begin renovating melonDSi

27 files changed, 537 insertions, 167 deletions

diff --git a/.github/FUNDING.yml b/.github/FUNDING.yml
new file mode 100644
index 0000000..bb83359
--- /dev/null
+++ b/.github/FUNDING.yml
@@ -0,0 +1 @@
+patreon: staplebutter
diff --git a/.gitignore b/.gitignore
index f63e239..bd1d485 100644
--- a/.gitignore
+++ b/.gitignore
@@ -6,3 +6,5 @@ obj
 melon_grc.c
 melon_grc.h
 cmake-build
+cmake-build-debug
+.idea
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6ebb69f..048dd44 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -22,6 +22,11 @@ endif()
 
 if(ENABLE_LTO)
 	add_compile_options(-O3 -flto)
+	set(CMAKE_AR "gcc-ar")
+	set(CMAKE_C_ARCHIVE_CREATE "<CMAKE_AR> qcs <TARGET> <LINK_FLAGS> <OBJECTS>")
+	set(CMAKE_C_ARCHIVE_FINISH   true)
+	set(CMAKE_CXX_ARCHIVE_CREATE "<CMAKE_AR> qcs <TARGET> <LINK_FLAGS> <OBJECTS>")
+	set(CMAKE_CXX_ARCHIVE_FINISH   true)
 endif()
 
 add_compile_options(-fno-pic)
diff --git a/README.md b/README.md
index 293af51..cf0afed 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 <h2 align="center"><b>melonDS</b></h2>
 <p align="center">
 <a href="http://melonds.kuribo64.net/" alt="melonDS website"><img src="https://img.shields.io/badge/website-melonds.kuribo64.net-%2331352e.svg"></a>
-<a href="http://melonds.kuribo64.net/downloads.php" alt="Release: 0.8.1"><img src="https://img.shields.io/badge/release-0.8.1-%235c913b.svg"></a>
+<a href="http://melonds.kuribo64.net/downloads.php" alt="Release: 0.8.3"><img src="https://img.shields.io/badge/release-0.8.3-%235c913b.svg"></a>
 <a href="https://www.gnu.org/licenses/gpl-3.0" alt="License: GPLv3"><img src="https://img.shields.io/badge/License-GPL%20v3-%23ff554d.svg"></a>
 <a href="https://kiwiirc.com/client/irc.badnik.net/?nick=IRC-Source_?#melonds" alt="IRC channel: #melonds"><img src="https://img.shields.io/badge/IRC%20chat-%23melonds-%23dd2e44.svg"></a>
 </p>
diff --git a/melon.rc b/melon.rc
index c1d773d..e2d8dc2 100644
--- a/melon.rc
+++ b/melon.rc
@@ -6,8 +6,8 @@
 
 //include version information in .exe, modify these values to match your needs
 1 VERSIONINFO
-FILEVERSION 0,8,1,0
-PRODUCTVERSION 0,8,1,0
+FILEVERSION 0,8,3,0
+PRODUCTVERSION 0,8,3,0
 FILETYPE VFT_APP
 {
   BLOCK "StringFileInfo"
@@ -15,14 +15,14 @@ FILETYPE VFT_APP
 		BLOCK "040904E4"
 		{
 			VALUE "CompanyName", "Melon Factory of Kuribo64"
-			VALUE "FileVersion", "0.8.1"
+			VALUE "FileVersion", "0.8.3"
 			VALUE "FileDescription", "DS emulator, sorta. also 1st quality melon."
 			VALUE "InternalName", "SDnolem"
 			VALUE "LegalCopyright", "2016-2019 Arisotura & co."
 			VALUE "LegalTrademarks", ""
 			VALUE "OriginalFilename", "zafkflzdasd.exe"
 			VALUE "ProductName", "melonDS"
-			VALUE "ProductVersion", "0.8.1"
+			VALUE "ProductVersion", "0.8.3"
 		}
 	}
   BLOCK "VarFileInfo"
diff --git a/src/ARMInterpreter_ALU.cpp b/src/ARMInterpreter_ALU.cpp
index f23d6b2..d60d8f8 100644
--- a/src/ARMInterpreter_ALU.cpp
+++ b/src/ARMInterpreter_ALU.cpp
@@ -106,12 +106,12 @@ namespace ARMInterpreter
     x = ROR(x, (s&0x1F));
 
 #define LSL_REG_S(x, s) \
-    if (s > 31)     { cpu->SetC(x & (1<<0));      x = 0; } \
-    else if (s > 0) { cpu->SetC(x & (1<<(32-s))); x <<= s; }
+    if (s > 31)     { cpu->SetC((s>32) ? 0 : (x & (1<<0))); x = 0; } \
+    else if (s > 0) { cpu->SetC(x & (1<<(32-s)));           x <<= s; }
 
 #define LSR_REG_S(x, s) \
-    if (s > 31)     { cpu->SetC(x & (1<<31));    x = 0; } \
-    else if (s > 0) { cpu->SetC(x & (1<<(s-1))); x >>= s; }
+    if (s > 31)     { cpu->SetC((s>32) ? 0 : (x & (1<<31))); x = 0; } \
+    else if (s > 0) { cpu->SetC(x & (1<<(s-1)));             x >>= s; }
 
 #define ASR_REG_S(x, s) \
     if (s > 31)     { cpu->SetC(x & (1<<31));    x = ((s32)x) >> 31; } \
@@ -134,7 +134,7 @@ namespace ARMInterpreter
 #define A_CALC_OP2_REG_SHIFT_REG(shiftop) \
     u32 b = cpu->R[cpu->CurInstr&0xF]; \
     if ((cpu->CurInstr&0xF)==15) b += 4; \
-    shiftop(b, cpu->R[(cpu->CurInstr>>8)&0xF]);
+    shiftop(b, (cpu->R[(cpu->CurInstr>>8)&0xF] & 0xFF));
 
 
 #define A_IMPLEMENT_ALU_OP(x,s) \
diff --git a/src/GPU.cpp b/src/GPU.cpp
index 0eb321c..13f2727 100644
--- a/src/GPU.cpp
+++ b/src/GPU.cpp
@@ -284,6 +284,7 @@ void SetDisplaySettings(bool accel)
     if (Framebuffer[1][0]) delete[] Framebuffer[1][0];
     if (Framebuffer[0][1]) delete[] Framebuffer[0][1];
     if (Framebuffer[1][1]) delete[] Framebuffer[1][1];
+
     Framebuffer[0][0] = new u32[fbsize];
     Framebuffer[1][0] = new u32[fbsize];
     Framebuffer[0][1] = new u32[fbsize];
diff --git a/src/GPU2D.cpp b/src/GPU2D.cpp
index e0da772..288ee58 100644
--- a/src/GPU2D.cpp
+++ b/src/GPU2D.cpp
@@ -246,6 +246,11 @@ u8 GPU2D::Read8(u32 addr)
     case 0x049: return WinCnt[1];
     case 0x04A: return WinCnt[2];
     case 0x04B: return WinCnt[3];
+
+    // there are games accidentally trying to read those
+    // those are write-only
+    case 0x04C:
+    case 0x04D: return 0;
     }
 
     printf("unknown GPU read8 %08X\n", addr);
@@ -299,10 +304,22 @@ void GPU2D::Write8(u32 addr, u8 val)
 
     switch (addr & 0x00000FFF)
     {
-    case 0x000: DispCnt = (DispCnt & 0xFFFFFF00) | val; return;
-    case 0x001: DispCnt = (DispCnt & 0xFFFF00FF) | (val << 8); return;
-    case 0x002: DispCnt = (DispCnt & 0xFF00FFFF) | (val << 16); return;
-    case 0x003: DispCnt = (DispCnt & 0x00FFFFFF) | (val << 24); return;
+    case 0x000:
+        DispCnt = (DispCnt & 0xFFFFFF00) | val;
+        if (Num) DispCnt &= 0xC0B1FFF7;
+        return;
+    case 0x001:
+        DispCnt = (DispCnt & 0xFFFF00FF) | (val << 8);
+        if (Num) DispCnt &= 0xC0B1FFF7;
+        return;
+    case 0x002:
+        DispCnt = (DispCnt & 0xFF00FFFF) | (val << 16);
+        if (Num) DispCnt &= 0xC0B1FFF7;
+        return;
+    case 0x003:
+        DispCnt = (DispCnt & 0x00FFFFFF) | (val << 24);
+        if (Num) DispCnt &= 0xC0B1FFF7;
+        return;
 
