directly set VRAMDirty for VRAM BG/OAM writes

2 files changed, 86 insertions, 69 deletions

diff --git a/src/GPU.cpp b/src/GPU.cpp
index d5465bb..d8e3d72 100644
--- a/src/GPU.cpp
+++ b/src/GPU.cpp
@@ -93,8 +93,7 @@ std::unique_ptr<GPU2D> GPU2D_B = {};
         we don't want to completely invalidate them every time they're unmapped and remapped
 
     For this reason we don't track the dirtyness per mapping region, but instead per VRAM bank
-    with VRAMDirty. Writes to LCDC go directly into VRAMDirty, while writes via other mapping regions
-    like BG or OBJ are first tracked in VRAMWritten_* and need to be flushed using SyncDirtyFlags.
+    with VRAMDirty.
 
     This is more or less a description of VRAMTrackingSet::DeriveState
         Each time before the memory is read two things could have happened
@@ -120,13 +119,6 @@ VRAMTrackingSet<8*1024, 8*1024> VRAMDirty_BOBJExtPal;
 VRAMTrackingSet<512*1024, 128*1024> VRAMDirty_Texture;
 VRAMTrackingSet<128*1024, 16*1024> VRAMDirty_TexPal;
 
-
-NonStupidBitField<512*1024/VRAMDirtyGranularity> VRAMWritten_ABG;
-NonStupidBitField<256*1024/VRAMDirtyGranularity> VRAMWritten_AOBJ;
-NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMWritten_BBG;
-NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMWritten_BOBJ;
-NonStupidBitField<256*1024/VRAMDirtyGranularity> VRAMWritten_ARM7;
-
 NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMDirty[9];
 
 u8 VRAMFlat_ABG[512*1024];
@@ -1028,8 +1020,6 @@ void StartHBlank(u32 line)
     DispStat[0] |= (1<<1);
     DispStat[1] |= (1<<1);
 
-    SyncDirtyFlags();
-
     if (VCount < 192)
     {
         // draw
@@ -1266,34 +1256,6 @@ template NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMTrackingSet<128*10
 template NonStupidBitField<256*1024/VRAMDirtyGranularity> VRAMTrackingSet<256*1024, 16*1024>::DeriveState(u32*);
 template NonStupidBitField<512*1024/VRAMDirtyGranularity> VRAMTrackingSet<512*1024, 16*1024>::DeriveState(u32*);
 
-template <u32 Size>
-void SyncDirtyFlags(u32* mappings, NonStupidBitField<Size>& writtenFlags)
-{
-    const u32 VRAMWrittenBitsPer16KB = 16*1024/VRAMDirtyGranularity;
-
-    for (typename NonStupidBitField<Size>::Iterator it = writtenFlags.Begin(); it != writtenFlags.End(); it++)
-    {
-        u32 mapping = mappings[*it / VRAMWrittenBitsPer16KB];
-        while (mapping != 0)
-        {
-            u32 num = __builtin_ctz(mapping);
-
-            VRAMDirty[num][*it & (VRAMMask[num] / VRAMDirtyGranularity)] = true;
-
-            mapping &= ~(1 << num);
-        }
-    }
-    writtenFlags.Clear();
-}
-
-void SyncDirtyFlags()
-{
-    SyncDirtyFlags(VRAMMap_ABG, VRAMWritten_ABG);
-    SyncDirtyFlags(VRAMMap_AOBJ, VRAMWritten_AOBJ);
-    SyncDirtyFlags(VRAMMap_BBG, VRAMWritten_BBG);
-    SyncDirtyFlags(VRAMMap_BOBJ, VRAMWritten_BOBJ);
-}
-
 template <u32 MappingGranularity, u32 Size>
 inline bool CopyLinearVRAM(u8* flat, u32* mappings, NonStupidBitField<Size>& dirty, u64 (*slowAccess)(u32 addr))
 {
diff --git a/src/GPU.h b/src/GPU.h
index 1e24051..3d9d13c 100644
--- a/src/GPU.h
+++ b/src/GPU.h
@@ -82,11 +82,6 @@ extern int Renderer;
 
 const u32 VRAMDirtyGranularity = 512;
 
-extern NonStupidBitField<512*1024/VRAMDirtyGranularity> VRAMWritten_ABG;
-extern NonStupidBitField<256*1024/VRAMDirtyGranularity> VRAMWritten_AOBJ;
-extern NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMWritten_BBG;
-extern NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMWritten_BOBJ;
-
 extern NonStupidBitField<128*1024/VRAMDirtyGranularity> VRAMDirty[9];
 
 template <u32 Size, u32 MappingGranularity>
@@ -350,15 +345,41 @@ void WriteVRAM_ABG(u32 addr, T val)
 {
     u32 mask = VRAMMap_ABG[(addr >> 14) & 0x1F];
 
