Bus: Reduce RAM write delay

src/core/bus.cpp

@@ -225,7 +225,8 @@ std::tuple<TickCount, TickCount, TickCount> Bus::CalculateMemoryTiming(MEMDELAY
   const TickCount byte_access_time = first;
   const TickCount halfword_access_time = mem_delay.data_bus_16bit ? first : (first + seq);
   const TickCount word_access_time = mem_delay.data_bus_16bit ? (first + seq) : (first + seq + seq + seq);
-  return std::tie(byte_access_time, halfword_access_time, word_access_time);
+  return std::tie(std::max(byte_access_time - 1, 0), std::max(halfword_access_time - 1, 0),
+                  std::max(word_access_time - 1, 0));
 }
 
 void Bus::RecalculateMemoryTimings()
@@ -238,14 +239,14 @@ void Bus::RecalculateMemoryTimings()
     CalculateMemoryTiming(m_MEMCTRL.spu_delay_size, m_MEMCTRL.common_delay);
 
   Log_TracePrintf("BIOS Memory Timing: %u bit bus, byte=%d, halfword=%d, word=%d",
-                  m_MEMCTRL.bios_delay_size.data_bus_16bit ? 16 : 8, m_bios_access_time[0], m_bios_access_time[1],
+                  m_MEMCTRL.bios_delay_size.data_bus_16bit ? 16 : 8, m_bios_access_time[0] + 1,
-                  m_bios_access_time[2]);
+                  m_bios_access_time[1] + 1, m_bios_access_time[2] + 1);
   Log_TracePrintf("CDROM Memory Timing: %u bit bus, byte=%d, halfword=%d, word=%d",
-                  m_MEMCTRL.cdrom_delay_size.data_bus_16bit ? 16 : 8, m_cdrom_access_time[0], m_cdrom_access_time[1],
+                  m_MEMCTRL.cdrom_delay_size.data_bus_16bit ? 16 : 8, m_cdrom_access_time[0] + 1,
-                  m_cdrom_access_time[2]);
+                  m_cdrom_access_time[1] + 1, m_cdrom_access_time[2] + 1);
   Log_TracePrintf("SPU Memory Timing: %u bit bus, byte=%d, halfword=%d, word=%d",
-                  m_MEMCTRL.spu_delay_size.data_bus_16bit ? 16 : 8, m_spu_access_time[0], m_spu_access_time[1],
+                  m_MEMCTRL.spu_delay_size.data_bus_16bit ? 16 : 8, m_spu_access_time[0] + 1, m_spu_access_time[1] + 1,
-                  m_spu_access_time[2]);
+                  m_spu_access_time[2] + 1);
 }
 
 TickCount Bus::DoInvalidAccess(MemoryAccessType type, MemoryAccessSize size, PhysicalMemoryAddress address, u32& value)

src/core/bus.h

@@ -134,10 +134,16 @@ private:
 
   enum : u32
   {
-    RAM_ACCESS_DELAY = 6, // Nocash docs say RAM takes 6 cycles to access.
     MEMCTRL_REG_COUNT = 9
   };
 
+  enum : TickCount
+  {
+    RAM_READ_ACCESS_DELAY = 5,  // Nocash docs say RAM takes 6 cycles to access. Subtract one because we already add a
+                                // tick for the instruction.
+    RAM_WRITE_ACCESS_DELAY = 0, // Writes are free unless we're executing more than 4 stores in a row.
+  };
+
   union MEMDELAY
   {
     u32 bits;

src/core/bus.inl

@@ -45,7 +45,7 @@ TickCount Bus::DoRAMAccess(u32 offset, u32& value)
   }
 
   // Nocash docs say RAM takes 6 cycles to access.
-  return RAM_ACCESS_DELAY;
+  return (type == MemoryAccessType::Read) ? RAM_READ_ACCESS_DELAY : RAM_WRITE_ACCESS_DELAY;
 }
 