     case 0x008: BGCnt[0] = (BGCnt[0] & 0xFF00) | val; return;
     case 0x009: BGCnt[0] = (BGCnt[0] & 0x00FF) | (val << 8); return;
@@ -381,8 +398,14 @@ void GPU2D::Write16(u32 addr, u16 val)
 
     switch (addr & 0x00000FFF)
     {
-    case 0x000: DispCnt = (DispCnt & 0xFFFF0000) | val; return;
-    case 0x002: DispCnt = (DispCnt & 0x0000FFFF) | (val << 16); return;
+    case 0x000:
+        DispCnt = (DispCnt & 0xFFFF0000) | val;
+        if (Num) DispCnt &= 0xC0B1FFF7;
+        return;
+    case 0x002:
+        DispCnt = (DispCnt & 0x0000FFFF) | (val << 16);
+        if (Num) DispCnt &= 0xC0B1FFF7;
+        return;
 
     case 0x008: BGCnt[0] = val; return;
     case 0x00A: BGCnt[1] = val; return;
@@ -514,6 +537,7 @@ void GPU2D::Write32(u32 addr, u32 val)
     {
     case 0x000:
         DispCnt = val;
+        if (Num) DispCnt &= 0xC0B1FFF7;
         return;
 
     case 0x028:
@@ -704,10 +728,6 @@ void GPU2D::DrawScanline(u32 line)
     // oddly that's not the case for GPU A
     if (Num && !Enabled) forceblank = true;
 
-    // forced blank
-    // (checkme: are there still things that can run under this mode? likely not)
-    if (DispCnt & (1<<7)) forceblank = true;
-
     if (forceblank)
     {
         for (int i = 0; i < 256; i++)
@@ -748,8 +768,10 @@ void GPU2D::DrawScanline(u32 line)
 
     case 1: // regular display
         {
-            for (int i = 0; i < stride; i+=2)
+            int i = 0;
+            for (; i < (stride & ~1); i+=2)
                 *(u64*)&dst[i] = *(u64*)&BGOBJLine[i];
+            if (stride & 1) dst[i] = BGOBJLine[i];
         }
         break;
 
@@ -1313,12 +1335,6 @@ void GPU2D::DrawScanlineBGMode(u32 line)
 
 void GPU2D::DrawScanlineBGMode6(u32 line)
 {
-    if (Num)
-    {
-        printf("GPU2D: MODE6 ON SUB GPU???\n");
-        return;
-    }
-
     for (int i = 3; i >= 0; i--)
     {
         if ((BGCnt[2] & 0x3) == i)
@@ -1332,8 +1348,36 @@ void GPU2D::DrawScanlineBGMode6(u32 line)
         {
             if (DispCnt & 0x0100)
             {
-                if (DispCnt & 0x8)
+                if ((!Num) && (DispCnt & 0x8))
+                    DrawBG_3D();
+            }
+        }
+        if ((DispCnt & 0x1000) && NumSprites)
+            InterleaveSprites(0x8000 | (i<<16));
+    }
+}
+
+void GPU2D::DrawScanlineBGMode7(u32 line)
+{
+    // mode 7 only has text-mode BG0 and BG1
+
+    for (int i = 3; i >= 0; i--)
+    {
+        if ((BGCnt[1] & 0x3) == i)
+        {
+            if (DispCnt & 0x0200)
+            {
+                DrawBG_Text(line, 1);
+            }
+        }
+        if ((BGCnt[0] & 0x3) == i)
+        {
+            if (DispCnt & 0x0100)
+            {
+                if ((!Num) && (DispCnt & 0x8))
                     DrawBG_3D();
+                else
+                    DrawBG_Text(line, 0);
             }
         }
         if ((DispCnt & 0x1000) && NumSprites)
@@ -1343,6 +1387,15 @@ void GPU2D::DrawScanlineBGMode6(u32 line)
 
 void GPU2D::DrawScanline_BGOBJ(u32 line)
 {
+    // forced blank disables BG/OBJ compositing
+    if (DispCnt & (1<<7))
+    {
+        for (int i = 0; i < 256; i++)
+            BGOBJLine[i] = 0xFF3F3F3F;
+
+        return;
+    }
+
     u64 backdrop;
     if (Num) backdrop = *(u16*)&GPU::Palette[0x400];
     else     backdrop = *(u16*)&GPU::Palette[0];
@@ -1364,7 +1417,6 @@ void GPU2D::DrawScanline_BGOBJ(u32 line)
     else
         memset(WindowMask, 0xFF, 256);
 
-    // TODO: what happens in mode 7? mode 6 on the sub engine?
     switch (DispCnt & 0x7)
     {
     case 0: DrawScanlineBGMode<0>(line); break;
@@ -1374,6 +1426,7 @@ void GPU2D::DrawScanline_BGOBJ(u32 line)
     case 4: DrawScanlineBGMode<4>(line); break;
     case 5: DrawScanlineBGMode<5>(line); break;
     case 6: DrawScanlineBGMode6(line); break;
+    case 7: DrawScanlineBGMode7(line); break;
     }
 
     // color special effects
@@ -2035,14 +2088,19 @@ void GPU2D::DrawBG_Large(u32 line) // BG is always BG2
     u32 tilesetaddr, tilemapaddr;
     u16* pal;
 
+    // large BG sizes:
+    // 0: 512x1024
+    // 1: 1024x512
+    // 2: 512x256
+    // 3: 512x512
     u32 xmask, ymask;
     u32 yshift;
     switch (bgcnt & 0xC000)
     {
     case 0x0000: xmask = 0x1FFFF; ymask = 0x3FFFF; yshift = 9; break;
     case 0x4000: xmask = 0x3FFFF; ymask = 0x1FFFF; yshift = 10; break;
-    case 0x8000: // TODO (most likely the second size bit is just ignored)
-    case 0xC000: printf("bad BG size for large BG: %04X\n", bgcnt); return;
+    case 0x8000: xmask = 0x1FFFF; ymask = 0x0FFFF; yshift = 9; break;
+    case 0xC000: xmask = 0x1FFFF; ymask = 0x1FFFF; yshift = 9; break;
     }
 
     u32 ofxmask, ofymask;
@@ -2134,21 +2192,19 @@ void GPU2D::DrawSprites(u32 line)
 
     const s32 spritewidth[16] =
     {
-        8, 16, 8, 0,
-        16, 32, 8, 0,
-        32, 32, 16, 0,
-        64, 64, 32, 0
+        8, 16, 8, 8,
+        16, 32, 8, 8,
+        32, 32, 16, 8,
+        64, 64, 32, 8
     };
     const s32 spriteheight[16] =
     {
-        8, 8, 16, 0,
-        16, 8, 32, 0,
-        32, 16, 32, 0,
-        64, 32, 64, 0
+        8, 8, 16, 8,
+        16, 8, 32, 8,
+        32, 16, 32, 8,
+        64, 32, 64, 8
     };
 
-    u32 nsprites = 0;
-
     for (int bgnum = 0x0C00; bgnum >= 0x0000; bgnum -= 0x0400)
     {
         for (int sprnum = 127; sprnum >= 0; sprnum--)
@@ -2287,8 +2343,10 @@ void GPU2D::DrawSprite_Rotscale(u16* attrib, u16* rotparams, u32 boundwidth, u32
         {
             if (DispCnt & 0x20)
             {
-                // TODO ("reserved")
-                printf("bad reserved mode\n");
+                // 'reserved'
+                // draws nothing
+
+                return;
             }
             else
             {
@@ -2520,8 +2578,10 @@ void GPU2D::DrawSprite_Normal(u16* attrib, u32 width, s32 xpos, s32 ypos)
         {
             if (DispCnt & 0x20)
             {
-                // TODO ("reserved")
-                printf("bad reserved mode\n");
+                // 'reserved'
+                // draws nothing
+
+                return;
             }
             else
             {
diff --git a/src/GPU2D.h b/src/GPU2D.h
index b9a2422..6ad9783 100644
--- a/src/GPU2D.h
+++ b/src/GPU2D.h
@@ -135,6 +135,7 @@ private:
 
     template<u32 bgmode> void DrawScanlineBGMode(u32 line);
     void DrawScanlineBGMode6(u32 line);
+    void DrawScanlineBGMode7(u32 line);
     void DrawScanline_BGOBJ(u32 line);
 
     static void DrawPixel_Normal(u32* dst, u16 color, u32 flag);
diff --git a/src/GPU3D_OpenGL.cpp b/src/GPU3D_OpenGL.cpp
index 0d86690..d616e1c 100644
--- a/src/GPU3D_OpenGL.cpp
+++ b/src/GPU3D_OpenGL.cpp
@@ -342,36 +342,11 @@ bool Init()
     SetupDefaultTexParams(FramebufferTex[7]);
 
