fpu improvements

src/backend/core/JIT/instructions.cpp

@@ -6,7 +6,7 @@
 
 namespace n64 {
 void JIT::add(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     movsx(eax, dword[rdi + offsetof(Registers, gpr[RS(instr)])]); // rs
     movsx(ecx, dword[rdi + offsetof(Registers, gpr[RT(instr)])]); // rt
     CodeGenerator::add(eax, ecx);
@@ -19,7 +19,7 @@ void JIT::addu(u32 instr) {
 }
 
 void JIT::addi(u32 instr) {
-  if(likely(RT(instr) != 0)) {
+  if (RT(instr) != 0) [[likely]] {
     movsx(eax, dword[rdi + offsetof(Registers, gpr[RS(instr)])]);
     mov(ecx, s32(s16(instr)));
     CodeGenerator::add(eax, ecx);
@@ -32,7 +32,7 @@ void JIT::addiu(u32 instr) {
 }
 
 void JIT::dadd(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RS(instr))); // rs
     mov(rcx, GPR(RT(instr))); // rt
     CodeGenerator::add(rax, rcx);
@@ -45,7 +45,7 @@ void JIT::daddu(u32 instr) {
 }
 
 void JIT::daddi(u32 instr) {
-  if(likely(RT(instr) != 0)) {
+  if (RT(instr) != 0) [[likely]] {
     mov(rax, GPR(RS(instr)));
     mov(rcx, s64(s16(instr)));
     CodeGenerator::add(rax, rcx);
@@ -665,7 +665,7 @@ void JIT::ori(u32 instr) {
 }
 
 void JIT::or_(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RS(instr)));
     mov(rcx, GPR(RT(instr)));
     CodeGenerator::or_(rax, rcx);
@@ -674,7 +674,7 @@ void JIT::or_(u32 instr) {
 }
 
 void JIT::nor(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RS(instr)));
     mov(rcx, GPR(RT(instr)));
     CodeGenerator::or_(rax, rcx);
@@ -702,7 +702,7 @@ void JIT::jalr(u32 instr) {
   mov(rax, GPR(RS(instr)));
   mov(qword[rdi + offsetof(Registers, nextPC)], rax);
 
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, qword[rdi + offsetof(Registers, pc)]);
     CodeGenerator::add(rax, 4);
     mov(GPR(RD(instr)), rax);
@@ -724,7 +724,7 @@ void JIT::sltiu(u32 instr) {
 }
 
 void JIT::slt(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RS(instr)));
     mov(rcx, GPR(RT(instr)));
     cmp(rax, rcx);
@@ -733,7 +733,7 @@ void JIT::slt(u32 instr) {
 }
 
 void JIT::sltu(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RS(instr)));
     mov(rcx, GPR(RT(instr)));
     cmp(rax, rcx);
@@ -750,7 +750,7 @@ void JIT::xori(u32 instr) {
 }
 
 void JIT::xor_(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RT(instr)));
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::xor_(rax, rcx);
@@ -767,7 +767,7 @@ void JIT::andi(u32 instr) {
 }
 
 void JIT::and_(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RT(instr)));
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(rax, rcx);
@@ -776,7 +776,7 @@ void JIT::and_(u32 instr) {
 }
 
 void JIT::sll(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     mov(rax, GPR(RT(instr)));
     sal(rax, sa);
@@ -786,7 +786,7 @@ void JIT::sll(u32 instr) {
 }
 
 void JIT::sllv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(cl, 0x1F);
     mov(rax, GPR(RT(instr)));
@@ -797,7 +797,7 @@ void JIT::sllv(u32 instr) {
 }
 
 void JIT::dsll32(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f) + 32;
     mov(rax, GPR(RT(instr)));
     sal(rax, sa);
@@ -806,7 +806,7 @@ void JIT::dsll32(u32 instr) {
 }
 
 void JIT::dsll(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     mov(rax, GPR(RT(instr)));
     sal(rax, sa);
@@ -815,7 +815,7 @@ void JIT::dsll(u32 instr) {
 }
 
 void JIT::dsllv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(cl, 63);
     mov(rax, GPR(RT(instr)));
@@ -825,7 +825,7 @@ void JIT::dsllv(u32 instr) {
 }
 
 void JIT::srl(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     mov(rax, GPR(RT(instr)));
     CodeGenerator::shr(rax, sa);
@@ -835,7 +835,7 @@ void JIT::srl(u32 instr) {
 }
 
 void JIT::srlv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(cl, 0x1F);
     mov(rax, GPR(RT(instr)));
@@ -846,7 +846,7 @@ void JIT::srlv(u32 instr) {
 }
 
 void JIT::dsrl(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     mov(rax, GPR(RT(instr)));
     CodeGenerator::shr(rax, sa);
@@ -855,7 +855,7 @@ void JIT::dsrl(u32 instr) {
 }
 
 void JIT::dsrlv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(cl, 63);
     mov(rax, GPR(RT(instr)));
@@ -865,7 +865,7 @@ void JIT::dsrlv(u32 instr) {
 }
 
 void JIT::dsrl32(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f) + 32;
     mov(rax, GPR(RT(instr)));
     CodeGenerator::shr(rax, sa);
@@ -874,7 +874,7 @@ void JIT::dsrl32(u32 instr) {
 }
 
 void JIT::sra(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     mov(rax, GPR(RT(instr)));
     sar(rax, sa);
@@ -884,7 +884,7 @@ void JIT::sra(u32 instr) {
 }
 
 void JIT::srav(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(cl, 0x1F);
     mov(rax, GPR(RT(instr)));
@@ -895,7 +895,7 @@ void JIT::srav(u32 instr) {
 }
 
 void JIT::dsra(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     mov(rax, GPR(RT(instr)));
     sar(rax, sa);
@@ -904,7 +904,7 @@ void JIT::dsra(u32 instr) {
 }
 
 void JIT::dsrav(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::and_(cl, 63);
     mov(rax, GPR(RT(instr)));
@@ -914,7 +914,7 @@ void JIT::dsrav(u32 instr) {
 }
 
 void JIT::dsra32(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f) + 32;
     mov(rax, GPR(RT(instr)));
     sar(rax, sa);
@@ -929,7 +929,7 @@ void JIT::jr(u32 instr) {
 }
 
 void JIT::dsub(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(rax, GPR(RT(instr)));
     mov(rcx, GPR(RS(instr)));
     CodeGenerator::sub(rcx, rax);
@@ -942,7 +942,7 @@ void JIT::dsubu(u32 instr) {
 }
 
 void JIT::sub(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     mov(eax, GPR(RT(instr)));
     mov(ecx, GPR(RS(instr)));
     CodeGenerator::sub(ecx, eax);
@@ -988,13 +988,13 @@ void JIT::mult(u32 instr) {
 }
 
 void JIT::mflo(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.lo;
   }
 }
 
 void JIT::mfhi(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.hi;
   }
 }

src/backend/core/interpreter/instructions.cpp

@@ -12,14 +12,14 @@ void Interpreter::add(u32 instr) {
   if(check_signed_overflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
-    if(likely(RD(instr) != 0)) {
+    if (RD(instr) != 0) [[likely]] {
       regs.gpr[RD(instr)] = s32(result);
     }
   }
 }
 
 void Interpreter::addu(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s32 rs = (s32)regs.gpr[RS(instr)];
     s32 rt = (s32)regs.gpr[RT(instr)];
     s32 result = rs + rt;
@@ -34,15 +34,19 @@ void Interpreter::addi(u32 instr) {
   if(check_signed_overflow(rs, imm, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = s32(result);
     }
+  }
 }
 
 void Interpreter::addiu(u32 instr) {
   s32 rs = (s32)regs.gpr[RS(instr)];
   s16 imm = (s16)(instr);
   s32 result = rs + imm;
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = result;
+  }
 }
 
 void Interpreter::dadd(u32 instr) {
@@ -52,14 +56,14 @@ void Interpreter::dadd(u32 instr) {
   if(check_signed_overflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
-    if(likely(RD(instr) != 0)) {
+    if (RD(instr) != 0) [[likely]] {
       regs.gpr[RD(instr)] = result;
     }
   }
 }
 
 void Interpreter::daddu(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s64 rs = regs.gpr[RS(instr)];
     s64 rt = regs.gpr[RT(instr)];
     regs.gpr[RD(instr)] = rs + rt;
@@ -73,14 +77,18 @@ void Interpreter::daddi(u32 instr) {
   if(check_signed_overflow(rs, imm, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = result;
     }
+  }
 }
 
 void Interpreter::daddiu(u32 instr) {
   s16 imm = (s16)(instr);
   s64 rs = regs.gpr[RS(instr)];
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = rs + imm;
+  }
 }
 
 void Interpreter::div(u32 instr) {
@@ -200,7 +208,9 @@ void Interpreter::bllink(u32 instr, bool cond) {
 void Interpreter::lui(u32 instr) {
   u64 val = s64((s16)instr);
   val <<= 16;
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = val;
+  }
 }
 
 void Interpreter::lb(u32 instr) {
@@ -210,13 +220,15 @@ void Interpreter::lb(u32 instr) {
     HandleTLBException(regs, address);
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
-    regs.gpr[RT(instr)] = (s8)mem.Read8(regs, paddr);
+    if (RT(instr) != 0) [[likely]] {
+      regs.gpr[RT(instr)] = (s8) mem.Read8(regs, paddr);
+    }
   }
 }
 
 void Interpreter::lh(u32 instr) {
   u64 address = regs.gpr[RS(instr)] + (s16)instr;
-  if ((address & 0b1) > 0) {
+  if (check_address_error(0b1, address)) {
     HandleTLBException(regs, address);
     FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
     return;
@@ -227,7 +239,9 @@ void Interpreter::lh(u32 instr) {
     HandleTLBException(regs, address);
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
-    regs.gpr[RT(instr)] = (s16)mem.Read16(regs, paddr);
+    if (RT(instr) != 0) [[likely]] {
+      regs.gpr[RT(instr)] = (s16) mem.Read16(regs, paddr);
+    }
   }
 }
 
