BPMemory: Refactor/consolidate TexUnit Addressing

Currently the logic for addressing the individual TexUnits is splattered all
across dolphin's codebase, this commit attempts to consolidate it all into a
single place and formalise it using our new TexUnitAddress struct.

7 files changed, 146 insertions, 59 deletions

diff --git a/Source/Core/VideoBackends/Software/DebugUtil.cpp b/Source/Core/VideoBackends/Software/DebugUtil.cpp
index 397442f8e8..c580461172 100644
--- a/Source/Core/VideoBackends/Software/DebugUtil.cpp
+++ b/Source/Core/VideoBackends/Software/DebugUtil.cpp
@@ -52,13 +52,8 @@ void Shutdown()
 
 static void SaveTexture(const std::string& filename, u32 texmap, s32 mip)
 {
-  FourTexUnits& texUnit = bpmem.tex[(texmap >> 2) & 1];
-  u8 subTexmap = texmap & 3;
-
-  TexImage0& ti0 = texUnit.texImage0[subTexmap];
-
-  u32 width = ti0.width + 1;
-  u32 height = ti0.height + 1;
+  u32 width = bpmem.tex.GetUnit(texmap).texImage0.width + 1;
+  u32 height = bpmem.tex.GetUnit(texmap).texImage0.height + 1;
 
   auto data = std::make_unique<u8[]>(width * height * 4);
 
@@ -80,10 +75,7 @@ void GetTextureRGBA(u8* dst, u32 texmap, s32 mip, u32 width, u32 height)
 
 static s32 GetMaxTextureLod(u32 texmap)
 {
-  FourTexUnits& texUnit = bpmem.tex[(texmap >> 2) & 1];
-  u8 subTexmap = texmap & 3;
-
-  u8 maxLod = texUnit.texMode1[subTexmap].max_lod;
+  u8 maxLod = bpmem.tex.GetUnit(texmap).texMode1.max_lod;
   u8 mip = maxLod >> 4;
   u8 fract = maxLod & 0xf;
 
diff --git a/Source/Core/VideoBackends/Software/Rasterizer.cpp b/Source/Core/VideoBackends/Software/Rasterizer.cpp
index de9925d596..df04ab5a31 100644
--- a/Source/Core/VideoBackends/Software/Rasterizer.cpp
+++ b/Source/Core/VideoBackends/Software/Rasterizer.cpp
@@ -163,13 +163,12 @@ static void InitSlope(Slope* slope, float f1, float f2, float f3, float DX31, fl
 
 static inline void CalculateLOD(s32* lodp, bool* linear, u32 texmap, u32 texcoord)
 {
-  const FourTexUnits& texUnit = bpmem.tex[(texmap >> 2) & 1];
-  const u8 subTexmap = texmap & 3;
+  auto texUnit = bpmem.tex.GetUnit(texmap);
 
   // LOD calculation requires data from the texture mode for bias, etc.
   // it does not seem to use the actual texture size
-  const TexMode0& tm0 = texUnit.texMode0[subTexmap];
-  const TexMode1& tm1 = texUnit.texMode1[subTexmap];
+  const TexMode0& tm0 = texUnit.texMode0;
+  const TexMode1& tm1 = texUnit.texMode1;
 
   float sDelta, tDelta;
   if (tm0.diag_lod == LODType::Diagonal)
diff --git a/Source/Core/VideoBackends/Software/TextureSampler.cpp b/Source/Core/VideoBackends/Software/TextureSampler.cpp
index f54002bdb5..064e9df41c 100644
--- a/Source/Core/VideoBackends/Software/TextureSampler.cpp
+++ b/Source/Core/VideoBackends/Software/TextureSampler.cpp
@@ -74,8 +74,8 @@ void Sample(s32 s, s32 t, s32 lod, bool linear, u8 texmap, u8* sample)
   bool mipLinear = false;
 
 #if (ALLOW_MIPMAP)
-  const FourTexUnits& texUnit = bpmem.tex[(texmap >> 2) & 1];
-  const TexMode0& tm0 = texUnit.texMode0[texmap & 3];
+  auto texUnit = bpmem.tex.GetUnit(texmap);
+  const TexMode0& tm0 = texUnit.texMode0;
 
   const s32 lodFract = lod & 0xf;
 