     // downscale framebuffer for antialiased mode
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[2]);
     SetupDefaultTexParams(FramebufferTex[2]);
 
     // downscale framebuffer for display capture (always 256x192)
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[3]);
     SetupDefaultTexParams(FramebufferTex[3]);
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, 256, 192, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[3], 0);
-
-    GLenum fbassign[2] = {GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1};
-
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[0], 0);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[4], 0);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[5], 0);
-    glDrawBuffers(2, fbassign);
-
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[1]);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[1], 0);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[4], 0);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[5], 0);
-    glDrawBuffers(2, fbassign);
-
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[2]);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[2], 0);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[6], 0);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[7], 0);
-    glDrawBuffers(2, fbassign);
-
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
 
     glEnable(GL_BLEND);
     glBlendEquationSeparate(GL_FUNC_ADD, GL_MAX);
@@ -475,11 +450,36 @@ void UpdateDisplaySettings()
         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB8UI, ScreenW, ScreenH, 0, GL_RGB_INTEGER, GL_UNSIGNED_BYTE, NULL);
     }
 
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[3]);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[3], 0);
+
+    GLenum fbassign[2] = {GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1};
+
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[0], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[4], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[5], 0);
+    glDrawBuffers(2, fbassign);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[1]);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[1], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[4], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[5], 0);
+    glDrawBuffers(2, fbassign);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[2]);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[2], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[6], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[7], 0);
+    glDrawBuffers(2, fbassign);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
+
     glBindBuffer(GL_PIXEL_PACK_BUFFER, PixelbufferID);
     glBufferData(GL_PIXEL_PACK_BUFFER, 256*192*4, NULL, GL_DYNAMIC_READ);
 
     //glLineWidth(scale);
-    glLineWidth(1.5);
+    //glLineWidth(1.5);
 }
 
 
@@ -743,6 +743,7 @@ void RenderSceneChunk(int y, int h)
     // pass 1: opaque pixels
 
     UseRenderShader(flags);
+    glLineWidth(1.0);
 
     glColorMaski(1, GL_TRUE, GL_TRUE, fogenable, GL_FALSE);
 
@@ -774,6 +775,7 @@ void RenderSceneChunk(int y, int h)
     if (RenderDispCnt & (1<<5))
     {
         UseRenderShader(flags | RenderFlag_Edge);
+        glLineWidth(1.5);
 
         glColorMaski(0, GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
         glColorMaski(1, GL_FALSE, GL_TRUE, GL_FALSE, GL_FALSE);
@@ -806,6 +808,7 @@ void RenderSceneChunk(int y, int h)
         glBlendFuncSeparate(GL_ONE, GL_ZERO, GL_ONE, GL_ONE);
 
     UseRenderShader(flags | RenderFlag_Trans);
+    glLineWidth(1.0);
 
     if (NumOpaqueFinalPolys > -1)
     {
diff --git a/src/GPU3D_OpenGL_shaders.h b/src/GPU3D_OpenGL_shaders.h
index afe7518..2545ee0 100644
--- a/src/GPU3D_OpenGL_shaders.h
+++ b/src/GPU3D_OpenGL_shaders.h
@@ -656,7 +656,7 @@ void main()
     int zshift = (attr >> 16) & 0x1F;
 
     vec4 fpos;
-    fpos.xy = (((vec2(vPosition.xy) + 0.5) * 2.0) / uScreenSize) - 1.0;
+    fpos.xy = (((vec2(vPosition.xy) ) * 2.0) / uScreenSize) - 1.0;
     fpos.z = (float(vPosition.z << zshift) / 8388608.0) - 1.0;
     fpos.w = float(vPosition.w) / 65536.0f;
     fpos.xyz *= fpos.w;
@@ -679,7 +679,7 @@ void main()
     int zshift = (attr >> 16) & 0x1F;
 
     vec4 fpos;
-    fpos.xy = (((vec2(vPosition.xy) + 0.5) * 2.0) / uScreenSize) - 1.0;
+    fpos.xy = (((vec2(vPosition.xy) ) * 2.0) / uScreenSize) - 1.0;
     fZ = float(vPosition.z << zshift) / 16777216.0;
     fpos.w = float(vPosition.w) / 65536.0f;
     fpos.xy *= fpos.w;
diff --git a/src/NDS.cpp b/src/NDS.cpp
index 6933e95..18e2ae1 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -72,7 +72,7 @@ u64 FrameStartTimestamp;
 
 int CurCPU;
 
-const s32 kMaxIterationCycles = 16;
+const s32 kMaxIterationCycles = 64;
 
 u32 ARM9ClockShift;
 
@@ -1516,7 +1516,7 @@ void DivDone(u32 param)
             if (den == 0)
             {
                 DivQuotient[0] = (num<0) ? 1:-1;
-                DivQuotient[1] = (num<0) ? -1:1;
+                DivQuotient[1] = (num<0) ? -1:0;
                 *(s64*)&DivRemainder[0] = num;
             }
             else if (num == -0x80000000 && den == -1)
@@ -1648,7 +1648,8 @@ void debug(u32 param)
     //    printf("VRAM %c: %02X\n", 'A'+i, GPU::VRAMCNT[i]);
 
     /*FILE*
-    shit = fopen("debug/card.bin", "wb");
+    shit = fopen("debug/party.bin", "wb");
+    fwrite(ARM9->ITCM, 0x8000, 1, shit);
     for (u32 i = 0x02000000; i < 0x02400000; i+=4)
     {
         u32 val = ARM7Read32(i);
@@ -2936,6 +2937,11 @@ void ARM9IOWrite32(u32 addr, u32 val)
 {
     switch (addr)
     {
+    case 0x04000004:
+        GPU::SetDispStat(0, val & 0xFFFF);
+        GPU::SetVCount(val >> 16);
+        return;
+
     case 0x04000060: GPU3D::Write32(addr, val); return;
     case 0x04000064:
     case 0x04000068: GPU::GPU2D_A->Write32(addr, val); return;
@@ -3500,6 +3506,11 @@ void ARM7IOWrite32(u32 addr, u32 val)
 {
     switch (addr)
     {
+    case 0x04000004:
+        GPU::SetDispStat(1, val & 0xFFFF);
+        GPU::SetVCount(val >> 16);
+        return;
+
     case 0x040000B0: DMAs[4]->SrcAddr = val; return;
     case 0x040000B4: DMAs[4]->DstAddr = val; return;
     case 0x040000B8: DMAs[4]->WriteCnt(val); return;
@@ -3583,6 +3594,8 @@ void ARM7IOWrite32(u32 addr, u32 val)
     case 0x04000210: IE[1] = val; UpdateIRQ(1); return;
     case 0x04000214: IF[1] &= ~val; UpdateIRQ(1); return;
 
+    case 0x04000304: PowerControl7 = val & 0xFFFF; return;
+
     case 0x04000308:
         if (ARM7BIOSProt == 0)
             ARM7BIOSProt = val & 0xFFFE;
diff --git a/src/OpenGLSupport.h b/src/OpenGLSupport.h
index fbc100c..a08bcdf 100644
--- a/src/OpenGLSupport.h
+++ b/src/OpenGLSupport.h
@@ -112,8 +112,6 @@
     func(GLGETUNIFORMLOCATION, glGetUniformLocation); \
     func(GLGETUNIFORMBLOCKINDEX, glGetUniformBlockIndex); \
      \
-    func(GLBINDIMAGETEXTURE, glBindImageTexture); \
-     \
     func(GLDRAWBUFFERS, glDrawBuffers); \
      \
     func(GLBLENDFUNCSEPARATE, glBlendFuncSeparate); \
diff --git a/src/SPU.cpp b/src/SPU.cpp
index d31a371..8e295c3 100644
--- a/src/SPU.cpp
+++ b/src/SPU.cpp
@@ -28,7 +28,6 @@
 // * channel hold
 // * 'length less than 4' glitch
 
-
 namespace SPU
 {
 
@@ -66,8 +65,8 @@ const u32 kSamplesPerRun = 1;
 
 const u32 OutputBufferSize = 2*1024;
 s16 OutputBuffer[2 * OutputBufferSize];
-u32 OutputReadOffset;
-u32 OutputWriteOffset;
+volatile u32 OutputReadOffset;
+volatile u32 OutputWriteOffset;
 
 
 u16 Cnt;
@@ -100,9 +99,7 @@ void DeInit()
 
 void Reset()
 {
-    memset(OutputBuffer, 0, 2*OutputBufferSize*2);
-    OutputReadOffset = 0;
-    OutputWriteOffset = OutputBufferSize;
+    InitOutput();
 