@@ -245,7 +259,9 @@ void Interpreter::lw(u32 instr) {
     HandleTLBException(regs, address);
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
-    regs.gpr[RT(instr)] = (s32)mem.Read32(regs, physical);
+    if(RT(instr) != 0) [[likely]] {
+      regs.gpr[RT(instr)] = (s32) mem.Read32(regs, physical);
+    }
   }
 }
 
@@ -257,12 +273,14 @@ void Interpreter::ll(u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     s32 result = mem.Read32(regs, physical);
-    if ((address & 0b11) > 0) {
+    if (check_address_error(0b11, address)) {
       FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
       return;
     }
 
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = result;
+    }
 
     regs.cop0.llbit = true;
     regs.cop0.LLAddr = physical >> 4;
@@ -280,8 +298,10 @@ void Interpreter::lwl(u32 instr) {
     u32 mask = 0xFFFFFFFF << shift;
     u32 data = mem.Read32(regs, paddr & ~3);
     s32 result = s32((regs.gpr[RT(instr)] & ~mask) | (data << shift));
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = result;
     }
+  }
 }
 
 void Interpreter::lwr(u32 instr) {
@@ -295,8 +315,10 @@ void Interpreter::lwr(u32 instr) {
     u32 mask = 0xFFFFFFFF >> shift;
     u32 data = mem.Read32(regs, paddr & ~3);
     s32 result = s32((regs.gpr[RT(instr)] & ~mask) | (data >> shift));
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = result;
     }
+  }
 }
 
 void Interpreter::ld(u32 instr) {
@@ -313,8 +335,10 @@ void Interpreter::ld(u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     s64 value = mem.Read64(regs, paddr);
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = value;
     }
+  }
 }
 
 void Interpreter::lld(u32 instr) {
@@ -329,10 +353,12 @@ void Interpreter::lld(u32 instr) {
     HandleTLBException(regs, address);
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
-    if ((address & 0b111) > 0) {
+    if (check_address_error(0b111, address)) {
       FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
     } else {
+      if (RT(instr) != 0) [[likely]] {
         regs.gpr[RT(instr)] = mem.Read64(regs, paddr);
+      }
       regs.cop0.llbit = true;
       regs.cop0.LLAddr = paddr >> 4;
     }
@@ -350,8 +376,10 @@ void Interpreter::ldl(u32 instr) {
     u64 mask = 0xFFFFFFFFFFFFFFFF << shift;
     u64 data = mem.Read64(regs, paddr & ~7);
     s64 result = (s64) ((regs.gpr[RT(instr)] & ~mask) | (data << shift));
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = result;
     }
+  }
 }
 
 void Interpreter::ldr(u32 instr) {
@@ -365,8 +393,10 @@ void Interpreter::ldr(u32 instr) {
     u64 mask = 0xFFFFFFFFFFFFFFFF >> shift;
     u64 data = mem.Read64(regs, paddr & ~7);
     s64 result = (s64) ((regs.gpr[RT(instr)] & ~mask) | (data >> shift));
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = result;
     }
+  }
 }
 
 void Interpreter::lbu(u32 instr) {
@@ -377,13 +407,15 @@ void Interpreter::lbu(u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     u8 value = mem.Read8(regs, paddr);
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = value;
     }
+  }
 }
 
 void Interpreter::lhu(u32 instr) {
   s64 address = regs.gpr[RS(instr)] + (s16)instr;
-  if ((address & 0b1) > 0) {
+  if (check_address_error(0b1, address)) {
     HandleTLBException(regs, address);
     FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
     return;
@@ -394,13 +426,15 @@ void Interpreter::lhu(u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     u16 value = mem.Read16(regs, paddr);
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = value;
     }
+  }
 }
 
 void Interpreter::lwu(u32 instr) {
   s64 address = regs.gpr[RS(instr)] + (s16)instr;
-  if ((address & 0b11) > 0) {
+  if (check_address_error(0b11, address)) {
     HandleTLBException(regs, address);
     FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
     return;
@@ -412,8 +446,10 @@ void Interpreter::lwu(u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     u32 value = mem.Read32(regs, paddr);
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = value;
     }
+  }
 }
 
 void Interpreter::sb(u32 instr) {
@@ -430,7 +466,7 @@ void Interpreter::sb(u32 instr) {
 void Interpreter::sc(u32 instr) {
   u64 address = regs.gpr[RS(instr)] + (s16)instr;
 
-  if ((address & 0b11) > 0) {
+  if (check_address_error(0b11, address)) {
     FireException(regs, ExceptionCode::AddressErrorStore, 0, true);
     return;
   }
@@ -443,11 +479,15 @@ void Interpreter::sc(u32 instr) {
       FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, STORE), 0, true);
     } else {
       mem.Write32(regs, paddr, regs.gpr[RT(instr)]);
+      if (RT(instr) != 0) [[likely]] {
         regs.gpr[RT(instr)] = 1;
       }
+    }
   } else {
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = 0;
     }
+  }
 }
 
 void Interpreter::scd(u32 instr) {
@@ -457,7 +497,7 @@ void Interpreter::scd(u32 instr) {
   }
 
   s64 address = regs.gpr[RS(instr)] + (s16)instr;
-  if ((address & 0b111) > 0) {
+  if (check_address_error(0b111, address)) {
     HandleTLBException(regs, address);
     FireException(regs, ExceptionCode::AddressErrorStore, 0, true);
     return;
@@ -471,11 +511,15 @@ void Interpreter::scd(u32 instr) {
       FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, STORE), 0, true);
     } else {
       mem.Write32(regs, paddr, regs.gpr[RT(instr)]);
+      if (RT(instr) != 0) [[likely]] {
         regs.gpr[RT(instr)] = 1;
       }
+    }
   } else {
+    if (RT(instr) != 0) [[likely]] {
       regs.gpr[RT(instr)] = 0;
     }
+  }
 }
 
 void Interpreter::sh(u32 instr) {
@@ -588,17 +632,19 @@ void Interpreter::swr(u32 instr) {
 void Interpreter::ori(u32 instr) {
   s64 imm = (u16)instr;
   s64 result = imm | regs.gpr[RS(instr)];
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = result;
+  }
 }
 
 void Interpreter::or_(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] | regs.gpr[RT(instr)];
   }
 }
 
 void Interpreter::nor(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = ~(regs.gpr[RS(instr)] | regs.gpr[RT(instr)]);
   }
 }
@@ -616,58 +662,71 @@ void Interpreter::jal(u32 instr) {
 }
 
 void Interpreter::jalr(u32 instr) {
-  branch(true, regs.gpr[RS(instr)]);
+  u64 addr = regs.gpr[RS(instr)];
-  if(likely(RD(instr) != 0)) {
+  if(check_address_error(0b11, addr)) {
+    FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
+  } else {
+    branch(true, addr);
+    if (RD(instr) != 0) [[likely]] {
       regs.gpr[RD(instr)] = regs.pc + 4;
     }
+  }
 }
 
 void Interpreter::slti(u32 instr) {
   s16 imm = instr;
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = regs.gpr[RS(instr)] < imm;
+  }
 }
 
 void Interpreter::sltiu(u32 instr) {
   s16 imm = instr;
-  regs.gpr[RT(instr)] = (u64)regs.gpr[RS(instr)] < imm;
+  if (RT(instr) != 0) [[likely]] {
+    regs.gpr[RT(instr)] = (u64) regs.gpr[RS(instr)] < imm;
+  }
 }
 
 void Interpreter::slt(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] < regs.gpr[RT(instr)];
   }
 }
 
 void Interpreter::sltu(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = (u64) regs.gpr[RS(instr)] < (u64) regs.gpr[RT(instr)];
   }
 }
 
 void Interpreter::xori(u32 instr) {
   s64 imm = (u16)instr;
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = regs.gpr[RS(instr)] ^ imm;
+  }
 }
 
 void Interpreter::xor_(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.gpr[RT(instr)] ^ regs.gpr[RS(instr)];
   }
 }
 
 void Interpreter::andi(u32 instr) {
   s64 imm = (u16)instr;
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = regs.gpr[RS(instr)] & imm;
+  }
 }
 
 void Interpreter::and_(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.gpr[RS(instr)] & regs.gpr[RT(instr)];
   }
 }
 
 void Interpreter::sll(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     s32 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = (s64) result;
@@ -675,7 +734,7 @@ void Interpreter::sll(u32 instr) {
 }
 
 void Interpreter::sllv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = (regs.gpr[RS(instr)]) & 0x1F;
     u32 rt = regs.gpr[RT(instr)];
     s32 result = rt << sa;
@@ -684,7 +743,7 @@ void Interpreter::sllv(u32 instr) {
 }
 
 void Interpreter::dsll32(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = regs.gpr[RT(instr)] << (sa + 32);
     regs.gpr[RD(instr)] = result;
@@ -692,7 +751,7 @@ void Interpreter::dsll32(u32 instr) {
 }
 
 void Interpreter::dsll(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = result;
@@ -700,7 +759,7 @@ void Interpreter::dsll(u32 instr) {
 }
 
 void Interpreter::dsllv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s64 sa = regs.gpr[RS(instr)] & 63;
     s64 result = regs.gpr[RT(instr)] << sa;
     regs.gpr[RD(instr)] = result;
@@ -708,7 +767,7 @@ void Interpreter::dsllv(u32 instr) {
 }
 
 void Interpreter::srl(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u32 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u32 result = rt >> sa;
@@ -717,7 +776,7 @@ void Interpreter::srl(u32 instr) {
 }
 