 template<MemoryAccessType type, MemoryAccessSize size>
@@ -92,12 +92,16 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
   else if (address < (EXP1_BASE + EXP1_SIZE))
   {
     if constexpr (type == MemoryAccessType::Read)
+    {
       value = DoReadEXP1(size, address & EXP1_MASK);
-    else
-      DoWriteEXP1(size, address & EXP1_MASK, value);
-
       return m_exp1_access_time[static_cast<u32>(size)];
     }
+    else
+    {
+      DoWriteEXP1(size, address & EXP1_MASK, value);
+      return 0;
+    }
+  }
   else if (address < MEMCTRL_BASE)
   {
     return DoInvalidAccess(type, size, address, value);
@@ -109,7 +113,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteMemoryControl(size, address & PAD_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (PAD_BASE + PAD_SIZE))
   {
@@ -118,7 +122,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWritePad(size, address & PAD_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (SIO_BASE + SIO_SIZE))
   {
@@ -127,7 +131,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteSIO(size, address & SIO_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (MEMCTRL2_BASE + MEMCTRL2_SIZE))
   {
@@ -136,7 +140,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteMemoryControl2(size, address & PAD_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (INTERRUPT_CONTROLLER_BASE + INTERRUPT_CONTROLLER_SIZE))
   {
@@ -145,7 +149,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteInterruptController(size, address & INTERRUPT_CONTROLLER_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (DMA_BASE + DMA_SIZE))
   {
@@ -154,7 +158,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteDMA(size, address & DMA_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (TIMERS_BASE + TIMERS_SIZE))
   {
@@ -163,7 +167,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteTimers(size, address & TIMERS_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < CDROM_BASE)
   {
@@ -172,12 +176,16 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
   else if (address < (CDROM_BASE + GPU_SIZE))
   {
     if constexpr (type == MemoryAccessType::Read)
+    {
       value = DoReadCDROM(size, address & CDROM_MASK);
-    else
-      DoWriteCDROM(size, address & CDROM_MASK, value);
-
       return m_cdrom_access_time[static_cast<u32>(size)];
     }
+    else
+    {
+      DoWriteCDROM(size, address & CDROM_MASK, value);
+      return 0;
+    }
+  }
   else if (address < (GPU_BASE + GPU_SIZE))
   {
     if constexpr (type == MemoryAccessType::Read)
@@ -185,7 +193,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteGPU(size, address & GPU_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < (MDEC_BASE + MDEC_SIZE))
   {
@@ -194,7 +202,7 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
     else
       DoWriteMDEC(size, address & MDEC_MASK, value);
 
-    return 1;
+    return 0;
   }
   else if (address < SPU_BASE)
   {
@@ -203,12 +211,16 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
   else if (address < (SPU_BASE + SPU_SIZE))
   {
     if constexpr (type == MemoryAccessType::Read)
+    {
       value = DoReadSPU(size, address & SPU_MASK);
-    else
-      DoWriteSPU(size, address & SPU_MASK, value);
-
       return m_spu_access_time[static_cast<u32>(size)];
     }
+    else
+    {
+      DoWriteSPU(size, address & SPU_MASK, value);
+      return 0;
+    }
+  }
   else if (address < EXP2_BASE)
   {
     return DoInvalidAccess(type, size, address, value);
@@ -216,12 +228,16 @@ TickCount Bus::DispatchAccess(PhysicalMemoryAddress address, u32& value)
   else if (address < (EXP2_BASE + EXP2_SIZE))
   {
     if constexpr (type == MemoryAccessType::Read)
+    {
       value = DoReadEXP2(size, address & EXP2_MASK);
-    else
-      DoWriteEXP2(size, address & EXP2_MASK, value);
-
       return m_exp2_access_time[static_cast<u32>(size)];
     }
+    else
+    {
+      DoWriteEXP2(size, address & EXP2_MASK, value);
+      return 0;
+    }
+  }
   else if (address < BIOS_BASE)
   {
     return DoInvalidAccess(type, size, address, value);

src/core/cpu_core.cpp

@@ -123,7 +123,7 @@ bool Core::ReadMemoryByte(VirtualMemoryAddress addr, u8* value)
     return false;
   }
 