     Cnt = 0;
     MasterVolume = 0;
@@ -632,7 +629,7 @@ void Mix(u32 samples)
                 else if (val > 0x7FFF)  val = 0x7FFF;
 
                 Capture[0]->Run(val);
-                if (!((Capture[0]->Cnt & (1<<7)))) break;
+                if (!(Capture[0]->Cnt & (1<<7))) break;
             }
         }
 
@@ -647,7 +644,7 @@ void Mix(u32 samples)
                 else if (val > 0x7FFF)  val = 0x7FFF;
 
                 Capture[1]->Run(val);
-                if (!((Capture[1]->Cnt & (1<<7)))) break;
+                if (!(Capture[1]->Cnt & (1<<7))) break;
             }
         }
 
@@ -737,12 +734,76 @@ void Mix(u32 samples)
         OutputBuffer[OutputWriteOffset + 1] = r >> 1;
         OutputWriteOffset += 2;
         OutputWriteOffset &= ((2*OutputBufferSize)-1);
+        if (OutputWriteOffset == OutputReadOffset)
+        {
+            //printf("!! SOUND FIFO OVERFLOW %d\n", OutputWriteOffset>>1);
+            // advance the read position too, to avoid losing the entire FIFO
+            OutputReadOffset += 2;
+            OutputReadOffset &= ((2*OutputBufferSize)-1);
+        }
     }
 
     NDS::ScheduleEvent(NDS::Event_SPU, true, 1024*kSamplesPerRun, Mix, kSamplesPerRun);
 }
 
 
+void TrimOutput()
+{
+    const int halflimit = (OutputBufferSize / 2);
+
+    int readpos = OutputWriteOffset - (halflimit*2);
+    if (readpos < 0) readpos += (OutputBufferSize*2);
+
+    OutputReadOffset = readpos;
+}
+
+void DrainOutput()
+{
+    OutputReadOffset = 0;
+    OutputWriteOffset = 0;
+}
+
+void InitOutput()
+{
+    memset(OutputBuffer, 0, 2*OutputBufferSize*2);
+    OutputReadOffset = 0;
+    OutputWriteOffset = OutputBufferSize;
+}
+
+int GetOutputSize()
+{
+    int ret;
+    if (OutputWriteOffset >= OutputReadOffset)
+        ret = OutputWriteOffset - OutputReadOffset;
+    else
+        ret = (OutputBufferSize*2) - OutputReadOffset + OutputWriteOffset;
+
+    ret >>= 1;
+    return ret;
+}
+
+void Sync(bool wait)
+{
+    // sync to audio output in case the core is running too fast
+    // * wait=true: wait until enough audio data has been played
+    // * wait=false: merely skip some audio data to avoid a FIFO overflow
+
+    const int halflimit = (OutputBufferSize / 2);
+
+    if (wait)
+    {
+        // TODO: less CPU-intensive wait?
+        while (GetOutputSize() > halflimit);
+    }
+    else if (GetOutputSize() > halflimit)
+    {
+        int readpos = OutputWriteOffset - (halflimit*2);
+        if (readpos < 0) readpos += (OutputBufferSize*2);
+
+        OutputReadOffset = readpos;
+    }
+}
+
 int ReadOutput(s16* data, int samples)
 {
     if (OutputReadOffset == OutputWriteOffset)
@@ -910,8 +971,8 @@ void Write16(u32 addr, u16 val)
             return;
         case 0xA: chan->SetLoopPos(val); return;
 
-        case 0xC: chan->SetLength((chan->Length & 0xFFFF0000) | val); return;
-        case 0xE: chan->SetLength((chan->Length & 0x0000FFFF) | (val << 16)); return;
+        case 0xC: chan->SetLength(((chan->Length >> 2) & 0xFFFF0000) | val); return;
+        case 0xE: chan->SetLength(((chan->Length >> 2) & 0x0000FFFF) | (val << 16)); return;
         }
     }
     else
diff --git a/src/SPU.h b/src/SPU.h
index 3cd5df3..53a8e0a 100644
--- a/src/SPU.h
+++ b/src/SPU.h
@@ -35,6 +35,11 @@ void SetBias(u16 bias);
 
 void Mix(u32 samples);
 
+void TrimOutput();
+void DrainOutput();
+void InitOutput();
+int GetOutputSize();
+void Sync(bool wait);
 int ReadOutput(s16* data, int samples);
 
 u8 Read8(u32 addr);
diff --git a/src/libui_sdl/CMakeLists.txt b/src/libui_sdl/CMakeLists.txt
index 63e9f48..64206bf 100644
--- a/src/libui_sdl/CMakeLists.txt
+++ b/src/libui_sdl/CMakeLists.txt
@@ -14,6 +14,10 @@ SET(SOURCES_LIBUI
 	OSD.cpp
 )
 
+if (WIN32)
+	set(CMAKE_RC_COMPILE_OBJECT "<CMAKE_RC_COMPILER> -i <SOURCE> -o <OBJECT>")
+endif()
+
 option(BUILD_SHARED_LIBS "Whether to build libui as a shared library or a static library" ON)
 set(BUILD_SHARED_LIBS OFF)
 add_subdirectory(libui)
@@ -39,11 +43,11 @@ if (UNIX)
 	ADD_DEFINITIONS(${GTK3_CFLAGS_OTHER})
 
 	add_custom_command(OUTPUT melon_grc.c
-		COMMAND glib-compile-resources --sourcedir="${CMAKE_SOURCE_DIR}"
-				--target="${CMAKE_CURRENT_BINARY_DIR}/melon_grc.c"
+		COMMAND glib-compile-resources --sourcedir=${CMAKE_SOURCE_DIR}
+				--target=${CMAKE_CURRENT_BINARY_DIR}/melon_grc.c
 				--generate-source "${CMAKE_SOURCE_DIR}/melon_grc.xml"
-		COMMAND glib-compile-resources --sourcedir="${CMAKE_SOURCE_DIR}"
-				--target="${CMAKE_CURRENT_BINARY_DIR}/melon_grc.h"
+		COMMAND glib-compile-resources --sourcedir=${CMAKE_SOURCE_DIR}
+				--target=${CMAKE_CURRENT_BINARY_DIR}/melon_grc.h
 				--generate-header "${CMAKE_SOURCE_DIR}/melon_grc.xml")
 
 	if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
diff --git a/src/libui_sdl/DlgInputConfig.cpp b/src/libui_sdl/DlgInputConfig.cpp
index 8882275..a80e2ec 100644
--- a/src/libui_sdl/DlgInputConfig.cpp
+++ b/src/libui_sdl/DlgInputConfig.cpp
@@ -125,21 +125,30 @@ void JoyMappingName(int id, char* str)
         return;
     }
 
-    if (id & 0x100)
+    bool hasbtn = ((id & 0xFFFF) != 0xFFFF);
+
+    if (hasbtn)
     {
-        int hatnum = ((id >> 4) & 0xF) + 1;
+        if (id & 0x100)
+        {
+            int hatnum = ((id >> 4) & 0xF) + 1;
 
-        switch (id & 0xF)
+            switch (id & 0xF)
+            {
+            case 0x1: sprintf(str, "Hat %d up", hatnum); break;
+            case 0x2: sprintf(str, "Hat %d right", hatnum); break;
+            case 0x4: sprintf(str, "Hat %d down", hatnum); break;
+            case 0x8: sprintf(str, "Hat %d left", hatnum); break;
+            }
+        }
+        else
         {
-        case 0x1: sprintf(str, "Hat %d up", hatnum); break;
-        case 0x2: sprintf(str, "Hat %d right", hatnum); break;
-        case 0x4: sprintf(str, "Hat %d down", hatnum); break;
-        case 0x8: sprintf(str, "Hat %d left", hatnum); break;
+            sprintf(str, "Button %d", (id & 0xFFFF) + 1);
         }
     }
     else
     {
-        sprintf(str, "Button %d", (id & 0xFFFF) + 1);
+        strcpy(str, "");
     }
 
     if (id & 0x10000)
@@ -151,9 +160,9 @@ void JoyMappingName(int id, char* str)
 