 void Interpreter::srlv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 sa = (regs.gpr[RS(instr)] & 0x1F);
     u32 rt = regs.gpr[RT(instr)];
     s32 result = rt >> sa;
@@ -726,7 +785,7 @@ void Interpreter::srlv(u32 instr) {
 }
 
 void Interpreter::dsrl(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u64 result = rt >> sa;
@@ -735,7 +794,7 @@ void Interpreter::dsrl(u32 instr) {
 }
 
 void Interpreter::dsrlv(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u8 amount = (regs.gpr[RS(instr)] & 63);
     u64 rt = regs.gpr[RT(instr)];
     u64 result = rt >> amount;
@@ -744,7 +803,7 @@ void Interpreter::dsrlv(u32 instr) {
 }
 
 void Interpreter::dsrl32(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     u64 result = rt >> (sa + 32);
@@ -753,7 +812,7 @@ void Interpreter::dsrl32(u32 instr) {
 }
 
 void Interpreter::sra(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s32 result = rt >> sa;
@@ -762,9 +821,9 @@ void Interpreter::sra(u32 instr) {
 }
 
 void Interpreter::srav(u32 instr) {
-  s64 rt = regs.gpr[RT(instr)];
+  if (RD(instr) != 0) [[likely]] {
-  if(likely(RD(instr) != 0)) {
     s64 rs = regs.gpr[RS(instr)];
+    s64 rt = regs.gpr[RT(instr)];
     u8 sa = rs & 0x1f;
     s32 result = rt >> sa;
     regs.gpr[RD(instr)] = result;
@@ -772,7 +831,7 @@ void Interpreter::srav(u32 instr) {
 }
 
 void Interpreter::dsra(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = rt >> sa;
@@ -781,7 +840,7 @@ void Interpreter::dsra(u32 instr) {
 }
 
 void Interpreter::dsrav(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s64 rt = regs.gpr[RT(instr)];
     s64 rs = regs.gpr[RS(instr)];
     s64 sa = rs & 63;
@@ -791,7 +850,7 @@ void Interpreter::dsrav(u32 instr) {
 }
 
 void Interpreter::dsra32(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     s64 rt = regs.gpr[RT(instr)];
     u8 sa = ((instr >> 6) & 0x1f);
     s64 result = rt >> (sa + 32);
@@ -801,7 +860,11 @@ void Interpreter::dsra32(u32 instr) {
 
 void Interpreter::jr(u32 instr) {
   s64 address = regs.gpr[RS(instr)];
+  if(check_address_error(0b11, address)) {
+    FireException(regs, ExceptionCode::AddressErrorLoad, 0, true);
+  } else {
     branch(true, address);
+  }
 }
 
 void Interpreter::dsub(u32 instr) {
@@ -811,14 +874,14 @@ void Interpreter::dsub(u32 instr) {
   if(check_signed_underflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
-    if(likely(RD(instr) != 0)) {
+    if (RD(instr) != 0) [[likely]] {
       regs.gpr[RD(instr)] = result;
     }
   }
 }
 
 void Interpreter::dsubu(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u64 rt = regs.gpr[RT(instr)];
     u64 rs = regs.gpr[RS(instr)];
     u64 result = rs - rt;
@@ -833,14 +896,14 @@ void Interpreter::sub(u32 instr) {
   if(check_signed_underflow(rs, rt, result)) {
     FireException(regs, ExceptionCode::Overflow, 0, true);
   } else {
-    if(likely(RD(instr) != 0)) {
+    if (RD(instr) != 0) [[likely]] {
       regs.gpr[RD(instr)] = result;
     }
   }
 }
 
 void Interpreter::subu(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     u32 rt = regs.gpr[RT(instr)];
     u32 rs = regs.gpr[RS(instr)];
     u32 result = rs - rt;
@@ -881,13 +944,13 @@ void Interpreter::mult(u32 instr) {
 }
 
 void Interpreter::mflo(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.lo;
   }
 }
 
 void Interpreter::mfhi(u32 instr) {
-  if(likely(RD(instr) != 0)) {
+  if (RD(instr) != 0) [[likely]] {
     regs.gpr[RD(instr)] = regs.hi;
   }
 }
@@ -912,7 +975,9 @@ void Interpreter::mtc2(u32 instr) {
 
 void Interpreter::mfc2(u32 instr) {
   s32 value = cop2Latch;
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = value;
+  }
 }
 
 void Interpreter::dmtc2(u32 instr) {
@@ -920,7 +985,9 @@ void Interpreter::dmtc2(u32 instr) {
 }
 
 void Interpreter::dmfc2(u32 instr) {
+  if (RT(instr) != 0) [[likely]] {
     regs.gpr[RT(instr)] = cop2Latch;
+  }
 }
 
 void Interpreter::ctc2(u32) {

src/backend/core/registers/Cop0.cpp

@@ -254,7 +254,7 @@ void FireException(Registers& regs, ExceptionCode code, int cop, bool useOldPC)
   bool old_exl = regs.cop0.status.exl;
   s64 pc = useOldPC ? regs.oldPC : regs.pc;
 
-  if(!regs.cop0.status.exl) {
+  if(!old_exl) {
     if(regs.prevDelaySlot) {
       regs.cop0.cause.branchDelay = true;
       pc -= 4;

src/backend/core/registers/Cop0.hpp

@@ -171,20 +171,20 @@ enum TLBError : u8 {
 };
 
 enum class ExceptionCode : u8 {
-  Interrupt,
+  Interrupt = 0,
-  TLBModification,
+  TLBModification = 1,
-  TLBLoad,
+  TLBLoad = 2,
-  TLBStore,
+  TLBStore = 3,
-  AddressErrorLoad,
+  AddressErrorLoad = 4,
-  AddressErrorStore,
+  AddressErrorStore = 5,
-  InstructionBusError,
+  InstructionBusError = 6,
-  DataBusError,
+  DataBusError = 7,
-  Syscall,
+  Syscall = 8,
-  Breakpoint,
+  Breakpoint = 9,
-  ReservedInstruction,
+  ReservedInstruction = 10,
-  CoprocessorUnusable,
+  CoprocessorUnusable = 11,
-  Overflow,
+  Overflow = 12,
-  Trap,
+  Trap = 13,
   FloatingPointError = 15,
   Watch = 23
 };

src/backend/core/registers/Cop1.cpp

@@ -11,8 +11,8 @@ Cop1::Cop1() {
 }
 
 void Cop1::Reset() {
-  fcr0 = 0;
+  fcr0 = 0xa00;
-  fcr31.raw = 0x01000800;
+  fcr31.write(0x01000800);
   memset(fgr, 0, 32 * sizeof(FGR));
 }
 
@@ -32,10 +32,6 @@ template void Cop1::decode<JIT>(JIT&, u32);
 
 void Cop1::decodeInterp(Interpreter &cpu, u32 instr) {
   Registers &regs = cpu.regs;
-  if(!regs.cop0.status.cu1) {
-    FireException(regs, ExceptionCode::CoprocessorUnusable, 1, true);
-    return;
-  }
 
