {
+            u32 old = PU_CodeRW;
+            PU_CodeRW = 0;
+            PU_CodeRW |= (val & 0x0003);
+            PU_CodeRW |= ((val & 0x000C) << 2);
+            PU_CodeRW |= ((val & 0x0030) << 4);
+            PU_CodeRW |= ((val & 0x00C0) << 6);
+            PU_CodeRW |= ((val & 0x0300) << 8);
+            PU_CodeRW |= ((val & 0x0C00) << 10);
+            PU_CodeRW |= ((val & 0x3000) << 12);
+            PU_CodeRW |= ((val & 0xC000) << 14);
+            u32 diff = old ^ PU_CodeRW;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (0xF<<(i*4))) UpdatePURegion(i);
+            }
+        }
         return;
 
     case 0x502: // data permissions
-        PU_DataRW = val;
-        printf("PU: DataRW=%08X\n", PU_DataRW);
-        UpdatePURegions();
+        {
+            u32 diff = PU_DataRW ^ val;
+            PU_DataRW = val;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (0xF<<(i*4))) UpdatePURegion(i);
+            }
+        }
         return;
 
     case 0x503: // code permissions
-        PU_CodeRW = val;
-        printf("PU: CodeRW=%08X\n", PU_CodeRW);
-        UpdatePURegions();
+        {
+            u32 diff = PU_CodeRW ^ val;
+            PU_CodeRW = val;
+            for (u32 i = 0; i < 8; i++)
+            {
+                if (diff & (0xF<<(i*4))) UpdatePURegion(i);
+            }
+        }
         return;
 
 
@@ -511,7 +549,8 @@ void ARMv5::CP15Write(u32 id, u32 val)
         printf("%s, ", val&1 ? "enabled":"disabled");
         printf("%08X-", val&0xFFFFF000);
         printf("%08X\n", (val&0xFFFFF000)+(2<<((val&0x3E)>>1)));
-        UpdatePURegions();
+        // TODO: smarter region update for this?
+        UpdatePURegions(true);
         return;
 
 
diff --git a/src/GPU2D.cpp b/src/GPU2D.cpp
index b510767..30a6d45 100644
--- a/src/GPU2D.cpp
+++ b/src/GPU2D.cpp
@@ -943,8 +943,8 @@ void GPU2D::VBlankEnd()
     BGMosaicYMax = BGMosaicSize[1];
     //OBJMosaicY = 0;
     //OBJMosaicYMax = OBJMosaicSize[1];
-    OBJMosaicY = 0;
-    OBJMosaicYCount = 0;
+    //OBJMosaicY = 0;
+    //OBJMosaicYCount = 0;
 
     if (Accelerated)
     {
@@ -2238,6 +2238,13 @@ void GPU2D::DrawBG_Large(u32 line) // BG is always BG2
     BGYRefInternal[0] += rotD;
 }
 
+// OBJ line buffer:
+// * bit0-15: color (bit15=1: direct color, bit15=0: palette index, bit12=0 to indicate extpal)
+// * bit16-17: BG-relative priority
+// * bit18: non-transparent sprite pixel exists here
+// * bit19: X mosaic should be applied here
+// * bit24-31: compositor flags
+
 void GPU2D::ApplySpriteMosaicX()
 {
     // apply X mosaic if needed
@@ -2255,7 +2262,7 @@ void GPU2D::ApplySpriteMosaicX()
             continue;
         }
 
-        if ((!(OBJLine[i-1] & 0x100000)) || (CurOBJXMosaicTable[i] == 0))
+        if ((OBJIndex[i] != OBJIndex[i-1]) || (CurOBJXMosaicTable[i] == 0))
             lastcolor = OBJLine[i];
         else
             OBJLine[i] = lastcolor;
@@ -2322,11 +2329,25 @@ void GPU2D::InterleaveSprites(u32 prio)
 
 void GPU2D::DrawSprites(u32 line)
 {
+    if (line == 0)
+    {
+        // reset those counters here
+        // TODO: find out when those are supposed to be reset
+        // it would make sense to reset them at the end of VBlank
+        // however, sprites are rendered one scanline in advance
+        // so they need to be reset a bit earlier
+
+        OBJMosaicY = 0;
+        OBJMosaicYCount = 0;
+    }
+
     NumSprites = 0;
     memset(OBJLine, 0, 256*4);
     memset(OBJWindow, 0, 256);
     if (!(DispCnt & 0x1000)) return;
 
