Optimizations (are they?)

src/backend/Debugger.cpp

@@ -178,7 +178,7 @@ size_t debugGetMemory(void* user_data, char* buffer, size_t length, u32 address,
   printf("Checking memory at address 0x%08X\n", address);
   int printed = 0;
   u32 paddr;
-  if(!n64::MapVAddr(debugger->core.CpuGetRegs(), LOAD, address, paddr)) {
+  if(!n64::MapVAddr(debugger->core.CpuGetRegs(), n64::LOAD, address, paddr)) {
     return 0;
   }
 

src/backend/MemoryRegions.hpp

@@ -13,6 +13,8 @@
 #define PIF_BOOTROM_DSIZE (PIF_BOOTROM_SIZE - 1)
 #define ISVIEWER_SIZE (0x13FFFFFF - 0x13FF0020)
 #define ISVIEWER_DSIZE (ISVIEWER_SIZE - 1)
+#define CART_SIZE 0xFC00000
+#define CART_DSIZE (CART_REGION_SIZE - 1)
 
 #define RDRAM_REGION_START 0
 #define RDRAM_REGION_END RDRAM_DSIZE
@@ -38,7 +40,7 @@
 #define RI_REGION 0x04700000 ... 0x047FFFFF
 #define SI_REGION 0x04800000 ... 0x048FFFFF
 #define SRAM_REGION 0x08000000 ... 0x0FFFFFFF
-#define CART_REGION 0x10000000 ... 0x1FBFFFFF
+#define CART_REGION (CART_REGION_START) ... (CART_REGION_END)
 #define PIF_ROM_REGION 0x1FC00000 ... 0x1FC007BF
 #define PIF_RAM_REGION PIF_RAM_REGION_START ... PIF_RAM_REGION_END
 

src/backend/core/Interpreter.cpp

@@ -22,8 +22,6 @@ inline void CheckCompareInterrupt(MI& mi, Registers& regs) {
 }
 
 void Interpreter::Step(Mem& mem) {
-  regs.gpr[0] = 0;
-
   CheckCompareInterrupt(mem.mmio.mi, regs);
 
   regs.prevDelaySlot = regs.delaySlot;

src/backend/core/Mem.cpp

@@ -12,10 +12,8 @@ Mem::Mem() {
 }
 
 void Mem::Reset() {
-  readPages.resize(PAGE_COUNT);
+  memset(readPages, 0, PAGE_COUNT);
-  writePages.resize(PAGE_COUNT);
+  memset(writePages, 0, PAGE_COUNT);
-  std::fill(readPages.begin(), readPages.end(), 0);
-  std::fill(writePages.begin(), writePages.end(), 0);
 
   for(int i = 0; i < RDRAM_SIZE / PAGE_SIZE; i++) {
     const auto addr = (i * PAGE_SIZE) & RDRAM_DSIZE;
@@ -24,9 +22,16 @@ void Mem::Reset() {
     writePages[i] = pointer;
   }
 
-  sram.resize(SRAM_SIZE);
+  if(sram) {
-  std::fill(sram.begin(), sram.end(), 0);
+    free(sram);
-  romMask = 0;
+  }
+
+  if(cart) {
+    free(cart);
+  }
+
+  cart = (u8*)calloc(CART_SIZE, 1);
+  sram = (u8*)calloc(SRAM_SIZE, 1);
   mmio.Reset();
 }
 
@@ -39,22 +44,20 @@ CartInfo Mem::LoadROM(const std::string& filename) {
   }
 
   file.seekg(0, std::ios::end);
-  auto size = file.tellg();
+  size_t size = file.tellg();
-  auto sizeAdjusted = Util::NextPow2(size);
+  size_t sizeAdjusted = Util::NextPow2(size);
   romMask = sizeAdjusted - 1;
   file.seekg(0, std::ios::beg);
 
-  std::fill(cart.begin(), cart.end(), 0);
+  file.read(reinterpret_cast<char*>(cart), size);
-  cart.resize(sizeAdjusted);
-  cart.insert(cart.begin(), std::istream_iterator<u8>(file), std::istream_iterator<u8>());
 
   file.close();
 
   CartInfo result{};
 
   u32 cicChecksum;
-  Util::SwapN64Rom(sizeAdjusted, cart.data(), result.crc, cicChecksum);
+  Util::SwapN64Rom(sizeAdjusted, cart, result.crc, cicChecksum);
-  memcpy(mmio.rsp.dmem, cart.data(), 0x1000);
+  memcpy(mmio.rsp.dmem, cart, 0x1000);
 