   u8 mask_sub = (instr >> 21) & 0x1F;
   u8 mask_fun = instr & 0x3F;
@@ -45,11 +41,11 @@ void Cop1::decodeInterp(Interpreter &cpu, u32 instr) {
     case 0x00: mfc1(regs, instr); break;
     case 0x01: dmfc1(regs, instr); break;
     case 0x02: cfc1(regs, instr); break;
-    case 0x03: FireException(regs, ExceptionCode::ReservedInstruction, 1, true); break;
+    case 0x03: unimplemented(regs); break;
     case 0x04: mtc1(regs, instr); break;
     case 0x05: dmtc1(regs, instr); break;
     case 0x06: ctc1(regs, instr); break;
-    case 0x07: FireException(regs, ExceptionCode::ReservedInstruction, 1, true); break;
+    case 0x07: unimplemented(regs); break;
     case 0x08:
       switch(mask_branch) {
         case 0: cpu.b(instr, !regs.cop1.fcr31.compare); break;
@@ -77,9 +73,6 @@ void Cop1::decodeInterp(Interpreter &cpu, u32 instr) {
         case 0x0D: truncws(regs, instr); break;
         case 0x0E: ceilws(regs, instr); break;
         case 0x0F: floorws(regs, instr); break;
-        case 0x20:
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
-          break;
         case 0x21: cvtds(regs, instr); break;
         case 0x24: cvtws(regs, instr); break;
         case 0x25: cvtls(regs, instr); break;
@@ -99,7 +92,7 @@ void Cop1::decodeInterp(Interpreter &cpu, u32 instr) {
         case 0x3D: ccond<float>(regs, instr, NGE); break;
         case 0x3E: ccond<float>(regs, instr, LE); break;
         case 0x3F: ccond<float>(regs, instr, NGT); break;
-        default: Util::panic("Unimplemented COP1 function S[{} {}] ({:08X}) ({:016X})", mask_fun >> 3, mask_fun & 7, instr, (u64)regs.oldPC);
+        default: unimplemented(regs);
       }
       break;
     case 0x11: // d
@@ -121,9 +114,6 @@ void Cop1::decodeInterp(Interpreter &cpu, u32 instr) {
         case 0x0E: ceilwd(regs, instr); break;
         case 0x0F: floorwd(regs, instr); break;
         case 0x20: cvtsd(regs, instr); break;
-        case 0x21:
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
-          break;
         case 0x24: cvtwd(regs, instr); break;
         case 0x25: cvtld(regs, instr); break;
         case 0x30: ccond<double>(regs, instr, F); break;
@@ -142,35 +132,21 @@ void Cop1::decodeInterp(Interpreter &cpu, u32 instr) {
         case 0x3D: ccond<double>(regs, instr, NGE); break;
         case 0x3E: ccond<double>(regs, instr, LE); break;
         case 0x3F: ccond<double>(regs, instr, NGT); break;
-        default: Util::panic("Unimplemented COP1 function D[{} {}] ({:08X}) ({:016X})", mask_fun >> 3, mask_fun & 7, instr, (u64)regs.oldPC);
+        default: unimplemented(regs);
       }
       break;
     case 0x14: // w
       switch(mask_fun) {
-        case 0x01: subw(regs, instr); break;
-        case 0x05: absw(regs, instr); break;
-        case 0x02: mulw(regs, instr); break;
-        case 0x06: movw(regs, instr); break;
         case 0x20: cvtsw(regs, instr); break;
         case 0x21: cvtdw(regs, instr); break;
-        case 0x24:
+        default: unimplemented(regs);
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
-          break;
-        default: Util::panic("Unimplemented COP1 function W[{} {}] ({:08X}) ({:016X})", mask_fun >> 3, mask_fun & 7, instr, (u64)regs.oldPC);
       }
       break;
     case 0x15: // l
       switch(mask_fun) {
-        case 0x01: subl(regs, instr); break;
-        case 0x05: absl(regs, instr); break;
-        case 0x02: mull(regs, instr); break;
-        case 0x06: movl(regs, instr); break;
         case 0x20: cvtsl(regs, instr); break;
         case 0x21: cvtdl(regs, instr); break;
-        case 0x24: case 0x25:
+        default: unimplemented(regs);
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
-          break;
-        default: Util::panic("Unimplemented COP1 function L[{} {}] ({:08X}) ({:016X})", mask_fun >> 3, mask_fun & 7, instr, (u64)regs.oldPC);
       }
       break;
     default: Util::panic("Unimplemented COP1 instruction {} {}", mask_sub >> 3, mask_sub & 7);
@@ -197,11 +173,11 @@ void Cop1::decodeJIT(JIT &cpu, u32 instr) {
     case 0x00: mfc1(regs, instr); break;
     case 0x01: dmfc1(regs, instr); break;
     case 0x02: cfc1(regs, instr); break;
-    case 0x03: FireException(regs, ExceptionCode::ReservedInstruction, 1, true); break;
+    case 0x03: unimplemented(regs); break;
     case 0x04: mtc1(regs, instr); break;
     case 0x05: dmtc1(regs, instr); break;
     case 0x06: ctc1(regs, instr); break;
-    case 0x07: FireException(regs, ExceptionCode::ReservedInstruction, 1, true); break;
+    case 0x07: unimplemented(regs); break;
     case 0x08:
       /*switch(mask_branch) {
         case 0: {
@@ -239,7 +215,7 @@ void Cop1::decodeJIT(JIT &cpu, u32 instr) {
         case 0x0E: ceilws(regs, instr); break;
         case 0x0F: floorws(regs, instr); break;
         case 0x20:
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
+          unimplemented(regs);
           break;
         case 0x21: cvtds(regs, instr); break;
         case 0x24: cvtws(regs, instr); break;
@@ -283,7 +259,7 @@ void Cop1::decodeJIT(JIT &cpu, u32 instr) {
         case 0x0F: floorwd(regs, instr); break;
         case 0x20: cvtsd(regs, instr); break;
         case 0x21:
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
+          unimplemented(regs);
           break;
         case 0x24: cvtwd(regs, instr); break;
         case 0x25: cvtld(regs, instr); break;
@@ -308,28 +284,20 @@ void Cop1::decodeJIT(JIT &cpu, u32 instr) {
       break;
     case 0x14: // w
       switch(mask_fun) {
-        case 0x01: subw(regs, instr); break;
-        case 0x05: absw(regs, instr); break;
-        case 0x02: mulw(regs, instr); break;
-        case 0x06: movw(regs, instr); break;
         case 0x20: cvtsw(regs, instr); break;
         case 0x21: cvtdw(regs, instr); break;
         case 0x24:
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
+          unimplemented(regs);
           break;
         default: Util::panic("Unimplemented COP1 function W[{} {}] ({:08X}) ({:016X})", mask_fun >> 3, mask_fun & 7, instr, (u64)regs.oldPC);
       }
       break;
     case 0x15: // l
       switch(mask_fun) {
-        case 0x01: subl(regs, instr); break;
-        case 0x05: absl(regs, instr); break;
-        case 0x02: mull(regs, instr); break;
-        case 0x06: movl(regs, instr); break;
         case 0x20: cvtsl(regs, instr); break;
         case 0x21: cvtdl(regs, instr); break;
         case 0x24: case 0x25:
-          FireException(regs, ExceptionCode::ReservedInstruction, 1, true);
+          unimplemented(regs);
           break;
         default: Util::panic("Unimplemented COP1 function L[{} {}] ({:08X}) ({:016X})", mask_fun >> 3, mask_fun & 7, instr, (u64)regs.oldPC);
       }

src/backend/core/registers/Cop1.hpp

@@ -36,7 +36,18 @@ union FCR31 {
     unsigned:14;
   } __attribute__((__packed__));
 
-  u32 raw;
+  u32 read() const {
+    return (fs << 24) | (compare << 23) | (cause << 12) | (enable << 7) | (flag << 2) | rounding_mode;
+  }
+
+  void write(u32 val) {
+    fs = (val & 0x01000000) >> 24;
+    compare = (val & 0x00800000) >> 23;
+    cause = (val & 0x0003f000) >> 12;
+    enable = (val & 0x00000f80) >> 7;
+    flag = (val & 0x0000007c) >> 2;
+    rounding_mode = val & 3;
+  }
 };
 
 enum CompConds {
@@ -70,92 +81,133 @@ struct Cop1 {
   friend struct Interpreter;
   friend struct JIT;
 
-  template <typename T>
+  template<typename T>
-  FORCE_INLINE void SetReg(Cop0& cop0, u8 index, T value) {
+  FORCE_INLINE T GetFGR_FR(Cop0& cop0, u8 r) {
-    if constexpr(sizeof(T) == 4) {
+    if constexpr (std::is_same_v<T, u32> || std::is_same_v<T, s32>) {
       if (cop0.status.fr) {
-        fgr[index].lo = value;
+        return fgr[r].lo;
       } else {
-        if (index & 1) {
+        if (r & 1) {
-          fgr[index & ~1].hi = value;
+          return fgr[r & ~1].hi;
         } else {
-          fgr[index].lo = value;
+          return fgr[r].lo;
         }
       }
-    } else if constexpr(sizeof(T) == 8) {
+    } else if constexpr (std::is_same_v<T, u64> || std::is_same_v<T, s64>) {
-      if(!cop0.status.fr) {
+      if (!cop0.status.fr) {
-        index &= ~1;
+        // When this bit is not set, accessing odd registers is not allowed.
+        r &= ~1;
       }
 
-      fgr[index].raw = value;
+      return fgr[r].raw;
+    }
+  }
+
+  template<typename T>
+  FORCE_INLINE void SetFGR_FR(Cop0& cop0, u8 r, T value) {
+    if constexpr (std::is_same_v<T, u32> || std::is_same_v<T, s32>) {
+      if (cop0.status.fr) {
+        fgr[r].lo = value;
+      } else {
+        if (r & 1) {
+          fgr[r & ~1].hi = value;
+        } else {
+          fgr[r].lo = value;
+        }
+      }
+    } else if constexpr (std::is_same_v<T, u64> || std::is_same_v<T, s64>) {
+      if (!cop0.status.fr) {
+        // When this bit is not set, accessing odd registers is not allowed.
+        r &= ~1;
+      }
+
+      fgr[r].raw = value;
+    }
+  }
+
+  template<typename T>
+  FORCE_INLINE void SetFGR(u8 r, T value) {
+    fgr[r].raw = value;
+  }
+
+  template<typename T>
+  FORCE_INLINE u64 GetFGR(u8 r) {
+    if constexpr (std::is_same_v<T, u32> || std::is_same_v<T, s32>) {
+      return fgr[r].lo;
+    } else if constexpr (std::is_same_v<T, u64> || std::is_same_v<T, s64>) {
+      return fgr[r].raw;
     }
   }
 
   template <typename T>
-  FORCE_INLINE T GetReg(Cop0& cop0, u8 index) {
+  FORCE_INLINE T GetFGR_FS(Cop0& cop0, u8 fs) {
-    if constexpr(sizeof(T) == 4) {
+    if constexpr (std::is_same_v<T, u32> || std::is_same_v<T, s32>) {
-      if(cop0.status.fr) {
+      if (!cop0.status.fr) {
-        return fgr[index].lo;
+        fs &= ~1;
-      } else {
-        if (index & 1) {
-          return fgr[index & ~1].hi;
-        } else {
-          return fgr[index].lo;
       }
+      return fgr[fs].lo;
+    } else if constexpr (std::is_same_v<T, float> || std::is_same_v<T, double>) {
+      if (!cop0.status.fr) {
+        fs &= ~1;
       }
-    } else if constexpr(sizeof(T) == 8) {
+      return GetFGR_Raw<T>(fs);
-      if(!cop0.status.fr) {
-        index &= ~1;
-      }
-
-      return fgr[index].raw;
     }
   }
 
   template <typename T>
-  FORCE_INLINE void SetCop1Reg(Cop0& cop0, u8 index, T value) {
+  FORCE_INLINE T GetFGR_Raw(u8 r) {
-    if constexpr (sizeof(T) == 4) {
+    if constexpr (std::is_same_v<T, float>) {
-      u32 raw;
+      static_assert(sizeof(float) == sizeof(u32), "float and u32 need to both be 32 bits for this to work.");
-      memcpy(&raw, &value, sizeof(T));
+      auto rawvalue = GetFGR<u32>(r);
-      SetReg<u32>(cop0, index, raw);
+      float floatvalue;
-    } else if constexpr (sizeof(T) == 8) {
+      memcpy(&floatvalue, &rawvalue, sizeof(float));
-      u64 raw;
+      return floatvalue;
-      memcpy(&raw, &value, sizeof(T));
+    } else if constexpr (std::is_same_v<T, double>) {
-      SetReg<u64>(cop0, index, raw);
+      static_assert(sizeof(double) == sizeof(u64), "double and u64 need to both be 64 bits for this to work.");
+      double doublevalue;
+      auto rawvalue = GetFGR<u64>(r);
+      memcpy(&doublevalue, &rawvalue, sizeof(double));
+      return doublevalue;
     }
   }
 