@@ -115,26 +115,25 @@ void Sample(s32 s, s32 t, s32 lod, bool linear, u8 texmap, u8* sample)
 
 void SampleMip(s32 s, s32 t, s32 mip, bool linear, u8 texmap, u8* sample)
 {
-  const FourTexUnits& texUnit = bpmem.tex[(texmap >> 2) & 1];
-  const u8 subTexmap = texmap & 3;
+  auto texUnit = bpmem.tex.GetUnit(texmap);
 
-  const TexMode0& tm0 = texUnit.texMode0[subTexmap];
-  const TexImage0& ti0 = texUnit.texImage0[subTexmap];
-  const TexTLUT& texTlut = texUnit.texTlut[subTexmap];
+  const TexMode0& tm0 = texUnit.texMode0;
+  const TexImage0& ti0 = texUnit.texImage0;
+  const TexTLUT& texTlut = texUnit.texTlut;
   const TextureFormat texfmt = ti0.format;
   const TLUTFormat tlutfmt = texTlut.tlut_format;
 
   const u8* imageSrc;
   const u8* imageSrcOdd = nullptr;
-  if (texUnit.texImage1[subTexmap].cache_manually_managed)
+  if (texUnit.texImage1.cache_manually_managed)
   {
-    imageSrc = &texMem[texUnit.texImage1[subTexmap].tmem_even * TMEM_LINE_SIZE];
+    imageSrc = &texMem[texUnit.texImage1.tmem_even * TMEM_LINE_SIZE];
     if (texfmt == TextureFormat::RGBA8)
-      imageSrcOdd = &texMem[texUnit.texImage2[subTexmap].tmem_odd * TMEM_LINE_SIZE];
+      imageSrcOdd = &texMem[texUnit.texImage2.tmem_odd * TMEM_LINE_SIZE];
   }
   else
   {
-    const u32 imageBase = texUnit.texImage3[subTexmap].image_base << 5;
+    const u32 imageBase = texUnit.texImage3.image_base << 5;
     imageSrc = Memory::GetPointer(imageBase);
   }
 
@@ -198,7 +197,7 @@ void SampleMip(s32 s, s32 t, s32 mip, bool linear, u8 texmap, u8* sample)
     WrapCoord(&imageSPlus1, tm0.wrap_s, image_width_minus_1 + 1);
     WrapCoord(&imageTPlus1, tm0.wrap_t, image_height_minus_1 + 1);
 
-    if (!(texfmt == TextureFormat::RGBA8 && texUnit.texImage1[subTexmap].cache_manually_managed))
+    if (!(texfmt == TextureFormat::RGBA8 && texUnit.texImage1.cache_manually_managed))
     {
       TexDecoder_DecodeTexel(sampledTex, imageSrc, imageS, imageT, image_width_minus_1, texfmt,
                              tlut, tlutfmt);
@@ -250,7 +249,7 @@ void SampleMip(s32 s, s32 t, s32 mip, bool linear, u8 texmap, u8* sample)
     WrapCoord(&imageS, tm0.wrap_s, image_width_minus_1 + 1);
     WrapCoord(&imageT, tm0.wrap_t, image_height_minus_1 + 1);
 
-    if (!(texfmt == TextureFormat::RGBA8 && texUnit.texImage1[subTexmap].cache_manually_managed))
+    if (!(texfmt == TextureFormat::RGBA8 && texUnit.texImage1.cache_manually_managed))
       TexDecoder_DecodeTexel(sample, imageSrc, imageS, imageT, image_width_minus_1, texfmt, tlut,
                              tlutfmt);
     else
diff --git a/Source/Core/VideoCommon/BPMemory.h b/Source/Core/VideoCommon/BPMemory.h
index 6c5b77220b..21636224c0 100644
--- a/Source/Core/VideoCommon/BPMemory.h
+++ b/Source/Core/VideoCommon/BPMemory.h
@@ -951,18 +951,6 @@ struct fmt::formatter<ZTex2>
   }
 };
 
-struct FourTexUnits
-{
-  TexMode0 texMode0[4];
-  TexMode1 texMode1[4];
-  TexImage0 texImage0[4];
-  TexImage1 texImage1[4];
-  TexImage2 texImage2[4];
-  TexImage3 texImage3[4];
-  TexTLUT texTlut[4];
-  u32 unknown[4];
-};
-
 // Geometry/other structs
 enum class CullMode : u32
 {
@@ -1979,6 +1967,8 @@ struct BPS_TmemConfig
   u32 texinvalidate;
 };
 