   SetCICType(result.cicType, cicChecksum);
   result.isPAL = IsROMPAL();
@@ -108,13 +111,13 @@ u8 Mem::Read8(n64::Registers &regs, u32 paddr) {
         int offs = 3 - (paddr & 3);
         return (w >> (offs * 8)) & 0xff;
       }
-      case 0x10000000 ... 0x1FBFFFFF:
+      case CART_REGION:
         paddr = (paddr + 2) & ~2;
         return cart[BYTE_ADDRESS(paddr) & romMask];
       case 0x1FC00000 ... 0x1FC007BF:
         return pifBootrom[BYTE_ADDRESS(paddr) - 0x1FC00000];
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
-        return pifRam[paddr - 0x1FC007C0];
+        return pifRam[paddr - PIF_RAM_REGION_START];
       case 0x00800000 ... 0x03FFFFFF:
       case 0x04200000 ... 0x042FFFFF:
       case 0x04900000 ... 0x0FFFFFFF:
@@ -136,7 +139,7 @@ u16 Mem::Read16(n64::Registers &regs, u32 paddr) {
   } else {
     switch (paddr) {
       case 0x00000000 ... 0x007FFFFF:
-        return Util::ReadAccess<u16>(mmio.rdp.rdram.data(), HALF_ADDRESS(paddr));
+        return Util::ReadAccess<u16>(mmio.rdp.rdram, HALF_ADDRESS(paddr));
       case 0x04000000 ... 0x0403FFFF:
         if (paddr  & 0x1000)
           return Util::ReadAccess<u16>(mmio.rsp.imem, HALF_ADDRESS(paddr) & IMEM_DSIZE);
@@ -149,11 +152,11 @@ u16 Mem::Read16(n64::Registers &regs, u32 paddr) {
         return mmio.Read(paddr);
       case 0x10000000 ... 0x1FBFFFFF:
         paddr = (paddr + 2) & ~3;
-        return Util::ReadAccess<u16>(cart.data(), HALF_ADDRESS(paddr) & romMask);
+        return Util::ReadAccess<u16>(cart, HALF_ADDRESS(paddr) & romMask);
       case 0x1FC00000 ... 0x1FC007BF:
         return Util::ReadAccess<u16>(pifBootrom, HALF_ADDRESS(paddr) - 0x1FC00000);
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
-        return be16toh(Util::ReadAccess<u16>(pifRam, paddr - 0x1FC007C0));
+        return be16toh(Util::ReadAccess<u16>(pifRam, paddr - PIF_RAM_REGION_START));
       case 0x00800000 ... 0x03FFFFFF:
       case 0x04200000 ... 0x042FFFFF:
       case 0x04900000 ... 0x0FFFFFFF:
@@ -175,7 +178,7 @@ u32 Mem::Read32(n64::Registers &regs, u32 paddr) {
   } else {
     switch(paddr) {
       case 0x00000000 ... 0x007FFFFF:
-        return Util::ReadAccess<u32>(mmio.rdp.rdram.data(), paddr);
+        return Util::ReadAccess<u32>(mmio.rdp.rdram, paddr);
       case 0x04000000 ... 0x0403FFFF:
         if(paddr  & 0x1000)
           return Util::ReadAccess<u32>(mmio.rsp.imem, paddr & IMEM_DSIZE);
@@ -185,11 +188,11 @@ u32 Mem::Read32(n64::Registers &regs, u32 paddr) {
       case 0x04300000 ...	0x044FFFFF: case 0x04500000 ... 0x048FFFFF:
         return mmio.Read(paddr);
       case 0x10000000 ... 0x1FBFFFFF:
-        return Util::ReadAccess<u32>(cart.data(), paddr & romMask);
+        return Util::ReadAccess<u32>(cart, paddr & romMask);
       case 0x1FC00000 ... 0x1FC007BF:
         return Util::ReadAccess<u32>(pifBootrom, paddr - 0x1FC00000);
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
-        return be32toh(Util::ReadAccess<u32>(pifRam, paddr - 0x1FC007C0));
+        return be32toh(Util::ReadAccess<u32>(pifRam, paddr - PIF_RAM_REGION_START));
       case 0x00800000 ... 0x03FFFFFF: case 0x04200000 ... 0x042FFFFF:
       case 0x04900000 ... 0x0FFFFFFF: case 0x1FC00800 ... 0xFFFFFFFF: return 0;
       default:
@@ -208,7 +211,7 @@ u64 Mem::Read64(n64::Registers &regs, u32 paddr) {
   } else {
     switch (paddr) {
       case 0x00000000 ... 0x007FFFFF:
-        return Util::ReadAccess<u64>(mmio.rdp.rdram.data(), paddr);
+        return Util::ReadAccess<u64>(mmio.rdp.rdram, paddr);
       case 0x04000000 ... 0x0403FFFF:
         if (paddr  & 0x1000)
           return Util::ReadAccess<u64>(mmio.rsp.imem, paddr & IMEM_DSIZE);
@@ -220,11 +223,11 @@ u64 Mem::Read64(n64::Registers &regs, u32 paddr) {
       case 0x04500000 ... 0x048FFFFF:
         return mmio.Read(paddr);
       case 0x10000000 ... 0x1FBFFFFF:
-        return Util::ReadAccess<u64>(cart.data(), paddr & romMask);
+        return Util::ReadAccess<u64>(cart, paddr & romMask);
       case 0x1FC00000 ... 0x1FC007BF:
         return Util::ReadAccess<u64>(pifBootrom, paddr - 0x1FC00000);
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
-        return be64toh(Util::ReadAccess<u64>(pifRam, paddr - 0x1FC007C0));
+        return be64toh(Util::ReadAccess<u64>(pifRam, paddr - PIF_RAM_REGION_START));
       case 0x00800000 ... 0x03FFFFFF:
       case 0x04200000 ... 0x042FFFFF:
       case 0x04900000 ... 0x0FFFFFFF:
@@ -283,9 +286,9 @@ void Mem::Write8(Registers& regs, u32 paddr, u32 val) {
         Util::panic("MMIO Write8!\n");
       case 0x10000000 ... 0x1FBFFFFF:
         break;
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
         val = val << (8 * (3 - (paddr & 3)));
-        paddr = (paddr - 0x1FC007C0) & ~3;
+        paddr = (paddr - PIF_RAM_REGION_START) & ~3;
         Util::WriteAccess<u32>(pifRam, paddr, htobe32(val));
         ProcessPIFCommands(pifRam, mmio.si.controller, *this);
         break;
@@ -317,7 +320,7 @@ void Mem::Write16(Registers& regs, u32 paddr, u32 val) {
   } else {
     switch (paddr) {
       case 0x00000000 ... 0x007FFFFF:
-        Util::WriteAccess<u16>(mmio.rdp.rdram.data(), HALF_ADDRESS(paddr), val);
+        Util::WriteAccess<u16>(mmio.rdp.rdram, HALF_ADDRESS(paddr), val);
         break;
       case 0x04000000 ... 0x0403FFFF:
         val = val << (16 * !(paddr & 2));
@@ -334,10 +337,10 @@ void Mem::Write16(Registers& regs, u32 paddr, u32 val) {
         Util::panic("MMIO Write16!\n");
       case 0x10000000 ... 0x1FBFFFFF:
         break;
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
         val = val << (16 * !(paddr & 2));
         paddr &= ~3;
-        Util::WriteAccess<u32>(pifRam, paddr - 0x1FC007C0, htobe32(val));
+        Util::WriteAccess<u32>(pifRam, paddr - PIF_RAM_REGION_START, htobe32(val));
         ProcessPIFCommands(pifRam, mmio.si.controller, *this);
         break;
       case 0x00800000 ... 0x03FFFFFF:
@@ -364,7 +367,7 @@ void Mem::Write32(Registers& regs, u32 paddr, u32 val) {
   } else {
     switch(paddr) {
       case 0x00000000 ... 0x007FFFFF:
-        Util::WriteAccess<u32>(mmio.rdp.rdram.data(), paddr, val);
+        Util::WriteAccess<u32>(mmio.rdp.rdram, paddr, val);
         break;
       case 0x04000000 ... 0x0403FFFF:
         if(paddr & 0x1000)
@@ -387,8 +390,8 @@ void Mem::Write32(Registers& regs, u32 paddr, u32 val) {
         Util::WriteAccess<u32>(isviewer, paddr - 0x13FF0020, htobe32(val));
         break;
       case 0x14000000 ... 0x1FBFFFFF: break;
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
-        Util::WriteAccess<u32>(pifRam, paddr - 0x1FC007C0, htobe32(val));
+        Util::WriteAccess<u32>(pifRam, paddr - PIF_RAM_REGION_START, htobe32(val));
         ProcessPIFCommands(pifRam, mmio.si.controller, *this);
         break;
       case 0x00800000 ... 0x03FFFFFF: case 0x04200000 ... 0x042FFFFF:
@@ -409,7 +412,7 @@ void Mem::Write64(Registers& regs, u32 paddr, u64 val) {
   } else {
     switch (paddr) {
       case 0x00000000 ... 0x007FFFFF:
-        Util::WriteAccess<u64>(mmio.rdp.rdram.data(), paddr, val);
+        Util::WriteAccess<u64>(mmio.rdp.rdram, paddr, val);
         break;
       case 0x04000000 ... 0x0403FFFF:
         val >>= 32;
@@ -425,8 +428,8 @@ void Mem::Write64(Registers& regs, u32 paddr, u64 val) {
         Util::panic("MMIO Write64!\n");
       case 0x10000000 ... 0x1FBFFFFF:
         break;
-      case 0x1FC007C0 ... 0x1FC007FF:
+      case PIF_RAM_REGION:
-        Util::WriteAccess<u64>(pifRam, paddr - 0x1FC007C0, htobe64(val));
+        Util::WriteAccess<u64>(pifRam, paddr - PIF_RAM_REGION_START, htobe64(val));
         ProcessPIFCommands(pifRam, mmio.si.controller, *this);
         break;
       case 0x00800000 ... 0x03FFFFFF:

src/backend/core/Mem.hpp

@@ -5,10 +5,10 @@
 #include <vector>
 #include <backend/RomHelpers.hpp>
 #include <log.hpp>
+#include <Registers.hpp>
+#include <algorithm>
 
 namespace n64 {
-struct Registers;
-
 struct CartInfo {
   bool isPAL;
   u32 cicType;
@@ -20,12 +20,14 @@ struct Dynarec;
 }
 
 struct Mem {
-  ~Mem() = default;
+  ~Mem() {
+    free(sram);
+  }
   Mem();
   void Reset();
   CartInfo LoadROM(const std::string&);
-  [[nodiscard]] auto GetRDRAM() -> u8* {
+  [[nodiscard]] auto GetRDRAM() const -> u8* {
-    return mmio.rdp.rdram.data();
+    return mmio.rdp.rdram;
   }
 
   u8 Read8(Registers&, u32);
@@ -47,7 +49,7 @@ struct Mem {
   inline void DumpRDRAM() const {
     FILE *fp = fopen("rdram.dump", "wb");
     u8 *temp = (u8*)calloc(RDRAM_SIZE, 1);
-    memcpy(temp, mmio.rdp.rdram.data(), RDRAM_SIZE);
+    memcpy(temp, mmio.rdp.rdram, RDRAM_SIZE);
     Util::SwapBuffer32(RDRAM_SIZE, temp);
     fwrite(temp, 1, RDRAM_SIZE, fp);
     free(temp);
@@ -73,20 +75,19 @@ struct Mem {
     free(temp);
     fclose(fp);
   }
-  std::vector<uintptr_t> writePages, readPages;
+  uintptr_t writePages[PAGE_COUNT], readPages[PAGE_COUNT];
 private:
   friend struct SI;
   friend struct PI;
   friend struct AI;
-  friend struct Cpu;
   friend struct RSP;
   friend struct Core;
-  std::vector<u8> cart, sram;
+  u8* sram, *cart;
   u8 pifBootrom[PIF_BOOTROM_SIZE]{};
   u8 isviewer[ISVIEWER_SIZE]{};
-  size_t romMask;
+  size_t romMask = 0;
 
-  void SetCICType(u32& cicType, u32 checksum) {
+  static void SetCICType(u32& cicType, u32 checksum) {
     switch(checksum) {
       case 0xEC8B1325: // 7102
         cicType = CIC_NUS_7102;
@@ -115,12 +116,15 @@ private:
 
   bool IsROMPAL() {
     static const char pal_codes[] = {'D', 'F', 'I', 'P', 'S', 'U', 'X', 'Y'};
-    for (int i = 0; i < 8; i++) {
+    return std::any_of(std::begin(pal_codes), std::end(pal_codes), [this](char a) {
-      if (cart[0x3e] == pal_codes[i]) {
+      return cart[0x3e] == a;
-        return true;
+    });
-      }
+    // for (char pal_code : pal_codes) {
-    }
+    //   if (cart[0x3e] == pal_code) {
-    return false;
+    //     return true;
+    //   }
+    // }
+    // return false;
   }
 };
 

src/backend/core/RDP.cpp

@@ -11,8 +11,10 @@ RDP::RDP() {
 
 void RDP::Reset() {
   dpc.status.raw = 0x80;
-  rdram.resize(RDRAM_SIZE);
+  if(rdram) {
-  std::fill(rdram.begin(), rdram.end(), 0);
+    free(rdram);
+  }
+  rdram = (u8*)calloc(RDRAM_SIZE, 1);
   memset(cmd_buf, 0, 0x100000);
 }
 
@@ -144,7 +146,7 @@ void RDP::RunCommand(MI& mi, Registers& regs, RSP& rsp) {
         return;
       }
       for (int i = 0; i < len; i += 4) {
-        u32 cmd = Util::ReadAccess<u32>(rdram.data(), current + i);
+        u32 cmd = Util::ReadAccess<u32>(rdram, current + i);
         cmd_buf[remaining_cmds + (i >> 2)] = cmd;
       }
     }

src/backend/core/RDP.hpp

@@ -57,7 +57,7 @@ struct RDP {
   RDP();
   void Reset();
 