   template <typename T>
-  FORCE_INLINE T GetCop1Reg(Cop0& cop0, u8 index) {
+  FORCE_INLINE void SetFGR_Raw(u8 r, T val) {
-    T value;
+    if constexpr (std::is_same_v<T, float>) {
-    if constexpr (sizeof(T) == 4) {
+      static_assert(sizeof(float) == sizeof(u32), "float and u32 need to both be 32 bits for this to work.");
-      u32 raw = GetReg<u32>(cop0, index);
+
-      memcpy(&value, &raw, sizeof(T));
+      u32 rawvalue;
-    } else if constexpr (sizeof(T) == 8) {
+      memcpy(&rawvalue, &val, sizeof(float));
-      u64 raw = GetReg<u64>(cop0, index);
+      SetFGR<u32>(r, rawvalue);
-      memcpy(&value, &raw, sizeof(T));
+    } else if constexpr (std::is_same_v<T, double>) {
+      static_assert(sizeof(double) == sizeof(u64), "double and u64 need to both be 64 bits for this to work.");
+
+      u64 rawvalue;
+      memcpy(&rawvalue, &val, sizeof(double));
+      SetFGR<u64>(r, rawvalue);
     }
-    return value;
+  }
+
+  template <typename T>
+  FORCE_INLINE T GetFGR_FT(u8 ft) {
+    return GetFGR_Raw<T>(ft);
   }
 private:
   void decodeInterp(Interpreter&, u32);
   void decodeJIT(JIT&, u32);
   void absd(Registers&, u32 instr);
   void abss(Registers&, u32 instr);
-  void absw(Registers&, u32 instr);
-  void absl(Registers&, u32 instr);
   void adds(Registers&, u32 instr);
   void addd(Registers&, u32 instr);
   void subs(Registers&, u32 instr);
   void subd(Registers&, u32 instr);
-  void subw(Registers&, u32 instr);
-  void subl(Registers&, u32 instr);
   void ceills(Registers&, u32 instr);
   void ceilws(Registers&, u32 instr);
   void ceilld(Registers&, u32 instr);
   void ceilwd(Registers&, u32 instr);
   void cfc1(Registers&, u32 instr) const;
   void ctc1(Registers&, u32 instr);
+  void unimplemented(Registers&);
   void roundls(Registers&, u32 instr);
   void roundld(Registers&, u32 instr);
   void roundws(Registers&, u32 instr);
@@ -180,12 +232,8 @@ private:
   void divd(Registers&, u32 instr);
   void muls(Registers&, u32 instr);
   void muld(Registers&, u32 instr);
-  void mulw(Registers&, u32 instr);
-  void mull(Registers&, u32 instr);
   void movs(Registers&, u32 instr);
   void movd(Registers&, u32 instr);
-  void movw(Registers&, u32 instr);
-  void movl(Registers&, u32 instr);
   void negs(Registers&, u32 instr);
   void negd(Registers&, u32 instr);
   void sqrts(Registers&, u32 instr);

src/backend/core/registers/cop/cop1instructions.cpp

@@ -7,6 +7,21 @@
 #include <cfenv>
 
 namespace n64 {
+FORCE_INLINE bool FireFPUException(Registers& regs) {
+  FCR31& fcr31 = regs.cop1.fcr31;
+  if(fcr31.cause & ((1 << 5) | fcr31.enable)) {
+    FireException(regs, ExceptionCode::FloatingPointError, 0, true);
+    return true;
+  }
+
+  return false;
+}
+
+#define CheckFPUException() do { if(FireFPUException(regs)) { return; } } while(0)
+#define CheckFPUUsable_PreserveCause() do { if(!regs.cop0.status.cu1) { FireException(regs, ExceptionCode::CoprocessorUnusable, 1, true); return; } } while(0)
+#define CheckFPUUsable() do { CheckFPUUsable_PreserveCause(); regs.cop1.fcr31.cause = 0; } while(0)
+#define CheckRound(a, b) do { if ((a) != (b)) { fcr31.cause_inexact_operation = true; if(!fcr31.enable_inexact_operation) { fcr31.flag_inexact_operation = true; } } CheckFPUException(); } while(0);
+
 FORCE_INLINE int PushRoundingMode(const FCR31& fcr31) {
   int og = fegetround();
   switch(fcr31.rounding_mode) {
@@ -19,85 +34,265 @@ FORCE_INLINE int PushRoundingMode(const FCR31& fcr31) {
   return og;
 }
 
+#define CheckCVTArg(f) do { SetCauseByArgCVT(regs, f); CheckFPUException(); } while(0)
+#define CheckArg(f) do { SetCauseByArg(regs, f); CheckFPUException(); } while(0)
+#define PUSHROUNDING int orig_round = PushRoundingMode(regs.cop1.fcr31)
+#define POPROUNDING fesetround(orig_round)
+#define OP_CheckExcept(op) do { PUSHROUNDING; feclearexcept(FE_ALL_EXCEPT); op; SetFPUCauseRaised(regs, fetestexcept(FE_ALL_EXCEPT)); POPROUNDING; } while(0)
+#define CVT_OP_CheckExcept(op) do { feclearexcept(FE_ALL_EXCEPT); op; SetFPUCauseCVTRaised(regs, fetestexcept(FE_ALL_EXCEPT)); CheckFPUException(); } while(0)
+
+#define OP(T, op) do { \
+  CheckFPUUsable(); \
+  auto fs = GetFGR_FS<T>(regs.cop0, FS(instr)); \
+  auto ft = GetFGR_FT<T>(FT(instr)); \
+  CheckArg(fs); \
+  CheckArg(ft); \
+  T result; \
+  OP_CheckExcept({result = (op);}); \
+  SetFGR_Raw<T>(FD(instr), result); \
+} while(0)
+
+template <typename T>
+FORCE_INLINE void SetCauseByArgCVT(Registers& regs, T f) {
+  if constexpr(sizeof(T) == 4) {
+    switch (std::fpclassify(f)) {
+      case FP_NAN:
+      case FP_INFINITE:
+      case FP_SUBNORMAL:
+        regs.cop1.fcr31.cause_unimplemented_operation = true;
+        break;
+
+      case FP_NORMAL:
+        // Check overflow
+        if (f >= 2147483648.0 || f < -2147483648.0) {
+          regs.cop1.fcr31.cause_unimplemented_operation = true;
+        }
+        break;
+
+      case FP_ZERO:
+        break; // Fine
+    }
+  } else if constexpr(sizeof(T) == 8) {
+    switch (std::fpclassify(f)) {
+      case FP_NAN:
+      case FP_INFINITE:
+      case FP_SUBNORMAL:
+        regs.cop1.fcr31.cause_unimplemented_operation = true;
+        break;
+
+      case FP_NORMAL:
+        // Check overflow
+        if (f >= 9007199254740992.000000 || f <= -9007199254740992.000000) {
+          regs.cop1.fcr31.cause_unimplemented_operation = true;
+        }
+        break;
+
+      case FP_ZERO:
+        break; // Fine
+    }
+  }
+}
+
+FORCE_INLINE void SetFPUCauseRaised(Registers& regs, int raised) {
+  if (raised == 0) {
+    return;
+  }
+
+  if (raised & FE_UNDERFLOW) {
+    if (!regs.cop1.fcr31.fs || regs.cop1.fcr31.enable_underflow || regs.cop1.fcr31.enable_inexact_operation) {
+      regs.cop1.fcr31.cause_unimplemented_operation = true;
+      return;
+    } else {
+      regs.cop1.fcr31.cause_underflow = true;
+      if(!regs.cop1.fcr31.enable_underflow) {
+        regs.cop1.fcr31.flag_underflow = true;
+      }
+    }
+  }
+
+  if (raised & FE_INEXACT) {
+    regs.cop1.fcr31.cause_inexact_operation = true;
+    if(!regs.cop1.fcr31.enable_inexact_operation) {
+      regs.cop1.fcr31.flag_inexact_operation = true;
+    }
+  }
+
+  if (raised & FE_DIVBYZERO) {
+    regs.cop1.fcr31.cause_division_by_zero = true;
+    if(!regs.cop1.fcr31.enable_division_by_zero) {
+      regs.cop1.fcr31.flag_division_by_zero = true;
+    }
+  }
+
+  if (raised & FE_OVERFLOW) {
+    regs.cop1.fcr31.cause_overflow = true;
+    if(!regs.cop1.fcr31.enable_overflow) {
+      regs.cop1.fcr31.flag_overflow = true;
+    }
+  }
+
+  if (raised & FE_INVALID) {
+    regs.cop1.fcr31.cause_invalid_operation = true;
+    if(!regs.cop1.fcr31.enable_invalid_operation) {
+      regs.cop1.fcr31.flag_invalid_operation = true;
+    }
+  }
+}
+
+FORCE_INLINE void SetFPUCauseCVTRaised(Registers& regs, int raised) {
+  if(raised & FE_INVALID) {
+    regs.cop1.fcr31.cause_unimplemented_operation = true;
+    return;
+  }
+
+  SetFPUCauseRaised(regs, raised);
+}
+
+template <typename T>
+FORCE_INLINE void SetCauseByArg(Registers& regs, T f) {
+  int c = std::fpclassify(f);
+  switch(c) {
+    case FP_NAN:
+      if(isqnan(f)) {
+        regs.cop1.fcr31.cause_invalid_operation = true;
+        if(!regs.cop1.fcr31.enable_invalid_operation) {
+          regs.cop1.fcr31.flag_invalid_operation = true;
+        }
+      } else {
+        regs.cop1.fcr31.cause_unimplemented_operation = true;
+      }
+      break;
+    case FP_SUBNORMAL:
+      regs.cop1.fcr31.cause_unimplemented_operation = true;
+      break;
+    case FP_INFINITE:
+    case FP_ZERO:
+    case FP_NORMAL:
+      break; // No-op, these are fine.
+    default:
+      Util::panic("Unknown floating point classification: {}", c);
+  }
+}
+
 #define PUSHROUNDINGMODE int og = PushRoundingMode(fcr31)
 #define POPROUNDINGMODE fesetround(og)
 