+union AllTexUnits;
+
 // The addressing of the texture units is a bit non-obvious.
 // This struct abstracts the complexity away.
 union TexUnitAddress
@@ -2019,9 +2009,118 @@ union TexUnitAddress
   }
 
   u32 GetUnitID() const { return UnitIdLow | (UnitIdHigh << 2); }
+
+private:
+  friend AllTexUnits;
+
+  size_t GetOffset() const { return FullAddress; }
+  size_t GetBPAddress() const { return FullAddress | 0x80; }
+
+  static constexpr size_t ComputeOffset(u32 unit_id)
+  {
+    // FIXME: Would be nice to construct a TexUnitAddress and get its offset,
+    // but that doesn't seem to be possible in c++17
+
+    // So we manually re-implement the calculation
+    return (unit_id & 3) | ((unit_id & 4) << 3);
+  }
 };
 static_assert(sizeof(TexUnitAddress) == sizeof(u32));
 
+// A view of the registers of a single TexUnit
+struct TexUnit
+{
+  TexMode0 texMode0;
+  u32 : 32;  // doing u32 : 96 is legal according to the standard, but msvc
+  u32 : 32;  // doesn't like it. So we stack multiple lines of u32 : 32;
+  u32 : 32;
+  TexMode1 texMode1;
+  u32 : 32;
+  u32 : 32;
+  u32 : 32;
+  TexImage0 texImage0;
+  u32 : 32;
+  u32 : 32;
+  u32 : 32;
+  TexImage1 texImage1;
+  u32 : 32;
+  u32 : 32;
+  u32 : 32;
+  TexImage2 texImage2;
+  u32 : 32;
+  u32 : 32;
+  u32 : 32;
+  TexImage3 texImage3;
+  u32 : 32;
+  u32 : 32;
+  u32 : 32;
+  TexTLUT texTlut;
+  u32 : 32;
+  u32 : 32;
+  u32 : 32;
+  u32 unknown;
+};
+static_assert(sizeof(TexUnit) == sizeof(u32) * 4 * 7 + sizeof(u32));
+
+union AllTexUnits
+{
+  std::array<u32, 8 * 8> AllRegisters;
+
+  const TexUnit& GetUnit(u32 UnitId) const
+  {
+    auto address = TexUnitAddress(UnitId);
+    const u32* ptr = &AllRegisters[address.GetOffset()];
+    return *reinterpret_cast<const TexUnit*>(ptr);
+  }
+
+private:
+  // For debuggers since GetUnit can be optimised out in release builds
+  template <u32 UnitId>
+  struct TexUnitPadding
+  {
+    static_assert(UnitId != 0, "Can't use 0 as sizeof(std::array<u32, 0>) != 0");
+    std::array<u32, TexUnitAddress::ComputeOffset(UnitId)> pad;
+  };
+
+  TexUnit tex0;
+  struct
+  {
+    TexUnitPadding<1> pad1;
+    TexUnit tex1;
+  };
+  struct
+  {
+    TexUnitPadding<2> pad2;
+    TexUnit tex2;
+  };
+  struct
+  {
+    TexUnitPadding<3> pad3;
+    TexUnit tex3;
+  };
+  struct
+  {
+    TexUnitPadding<4> pad4;
+    TexUnit tex4;
+  };
+  struct
+  {
+    TexUnitPadding<5> pad5;
+    TexUnit tex5;
+  };
+  struct
+  {
+    TexUnitPadding<6> pad6;
+    TexUnit tex6;
+  };
+  struct
+  {
+    TexUnitPadding<7> pad7;
+    TexUnit tex7;
+  };
+};
+static_assert(sizeof(AllTexUnits) == 8 * 8 * sizeof(u32));
+
 // All of BP memory
 
 struct BPCmd
@@ -2086,7 +2185,7 @@ struct BPMemory
   FieldMode fieldmode;                // 68
   u32 unknown10[7];                   // 69-6F
   u32 unknown11[16];                  // 70-7F
-  FourTexUnits tex[2];                // 80-bf
+  AllTexUnits tex;                    // 80-bf
   TevStageCombiner combiners[16];     // 0xC0-0xDF
   TevReg tevregs[4];                  // 0xE0
   FogRangeParams fogRange;            // 0xE8
diff --git a/Source/Core/VideoCommon/RenderState.cpp b/Source/Core/VideoCommon/RenderState.cpp
index 291aa9ce91..358c99974e 100644
--- a/Source/Core/VideoCommon/RenderState.cpp
+++ b/Source/Core/VideoCommon/RenderState.cpp
@@ -224,9 +224,9 @@ BlendingState& BlendingState::operator=(const BlendingState& rhs)
 
 void SamplerState::Generate(const BPMemory& bp, u32 index)
 {
-  const FourTexUnits& tex = bpmem.tex[index / 4];
-  const TexMode0& tm0 = tex.texMode0[index % 4];
-  const TexMode1& tm1 = tex.texMode1[index % 4];
+  auto tex = bp.tex.GetUnit(index);
+  const TexMode0& tm0 = tex.texMode0;
+  const TexMode1& tm1 = tex.texMode1;
 