+    memset(OBJIndex, 0xFF, 256);
+
     u16* oam = (u16*)&GPU::OAM[Num ? 0x400 : 0];
 
     const s32 spritewidth[16] =
@@ -2355,6 +2376,15 @@ void GPU2D::DrawSprites(u32 line)
 
             bool iswin = (((attrib[0] >> 10) & 0x3) == 2);
 
+            u32 sprline;
+            if ((attrib[0] & 0x1000) && !iswin)
+            {
+                // apply Y mosaic
+                sprline = OBJMosaicY;
+            }
+            else
+                sprline = line;
+
             if (attrib[0] & 0x0100)
             {
                 u32 sizeparam = (attrib[0] >> 14) | ((attrib[1] & 0xC000) >> 12);
@@ -2370,7 +2400,7 @@ void GPU2D::DrawSprites(u32 line)
                 }
 
                 u32 ypos = attrib[0] & 0xFF;
-                ypos = (line - ypos) & 0xFF;
+                ypos = (sprline - ypos) & 0xFF;
                 if (ypos >= (u32)boundheight)
                     continue;
 
@@ -2380,7 +2410,7 @@ void GPU2D::DrawSprites(u32 line)
 
                 u32 rotparamgroup = (attrib[1] >> 9) & 0x1F;
 
-                DoDrawSprite(Rotscale, attrib, &oam[(rotparamgroup*16) + 3], boundwidth, boundheight, width, height, xpos, ypos);
+                DoDrawSprite(Rotscale, sprnum, boundwidth, boundheight, width, height, xpos, ypos);
 
                 NumSprites++;
             }
@@ -2394,7 +2424,7 @@ void GPU2D::DrawSprites(u32 line)
                 s32 height = spriteheight[sizeparam];
 
                 u32 ypos = attrib[0] & 0xFF;
-                ypos = (line - ypos) & 0xFF;
+                ypos = (sprline - ypos) & 0xFF;
                 if (ypos >= (u32)height)
                     continue;
 
@@ -2402,11 +2432,7 @@ void GPU2D::DrawSprites(u32 line)
                 if (xpos <= -width)
                     continue;
 
-                // yflip
-                if (attrib[1] & 0x2000)
-                    ypos = height-1 - ypos;
-
-                DoDrawSprite(Normal, attrib, width, xpos, ypos);
+                DoDrawSprite(Normal, sprnum, width, height, xpos, ypos);
 
                 NumSprites++;
             }
@@ -2415,8 +2441,12 @@ void GPU2D::DrawSprites(u32 line)
 }
 
 template<bool window>
-void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32 boundheight, u32 width, u32 height, s32 xpos, s32 ypos)
+void GPU2D::DrawSprite_Rotscale(u32 num, u32 boundwidth, u32 boundheight, u32 width, u32 height, s32 xpos, s32 ypos)
 {
+    u16* oam = (u16*)&GPU::OAM[Num ? 0x400 : 0];
+    u16* attrib = &oam[num * 4];
+    u16* rotparams = &oam[(((attrib[1] >> 9) & 0x1F) * 16) + 3];
+
     u32 pixelattr = ((attrib[2] & 0x0C00) << 6) | 0xC0000;
     u32 tilenum = attrib[2] & 0x03FF;
     u32 spritemode = window ? 0 : ((attrib[0] >> 10) & 0x3);
@@ -2429,9 +2459,6 @@ void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32
     if ((attrib[0] & 0x1000) && !window)
     {
         // apply Y mosaic
-        ypos = OBJMosaicY - (attrib[0] & 0xFF);
-        if (ypos < 0) ypos = 0;
-
         pixelattr |= 0x100000;
     }
 
@@ -2509,12 +2536,15 @@ void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32
                 if (color & 0x8000)
                 {
                     if (window) OBJWindow[xpos] = 1;
-                    else        OBJLine[xpos] = color | pixelattr;
+                    else      { OBJLine[xpos] = color | pixelattr; OBJIndex[xpos] = num; }
                 }
                 else if (!window)
                 {
                     if (OBJLine[xpos] == 0)
+                    {
                         OBJLine[xpos] = pixelattr & 0x180000;
+                        OBJIndex[xpos] = num;
+                    }
                 }
             }
 