-  std::vector<u8> rdram;
+  u8* rdram;
   [[nodiscard]] auto Read(u32 addr) const -> u32;
   void Write(MI& mi, Registers& regs, RSP& rsp, u32 addr, u32 val);
   void WriteStatus(MI& mi, Registers& regs, RSP& rsp, u32 val);

src/backend/core/dynarec/instructions.cpp

@@ -13,16 +13,20 @@ void add(Registers& regs, u32 instr) {
   if(check_signed_overflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = s32(result);
     }
   }
+}
 
 void addu(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s32 rs = (s32)regs.gpr[RS(instr)];
     s32 rt = (s32)regs.gpr[RT(instr)];
     s32 result = rs + rt;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void addi(Registers& regs, u32 instr) {
   u32 rs = regs.gpr[RS(instr)];
@@ -49,15 +53,19 @@ void dadd(Registers& regs, u32 instr) {
   if(check_signed_overflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = result;
     }
   }
+}
 
 void daddu(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rs = regs.gpr[RS(instr)];
     s64 rt = regs.gpr[RT(instr)];
     regs.gpr[RD(instr)] = rs + rt;
   }
+}
 
 void daddi(Registers& regs, u32 instr) {
   u64 imm = s64(s16(instr));
@@ -545,12 +553,16 @@ void ori(Registers& regs, u32 instr) {
 }
 
 void or_(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] | regs.gpr[RT(instr)];
   }
+}
 
 void nor(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = ~(regs.gpr[RS(instr)] | regs.gpr[RT(instr)]);
   }
+}
 
 void j(Registers& regs, u32 instr) {
   u64 target = (instr & 0x3ffffff) << 2;
@@ -566,8 +578,10 @@ void jal(Registers& regs, u32 instr) {
 
 void jalr(Registers& regs, u32 instr) {
   branch(regs, true, regs.gpr[RS(instr)]);
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.pc + 4;
   }
+}
 
 void slti(Registers& regs, u32 instr) {
   regs.gpr[RT(instr)] = regs.gpr[RS(instr)] < se_imm(instr);
@@ -578,12 +592,16 @@ void sltiu(Registers& regs, u32 instr) {
 }
 
 void slt(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] < regs.gpr[RT(instr)];
   }
+}
 
 void sltu(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = (u64) regs.gpr[RS(instr)] < (u64) regs.gpr[RT(instr)];
   }
+}
 
 void xori(Registers& regs, u32 instr) {
   s64 imm = (u16)instr;
@@ -591,8 +609,10 @@ void xori(Registers& regs, u32 instr) {
 }
 
 void xor_(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RT(instr)] ^ regs.gpr[RS(instr)];
   }
+}
 
 void andi(Registers& regs, u32 instr) {
   s64 imm = (u16)instr;
@@ -600,111 +620,143 @@ void andi(Registers& regs, u32 instr) {
 }
 
 void and_(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] & regs.gpr[RT(instr)];
   }
+}
 
 void sll(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = ((instr >> 6) & 0x1f);
     s32 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = (s64) result;
   }
+}
 
 void sllv(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = (regs.gpr[RS(instr)]) & 0x1F;
     u32 rt = regs.gpr[RT(instr)];
     s32 result = rt << sa;
     regs.gpr[RD(instr)] = (s64) result;
   }
+}
 
 void dsll32(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = regs.gpr[RT(instr)] << (sa + 32);
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void dsll(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void dsllv(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 sa = regs.gpr[RS(instr)] & 63;
     s64 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void srl(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u32 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u32 result = rt >> sa;
     regs.gpr[RD(instr)] = (s32) result;
   }
+}
 
 void srlv(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = (regs.gpr[RS(instr)] & 0x1F);
     u32 rt = regs.gpr[RT(instr)];
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = (s64) result;
   }
+}
 
 void dsrl(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u64 result = rt >> sa;
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void dsrlv(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 amount = (regs.gpr[RS(instr)] & 63);
     u64 rt = regs.gpr[RT(instr)];
     u64 result = rt >> amount;
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void dsrl32(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u64 result = rt >> (sa + 32);
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void sra(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void srav(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     s64 rs = regs.gpr[RS(instr)];
     u8 sa = rs & 0x1f;
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void dsra(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void dsrav(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     s64 rs = regs.gpr[RS(instr)];
     s64 sa = rs & 63;
     s64 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void dsra32(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = rt >> (sa + 32);
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void jr(Registers& regs, u32 instr) {
   s64 address = regs.gpr[RS(instr)];
@@ -723,16 +775,20 @@ void dsub(Registers& regs, u32 instr) {
   if(check_signed_underflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = result;
     }
   }
+}
 
 void dsubu(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u64 rt = regs.gpr[RT(instr)];
     u64 rs = regs.gpr[RS(instr)];
     u64 result = rs - rt;
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void sub(Registers& regs, u32 instr) {
   s32 rt = regs.gpr[RT(instr)];
@@ -741,16 +797,20 @@ void sub(Registers& regs, u32 instr) {
   if(check_signed_underflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = result;
     }
   }
+}
 
 void subu(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u32 rt = regs.gpr[RT(instr)];
     u32 rs = regs.gpr[RS(instr)];
     u32 result = rs - rt;
     regs.gpr[RD(instr)] = (s64) ((s32) result);
   }
+}
 
 void dmultu(Registers& regs, u32 instr) {
   u64 rt = regs.gpr[RT(instr)];
@@ -785,12 +845,16 @@ void mult(Registers& regs, u32 instr) {
 }
 
 void mflo(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.lo;
   }
+}
 
 void mfhi(Registers& regs, u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.hi;
   }
+}
 
 void mtlo(Registers& regs, u32 instr) {
   regs.lo = regs.gpr[RS(instr)];

src/backend/core/interpreter/instructions.cpp

@@ -12,16 +12,20 @@ void Interpreter::add(u32 instr) {
   if(check_signed_overflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = s32(result);
     }
   }
+}
 
 void Interpreter::addu(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s32 rs = (s32)regs.gpr[RS(instr)];
     s32 rt = (s32)regs.gpr[RT(instr)];
     s32 result = rs + rt;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::addi(u32 instr) {
   u32 rs = regs.gpr[RS(instr)];
@@ -48,15 +52,19 @@ void Interpreter::dadd(u32 instr) {
   if(check_signed_overflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = result;
     }
   }
+}
 
 void Interpreter::daddu(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rs = regs.gpr[RS(instr)];
     s64 rt = regs.gpr[RT(instr)];
     regs.gpr[RD(instr)] = rs + rt;
   }
+}
 
 void Interpreter::daddi(u32 instr) {
   u64 imm = s64(s16(instr));
@@ -583,12 +591,16 @@ void Interpreter::ori(u32 instr) {
 }
 