         switch ((id >> 20) & 0xF)
         {
-        case 0: sprintf(str, "%s / Axis %d +", tmp, axisnum); break;
-        case 1: sprintf(str, "%s / Axis %d -", tmp, axisnum); break;
-        case 2: sprintf(str, "%s / Trigger %d", tmp, axisnum); break;
+        case 0: sprintf(str, "%s%sAxis %d +", tmp, hasbtn?" / ":"", axisnum); break;
+        case 1: sprintf(str, "%s%sAxis %d -", tmp, hasbtn?" / ":"", axisnum); break;
+        case 2: sprintf(str, "%s%sTrigger %d", tmp, hasbtn?" / ":"", axisnum); break;
         }
     }
 }
@@ -223,10 +232,12 @@ int OnAreaKeyEvent(uiAreaHandler* handler, uiArea* area, uiAreaKeyEvent* evt)
         // set key.
         if (evt->Scancode != 0x1 || evt->Modifiers != 0) // ESC
         {
+            int mod = (dlg->type == 0) ? 0 : evt->Modifiers;
+
             if (evt->Scancode == 0xE && evt->Modifiers == 0) // backspace
                 dlg->keymap[dlg->pollid] = -1;
             else
-                dlg->keymap[dlg->pollid] = evt->Scancode | (evt->Modifiers << 16);
+                dlg->keymap[dlg->pollid] = evt->Scancode | (mod << 16);
         }
 
         char keyname[64];
@@ -287,7 +298,7 @@ Uint32 JoyPoll(Uint32 interval, void* param)
     }
 
     int oldmap;
-    if (dlg->joymap[id] == -1) oldmap = 0;
+    if (dlg->joymap[id] == -1) oldmap = 0xFFFF;
     else                       oldmap = dlg->joymap[id];
 
     int nbuttons = SDL_JoystickNumButtons(joy);
diff --git a/src/libui_sdl/DlgVideoSettings.cpp b/src/libui_sdl/DlgVideoSettings.cpp
index 7d876e2..ff88ba8 100644
--- a/src/libui_sdl/DlgVideoSettings.cpp
+++ b/src/libui_sdl/DlgVideoSettings.cpp
@@ -39,12 +39,14 @@ uiWindow* win;
 
 uiRadioButtons* rbRenderer;
 uiCheckbox* cbGLDisplay;
+uiCheckbox* cbVSync;
 uiCheckbox* cbThreaded3D;
 uiCombobox* cbResolution;
 uiCheckbox* cbAntialias;
 
 int old_renderer;
 int old_gldisplay;
+int old_vsync;
 int old_threaded3D;
 int old_resolution;
 int old_antialias;
@@ -89,6 +91,11 @@ void RevertSettings()
     {
         Config::ScreenUseGL = old_gldisplay;
     }
+    if (old_vsync != Config::ScreenVSync)
+    {
+        Config::ScreenVSync = old_vsync;
+        //ApplyNewSettings(4);
+    }
     if (old_usegl != new_usegl)
     {
         apply2 = true;
@@ -130,21 +137,33 @@ void OnRendererChanged(uiRadioButtons* rb, void* blarg)
     UpdateControls();
 
     bool new_usegl = (Config::ScreenUseGL != 0) || (Config::_3DRenderer != 0);
+
+    if (new_usegl) uiControlEnable(uiControl(cbVSync));
+    else           uiControlDisable(uiControl(cbVSync));
+
     if (new_usegl != old_usegl)
         ApplyNewSettings(2);
     else
         ApplyNewSettings(3);
-    
+
     uiControlSetFocus(uiControl(win));
 }
 
 void OnGLDisplayChanged(uiCheckbox* cb, void* blarg)
 {
     Config::ScreenUseGL = uiCheckboxChecked(cb);
+    if (Config::ScreenUseGL) uiControlEnable(uiControl(cbVSync));
+    else                     uiControlDisable(uiControl(cbVSync));
     ApplyNewSettings(2);
     uiControlSetFocus(uiControl(win));
 }
 
+void OnVSyncChanged(uiCheckbox* cb, void* blarg)
+{
+    Config::ScreenVSync = uiCheckboxChecked(cb);
+    //ApplyNewSettings(4);
+}
+
 void OnThreaded3DChanged(uiCheckbox* cb, void* blarg)
 {
     Config::Threaded3D = uiCheckboxChecked(cb);
@@ -232,6 +251,10 @@ void Open()
         cbGLDisplay = uiNewCheckbox("OpenGL display");
         uiCheckboxOnToggled(cbGLDisplay, OnGLDisplayChanged, NULL);
         uiBoxAppend(in_ctrl, uiControl(cbGLDisplay), 0);
+
+        cbVSync = uiNewCheckbox("VSync");
+        uiCheckboxOnToggled(cbVSync, OnVSyncChanged, NULL);
+        uiBoxAppend(in_ctrl, uiControl(cbVSync), 0);
     }
 
     {
@@ -300,17 +323,24 @@ void Open()
 
     old_renderer = Config::_3DRenderer;
     old_gldisplay = Config::ScreenUseGL;
+    old_vsync = Config::ScreenVSync;
     old_threaded3D = Config::Threaded3D;
     old_resolution = Config::GL_ScaleFactor;
     old_antialias = Config::GL_Antialias;
 
     uiCheckboxSetChecked(cbGLDisplay, Config::ScreenUseGL);
+    uiCheckboxSetChecked(cbVSync, Config::ScreenVSync);
     uiCheckboxSetChecked(cbThreaded3D, Config::Threaded3D);
     uiComboboxSetSelected(cbResolution, Config::GL_ScaleFactor-1);
     //uiCheckboxSetChecked(cbAntialias, Config::GL_Antialias);
     uiRadioButtonsSetSelected(rbRenderer, Config::_3DRenderer);
     UpdateControls();
 
+    if (Config::ScreenUseGL || Config::_3DRenderer != 0)
+        uiControlEnable(uiControl(cbVSync));
+    else
+        uiControlDisable(uiControl(cbVSync));
+
     uiControlShow(uiControl(win));
 }
 
diff --git a/src/libui_sdl/PlatformConfig.cpp b/src/libui_sdl/PlatformConfig.cpp
index 7f45b59..c8ec19f 100644
--- a/src/libui_sdl/PlatformConfig.cpp
+++ b/src/libui_sdl/PlatformConfig.cpp
@@ -43,9 +43,11 @@ int ScreenSizing;
 int ScreenFilter;
 
 int ScreenUseGL;
+int ScreenVSync;
 int ScreenRatio;
 
 int LimitFPS;
+int AudioSync;
 int ShowOSD;
 
 int DirectBoot;
@@ -118,9 +120,11 @@ ConfigEntry PlatformConfigFile[] =
     {"ScreenFilter",   0, &ScreenFilter,   1, NULL, 0},
 
     {"ScreenUseGL",     0, &ScreenUseGL,     1, NULL, 0},
+    {"ScreenVSync",     0, &ScreenVSync,     0, NULL, 0},
     {"ScreenRatio",     0, &ScreenRatio,     0, NULL, 0},
 
-    {"LimitFPS", 0, &LimitFPS, 1, NULL, 0},
+    {"LimitFPS", 0, &LimitFPS, 0, NULL, 0},
+    {"AudioSync", 0, &AudioSync, 1, NULL, 0},
     {"ShowOSD", 0, &ShowOSD, 1, NULL, 0},
 
     {"DirectBoot", 0, &DirectBoot, 1, NULL, 0},
diff --git a/src/libui_sdl/PlatformConfig.h b/src/libui_sdl/PlatformConfig.h
index 2c59e5d..842e72a 100644
--- a/src/libui_sdl/PlatformConfig.h
+++ b/src/libui_sdl/PlatformConfig.h
@@ -54,9 +54,11 @@ extern int ScreenSizing;
 extern int ScreenFilter;
 
 extern int ScreenUseGL;
+extern int ScreenVSync;
 extern int ScreenRatio;
 
 extern int LimitFPS;
+extern int AudioSync;
 extern int ShowOSD;
 
 extern int DirectBoot;
diff --git a/src/libui_sdl/libui/ui.h b/src/libui_sdl/libui/ui.h
index e15c127..03aef5d 100644
--- a/src/libui_sdl/libui/ui.h
+++ b/src/libui_sdl/libui/ui.h
@@ -620,6 +620,7 @@ _UI_EXTERN void *uiGLGetProcAddress(const char* proc);
 _UI_EXTERN int uiGLGetFramebuffer(uiGLContext* ctx);
 _UI_EXTERN float uiGLGetFramebufferScale(uiGLContext* ctx);
 _UI_EXTERN void uiGLSwapBuffers(uiGLContext* ctx);
+_UI_EXTERN void uiGLSetVSync(int sync);
 