+#define any_unordered(fs, ft) (std::isnan(fs) || std::isnan(ft))
+
+#define F_TO_U32(f) (*((u32*)(&(f))))
+#define D_TO_U64(d) (*((u64*)(&(d))))
+#define U64_TO_D(d) (*((double*)(&(d))))
+
+template <typename T>
+FORCE_INLINE bool isnan(T f) {
+  if constexpr(std::is_same<T, float>::value) {
+    u32 v = F_TO_U32(f);
+    return ((v & 0x7F800000) == 0x7F800000) && ((v & 0x7FFFFF) != 0);
+  } else if constexpr(std::is_same<T, double>::value) {
+    u64 v = D_TO_U64(f);
+    return ((v & 0x7FF0000000000000) == 0x7FF0000000000000) && ((v & 0xFFFFFFFFFFFFF) != 0);
+  } else {
+    Util::panic("Invalid float type in isnan");
+  }
+}
+
+template <typename T>
+FORCE_INLINE bool isqnan(T f) {
+  if constexpr(std::is_same<T, float>::value) {
+    u32 v = F_TO_U32(f);
+    return (v & 0x7FC00000) == 0x7FC00000;
+  } else if constexpr(std::is_same<T, double>::value) {
+    u64 v = D_TO_U64(f);
+    return (v & 0x7FF8000000000000) == 0x7FF8000000000000;
+  } else {
+    Util::panic("Invalid float type in isqnan");
+  }
+}
+
 #define checknanregs(fs, ft) do { \
-  if(std::isnan(fs) || std::isnan(ft)) { \
+  if(isnan(fs) || isnan(ft)) {                                       \
-    regs.cop1.fcr31.flag_invalid_operation = true; \
     regs.cop1.fcr31.cause_invalid_operation = true;                  \
-    FireException(regs, ExceptionCode::FloatingPointError, 1, true); \
+    if(!regs.cop1.fcr31.enable_invalid_operation) {                  \
-    return; \
+      regs.cop1.fcr31.flag_invalid_operation = true;                 \
+    }                                                                \
+    CheckFPUException();                                             \
+  }                                                                  \
+} while(0)
+
+#define checkqnanregs(fs, ft) do { \
+  if(isqnan(fs) || isqnan(ft)) {                                     \
+    regs.cop1.fcr31.cause_invalid_operation = true;                  \
+    if(!regs.cop1.fcr31.enable_invalid_operation) {                  \
+      regs.cop1.fcr31.flag_invalid_operation = true;                 \
+    }                                                                \
+    CheckFPUException();                                             \
   }                                                                  \
 } while(0)
 
 void Cop1::absd(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  OP(double, std::abs(fs));
-  SetCop1Reg<double>(regs.cop0, FD(instr), std::abs(fs));
 }
 
 void Cop1::abss(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  OP(float, std::abs(fs));
-  SetCop1Reg<float>(regs.cop0, FD(instr), std::abs(fs));
-}
-
-void Cop1::absw(Registers& regs, u32 instr) {
-  s32 fs = GetReg<s32>(regs.cop0, FS(instr));
-  SetReg<u32>(regs.cop0, FD(instr), std::abs(fs));
-}
-
-void Cop1::absl(Registers& regs, u32 instr) {
-  s64 fs = GetReg<s64>(regs.cop0, FS(instr));
-  SetReg(regs.cop0, FD(instr), std::abs(fs));
 }
 
 void Cop1::adds(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  OP(float, fs + ft);
-  float ft = GetCop1Reg<float>(regs.cop0, FT(instr));
-  checknanregs(fs, ft);
-  float result = fs + ft;
-  SetCop1Reg<float>(regs.cop0, FD(instr), result);
 }
 
 void Cop1::addd(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  OP(double, fs + ft);
-  double ft = GetCop1Reg<double>(regs.cop0, FT(instr));
-  checknanregs(fs, ft);
-  double result = fs + ft;
-  SetCop1Reg<double>(regs.cop0, FD(instr), result);
 }
 
 void Cop1::ceills(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  s64 result = std::ceil(fs);
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-  SetReg<u64>(regs.cop0, FD(instr), result);
+  CheckCVTArg(fs);
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::ceil(fs); });
+  CheckRound(result, fs);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::ceilws(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  s32 result = std::ceil(fs);
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-  SetReg<u32>(regs.cop0, FD(instr), result);
+  CheckCVTArg(fs);
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::ceil(fs); });
+  CheckRound(result, fs);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::ceilld(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  s64 result = std::ceil(fs);
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-  SetReg<u64>(regs.cop0, FD(instr), result);
+  CheckCVTArg(fs);
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::ceil(fs); });
+  CheckRound(result, fs);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::ceilwd(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  s32 result = std::ceil(fs);
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-  SetReg<u32>(regs.cop0, FD(instr), result);
+  CheckCVTArg(fs);
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::ceil(fs); });
+  CheckRound(result, fs);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::cfc1(Registers& regs, u32 instr) const {
+  CheckFPUUsable_PreserveCause();
   u8 fd = RD(instr);
   s32 val = 0;
   switch(fd) {
     case 0: val = fcr0; break;
     case 31:
-      val = fcr31.raw;
+      val = fcr31.read();
       break;
     default: Util::panic("Undefined CFC1 with rd != 0 or 31");
   }
@@ -105,349 +300,327 @@ void Cop1::cfc1(Registers& regs, u32 instr) const {
 }
 
 void Cop1::ctc1(Registers& regs, u32 instr) {
+  CheckFPUUsable_PreserveCause();
   u8 fs = RD(instr);
   u32 val = regs.gpr[RT(instr)];
   switch(fs) {
     case 0: break;
     case 31: {
-      val &= 0x183ffff;
+      fcr31.write(val);
-      fcr31.raw = val;
+      FireFPUException(regs);
     } break;
     default: Util::panic("Undefined CTC1 with rd != 0 or 31");
   }
 }
 
 void Cop1::cvtds(Registers& regs, u32 instr) {
-  SetCop1Reg<double>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-    FD(instr),
+  CheckCVTArg(fs);
-    GetCop1Reg<float>(
+  double result;
-            regs.cop0,
+  CVT_OP_CheckExcept({ result = double(fs); });
-      FS(instr)
+  SetFGR_Raw(FD(instr), result);
-    )
-  );
 }
 
 void Cop1::cvtsd(Registers& regs, u32 instr) {
-  SetCop1Reg<float>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-    FD(instr),
+  CheckCVTArg(fs);
-    GetCop1Reg<double>(
+  float result;
-            regs.cop0,
+  CVT_OP_CheckExcept({ result = float(fs); });
-      FS(instr)
+  SetFGR_Raw(FD(instr), result);
-    )
-  );
 }
 
 void Cop1::cvtwd(Registers& regs, u32 instr) {
-  SetReg<u32>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-    FD(instr),
+  CheckCVTArg(fs);
-    GetCop1Reg<double>(
+  s32 result;
-            regs.cop0,
+  CVT_OP_CheckExcept({ result = s32(fs); });
-      FS(instr)
+  SetFGR(FD(instr), result);
-    )
-  );
 }
 
 void Cop1::cvtws(Registers& regs, u32 instr) {
-  SetReg<u32>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-    FD(instr),
+  CheckCVTArg(fs);
-    GetCop1Reg<float>(
+  s32 result;
-            regs.cop0,
+  CVT_OP_CheckExcept({ result = s32(fs); });
-      FS(instr)
+  SetFGR(FD(instr), result);
-    )
-  );
 }
 
 void Cop1::cvtls(Registers& regs, u32 instr) {
-  SetReg<u64>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-    FD(instr),
+  CheckCVTArg(fs);
-    GetCop1Reg<float>(
+  s64 result;
-            regs.cop0,
+  CVT_OP_CheckExcept({ result = s64(fs); });
-      FS(instr)
+  SetFGR(FD(instr), result);
-    )
-  );
 }
 
 void Cop1::cvtsl(Registers& regs, u32 instr) {
-  SetCop1Reg<float>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FR<s64>(regs.cop0, FS(instr));
-    FD(instr),
+  if (fs >= (s64)0x0080000000000000 || fs < s64(0xff80'0000'0000'0000)) {
-    (s64)GetReg<u64>(
+    fcr31.cause_unimplemented_operation = true;
-            regs.cop0,
+    CheckFPUException();
-      FS(instr)
+  }
-    )
+  float result;
-  );
+  OP_CheckExcept({ result = float(fs); });
+  SetFGR_Raw(FD(instr), result);
 }
 
 void Cop1::cvtdw(Registers& regs, u32 instr) {
-  SetCop1Reg<double>(
+  CheckFPUUsable();
-    regs.cop0,
+  auto fs = GetFGR_FS<s32>(regs.cop0, FS(instr));
-    FD(instr),
+  double result;
-    (s32)GetReg<u32>(
+  OP_CheckExcept({ result = double(fs); });
-      regs.cop0,
+  SetFGR_Raw(FD(instr), result);
-      FS(instr)
-    )
-  );
 }
 