 void Interpreter::or_(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] | regs.gpr[RT(instr)];
   }
+}
 
 void Interpreter::nor(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = ~(regs.gpr[RS(instr)] | regs.gpr[RT(instr)]);
   }
+}
 
 void Interpreter::j(u32 instr) {
   s32 target = (instr & 0x3ffffff) << 2;
@@ -604,8 +616,10 @@ void Interpreter::jal(u32 instr) {
 
 void Interpreter::jalr(u32 instr) {
   branch(true, regs.gpr[RS(instr)]);
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.pc + 4;
   }
+}
 
 void Interpreter::slti(u32 instr) {
   s16 imm = instr;
@@ -618,12 +632,16 @@ void Interpreter::sltiu(u32 instr) {
 }
 
 void Interpreter::slt(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] < regs.gpr[RT(instr)];
   }
+}
 
 void Interpreter::sltu(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = (u64) regs.gpr[RS(instr)] < (u64) regs.gpr[RT(instr)];
   }
+}
 
 void Interpreter::xori(u32 instr) {
   s64 imm = (u16)instr;
@@ -631,8 +649,10 @@ void Interpreter::xori(u32 instr) {
 }
 
 void Interpreter::xor_(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RT(instr)] ^ regs.gpr[RS(instr)];
   }
+}
 
 void Interpreter::andi(u32 instr) {
   s64 imm = (u16)instr;
@@ -640,111 +660,143 @@ void Interpreter::andi(u32 instr) {
 }
 
 void Interpreter::and_(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] & regs.gpr[RT(instr)];
   }
+}
 
 void Interpreter::sll(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = ((instr >> 6) & 0x1f);
     s32 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = (s64) result;
   }
+}
 
 void Interpreter::sllv(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = (regs.gpr[RS(instr)]) & 0x1F;
     u32 rt = regs.gpr[RT(instr)];
     s32 result = rt << sa;
     regs.gpr[RD(instr)] = (s64) result;
   }
+}
 
 void Interpreter::dsll32(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = regs.gpr[RT(instr)] << (sa + 32);
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::dsll(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::dsllv(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 sa = regs.gpr[RS(instr)] & 63;
     s64 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::srl(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u32 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u32 result = rt >> sa;
     regs.gpr[RD(instr)] = (s32) result;
   }
+}
 
 void Interpreter::srlv(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 sa = (regs.gpr[RS(instr)] & 0x1F);
     u32 rt = regs.gpr[RT(instr)];
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = (s64) result;
   }
+}
 
 void Interpreter::dsrl(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u64 result = rt >> sa;
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void Interpreter::dsrlv(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u8 amount = (regs.gpr[RS(instr)] & 63);
     u64 rt = regs.gpr[RT(instr)];
     u64 result = rt >> amount;
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void Interpreter::dsrl32(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u64 result = rt >> (sa + 32);
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void Interpreter::sra(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::srav(u32 instr) {
   s64 rt = regs.gpr[RT(instr)];
+  if(likely(RD(instr) != 0)) {
     s64 rs = regs.gpr[RS(instr)];
     u8 sa = rs & 0x1f;
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::dsra(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::dsrav(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     s64 rs = regs.gpr[RS(instr)];
     s64 sa = rs & 63;
     s64 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::dsra32(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = rt >> (sa + 32);
     regs.gpr[RD(instr)] = result;
   }
+}
 
 void Interpreter::jr(u32 instr) {
   s64 address = regs.gpr[RS(instr)];
@@ -758,16 +810,20 @@ void Interpreter::dsub(u32 instr) {
   if(check_signed_underflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = result;
     }
   }
+}
 
 void Interpreter::dsubu(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u64 rt = regs.gpr[RT(instr)];
     u64 rs = regs.gpr[RS(instr)];
     u64 result = rs - rt;
     regs.gpr[RD(instr)] = s64(result);
   }
+}
 
 void Interpreter::sub(u32 instr) {
   s32 rt = regs.gpr[RT(instr)];
@@ -776,16 +832,20 @@ void Interpreter::sub(u32 instr) {
   if(check_signed_underflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if(likely(RD(instr) != 0)) {
       regs.gpr[RD(instr)] = result;
     }
   }
+}
 
 void Interpreter::subu(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     u32 rt = regs.gpr[RT(instr)];
     u32 rs = regs.gpr[RS(instr)];
     u32 result = rs - rt;
     regs.gpr[RD(instr)] = (s64) ((s32) result);
   }
+}
 
 void Interpreter::dmultu(u32 instr) {
   u64 rt = regs.gpr[RT(instr)];
@@ -820,12 +880,16 @@ void Interpreter::mult(u32 instr) {
 }
 
 void Interpreter::mflo(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.lo;
   }
+}
 
 void Interpreter::mfhi(u32 instr) {
+  if(likely(RD(instr) != 0)) {
     regs.gpr[RD(instr)] = regs.hi;
   }
+}
 
 void Interpreter::mtlo(u32 instr) {
   regs.lo = regs.gpr[RS(instr)];

src/backend/core/mmio/AI.cpp

@@ -82,7 +82,7 @@ void AI::Step(Mem& mem, Registers& regs, int cpuCycles, float volumeL, float vol
 
     if(dmaLen[0] && dmaEnable) {u32 addrHi = ((dmaAddr[0] >> 13) + dmaAddrCarry) & 0x7FF;
       dmaAddr[0] = (addrHi << 13) | (dmaAddr[0] & 0x1FFF);
-      u32 data = Util::ReadAccess<u32>(mem.mmio.rdp.rdram.data(), dmaAddr[0] & RDRAM_DSIZE);
+      u32 data = Util::ReadAccess<u32>(mem.mmio.rdp.rdram, dmaAddr[0] & RDRAM_DSIZE);
       s16 l = s16(data >> 16);
       s16 r = s16(data);
 

src/backend/core/registers/Cop0.cpp

@@ -69,7 +69,7 @@ u64 Cop0::GetReg64(u8 addr) {
     case 23: case 24: case 25:
     case 31: return openbus;
     default:
-      Util::panic("Unsupported word read from COP0 register {}\n", index);
+      Util::panic("Unsupported dword read from COP0 register {}\n", index);
   }
 }
 