 
 _UI_ENUM(uiModifiers) {
diff --git a/src/libui_sdl/libui/unix/gl.c b/src/libui_sdl/libui/unix/gl.c
index e1665ee..e15cf4f 100644
--- a/src/libui_sdl/libui/unix/gl.c
+++ b/src/libui_sdl/libui/unix/gl.c
@@ -193,6 +193,11 @@ void uiGLSwapBuffers(uiGLContext* ctx)
     ctx->backbuffer = ctx->backbuffer ? 0 : 1;
 }
 
+void uiGLSetVSync(int sync)
+{
+	// TODO
+}
+
 void uiGLMakeContextCurrent(uiGLContext* ctx)
 {
 	if (!ctx)
diff --git a/src/libui_sdl/libui/windows/CMakeLists.txt b/src/libui_sdl/libui/windows/CMakeLists.txt
index 9d5313a..24d4ad9 100644
--- a/src/libui_sdl/libui/windows/CMakeLists.txt
+++ b/src/libui_sdl/libui/windows/CMakeLists.txt
@@ -73,7 +73,7 @@ macro(_handle_static)
 	add_custom_command(
 		TARGET libui POST_BUILD
 		COMMAND
-			${CMAKE_COMMAND} -E copy $<TARGET_PROPERTY:libui,BINARY_DIR>/CMakeFiles/libui.dir/windows/resources.rc.* ${_LIBUI_STATIC_RES}
+			${CMAKE_COMMAND} -E copy $<TARGET_PROPERTY:libui,BINARY_DIR>/CMakeFiles/libui.dir/windows/resources.rc.obj ${_LIBUI_STATIC_RES}
 		COMMENT "Copying libui.res")
 endmacro()
 
diff --git a/src/libui_sdl/libui/windows/gl.cpp b/src/libui_sdl/libui/windows/gl.cpp
index c621721..07ef19b 100644
--- a/src/libui_sdl/libui/windows/gl.cpp
+++ b/src/libui_sdl/libui/windows/gl.cpp
@@ -3,6 +3,7 @@
 #include "area.hpp"
 
 #include <GL/gl.h>
+#include <GL/glext.h>
 #include <GL/wglext.h>
 
 struct uiGLContext
@@ -159,3 +160,37 @@ float uiGLGetFramebufferScale(uiGLContext* ctx)
     // TODO
     return 1;
 }
+
+void uiGLSetVSync(int sync)
+{
+	static PFNWGLSWAPINTERVALEXTPROC _wglSwapIntervalEXT = NULL;
+	static bool symloaded = false;
+
+	if (!symloaded)
+    {
+        PFNGLGETSTRINGIPROC _glGetStringi = (PFNGLGETSTRINGIPROC)wglGetProcAddress("glGetStringi");
+        if (_glGetStringi == NULL) return;
+
+        GLint numext;
+        glGetIntegerv(GL_NUM_EXTENSIONS, &numext);
+
+        bool hasswapctrl = false;
+        for (GLint i = 0; i < numext; i++)
+        {
+            const char* ext = (const char*)_glGetStringi(GL_EXTENSIONS, i);
+            if (!stricmp(ext, "WGL_EXT_swap_control"))
+            {
+                hasswapctrl = true;
+                break;
+            }
+        }
+
+        if (hasswapctrl)
+            _wglSwapIntervalEXT = (PFNWGLSWAPINTERVALEXTPROC)wglGetProcAddress("wglSwapIntervalEXT");
+
+        symloaded = true;
+    }
+
+    if (_wglSwapIntervalEXT)
+        _wglSwapIntervalEXT(sync);
+}
diff --git a/src/libui_sdl/libui/windows/window.cpp b/src/libui_sdl/libui/windows/window.cpp
index 18d1171..a8f7f23 100644
--- a/src/libui_sdl/libui/windows/window.cpp
+++ b/src/libui_sdl/libui/windows/window.cpp
@@ -95,7 +95,7 @@ static LRESULT CALLBACK windowWndProc(HWND hwnd, UINT uMsg, WPARAM wParam, LPARA
 		// not a menu
 		if (lParam != 0)
 			break;
-		if (HIWORD(wParam) != 0)
+		if (HIWORD(wParam) != 0 || LOWORD(wParam) <= IDCANCEL)
 			break;
 		runMenuEvent(LOWORD(wParam), uiWindow(w));
 		return 0;
diff --git a/src/libui_sdl/main.cpp b/src/libui_sdl/main.cpp
index 34e838c..d3f9024 100644
--- a/src/libui_sdl/main.cpp
+++ b/src/libui_sdl/main.cpp
@@ -93,6 +93,7 @@ uiMenuItem* MenuItem_ScreenSizing[4];
 
 uiMenuItem* MenuItem_ScreenFilter;
 uiMenuItem* MenuItem_LimitFPS;
+uiMenuItem* MenuItem_AudioSync;
 uiMenuItem* MenuItem_ShowOSD;
 
 SDL_Thread* EmuThread;
@@ -129,6 +130,8 @@ bool GL_ScreenSizeDirty;
 
 int GL_3DScale;
 
+bool GL_VSyncStatus;
+
 int ScreenGap = 0;
 int ScreenLayout = 0;
 int ScreenSizing = 0;
@@ -158,8 +161,13 @@ bool LidStatus;
 int JoystickID;
 SDL_Joystick* Joystick;
 
+int AudioFreq;
+float AudioSampleFrac;
 SDL_AudioDeviceID AudioDevice, MicDevice;
 
+SDL_cond* AudioSync;
+SDL_mutex* AudioSyncLock;
+
 u32 MicBufferLength = 2048;
 s16 MicBuffer[2048];
 u32 MicBufferReadPos, MicBufferWritePos;
@@ -236,6 +244,8 @@ bool GLScreen_InitOSDShader(GLuint* shader)
 
 bool GLScreen_Init()
 {
+    GL_VSyncStatus = Config::ScreenVSync;
+
     // TODO: consider using epoxy?
     if (!OpenGL_Init())
         return false;
@@ -298,6 +308,13 @@ void GLScreen_DeInit()
 
 void GLScreen_DrawScreen()
 {
+    bool vsync = Config::ScreenVSync && !HotkeyDown(HK_FastForward);
+    if (vsync != GL_VSyncStatus)
+    {
+        GL_VSyncStatus = vsync;
+        uiGLSetVSync(vsync);
+    }
+
     float scale = uiGLGetFramebufferScale(GLContext);
 
     glBindFramebuffer(GL_FRAMEBUFFER, uiGLGetFramebuffer(GLContext));
@@ -560,26 +577,42 @@ void MicLoadWav(char* name)
 
 void AudioCallback(void* data, Uint8* stream, int len)
 {
-    // resampling:
-    // buffer length is 1024 samples
-    // which is 710 samples at the original sample rate
+    len /= (sizeof(s16) * 2);
+
+    // resample incoming audio to match the output sample rate
 
-    s16 buf_in[710*2];
+    float f_len_in = (len * 32823.6328125) / (float)AudioFreq;
+    f_len_in += AudioSampleFrac;
+    int len_in = (int)floor(f_len_in);
+    AudioSampleFrac = f_len_in - len_in;
+
+    s16 buf_in[1024*2];
     s16* buf_out = (s16*)stream;
 
-    int num_in = SPU::ReadOutput(buf_in, 710);
-    int num_out = 1024;
+    int num_in;
+    int num_out = len;
+
+    SDL_LockMutex(AudioSyncLock);
+    num_in = SPU::ReadOutput(buf_in, len_in);
+    SDL_CondSignal(AudioSync);
+    SDL_UnlockMutex(AudioSyncLock);
+
+    if (num_in < 1)
+    {
+        memset(stream, 0, len*sizeof(s16)*2);
+        return;
+    }
 
     int margin = 6;
-    if (num_in < 710-margin)
+    if (num_in < len_in-margin)
     {
         int last = num_in-1;
         if (last < 0) last = 0;
 
-        for (int i = num_in; i < 710-margin; i++)
+        for (int i = num_in; i < len_in-margin; i++)
             ((u32*)buf_in)[i] = ((u32*)buf_in)[last];
 
-        num_in = 710-margin;
+        num_in = len_in-margin;
     }
 
     float res_incr = num_in / (float)num_out;
@@ -588,12 +621,22 @@ void AudioCallback(void* data, Uint8* stream, int len)
 
     int volume = Config::AudioVolume;
 
-    for (int i = 0; i < 1024; i++)
+    for (int i = 0; i < len; i++)
     {
-        // TODO: interp!!
         buf_out[i*2  ] = (buf_in[res_pos*2  ] * volume) >> 8;
         buf_out[i*2+1] = (buf_in[res_pos*2+1] * volume) >> 8;
 