 void Cop1::cvtsw(Registers& regs, u32 instr) {
-  SetCop1Reg<float>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FS<s32>(regs.cop0, FS(instr));
-    FD(instr),
+  float result;
-    (s32)GetReg<u32>(
+  OP_CheckExcept({ result = float(fs); });
-            regs.cop0,
+  SetFGR_Raw(FD(instr), result);
-      FS(instr)
-    )
-  );
 }
 
 void Cop1::cvtdl(Registers& regs, u32 instr) {
-  SetCop1Reg<double>(
+  CheckFPUUsable();
-        regs.cop0,
+  auto fs = GetFGR_FR<s64>(regs.cop0, FS(instr));
-    FD(instr),
+  if (fs >= (s64)0x0080000000000000 || fs < s64(0xff80'0000'0000'0000)) {
-    (s64)GetReg<u64>(
+    fcr31.cause_unimplemented_operation = true;
-            regs.cop0,
+    CheckFPUException();
-      FS(instr)
+  }
-    )
+  double result;
-  );
+  OP_CheckExcept({ result = double(fs); });
+  SetFGR_Raw(FD(instr), result);
 }
 
 void Cop1::cvtld(Registers& regs, u32 instr) {
-  SetReg<u64>(
+  CheckFPUUsable();
-    regs.cop0,
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-    FD(instr),
+  CheckCVTArg(fs);
-    GetCop1Reg<double>(
+  s64 result;
-      regs.cop0,
+  PUSHROUNDING;
-      FS(instr)
+  CVT_OP_CheckExcept({ result = std::rint(fs); });
-    )
+  POPROUNDING;
-  );
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 template <typename T>
-inline bool CalculateCondition(Registers& regs, T fs, T ft, CompConds cond) {
+inline bool CalculateCondition(T fs, T ft, CompConds cond) {
   switch(cond) {
-    case F: return false;
+    case F: case SF: return false;
-    case UN: return std::isnan(fs) || std::isnan(ft);
+    case UN: case NGLE: return any_unordered(fs, ft);
-    case EQ: return fs == ft;
+    case EQ: case SEQ: return fs == ft;
-    case UEQ: return (std::isnan(fs) || std::isnan(ft)) || (fs == ft);
+    case UEQ: case NGL: return fs == ft || any_unordered(fs, ft);
-    case OLT: return (!std::isnan(fs) && !std::isnan(ft)) && (fs < ft);
+    case OLT: case LT: return fs < ft;
-    case ULT: return (std::isnan(fs) || std::isnan(ft)) || (fs < ft);
+    case ULT: case NGE: return fs < ft || any_unordered(fs, ft);
-    case OLE: return (!std::isnan(fs) && !std::isnan(ft)) && (fs <= ft);
+    case OLE: case LE: return fs <= ft;
-    case ULE: return (std::isnan(fs) || std::isnan(ft)) || (fs <= ft);
+    case ULE: case NGT: return fs <= ft || any_unordered(fs, ft);
-    default:
-      if(std::isnan(fs) || std::isnan(ft)) {
-        regs.cop1.fcr31.flag_invalid_operation = true;
-        regs.cop1.fcr31.cause_invalid_operation = true;
-        FireException(regs, ExceptionCode::FloatingPointError, 1, true);
-        return false;
   }
+}
 
-      return CalculateCondition(regs, fs, ft, static_cast<CompConds>(cond - 8));
+template <typename T>
+inline void CheckInvalidRegs(Registers& regs, T fs, T ft, CompConds cond) {
+  switch(cond) {
+    case F ... ULE: checkqnanregs(fs, ft); break;
+    case SF ... NGT: checknanregs(fs, ft); break;
   }
 }
 
 template <typename T>
 void Cop1::ccond(Registers& regs, u32 instr, CompConds cond) {
-  T fs = GetCop1Reg<T>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  T ft = GetCop1Reg<T>(regs.cop0, FT(instr));
+  T fs = GetFGR_FS<T>(regs.cop0, FS(instr));
-
+  T ft = GetFGR_FT<T>(FT(instr));
-  fcr31.compare = CalculateCondition(regs, fs, ft, cond);
+  CheckInvalidRegs(regs, fs, ft, cond);
+  fcr31.compare = CalculateCondition(fs, ft, cond);
 }
 
 template void Cop1::ccond<float>(Registers& regs, u32 instr, CompConds cond);
 template void Cop1::ccond<double>(Registers& regs, u32 instr, CompConds cond);
 
 void Cop1::divs(Registers &regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  OP(float, fs / ft);
-  float ft = GetCop1Reg<float>(regs.cop0, FT(instr));
-  SetCop1Reg<float>(regs.cop0, FD(instr), fs / ft);
 }
 
 void Cop1::divd(Registers &regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  OP(double, fs / ft);
-  double ft = GetCop1Reg<double>(regs.cop0, FT(instr));
-  SetCop1Reg<double>(regs.cop0, FD(instr), fs / ft);
 }
 
 void Cop1::muls(Registers &regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  OP(float, fs * ft);
-  float ft = GetCop1Reg<float>(regs.cop0, FT(instr));
-  SetCop1Reg<float>(regs.cop0, FD(instr), fs * ft);
 }
 
 void Cop1::muld(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  OP(double, fs * ft);
-  double ft = GetCop1Reg<double>(regs.cop0, FT(instr));
-  SetCop1Reg<double>(regs.cop0, FD(instr), fs * ft);
-}
-
-void Cop1::mulw(Registers &regs, u32 instr) {
-  u32 fs = GetReg<u32>(regs.cop0, FS(instr));
-  u32 ft = GetReg<u32>(regs.cop0, FT(instr));
-  SetReg<u32>(regs.cop0, FD(instr), fs * ft);
-}
-
-void Cop1::mull(Registers &regs, u32 instr) {
-  u64 fs = GetReg<u64>(regs.cop0, FS(instr));
-  u64 ft = GetReg<u64>(regs.cop0, FT(instr));
-  SetReg<u64>(regs.cop0, FD(instr), fs * ft);
 }
 
 void Cop1::subs(Registers &regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  OP(float, fs - ft);
-  float ft = GetCop1Reg<float>(regs.cop0, FT(instr));
-  SetCop1Reg<float>(regs.cop0, FD(instr), fs - ft);
 }
 
 void Cop1::subd(Registers &regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  OP(double, fs - ft);
-  double ft = GetCop1Reg<double>(regs.cop0, FT(instr));
-  SetCop1Reg<double>(regs.cop0, FD(instr), fs - ft);
-}
-
-void Cop1::subw(Registers &regs, u32 instr) {
-  u32 fs = GetReg<u32>(regs.cop0, FS(instr));
-  u32 ft = GetReg<u32>(regs.cop0, FT(instr));
-  SetReg<u32>(regs.cop0, FD(instr), fs - ft);
-}
-
-void Cop1::subl(Registers &regs, u32 instr) {
-  u64 fs = GetReg<u64>(regs.cop0, FS(instr));
-  u64 ft = GetReg<u64>(regs.cop0, FT(instr));
-  SetReg<u64>(regs.cop0, FD(instr), fs - ft);
 }
 
 void Cop1::movs(Registers& regs, u32 instr) {
-  SetCop1Reg<float>(
+  CheckFPUUsable_PreserveCause();
-    regs.cop0,
+  auto val = GetFGR_FR<u64>(regs.cop0, FS(instr));
-    FD(instr),
+  SetFGR(FD(instr), val);
-    GetCop1Reg<float>(
-      regs.cop0,
-      FS(instr)
-    )
-  );
 }
 
 void Cop1::movd(Registers& regs, u32 instr) {
-  SetCop1Reg<double>(
+  CheckFPUUsable_PreserveCause();
-    regs.cop0,
+  auto val = GetFGR_FS<double>(regs.cop0, FS(instr));
-    FD(instr),
+  SetFGR_Raw(FD(instr), val);
-    GetCop1Reg<double>(
-      regs.cop0,
-      FS(instr)
-    )
-  );
-}
-
-void Cop1::movw(Registers& regs, u32 instr) {
-  SetReg<u32>(
-    regs.cop0,
-    FD(instr),
-    GetReg<u32>(
-      regs.cop0,
-      FS(instr)
-    )
-  );
-}
-
-void Cop1::movl(Registers& regs, u32 instr) {
-  SetReg<u64>(
-    regs.cop0,
-    FD(instr),
-    GetReg<u64>(
-      regs.cop0,
-      FS(instr)
-    )
-  );
 }
 
 void Cop1::negs(Registers &regs, u32 instr) {
-  SetCop1Reg<float>(
+  OP(float, -fs);
-    regs.cop0,
-    FD(instr),
-    -GetCop1Reg<float>(
-      regs.cop0,
-      FS(instr)
-    )
-  );
 }
 
 void Cop1::negd(Registers &regs, u32 instr) {
-  SetCop1Reg<double>(
+  OP(double, -ft);
-    regs.cop0,
-    FD(instr),
-    -GetCop1Reg<double>(
-      regs.cop0,
-      FS(instr)
-    )
-  );
 }
 
 void Cop1::sqrts(Registers &regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  OP(float, std::sqrt(fs));
-  SetCop1Reg<float>(regs.cop0, FD(instr), std::sqrt(fs));
 }
 
 void Cop1::sqrtd(Registers &regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  OP(double, std::sqrt(fs));
-  SetCop1Reg<double>(regs.cop0, FD(instr), std::sqrt(fs));
 }
 