-  AddTicks(cycles - 1);
+  AddTicks(cycles);
   return true;
 }
 
@@ -141,7 +141,7 @@ bool Core::ReadMemoryHalfWord(VirtualMemoryAddress addr, u16* value)
     return false;
   }
 
-  AddTicks(cycles - 1);
+  AddTicks(cycles);
   return true;
 }
 
@@ -157,7 +157,7 @@ bool Core::ReadMemoryWord(VirtualMemoryAddress addr, u32* value)
     return false;
   }
 
-  AddTicks(cycles - 1);
+  AddTicks(cycles);
   return true;
 }
 
@@ -171,7 +171,7 @@ bool Core::WriteMemoryByte(VirtualMemoryAddress addr, u8 value)
     return false;
   }
 
-  AddTicks(cycles - 1);
+  DebugAssert(cycles == 0);
   return true;
 }
 
@@ -188,8 +188,8 @@ bool Core::WriteMemoryHalfWord(VirtualMemoryAddress addr, u16 value)
     return false;
   }
 
-  AddTicks(cycles - 1);
+  DebugAssert(cycles == 0);
-  return cycles;
+  return true;
 }
 
 bool Core::WriteMemoryWord(VirtualMemoryAddress addr, u32 value)
@@ -204,7 +204,7 @@ bool Core::WriteMemoryWord(VirtualMemoryAddress addr, u32 value)
     return false;
   }
 
-  AddTicks(cycles - 1);
+  DebugAssert(cycles == 0);
   return true;
 }
 

src/core/cpu_core.inl

@@ -15,14 +15,14 @@ TickCount Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
       if constexpr (type == MemoryAccessType::Write)
       {
         if (m_cop0_regs.sr.Isc)
-          return 1;
+          return 0;
       }
 
       const PhysicalMemoryAddress phys_addr = address & UINT32_C(0x1FFFFFFF);
       if ((phys_addr & DCACHE_LOCATION_MASK) == DCACHE_LOCATION)
       {
         DoScratchpadAccess<type, size>(phys_addr, value);
-        return 1;
+        return 0;
       }
 
       return m_bus->DispatchAccess<type, size>(phys_addr, value);
@@ -41,14 +41,14 @@ TickCount Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
       if constexpr (type == MemoryAccessType::Write)
       {
         if (m_cop0_regs.sr.Isc)
-          return 1;
+          return 0;
       }
 
       const PhysicalMemoryAddress phys_addr = address & UINT32_C(0x1FFFFFFF);
       if ((phys_addr & DCACHE_LOCATION_MASK) == DCACHE_LOCATION)
       {
         DoScratchpadAccess<type, size>(phys_addr, value);
-        return 1;
+        return 0;
       }
 
       return m_bus->DispatchAccess<type, size>(phys_addr, value);
@@ -72,7 +72,7 @@ TickCount Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
         else
           WriteCacheControl(value);
 
-        return 1;
+        return 0;
       }
       else
       {

src/core/cpu_recompiler_thunks.cpp

@@ -14,7 +14,7 @@ u64 Thunks::ReadMemoryByte(Core* cpu, u32 address)
     return UINT64_C(0xFFFFFFFFFFFFFFFF);
   }
 
-  cpu->AddTicks(cycles - 1);
+  cpu->AddTicks(cycles);
   return ZeroExtend64(temp);
 }
 
@@ -31,7 +31,7 @@ u64 Thunks::ReadMemoryHalfWord(Core* cpu, u32 address)
     return UINT64_C(0xFFFFFFFFFFFFFFFF);
   }
 
-  cpu->AddTicks(cycles - 1);
+  cpu->AddTicks(cycles);
   return ZeroExtend64(temp);
 }
 
@@ -48,7 +48,7 @@ u64 Thunks::ReadMemoryWord(Core* cpu, u32 address)
     return UINT64_C(0xFFFFFFFFFFFFFFFF);
   }
 
-  cpu->AddTicks(cycles - 1);
+  cpu->AddTicks(cycles);
   return ZeroExtend64(temp);
 }