   // GX can configure the mip filter to none. However, D3D and Vulkan can't express this in their
   // sampler states. Therefore, we set the min/max LOD to zero if this option is used.
diff --git a/Source/Core/VideoCommon/TextureCacheBase.cpp b/Source/Core/VideoCommon/TextureCacheBase.cpp
index 766c017915..e113f15832 100644
--- a/Source/Core/VideoCommon/TextureCacheBase.cpp
+++ b/Source/Core/VideoCommon/TextureCacheBase.cpp
@@ -970,8 +970,7 @@ void TextureCacheBase::DumpTexture(TCacheEntry* entry, std::string basename, uns
 static void SetSamplerState(u32 index, float custom_tex_scale, bool custom_tex,
                             bool has_arbitrary_mips)
 {
-  const FourTexUnits& tex = bpmem.tex[index / 4];
-  const TexMode0& tm0 = tex.texMode0[index % 4];
+  const TexMode0& tm0 = bpmem.tex.GetUnit(index).texMode0;
 
   SamplerState state = {};
   state.Generate(bpmem, index);
diff --git a/Source/Core/VideoCommon/TextureInfo.cpp b/Source/Core/VideoCommon/TextureInfo.cpp
index 631e91c6ee..6959b95534 100644
--- a/Source/Core/VideoCommon/TextureInfo.cpp
+++ b/Source/Core/VideoCommon/TextureInfo.cpp
@@ -14,30 +14,29 @@
 
 TextureInfo TextureInfo::FromStage(u32 stage)
 {
-  const FourTexUnits& tex = bpmem.tex[stage >> 2];
-  const u32 id = stage & 3;
+  const auto tex = bpmem.tex.GetUnit(stage);
 
-  const auto texture_format = tex.texImage0[id].format;
-  const auto tlut_format = tex.texTlut[id].tlut_format;
+  const auto texture_format = tex.texImage0.format;
+  const auto tlut_format = tex.texTlut.tlut_format;
 
-  const auto width = tex.texImage0[id].width + 1;
-  const auto height = tex.texImage0[id].height + 1;
+  const auto width = tex.texImage0.width + 1;
+  const auto height = tex.texImage0.height + 1;
 
-  const u32 address = (tex.texImage3[id].image_base /* & 0x1FFFFF*/) << 5;
+  const u32 address = (tex.texImage3.image_base /* & 0x1FFFFF*/) << 5;
 
-  const u32 tlutaddr = tex.texTlut[id].tmem_offset << 9;
+  const u32 tlutaddr = tex.texTlut.tmem_offset << 9;
   const u8* tlut_ptr = &texMem[tlutaddr];
 
   std::optional<u32> mip_count;
-  const bool has_mipmaps = SamplerCommon::AreBpTexMode0MipmapsEnabled(tex.texMode0[id]);
+  const bool has_mipmaps = SamplerCommon::AreBpTexMode0MipmapsEnabled(tex.texMode0);
   if (has_mipmaps)
   {
-    mip_count = (tex.texMode1[id].max_lod + 0xf) / 0x10;
+    mip_count = (tex.texMode1.max_lod + 0xf) / 0x10;
   }
 
-  const bool from_tmem = tex.texImage1[id].cache_manually_managed != 0;
-  const u32 tmem_address_even = from_tmem ? tex.texImage1[id].tmem_even * TMEM_LINE_SIZE : 0;
-  const u32 tmem_address_odd = from_tmem ? tex.texImage2[id].tmem_odd * TMEM_LINE_SIZE : 0;
+  const bool from_tmem = tex.texImage1.cache_manually_managed != 0;
+  const u32 tmem_address_even = from_tmem ? tex.texImage1.tmem_even * TMEM_LINE_SIZE : 0;
+  const u32 tmem_address_odd = from_tmem ? tex.texImage2.tmem_odd * TMEM_LINE_SIZE : 0;
 
   if (from_tmem)
   {