 void Cop1::roundls(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  PUSHROUNDINGMODE;
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-  SetReg<u64>(regs.cop0, FD(instr), (s32)std::nearbyint(fs));
+  CheckCVTArg(fs);
-  POPROUNDINGMODE;
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::nearbyint(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::roundld(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  PUSHROUNDINGMODE;
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-  SetReg<u64>(regs.cop0, FD(instr), (s64)std::nearbyint(fs));
+  CheckCVTArg(fs);
-  POPROUNDINGMODE;
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::nearbyint(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::roundws(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  PUSHROUNDINGMODE;
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
-  SetReg<u32>(regs.cop0, FD(instr), (s32)std::nearbyint(fs));
+  CheckCVTArg(fs);
-  POPROUNDINGMODE;
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::nearbyint(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::roundwd(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  PUSHROUNDINGMODE;
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
-  SetReg<u32>(regs.cop0, FD(instr), (s32)std::nearbyint(fs));
+  CheckCVTArg(fs);
-  POPROUNDINGMODE;
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::nearbyint(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::floorls(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  SetReg<u64>(regs.cop0, FD(instr), (s64)std::floor(fs));
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::floor(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::floorld(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  SetReg<u64>(regs.cop0, FD(instr), (s64)std::floor(fs));
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::floor(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::floorws(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  SetReg<u32>(regs.cop0, FD(instr), (s64)std::floor(fs));
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::floor(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 void Cop1::floorwd(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
+  CheckFPUUsable();
-  SetReg<u32>(regs.cop0, FD(instr), (s64)std::floor(fs));
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::floor(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
+}
+
+void Cop1::truncws(Registers& regs, u32 instr) {
+  CheckFPUUsable();
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::trunc(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
+}
+
+void Cop1::truncwd(Registers& regs, u32 instr) {
+  CheckFPUUsable();
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s32 result;
+  CVT_OP_CheckExcept({ result = std::trunc(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
+}
+
+void Cop1::truncls(Registers& regs, u32 instr) {
+  CheckFPUUsable();
+  auto fs = GetFGR_FS<float>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::trunc(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
+}
+
+void Cop1::truncld(Registers& regs, u32 instr) {
+  CheckFPUUsable();
+  auto fs = GetFGR_FS<double>(regs.cop0, FS(instr));
+  CheckCVTArg(fs);
+  s64 result;
+  CVT_OP_CheckExcept({ result = std::trunc(fs); });
+  CheckRound(fs, result);
+  SetFGR(FD(instr), result);
 }
 
 template<class T>
 void Cop1::lwc1(T &cpu, Mem &mem, u32 instr) {
   if constexpr(std::is_same_v<decltype(cpu), Interpreter&>) {
+    Registers& regs = cpu.regs;
+    CheckFPUUsable_PreserveCause();
     lwc1Interp(cpu.regs, mem, instr);
   } else if constexpr (std::is_same_v<decltype(cpu), JIT&>) {
     lwc1JIT(cpu, mem, instr);
@@ -462,6 +635,8 @@ template void Cop1::lwc1<JIT>(JIT&, Mem&, u32);
 template<class T>
 void Cop1::swc1(T &cpu, Mem &mem, u32 instr) {
   if constexpr(std::is_same_v<decltype(cpu), Interpreter&>) {
+    Registers& regs = cpu.regs;
+    CheckFPUUsable_PreserveCause();
     swc1Interp(cpu.regs, mem, instr);
   } else if constexpr (std::is_same_v<decltype(cpu), JIT&>) {
     swc1JIT(cpu, mem, instr);
@@ -476,6 +651,8 @@ template void Cop1::swc1<JIT>(JIT&, Mem&, u32);
 template<class T>
 void Cop1::ldc1(T &cpu, Mem &mem, u32 instr) {
   if constexpr(std::is_same_v<decltype(cpu), Interpreter&>) {
+    Registers& regs = cpu.regs;
+    CheckFPUUsable_PreserveCause();
     ldc1Interp(cpu.regs, mem, instr);
   } else if constexpr (std::is_same_v<decltype(cpu), JIT&>) {
     ldc1JIT(cpu, mem, instr);
@@ -490,6 +667,8 @@ template void Cop1::ldc1<JIT>(JIT&, Mem&, u32);
 template<class T>
 void Cop1::sdc1(T &cpu, Mem &mem, u32 instr) {
   if constexpr(std::is_same_v<decltype(cpu), Interpreter&>) {
+    Registers& regs = cpu.regs;
+    CheckFPUUsable_PreserveCause();
     sdc1Interp(cpu.regs, mem, instr);
   } else if constexpr (std::is_same_v<decltype(cpu), JIT&>) {
     sdc1JIT(cpu, mem, instr);
@@ -515,7 +694,7 @@ void Cop1::lwc1Interp(Registers& regs, Mem& mem, u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     u32 data = mem.Read32(regs, physical);
-    SetReg<u32>(regs.cop0, FT(instr), data);
+    SetFGR_FR<u32>(regs.cop0, FT(instr), data);
   }
 }
 
@@ -532,10 +711,16 @@ void Cop1::swc1Interp(Registers& regs, Mem& mem, u32 instr) {
     HandleTLBException(regs, addr);
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, STORE), 0, true);
   } else {
-    mem.Write32(regs, physical, GetReg<u32>(regs.cop0, FT(instr)));
+    mem.Write32(regs, physical, GetFGR_FR<u32>(regs.cop0, FT(instr)));
   }
 }
 
+void Cop1::unimplemented(Registers& regs) {
+  CheckFPUUsable();
+  fcr31.cause_unimplemented_operation = true;
+  FireFPUException(regs);
+}
+
 void Cop1::ldc1Interp(Registers& regs, Mem& mem, u32 instr) {
   if(!regs.cop0.status.cu1) {
     FireException(regs, ExceptionCode::CoprocessorUnusable, 1, true);
@@ -550,7 +735,7 @@ void Cop1::ldc1Interp(Registers& regs, Mem& mem, u32 instr) {
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, LOAD), 0, true);
   } else {
     u64 data = mem.Read64(regs, physical);
-    SetReg<u64>(regs.cop0, FT(instr), data);
+    SetFGR_FR<u64>(regs.cop0, FT(instr), data);
   }
 }
 
@@ -567,48 +752,28 @@ void Cop1::sdc1Interp(Registers& regs, Mem& mem, u32 instr) {
     HandleTLBException(regs, addr);
     FireException(regs, GetTLBExceptionCode(regs.cop0.tlbError, STORE), 0, true);
   } else {
-    mem.Write64(regs, physical, GetReg<u64>(regs.cop0, FT(instr)));
+    mem.Write64(regs, physical, GetFGR_FR<u64>(regs.cop0, FT(instr)));
   }
 }
 
-void Cop1::truncws(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
-  s32 result = (s32)std::trunc(fs);
-  SetReg<u32>(regs.cop0, FD(instr), result);
-}
-
-void Cop1::truncwd(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
-  s32 result = (s32)std::trunc(fs);
-  SetReg<u32>(regs.cop0, FD(instr), result);
-}
-
-void Cop1::truncls(Registers& regs, u32 instr) {
-  float fs = GetCop1Reg<float>(regs.cop0, FS(instr));
-  s64 result = (s64)std::trunc(fs);
-  SetReg<u64>(regs.cop0, FD(instr), result);
-}
-
-void Cop1::truncld(Registers& regs, u32 instr) {
-  double fs = GetCop1Reg<double>(regs.cop0, FS(instr));
-  s64 result = (s64)std::trunc(fs);
-  SetReg<u64>(regs.cop0, FD(instr), result);
-}
-
 void Cop1::mfc1(Registers& regs, u32 instr) {
-  regs.gpr[RT(instr)] = (s32)GetReg<u32>(regs.cop0, FS(instr));
+  CheckFPUUsable_PreserveCause();
+  regs.gpr[RT(instr)] = (s32)GetFGR_FR<u32>(regs.cop0, FS(instr));
 }
 
 void Cop1::dmfc1(Registers& regs, u32 instr) {
-  regs.gpr[RT(instr)] = (s64)GetReg<u64>(regs.cop0, FS(instr));
+  CheckFPUUsable_PreserveCause();
+  regs.gpr[RT(instr)] = (s64)GetFGR_FR<u64>(regs.cop0, FS(instr));
 }
 
 void Cop1::mtc1(Registers& regs, u32 instr) {
-  SetReg<u32>(regs.cop0, FS(instr), regs.gpr[RT(instr)]);
+  CheckFPUUsable_PreserveCause();
+  SetFGR_FR<u32>(regs.cop0, FS(instr), regs.gpr[RT(instr)]);
 }
 
 void Cop1::dmtc1(Registers& regs, u32 instr) {
-  SetReg<u64>(regs.cop0, FS(instr), regs.gpr[RT(instr)]);
+  CheckFPUUsable_PreserveCause();
+  SetFGR_FR<u64>(regs.cop0, FS(instr), regs.gpr[RT(instr)]);
 }
 
 }

src/common.hpp

@@ -56,9 +56,6 @@ static FORCE_INLINE constexpr u32 GetVideoFrequency(bool pal) {
 #define ELEMENT_INDEX(i) (7 - (i))
 #define BYTE_INDEX(i)   (15 - (i))
 
-#define unlikely(exp) __builtin_expect(exp, 0)
-#define likely(exp) __builtin_expect(exp, 1)
-
 #if defined(WIN32) || defined(_WIN32) || defined(__WIN32) && !defined(__CYGWIN__)
 #define ABI_WINDOWS
 #else

src/frontend/imgui/Settings.cpp

@@ -29,10 +29,10 @@ Settings::Settings(n64::Core& core) {
     settings = json::parse(settingsFile);
 
     checkjsonentry(oldVolumeL, float, "audio", "volumeL", 0.5);
-    volumeL = oldVolumeL;
     checkjsonentry(oldVolumeR, float, "audio", "volumeR", 0.5);
-    volumeR = oldVolumeR;
     checkjsonentry(mute, bool, "audio", "mute", false);
+    volumeL = mute ? 0 : oldVolumeL;
+    volumeR = mute ? 0 : oldVolumeR;
     checkjsonentry(lockChannels, bool, "audio", "lockChannels", true);
     checkjsonentry(jit, bool, "cpu", "enableJIT", false);
   } else {