@@ -2563,12 +2593,15 @@ void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32
                     if (color)
                     {
                         if (window) OBJWindow[xpos] = 1;
-                        else        OBJLine[xpos] = color | pixelattr;
+                        else      { OBJLine[xpos] = color | pixelattr; OBJIndex[xpos] = num; }
                     }
                     else if (!window)
                     {
                         if (OBJLine[xpos] == 0)
+                        {
                             OBJLine[xpos] = pixelattr & 0x180000;
+                            OBJIndex[xpos] = num;
+                        }
                     }
                 }
 
@@ -2604,12 +2637,15 @@ void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32
                     if (color)
                     {
                         if (window) OBJWindow[xpos] = 1;
-                        else        OBJLine[xpos] = color | pixelattr;
+                        else      { OBJLine[xpos] = color | pixelattr; OBJIndex[xpos] = num; }
                     }
                     else if (!window)
                     {
                         if (OBJLine[xpos] == 0)
+                        {
                             OBJLine[xpos] = pixelattr & 0x180000;
+                            OBJIndex[xpos] = num;
+                        }
                     }
                 }
 
@@ -2623,8 +2659,11 @@ void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32
 }
 
 template<bool window>
-void GPU2D::DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos)
+void GPU2D::DrawSprite_Normal(u32 num, u32 width, u32 height, s32 xpos, s32 ypos)
 {
+    u16* oam = (u16*)&GPU::OAM[Num ? 0x400 : 0];
+    u16* attrib = &oam[num * 4];
+
     u32 pixelattr = ((attrib[2] & 0x0C00) << 6) | 0xC0000;
     u32 tilenum = attrib[2] & 0x03FF;
     u32 spritemode = window ? 0 : ((attrib[0] >> 10) & 0x3);
@@ -2634,12 +2673,13 @@ void GPU2D::DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos)
     if ((attrib[0] & 0x1000) && !window)
     {
         // apply Y mosaic
-        ypos = OBJMosaicY - (attrib[0] & 0xFF);
-        if (ypos < 0) ypos = 0;
-
         pixelattr |= 0x100000;
     }
 
+    // yflip
+    if (attrib[1] & 0x2000)
+        ypos = height-1 - ypos;
+
     u32 xoff;
     u32 xend = width;
     if (xpos >= 0)
@@ -2719,12 +2759,15 @@ void GPU2D::DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos)
             if (color & 0x8000)
             {
                 if (window) OBJWindow[xpos] = 1;
-                else        OBJLine[xpos] = color | pixelattr;
+                else      { OBJLine[xpos] = color | pixelattr; OBJIndex[xpos] = num; }
             }
             else if (!window)
             {
                 if (OBJLine[xpos] == 0)
+                {
                     OBJLine[xpos] = pixelattr & 0x180000;
+                    OBJIndex[xpos] = num;
+                }
             }
 
             xoff++;
@@ -2786,12 +2829,15 @@ void GPU2D::DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos)
                 if (color)
                 {
                     if (window) OBJWindow[xpos] = 1;
-                    else        OBJLine[xpos] = color | pixelattr;
+                    else      { OBJLine[xpos] = color | pixelattr; OBJIndex[xpos] = num; }
                 }
                 else if (!window)
                 {
                     if (OBJLine[xpos] == 0)
+                    {
                         OBJLine[xpos] = pixelattr & 0x180000;
+                        OBJIndex[xpos] = num;
+                    }
                 }
 
                 xoff++;
@@ -2847,12 +2893,15 @@ void GPU2D::DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos)
                 if (color)
                 {
                     if (window) OBJWindow[xpos] = 1;
-                    else        OBJLine[xpos] = color | pixelattr;
+                    else      { OBJLine[xpos] = color | pixelattr; OBJIndex[xpos] = num; }
                 }
                 else if (!window)
                 {
                     if (OBJLine[xpos] == 0)
+                    {
                         OBJLine[xpos] = pixelattr & 0x180000;
+                        OBJIndex[xpos] = num;
+                    }
                 }
 
                 xoff++;
diff --git a/src/GPU2D.h b/src/GPU2D.h
index 8c6ec27..6675ce7 100644
--- a/src/GPU2D.h
+++ b/src/GPU2D.h
@@ -78,6 +78,7 @@ private:
     u8 WindowMask[256] __attribute__((aligned (8)));
     u32 OBJLine[256] __attribute__((aligned (8)));
     u8 OBJWindow[256] __attribute__((aligned (8)));
+    u8 OBJIndex[256] __attribute__((aligned (8)));
 
     u32 NumSprites;
 