@@ -131,7 +131,7 @@ void Cop0::SetReg32(u8 addr, u32 value) {
     case 23: case 24: case 25:
     case 31: break;
     default:
-      Util::panic("Unsupported word read from COP0 register {}\n", index);
+      Util::panic("Unsupported word write from COP0 register {}\n", index);
   }
 }
 
@@ -159,7 +159,7 @@ void Cop0::SetReg64(u8 addr, u64 value) {
     case COP0_REG_LLADDR: LLAddr = value; break;
     case COP0_REG_ERROREPC: ErrorEPC = (s64)value; break;
     default:
-      Util::panic("Unsupported word write to COP0 register {}\n", addr);
+      Util::panic("Unsupported dword write to COP0 register {}\n", addr);
   }
 }
 

src/backend/core/registers/Cop0.hpp

@@ -268,6 +268,10 @@ private:
 struct Registers;
 enum class ExceptionCode : u8;
 
+enum TLBAccessType {
+  LOAD, STORE
+};
+
 TLBEntry* TLBTryMatch(Registers& regs, u64 vaddr, int* match);
 bool ProbeTLB(Registers& regs, TLBAccessType access_type, u64 vaddr, u32& paddr, int* match);
 void HandleTLBException(Registers& regs, u64 vaddr);

src/backend/core/rsp/instructions.cpp

@@ -80,6 +80,7 @@ inline void SetCop0Reg(Registers& regs, Mem& mem, u8 index, u32 val) {
   }
 }
 
+ARCH_TARGET("sse3", "avx2", "default")
 inline VPR Broadcast(const VPR& vt, int l0, int l1, int l2, int l3, int l4, int l5, int l6, int l7) {
   VPR vte{};
   vte.element[ELEMENT_INDEX(0)] = vt.element[ELEMENT_INDEX(l0)];
@@ -93,6 +94,7 @@ inline VPR Broadcast(const VPR& vt, int l0, int l1, int l2, int l3, int l4, int
   return vte;
 }
 
+ARCH_TARGET("sse3", "avx2", "default")
 inline VPR GetVTE(const VPR& vt, u8 e) {
   VPR vte{};
   e &= 0xf;
@@ -692,6 +694,7 @@ inline u16 unsignedClamp(s64 val) {
   return val;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vabs(u32 instr) {
   VPR& vs = vpr[VS(instr)];
   VPR& vd = vpr[VD(instr)];
@@ -716,6 +719,7 @@ void RSP::vabs(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vadd(u32 instr) {
   VPR& vs = vpr[VS(instr)];
   VPR& vd = vpr[VD(instr)];
@@ -730,6 +734,7 @@ void RSP::vadd(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vaddc(u32 instr) {
   VPR& vs = vpr[VS(instr)];
   VPR& vd = vpr[VD(instr)];
@@ -744,6 +749,7 @@ void RSP::vaddc(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vch(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -776,6 +782,7 @@ void RSP::vch(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vcr(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -808,6 +815,7 @@ void RSP::vcr(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vcl(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -843,6 +851,7 @@ void RSP::vcl(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmov(u32 instr) {
   u8 e = E2(instr), vs = VS(instr) & 7;
   VPR& vd = vpr[VD(instr)];
@@ -885,6 +894,7 @@ inline bool IsSignExtension(s16 hi, s16 lo) {
   return false;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmulf(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -906,6 +916,7 @@ void RSP::vmulf(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmulq(u32 instr) {
   VPR& vs = vpr[VS(instr)];
   VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
@@ -924,6 +935,7 @@ void RSP::vmulq(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmulu(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -945,6 +957,7 @@ void RSP::vmulu(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmudl(u32 instr) {
   u8 e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -971,6 +984,7 @@ void RSP::vmudl(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmudh(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -989,6 +1003,7 @@ void RSP::vmudh(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmudm(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1005,6 +1020,7 @@ void RSP::vmudm(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmudn(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1030,6 +1046,7 @@ void RSP::vmudn(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmadh(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1052,6 +1069,7 @@ void RSP::vmadh(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmadl(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1080,6 +1098,7 @@ void RSP::vmadl(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmadm(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1099,6 +1118,7 @@ void RSP::vmadm(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmadn(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1124,6 +1144,7 @@ void RSP::vmadn(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmacf(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR& vs = vpr[VS(instr)];
@@ -1145,6 +1166,7 @@ void RSP::vmacf(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmacu(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR& vs = vpr[VS(instr)];
@@ -1165,6 +1187,7 @@ void RSP::vmacu(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmacq(u32 instr) {
   VPR& vd = vpr[VD(instr)];
 
@@ -1182,6 +1205,7 @@ void RSP::vmacq(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::veq(u32 instr) {
   int e = E2(instr);
   VPR& vd = vpr[VD(instr)];
@@ -1196,6 +1220,7 @@ void RSP::veq(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vne(u32 instr) {
   int e = E2(instr);
   VPR& vd = vpr[VD(instr)];
@@ -1210,6 +1235,7 @@ void RSP::vne(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vge(u32 instr) {
   int e = E2(instr);
   VPR& vd = vpr[VD(instr)];
@@ -1226,6 +1252,7 @@ void RSP::vge(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vlt(u32 instr) {
   int e = E2(instr);
   VPR& vd = vpr[VD(instr)];
@@ -1287,6 +1314,7 @@ inline u32 rsq(u32 input) {
   return result ^ mask;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrcpl(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR& vt = vpr[VT(instr)];
@@ -1313,6 +1341,7 @@ void RSP::vrcpl(u32 instr) {
   vd.element[ELEMENT_INDEX(de)] = result;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrcp(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR& vt = vpr[VT(instr)];
@@ -1330,6 +1359,7 @@ void RSP::vrcp(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrsq(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR& vt = vpr[VT(instr)];
@@ -1354,6 +1384,7 @@ static inline s64 sclip(s64 x, u32 bits) {
   return ((x & m) ^ b) - b;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrndn(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
@@ -1385,6 +1416,7 @@ void RSP::vrndn(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrndp(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
@@ -1416,6 +1448,7 @@ void RSP::vrndp(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrsql(u32 instr) {
   VPR& vd = vpr[VD(instr)];
   VPR& vt = vpr[VT(instr)];
@@ -1441,6 +1474,7 @@ void RSP::vrsql(u32 instr) {
   vd.element[ELEMENT_INDEX(de)] = result;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vrcph(u32 instr) {
   int e = E2(instr) & 7;
   int de = DE(instr) & 7;
@@ -1457,6 +1491,7 @@ void RSP::vrcph(u32 instr) {
   divInLoaded = true;
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vsar(u32 instr) {
   u8 e = E2(instr);
   switch(e) {
@@ -1483,6 +1518,7 @@ void RSP::vsar(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vsubc(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1499,6 +1535,7 @@ void RSP::vsubc(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vsub(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1515,6 +1552,7 @@ void RSP::vsub(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vmrg(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1529,6 +1567,7 @@ void RSP::vmrg(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vxor(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1541,6 +1580,7 @@ void RSP::vxor(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vnxor(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1553,6 +1593,7 @@ void RSP::vnxor(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vand(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1565,6 +1606,7 @@ void RSP::vand(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vnand(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1577,6 +1619,7 @@ void RSP::vnand(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vnor(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1589,6 +1632,7 @@ void RSP::vnor(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vor(u32 instr) {
   int e = E2(instr);
   VPR& vs = vpr[VS(instr)];
@@ -1601,6 +1645,7 @@ void RSP::vor(u32 instr) {
   }
 }
 