-    VRAMWritten_ABG[(addr & 0x7FFFF) / VRAMDirtyGranularity] = true;
-
-    if (mask & (1<<0)) *(T*)&VRAM_A[addr & 0x1FFFF] = val;
-    if (mask & (1<<1)) *(T*)&VRAM_B[addr & 0x1FFFF] = val;
-    if (mask & (1<<2)) *(T*)&VRAM_C[addr & 0x1FFFF] = val;
-    if (mask & (1<<3)) *(T*)&VRAM_D[addr & 0x1FFFF] = val;
-    if (mask & (1<<4)) *(T*)&VRAM_E[addr & 0xFFFF] = val;
-    if (mask & (1<<5)) *(T*)&VRAM_F[addr & 0x3FFF] = val;
-    if (mask & (1<<6)) *(T*)&VRAM_G[addr & 0x3FFF] = val;
+    if (mask & (1<<0))
+    {
+        VRAMDirty[0][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_A[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<1))
+    {
+        VRAMDirty[1][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_B[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<2))
+    {
+        VRAMDirty[2][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_C[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<3))
+    {
+        VRAMDirty[3][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_D[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<4))
+    {
+        VRAMDirty[4][(addr & 0xFFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_E[addr & 0xFFFF] = val;
+    }
+    if (mask & (1<<5))
+    {
+        VRAMDirty[5][(addr & 0x3FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_F[addr & 0x3FFF] = val;
+    }
+    if (mask & (1<<6))
+    {
+        VRAMDirty[6][(addr & 0x3FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_G[addr & 0x3FFF] = val;
+    }
 }
 
 
@@ -385,13 +406,31 @@ void WriteVRAM_AOBJ(u32 addr, T val)
 {
     u32 mask = VRAMMap_AOBJ[(addr >> 14) & 0xF];
 
-    VRAMWritten_AOBJ[(addr & 0x3FFFF) / VRAMDirtyGranularity] = true;
-
-    if (mask & (1<<0)) *(T*)&VRAM_A[addr & 0x1FFFF] = val;
-    if (mask & (1<<1)) *(T*)&VRAM_B[addr & 0x1FFFF] = val;
-    if (mask & (1<<4)) *(T*)&VRAM_E[addr & 0xFFFF] = val;
-    if (mask & (1<<5)) *(T*)&VRAM_F[addr & 0x3FFF] = val;
-    if (mask & (1<<6)) *(T*)&VRAM_G[addr & 0x3FFF] = val;
+    if (mask & (1<<0))
+    {
+        VRAMDirty[0][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_A[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<1))
+    {
+        VRAMDirty[1][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_B[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<4))
+    {
+        VRAMDirty[4][(addr & 0xFFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_E[addr & 0xFFFF] = val;
+    }
+    if (mask & (1<<5))
+    {
+        VRAMDirty[5][(addr & 0x3FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_F[addr & 0x3FFF] = val;
+    }
+    if (mask & (1<<6))
+    {
+        VRAMDirty[6][(addr & 0x3FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_G[addr & 0x3FFF] = val;
+    }
 }
 
 
@@ -416,11 +455,21 @@ void WriteVRAM_BBG(u32 addr, T val)
 {
     u32 mask = VRAMMap_BBG[(addr >> 14) & 0x7];
 
-    VRAMWritten_BBG[(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
-
-    if (mask & (1<<2)) *(T*)&VRAM_C[addr & 0x1FFFF] = val;
-    if (mask & (1<<7)) *(T*)&VRAM_H[addr & 0x7FFF] = val;
-    if (mask & (1<<8)) *(T*)&VRAM_I[addr & 0x3FFF] = val;
+    if (mask & (1<<2))
+    {
+        VRAMDirty[2][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_C[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<7))
+    {
+        VRAMDirty[7][(addr & 0x7FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_H[addr & 0x7FFF] = val;
+    }
+    if (mask & (1<<8))
+    {
+        VRAMDirty[8][(addr & 0x3FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_I[addr & 0x3FFF] = val;
+    }
 }
 
 
@@ -444,10 +493,16 @@ void WriteVRAM_BOBJ(u32 addr, T val)
 {
     u32 mask = VRAMMap_BOBJ[(addr >> 14) & 0x7];
 
-    VRAMWritten_BOBJ[(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
-
-    if (mask & (1<<3)) *(T*)&VRAM_D[addr & 0x1FFFF] = val;
-    if (mask & (1<<8)) *(T*)&VRAM_I[addr & 0x3FFF] = val;
+    if (mask & (1<<3))
+    {
+        VRAMDirty[3][(addr & 0x1FFFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_D[addr & 0x1FFFF] = val;
+    }
+    if (mask & (1<<8))
+    {
+        VRAMDirty[8][(addr & 0x3FFF) / VRAMDirtyGranularity] = true;
+        *(T*)&VRAM_I[addr & 0x3FFF] = val;
+    }
 }
 
 template<typename T>