+        /*s16 s_l = buf_in[res_pos*2  ];
+        s16 s_r = buf_in[res_pos*2+1];
+
+        float a = res_timer;
+        float b = 1.0 - a;
+        s_l = (s_l * a) + (buf_in[(res_pos-1)*2  ] * b);
+        s_r = (s_r * a) + (buf_in[(res_pos-1)*2+1] * b);
+
+        buf_out[i*2  ] = (s_l * volume) >> 8;
+        buf_out[i*2+1] = (s_r * volume) >> 8;*/
+
         res_timer += res_incr;
         while (res_timer >= 1.0)
         {
@@ -710,26 +753,31 @@ bool JoystickButtonDown(int val)
 {
     if (val == -1) return false;
 
-    if (val & 0x100)
+    bool hasbtn = ((val & 0xFFFF) != 0xFFFF);
+
+    if (hasbtn)
     {
-        int hatnum = (val >> 4) & 0xF;
-        int hatdir = val & 0xF;
-        Uint8 hatval = SDL_JoystickGetHat(Joystick, hatnum);
+        if (val & 0x100)
+        {
+            int hatnum = (val >> 4) & 0xF;
+            int hatdir = val & 0xF;
+            Uint8 hatval = SDL_JoystickGetHat(Joystick, hatnum);
 
-        bool pressed = false;
-        if      (hatdir == 0x1) pressed = (hatval & SDL_HAT_UP);
-        else if (hatdir == 0x4) pressed = (hatval & SDL_HAT_DOWN);
-        else if (hatdir == 0x2) pressed = (hatval & SDL_HAT_RIGHT);
-        else if (hatdir == 0x8) pressed = (hatval & SDL_HAT_LEFT);
+            bool pressed = false;
+            if      (hatdir == 0x1) pressed = (hatval & SDL_HAT_UP);
+            else if (hatdir == 0x4) pressed = (hatval & SDL_HAT_DOWN);
+            else if (hatdir == 0x2) pressed = (hatval & SDL_HAT_RIGHT);
+            else if (hatdir == 0x8) pressed = (hatval & SDL_HAT_LEFT);
 
-        if (pressed) return true;
-    }
-    else
-    {
-        int btnnum = val & 0xFFFF;
-        Uint8 btnval = SDL_JoystickGetButton(Joystick, btnnum);
+            if (pressed) return true;
+        }
+        else
+        {
+            int btnnum = val & 0xFFFF;
+            Uint8 btnval = SDL_JoystickGetButton(Joystick, btnnum);
 
-        if (btnval) return true;
+            if (btnval) return true;
+        }
     }
 
     if (val & 0x10000)
@@ -833,6 +881,7 @@ bool JoyButtonHeld(int btnid, int njoybuttons, Uint8* joybuttons, Uint32 hat)
 
 void UpdateWindowTitle(void* data)
 {
+    if (EmuStatus == 0) return;
     uiWindowSetTitle(MainWindow, (const char*)data);
 }
 
@@ -875,6 +924,10 @@ int EmuThreadFunc(void* burp)
     u32 lasttick = starttick;
     u32 lastmeasuretick = lasttick;
     u32 fpslimitcount = 0;
+    u64 perfcount = SDL_GetPerformanceCounter();
+    u64 perffreq = SDL_GetPerformanceFrequency();
+    float samplesleft = 0;
+    u32 nsamples = 0;
     char melontitle[100];
 
     while (EmuRunning != 0)
@@ -886,6 +939,7 @@ int EmuThreadFunc(void* burp)
             Config::LimitFPS = !Config::LimitFPS;
             uiQueueMain(UpdateFPSLimit, NULL);
         }
+        // TODO: similar hotkeys for video/audio sync?
 
         if (HotkeyPressed(HK_Pause)) uiQueueMain(TogglePause, NULL);
         if (HotkeyPressed(HK_Reset)) uiQueueMain(Reset, NULL);
@@ -954,25 +1008,46 @@ int EmuThreadFunc(void* burp)
             }
             uiAreaQueueRedrawAll(MainDrawArea);
 
-            // framerate limiter based off SDL2_gfx
-            float framerate = (1000.0f * nlines) / (60.0f * 263.0f);
-
-            fpslimitcount++;
-            u32 curtick = SDL_GetTicks();
-            u32 delay = curtick - lasttick;
-            lasttick = curtick;
+            bool fastforward = HotkeyDown(HK_FastForward);
 
-            bool limitfps = Config::LimitFPS && !HotkeyDown(HK_FastForward);
-
-            u32 wantedtick = starttick + (u32)((float)fpslimitcount * framerate);
-            if (curtick < wantedtick && limitfps)
+            if (Config::AudioSync && !fastforward)
             {
-                SDL_Delay(wantedtick - curtick);
+                SDL_LockMutex(AudioSyncLock);
+                while (SPU::GetOutputSize() > 1024)
+                {
+                    int ret = SDL_CondWaitTimeout(AudioSync, AudioSyncLock, 500);
+                    if (ret == SDL_MUTEX_TIMEDOUT) break;
+                }
+                SDL_UnlockMutex(AudioSyncLock);
             }
             else
             {
-                fpslimitcount = 0;
-                starttick = curtick;
+                // ensure the audio FIFO doesn't overflow
+                //SPU::TrimOutput();
+            }
+
+            float framerate = (1000.0f * nlines) / (60.0f * 263.0f);
+
+            {
+                u32 curtick = SDL_GetTicks();
+                u32 delay = curtick - lasttick;
+
+                bool limitfps = Config::LimitFPS && !fastforward;
+                if (limitfps)
+                {
+                    float wantedtickF = starttick + (framerate * (fpslimitcount+1));
+                    u32 wantedtick = (u32)ceil(wantedtickF);
+                    if (curtick < wantedtick) SDL_Delay(wantedtick - curtick);
+
+                    lasttick = SDL_GetTicks();
+                    fpslimitcount++;
+                    if ((abs(wantedtickF - (float)wantedtick) < 0.001312) || (fpslimitcount > 60))
+                    {
+                        fpslimitcount = 0;
+                        nsamples = 0;
+                        starttick = lasttick;
+                    }
+                }
             }
 
             nframes++;
@@ -1044,6 +1119,12 @@ int EmuThreadFunc(void* burp)
     return 44203;
 }
 
+void StopEmuThread()
+{
+    EmuRunning = 0;
+    SDL_WaitThread(EmuThread, NULL);
+}
+
 
 void OnAreaDraw(uiAreaHandler* handler, uiArea* area, uiAreaDrawParams* params)
 {
@@ -1155,13 +1236,13 @@ void OnAreaDragBroken(uiAreaHandler* handler, uiArea* area)
 {
 }
 
-bool EventMatchesKey(uiAreaKeyEvent* evt, int val)
+bool EventMatchesKey(uiAreaKeyEvent* evt, int val, bool checkmod)
 {
     if (val == -1) return false;
 
     int key = val & 0xFFFF;
     int mod = val >> 16;
-    return evt->Scancode == key && evt->Modifiers == mod;
+    return evt->Scancode == key && (!checkmod || evt->Modifiers == mod);
 }
 
 int OnAreaKeyEvent(uiAreaHandler* handler, uiArea* area, uiAreaKeyEvent* evt)
@@ -1175,11 +1256,11 @@ int OnAreaKeyEvent(uiAreaHandler* handler, uiArea* area, uiAreaKeyEvent* evt)
     if (evt->Up)
     {
         for (int i = 0; i < 12; i++)
-            if (EventMatchesKey(evt, Config::KeyMapping[i]))
+            if (EventMatchesKey(evt, Config::KeyMapping[i], false))
                 KeyInputMask |= (1<<i);
 
         for (int i = 0; i < HK_MAX; i++)
-            if (EventMatchesKey(evt, Config::HKKeyMapping[i]))
+            if (EventMatchesKey(evt, Config::HKKeyMapping[i], true))
                 KeyHotkeyMask &= ~(1<<i);
     }
     else if (!evt->Repeat)
@@ -1202,16 +1283,16 @@ int OnAreaKeyEvent(uiAreaHandler* handler, uiArea* area, uiAreaKeyEvent* evt)
         }
 
         for (int i = 0; i < 12; i++)
-            if (EventMatchesKey(evt, Config::KeyMapping[i]))
+            if (EventMatchesKey(evt, Config::KeyMapping[i], false))
                 KeyInputMask &= ~(1<<i);
 
         for (int i = 0; i < HK_MAX; i++)
-            if (EventMatchesKey(evt, Config::HKKeyMapping[i]))
+            if (EventMatchesKey(evt, Config::HKKeyMapping[i], true))
                 KeyHotkeyMask |= (1<<i);
 