+ARCH_TARGET("sse4.2", "avx2", "default")
 void RSP::vzero(u32 instr) {
   VPR& vs = vpr[VS(instr)];
   VPR vte = GetVTE(vpr[VT(instr)], E2(instr));

src/common.hpp

@@ -34,7 +34,6 @@ using s128 = __int128_t;
 #define BYTE_INDEX(i)   (15 - (i))
 
 #define SI_DMA_DELAY (65536 * 2)
-
+#define ARCH_TARGET(...) __attribute__ ((target_clones (__VA_ARGS__)))
-enum TLBAccessType {
+#define unlikely(exp) __builtin_expect(exp, 0)
-  LOAD, STORE
+#define likely(exp) __builtin_expect(exp, 1)
-};

src/frontend/imgui/GameList.cpp

@@ -1,167 +0,0 @@
-#include <GameList.hpp>
-#include <filesystem>
-#include <imgui.h>
-#include <fstream>
-#include <nlohmann/json.hpp>
-#include <fmt/format.h>
-#include <RomHelpers.hpp>
-#include <File.hpp>
-#include <thread>
-#include <algorithm>
-
-using namespace nlohmann;
-namespace fs = std::filesystem;
-
-GameList::GameList(const std::string& path) {
-  Create(path);
-}
-
-void GameList::Create(const std::string &path) {
-  threadDone = false;
-  if(!path.empty()) {
-    std::thread searchThread([path, this]() {
-      std::ifstream gameDbFile("resources/db.json");
-      json gameDb = json::parse(gameDbFile);
-      std::vector<u8> rom{};
-      for(const auto& p : fs::recursive_directory_iterator{path}) {
-        const auto filename = p.path().string();
-        if(threadDone) {
-          gamesList.clear();
-          break;
-        }
-        if(p.path().extension() == ".n64" || p.path().extension() == ".z64" || p.path().extension() == ".v64" ||
-           p.path().extension() == ".N64" || p.path().extension() == ".Z64" || p.path().extension() == ".V64") {
-          std::ifstream file(filename, std::ios::binary);
-          file.unsetf(std::ios::skipws);
-
-          if(!file.is_open()) {
-            Util::panic("Unable to open {}!", filename);
-          }
-
-          file.seekg(0, std::ios::end);
-          auto size = file.tellg();
-          auto sizeAdjusted = Util::NextPow2(size);
-          file.seekg(0, std::ios::beg);
-
-          std::fill(rom.begin(), rom.end(), 0);
-          rom.resize(sizeAdjusted);
-          rom.insert(rom.begin(), std::istream_iterator<u8>(file), std::istream_iterator<u8>());
-          file.close();
-
-          u32 crc{};
-          Util::GetRomCRC(sizeAdjusted, rom.data(), crc);
-
-          bool found = false;
-
-          for(const auto& item : gameDb["items"]) {
-            const auto& crcEntry = item["crc"];
-            if(!crcEntry.empty()) {
-              if(crcEntry.get<std::string>() == fmt::format("{:08X}", crc)) {
-                found = true;
-                gamesList.push_back(GameInfo{
-                  item["name"].get<std::string>(),
-                  item["region"].get<std::string>(),
-                  fmt::format("{:.2f} MiB", float(size) / 1024 / 1024),
-                  "Good",
-                  p.path().string()
-                });
-              }
-            }
-          }
-
-          if(!found) {
-            gamesList.push_back(GameInfo{
-              p.path().stem().string(),
-              "Unknown",
-              fmt::format("{:.2f} MiB", float(size) / 1024 / 1024),
-              "Unknown",
-              p.path().string()
-            });
-          }
-        }
-      };
-
-      gameDbFile.close();
-      threadDone = true;
-    });
-
-    searchThread.detach();
-  }
-}
-
-bool GameList::RenderWidget(float mainMenuBarHeight, std::string& rom) {
-  const auto windowSize = ImGui::GetIO().DisplaySize;
-  ImGui::SetNextWindowPos(ImVec2(0, mainMenuBarHeight));
-  ImGui::SetNextWindowSize(ImVec2(windowSize.x, windowSize.y - mainMenuBarHeight));
-  ImGui::PushStyleVar(ImGuiStyleVar_WindowPadding, ImVec2(0.f, 0.f));
-  ImGui::PushStyleVar(ImGuiStyleVar_WindowBorderSize, 0.f);
-
-  ImGui::Begin(
-    "Games list",
-    nullptr,
-    ImGuiWindowFlags_NoTitleBar | ImGuiWindowFlags_NoCollapse | ImGuiWindowFlags_NoResize |
-    ImGuiWindowFlags_NoBringToFrontOnFocus
-  );
-
-  static ImGuiTableFlags flags =
-    ImGuiTableFlags_Resizable | ImGuiTableFlags_Reorderable | ImGuiTableFlags_Hideable | ImGuiTableFlags_Sortable | ImGuiTableFlags_SortMulti
-    | ImGuiTableFlags_RowBg | ImGuiTableFlags_BordersOuter | ImGuiTableFlags_BordersV | ImGuiTableFlags_NoBordersInBody
-    | ImGuiTableFlags_ScrollY;
-
-  bool toOpen = false;
-  if (ImGui::BeginTable("Games List", 4, flags)) {
-    ImGui::TableSetupColumn("Title");
-    ImGui::TableSetupColumn("Region");
-    ImGui::TableSetupColumn("Status");
-    ImGui::TableSetupColumn("Size");
-    ImGui::TableSetupScrollFreeze(0, 1); // Make row always visible
-    ImGui::TableHeadersRow();
-
-    int i = 0;
-
-    for (const auto& entry : gamesList) {
-      ImGui::TableNextRow(ImGuiTableRowFlags_None);
-      ImGui::PushID(i);
-      ImGui::PushStyleVar(ImGuiStyleVar_SelectableTextAlign, ImVec2(0.0f, 0.5f));
-      ImGui::TableSetColumnIndex(0);
-
-      if (ImGui::Selectable(entry.name.c_str(), false, ImGuiSelectableFlags_SpanAllColumns | ImGuiSelectableFlags_AllowItemOverlap, ImVec2(0.0f, 20.f))) {
-        toOpen = true;
-        rom = entry.path;
-      }
-
-      ImGui::TableSetColumnIndex(1);
-
-      if (ImGui::Selectable(entry.region.c_str(), false, ImGuiSelectableFlags_SpanAllColumns | ImGuiSelectableFlags_AllowItemOverlap, ImVec2(0.0f, 20.f))) {
-        toOpen = true;
-        rom = entry.path;
-      }
-
-      ImGui::TableSetColumnIndex(2);
-
-      if (ImGui::Selectable(entry.status.c_str(), false, ImGuiSelectableFlags_SpanAllColumns | ImGuiSelectableFlags_AllowItemOverlap, ImVec2(0.0f, 20.f))) {
-        toOpen = true;
-        rom = entry.path;
-      }
-
-      ImGui::TableSetColumnIndex(3);
-
-      if (ImGui::Selectable(entry.size.c_str(), false, ImGuiSelectableFlags_SpanAllColumns | ImGuiSelectableFlags_AllowItemOverlap, ImVec2(0.0f, 20.f))) {
-        toOpen = true;
-        rom = entry.path;
-      }
-
-      ImGui::PopStyleVar();
-      ImGui::PopID();
-      i++;
-    }
-
-    ImGui::EndTable();
-  }
-
-  ImGui::End();
-  ImGui::PopStyleVar();
-  ImGui::PopStyleVar();
-
-  return toOpen;
-}