@@ -156,8 +157,8 @@ private:
 
     void ApplySpriteMosaicX();
     void InterleaveSprites(u32 prio);
-    template<bool window> void DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32 boundheight, u32 width, u32 height, s32 xpos, s32 ypos);
-    template<bool window> void DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos);
+    template<bool window> void DrawSprite_Rotscale(u32 num, u32 boundwidth, u32 boundheight, u32 width, u32 height, s32 xpos, s32 ypos);
+    template<bool window> void DrawSprite_Normal(u32 num, u32 width, u32 height, s32 xpos, s32 ypos);
 
     void DoCapture(u32 line, u32 width);
 
diff --git a/src/GPU3D_OpenGL.cpp b/src/GPU3D_OpenGL.cpp
index d616e1c..6abb1ab 100644
--- a/src/GPU3D_OpenGL.cpp
+++ b/src/GPU3D_OpenGL.cpp
@@ -807,7 +807,6 @@ void RenderSceneChunk(int y, int h)
     else
         glBlendFuncSeparate(GL_ONE, GL_ZERO, GL_ONE, GL_ONE);
 
-    UseRenderShader(flags | RenderFlag_Trans);
     glLineWidth(1.0);
 
     if (NumOpaqueFinalPolys > -1)
@@ -843,6 +842,8 @@ void RenderSceneChunk(int y, int h)
                 }
                 else if (rp->PolyData->Translucent)
                 {
+                    UseRenderShader(flags | RenderFlag_Trans);
+
                     // zorp
                     glDepthFunc(GL_LESS);
 
diff --git a/src/libui_sdl/main.cpp b/src/libui_sdl/main.cpp
index 08d38b2..3d894f4 100644
--- a/src/libui_sdl/main.cpp
+++ b/src/libui_sdl/main.cpp
@@ -2662,6 +2662,45 @@ int main(int argc, char** argv)
         SDL_Quit();
         return 0;
     }
+    if (!Platform::LocalFileExists("firmware.bin.bak"))
+    {
+        // verify the firmware
+        //
+        // there are dumps of an old hacked firmware floating around on the internet
+        // and those are problematic
+        // the hack predates WFC, and, due to this, any game that alters the WFC
+        // access point data will brick that firmware due to it having critical
+        // data in the same area. it has the same problem on hardware.
+        //
+        // but this should help stop users from reporting that issue over and over
+        // again, when the issue is not from melonDS but from their firmware dump.
+        //
+        // I don't know about all the firmware hacks in existence, but the one I
+        // looked at has 0x180 bytes from the header repeated at 0x3FC80, but
+        // bytes 0x0C-0x14 are different.
+
+        FILE* f = Platform::OpenLocalFile("firmware.bin", "rb");
+        u8 chk1[0x180], chk2[0x180];
+
+        fseek(f, 0, SEEK_SET);
+        fread(chk1, 1, 0x180, f);
+        fseek(f, -0x380, SEEK_END);
+        fread(chk2, 1, 0x180, f);
+
+        memset(&chk1[0x0C], 0, 8);
+        memset(&chk2[0x0C], 0, 8);
+
+        fclose(f);
+
+        if (!memcmp(chk1, chk2, 0x180))
+        {
+            uiMsgBoxError(NULL,
+                          "Problematic firmware dump",
+                          "You are using an old hacked firmware dump.\n"
+                          "Firmware boot will stop working if you run any game that alters WFC settings.\n\n"
+                          "Note that the issue is not from melonDS, it would also happen on an actual DS.");
+        }
+    }
     {
         FILE* f = Platform::OpenLocalFile("romlist.bin", "rb");
         if (f)