         // REMOVE ME
-        if (evt->Scancode == 0x57) // F11
-            NDS::debug(0);
+        //if (evt->Scancode == 0x57) // F11
+        //    NDS::debug(0);
     }
 
     return 1;
@@ -1486,6 +1567,8 @@ void Run()
     EmuRunning = 1;
     RunningSomething = true;
 
+    SPU::InitOutput();
+    AudioSampleFrac = 0;
     SDL_PauseAudioDevice(AudioDevice, 0);
     SDL_PauseAudioDevice(MicDevice, 0);
 
@@ -1524,6 +1607,7 @@ void TogglePause(void* blarg)
         EmuRunning = 2;
         uiMenuItemSetChecked(MenuItem_Pause, 1);
 
+        SPU::DrainOutput();
         SDL_PauseAudioDevice(AudioDevice, 1);
         SDL_PauseAudioDevice(MicDevice, 1);
 
@@ -1535,6 +1619,8 @@ void TogglePause(void* blarg)
         EmuRunning = 1;
         uiMenuItemSetChecked(MenuItem_Pause, 0);
 
+        SPU::InitOutput();
+        AudioSampleFrac = 0;
         SDL_PauseAudioDevice(AudioDevice, 0);
         SDL_PauseAudioDevice(MicDevice, 0);
 
@@ -1585,6 +1671,7 @@ void Stop(bool internal)
 
     uiAreaQueueRedrawAll(MainDrawArea);
 
+    SPU::DrainOutput();
     SDL_PauseAudioDevice(AudioDevice, 1);
     SDL_PauseAudioDevice(MicDevice, 1);
 
@@ -1849,6 +1936,7 @@ int OnCloseWindow(uiWindow* window, void* blarg)
     while (EmuStatus != 3);
 
     CloseAllDialogs();
+    StopEmuThread();
     uiQuit();
     return 1;
 }
@@ -1886,6 +1974,7 @@ void OnCloseByMenu(uiMenuItem* item, uiWindow* window, void* blarg)
     while (EmuStatus != 3);
 
     CloseAllDialogs();
+    StopEmuThread();
     DestroyMainWindow();
     uiQuit();
 }
@@ -2143,6 +2232,13 @@ void OnSetLimitFPS(uiMenuItem* item, uiWindow* window, void* blarg)
     else          Config::LimitFPS = false;
 }
 
+void OnSetAudioSync(uiMenuItem* item, uiWindow* window, void* blarg)
+{
+    int chk = uiMenuItemChecked(item);
+    if (chk != 0) Config::AudioSync = true;
+    else          Config::AudioSync = false;
+}
+
 void OnSetShowOSD(uiMenuItem* item, uiWindow* window, void* blarg)
 {
     int chk = uiMenuItemChecked(item);
@@ -2206,6 +2302,19 @@ void ApplyNewSettings(int type)
         GPU3D::InitRenderer(Screen_UseGL);
         if (Screen_UseGL) uiGLMakeContextCurrent(NULL);
     }
+    /*else if (type == 4) // vsync
+    {
+        if (Screen_UseGL)
+        {
+            uiGLMakeContextCurrent(GLContext);
+            uiGLSetVSync(Config::ScreenVSync);
+            uiGLMakeContextCurrent(NULL);
+        }
+        else
+        {
+            // TODO eventually: VSync for non-GL screen?
+        }
+    }*/
 
     EmuRunning = prevstatus;
 }
@@ -2374,11 +2483,16 @@ void CreateMainWindowMenu()
     MenuItem_ScreenFilter = uiMenuAppendCheckItem(menu, "Screen filtering");
     uiMenuItemOnClicked(MenuItem_ScreenFilter, OnSetScreenFiltering, NULL);
 
+    MenuItem_ShowOSD = uiMenuAppendCheckItem(menu, "Show OSD");
+    uiMenuItemOnClicked(MenuItem_ShowOSD, OnSetShowOSD, NULL);
+
+    uiMenuAppendSeparator(menu);
+
     MenuItem_LimitFPS = uiMenuAppendCheckItem(menu, "Limit framerate");
     uiMenuItemOnClicked(MenuItem_LimitFPS, OnSetLimitFPS, NULL);
 
-    MenuItem_ShowOSD = uiMenuAppendCheckItem(menu, "Show OSD");
-    uiMenuItemOnClicked(MenuItem_ShowOSD, OnSetShowOSD, NULL);
+    MenuItem_AudioSync = uiMenuAppendCheckItem(menu, "Audio sync");
+    uiMenuItemOnClicked(MenuItem_AudioSync, OnSetAudioSync, NULL);
 }
 
 void CreateMainWindow(bool opengl)
@@ -2415,6 +2529,7 @@ void CreateMainWindow(bool opengl)
     if (opengl_good)
     {
         uiGLMakeContextCurrent(GLContext);
+        uiGLSetVSync(Config::ScreenVSync);
         if (!GLScreen_Init()) opengl_good = false;
         if (opengl_good)
         {
@@ -2539,7 +2654,6 @@ int main(int argc, char** argv)
         SDL_Quit();
         return 0;
     }
-
     {
         FILE* f = Platform::OpenLocalFile("romlist.bin", "rb");
         if (f)
@@ -2556,6 +2670,13 @@ int main(int argc, char** argv)
                               "You should use the latest version of romlist.bin (provided in melonDS release packages).");
             }
         }
+        else
+        {
+        	uiMsgBoxError(NULL,
+        			     "romlist.bin not found.",
+        			     "Save memory type detection will not work correctly.\n\n"
+				         "You should use the latest version of romlist.bin (provided in melonDS release packages).");
+        }
     }
 
     CreateMainWindowMenu();
@@ -2613,22 +2734,29 @@ int main(int argc, char** argv)
 
     uiMenuItemSetChecked(MenuItem_ScreenFilter, Config::ScreenFilter==1);
     uiMenuItemSetChecked(MenuItem_LimitFPS, Config::LimitFPS==1);
+    uiMenuItemSetChecked(MenuItem_AudioSync, Config::AudioSync==1);
     uiMenuItemSetChecked(MenuItem_ShowOSD, Config::ShowOSD==1);
 
+    AudioSync = SDL_CreateCond();
+    AudioSyncLock = SDL_CreateMutex();
+
+    AudioFreq = 48000; // TODO: make configurable?
     SDL_AudioSpec whatIwant, whatIget;
     memset(&whatIwant, 0, sizeof(SDL_AudioSpec));
-    whatIwant.freq = 47340;
+    whatIwant.freq = AudioFreq;
     whatIwant.format = AUDIO_S16LSB;
     whatIwant.channels = 2;
     whatIwant.samples = 1024;
     whatIwant.callback = AudioCallback;
-    AudioDevice = SDL_OpenAudioDevice(NULL, 0, &whatIwant, &whatIget, 0);
+    AudioDevice = SDL_OpenAudioDevice(NULL, 0, &whatIwant, &whatIget, SDL_AUDIO_ALLOW_FREQUENCY_CHANGE);
     if (!AudioDevice)
     {
         printf("Audio init failed: %s\n", SDL_GetError());
     }
     else
     {
+        AudioFreq = whatIget.freq;
+        printf("Audio output frequency: %d Hz\n", AudioFreq);
         SDL_PauseAudioDevice(AudioDevice, 1);
     }
 
@@ -2683,13 +2811,13 @@ int main(int argc, char** argv)
 
     uiMain();
 
-    EmuRunning = 0;
-    SDL_WaitThread(EmuThread, NULL);
-
     if (Joystick) SDL_JoystickClose(Joystick);
     if (AudioDevice) SDL_CloseAudioDevice(AudioDevice);
     if (MicDevice)   SDL_CloseAudioDevice(MicDevice);
 
+    SDL_DestroyCond(AudioSync);
+    SDL_DestroyMutex(AudioSyncLock);
+
     if (MicWavBuffer) delete[] MicWavBuffer;
 
     if (ScreenBitmap[0]) uiDrawFreeBitmap(ScreenBitmap[0]);
diff --git a/src/version.h b/src/version.h
index bb97c3c..0b91540 100644
--- a/src/version.h
+++ b/src/version.h
@@ -19,7 +19,7 @@
 #ifndef VERSION_H
 #define VERSION_H
 
-#define MELONDS_VERSION    "0.8.1-DSi"
+#define MELONDS_VERSION    "0.8.3-DSi"
 
 #define MELONDS_URL        "http://melonds.kuribo64.net/"