src/frontend/imgui/GameList.hpp

@@ -1,21 +0,0 @@
-#pragma once
-#include <vector>
-#include <string>
-#include <atomic>
-
-struct GameInfo {
-  std::string name, region, size, status, path;
-};
-
-struct GameList {
-  GameList(const std::string&);
-  ~GameList() = default;
-
-  void Create(const std::string&);
-  bool RenderWidget(float, std::string&);
-
-  [[nodiscard]] std::vector<GameInfo> GetGamesList() const { return gamesList; }
-  std::atomic_bool threadDone = false;
-private:
-  std::vector<GameInfo> gamesList{}, notMatch{};
-};

src/frontend/imgui/Window.cpp

@@ -12,7 +12,7 @@
 VkInstance instance{};
 namespace fs = std::filesystem;
 
-Window::Window(n64::Core& core) : settings(core), gameList(settings.GetGamesDir()) {
+Window::Window(n64::Core& core) : settings(core) {
   InitSDL();
   InitParallelRDP(core.mem.GetRDRAM(), window);
   InitImgui();
@@ -156,7 +156,6 @@ ImDrawData* Window::Present(n64::Core& core) {
 
 void Window::LoadROM(n64::Core& core, const std::string &path) {
   if(!path.empty()) {
-    gameList.threadDone = true;
     n64::CartInfo cartInfo = core.LoadROM(path);
     std::ifstream gameDbFile("resources/db.json");
     json gameDb = json::parse(gameDbFile);
@@ -181,7 +180,6 @@ void Window::LoadROM(n64::Core& core, const std::string &path) {
     Util::UpdateRPC(Util::Playing, gameName);
     windowTitle = "Gadolinium - " + gameName;
     shadowWindowTitle = windowTitle;
-    renderGameList = false;
 
     SDL_SetWindowTitle(window, windowTitle.c_str());
     gameDbFile.close();
@@ -190,7 +188,6 @@ void Window::LoadROM(n64::Core& core, const std::string &path) {
 
 void Window::RenderMainMenuBar(n64::Core &core) {
   ImGui::BeginMainMenuBar();
-  mainMenuBarHeight = ImGui::GetWindowSize().y;
 
   if (ImGui::BeginMenu("File")) {
     if (ImGui::MenuItem("Open", "O")) {
@@ -222,12 +219,10 @@ void Window::RenderMainMenuBar(n64::Core &core) {
       LoadROM(core, core.rom);
     }
     if (ImGui::MenuItem("Stop")) {
-      renderGameList = true;
       windowTitle = "Gadolinium";
       core.rom.clear();
       Util::UpdateRPC(Util::Idling);
       SDL_SetWindowTitle(window, windowTitle.c_str());
-      gameList.Create(settings.GetGamesDir());
       core.Stop();
     }
     if (ImGui::MenuItem(core.pause ? "Resume" : "Pause", nullptr, false, core.romLoaded)) {
@@ -267,13 +262,6 @@ void Window::Render(n64::Core& core) {
     RenderMainMenuBar(core);
   }
 
-  static std::string rom{};
-  if(renderGameList && gameList.RenderWidget(mainMenuBarHeight, rom)) {
-    LoadROM(core, rom);
-    renderGameList = false;
-  }
-
-  mainMenuBarHeight = 0;
   settings.RenderWidget(showSettings);
 
   ImGui::PopFont();

src/frontend/imgui/Window.hpp

@@ -7,7 +7,6 @@
 #include <backend/Core.hpp>
 #include <vector>
 #include <frontend/imgui/Settings.hpp>
-#include <frontend/imgui/GameList.hpp>
 
 struct Window {
   explicit Window(n64::Core& core);
@@ -17,12 +16,9 @@ struct Window {
   [[nodiscard]] bool gotClosed(SDL_Event event);
   ImFont *uiFont{};
   Settings settings;
-  GameList gameList;
   void LoadROM(n64::Core& core, const std::string& path);
 private:
-  bool renderGameList = true;
   bool showSettings = false;
-  float mainMenuBarHeight = 0;
   SDL_Window* window{};
   std::string windowTitle{"Gadolinium"};
   std::string shadowWindowTitle{windowTitle};