CPU: Move interrupt check out of inner-most exec loop

src/core/cpu_code_cache.cpp

@@ -132,18 +132,17 @@ static void ExecuteImpl()
   g_state.frame_done = false;
   while (!g_state.frame_done)
   {
+    if (HasPendingInterrupt())
+    {
+      SafeReadInstruction(g_state.regs.pc, &g_state.next_instruction.bits);
+      DispatchInterrupt();
+    }
+
     TimingEvents::UpdateCPUDowncount();
 
     next_block_key = GetNextBlockKey();
     while (g_state.pending_ticks < g_state.downcount)
     {
-      if (HasPendingInterrupt())
-      {
-        SafeReadInstruction(g_state.regs.pc, &g_state.next_instruction.bits);
-        DispatchInterrupt();
-        next_block_key = GetNextBlockKey();
-      }
-
       CodeBlock* block = LookupBlock(next_block_key);
       if (!block)
       {
@@ -153,6 +152,7 @@ static void ExecuteImpl()
       }
 
     reexecute_block:
+      Assert(!(HasPendingInterrupt()));
 
 #if 0
       const u32 tick = TimingEvents::GetGlobalTickCounter() + CPU::GetPendingTicks();
@@ -171,7 +171,7 @@ static void ExecuteImpl()
 
       if (g_state.pending_ticks >= g_state.downcount)
         break;
-      else if (HasPendingInterrupt() || !USE_BLOCK_LINKING)
+      else if (!USE_BLOCK_LINKING)
         continue;
 
       next_block_key = GetNextBlockKey();
@@ -242,10 +242,6 @@ void ExecuteRecompiler()
 {
   g_state.frame_done = false;
   while (!g_state.frame_done)
-  {
-    TimingEvents::UpdateCPUDowncount();
-
-    while (g_state.pending_ticks < g_state.downcount)
   {
     if (HasPendingInterrupt())
     {
@@ -253,6 +249,10 @@ void ExecuteRecompiler()
       DispatchInterrupt();
     }
 
+    TimingEvents::UpdateCPUDowncount();
+
+    while (g_state.pending_ticks < g_state.downcount)
+    {
       const u32 pc = g_state.regs.pc;
       g_state.current_instruction_pc = pc;
       const u32 fast_map_index = GetFastMapIndex(pc);

src/core/cpu_core.cpp

@@ -229,7 +229,16 @@ void RaiseException(u32 CAUSE_bits, u32 EPC)
 void SetExternalInterrupt(u8 bit)
 {
   g_state.cop0_regs.cause.Ip |= static_cast<u8>(1u << bit);
+
+  if (g_settings.cpu_execution_mode == CPUExecutionMode::Interpreter)
+  {
     g_state.interrupt_delay = 1;
+  }
+  else
+  {
+    g_state.interrupt_delay = 0;
+    CheckForPendingInterrupt();
+  }
 }
 
 void ClearExternalInterrupt(u8 bit)
@@ -395,6 +404,7 @@ ALWAYS_INLINE_RELEASE static void WriteCop0Reg(Cop0Reg reg, u32 value)
       g_state.cop0_regs.sr.bits =
         (g_state.cop0_regs.sr.bits & ~Cop0Registers::SR::WRITE_MASK) | (value & Cop0Registers::SR::WRITE_MASK);
       Log_DebugPrintf("COP0 SR <- %08X (now %08X)", value, g_state.cop0_regs.sr.bits);
+      CheckForPendingInterrupt();
     }
     break;
 
@@ -403,6 +413,7 @@ ALWAYS_INLINE_RELEASE static void WriteCop0Reg(Cop0Reg reg, u32 value)
       g_state.cop0_regs.cause.bits =
         (g_state.cop0_regs.cause.bits & ~Cop0Registers::CAUSE::WRITE_MASK) | (value & Cop0Registers::CAUSE::WRITE_MASK);
       Log_DebugPrintf("COP0 CAUSE <- %08X (now %08X)", value, g_state.cop0_regs.cause.bits);
+      CheckForPendingInterrupt();
     }
     break;
 
@@ -1216,6 +1227,7 @@ restart_instruction:
             // restore mode
             g_state.cop0_regs.sr.mode_bits =
               (g_state.cop0_regs.sr.mode_bits & UINT32_C(0b110000)) | (g_state.cop0_regs.sr.mode_bits >> 2);
+            CheckForPendingInterrupt();
           }
           break;
 
@@ -1365,6 +1377,20 @@ restart_instruction:
   }
 }
 
+void DispatchInterrupt()
+{
+  // If the instruction we're about to execute is a GTE instruction, delay dispatching the interrupt until the next
+  // instruction. For some reason, if we don't do this, we end up with incorrectly sorted polygons and flickering..
+  if (g_state.next_instruction.op == InstructionOp::cop2 && !g_state.next_instruction.cop.IsCommonInstruction())
+    GTE::ExecuteInstruction(g_state.next_instruction.bits);
+
+  // Interrupt raising occurs before the start of the instruction.
+  RaiseException(
+    Cop0Registers::CAUSE::MakeValueForException(Exception::INT, g_state.next_instruction_is_branch_delay_slot,
+                                                g_state.branch_was_taken, g_state.next_instruction.cop.cop_n),
+    g_state.regs.pc);
+}
+
 template<PGXPMode pgxp_mode>
 static void ExecuteImpl()
 {
@@ -1375,9 +1401,10 @@ static void ExecuteImpl()
 
     while (g_state.pending_ticks < g_state.downcount)
     {
-      if (HasPendingInterrupt())
+      if (HasPendingInterrupt() && !g_state.interrupt_delay)
         DispatchInterrupt();
 
+      g_state.interrupt_delay = false;
       g_state.pending_ticks++;
 
       // now executing the instruction we previously fetched

src/core/cpu_core_private.h

@@ -8,33 +8,20 @@ namespace CPU {
 void RaiseException(Exception excode);
 void RaiseException(u32 CAUSE_bits, u32 EPC);
 
-ALWAYS_INLINE static bool HasPendingInterrupt()
+ALWAYS_INLINE bool HasPendingInterrupt()
 {
-  // const bool do_interrupt = g_state.m_cop0_regs.sr.IEc && ((g_state.m_cop0_regs.cause.Ip & g_state.m_cop0_regs.sr.Im)
+  return g_state.cop0_regs.sr.IEc &&
-  // != 0);
-  const bool do_interrupt = g_state.cop0_regs.sr.IEc &&
                             (((g_state.cop0_regs.cause.bits & g_state.cop0_regs.sr.bits) & (UINT32_C(0xFF) << 8)) != 0);
-
-  const bool interrupt_delay = g_state.interrupt_delay;
-  g_state.interrupt_delay = false;
-
-  return do_interrupt && !interrupt_delay;
 }
 
-ALWAYS_INLINE static void DispatchInterrupt()
+ALWAYS_INLINE void CheckForPendingInterrupt()
 {
-  // If the instruction we're about to execute is a GTE instruction, delay dispatching the interrupt until the next
+  if (HasPendingInterrupt())
-  // instruction. For some reason, if we don't do this, we end up with incorrectly sorted polygons and flickering..
+    g_state.downcount = 0;
-  if (g_state.next_instruction.IsCop2Instruction())
-    return;
-
-  // Interrupt raising occurs before the start of the instruction.
-  RaiseException(
-    Cop0Registers::CAUSE::MakeValueForException(Exception::INT, g_state.next_instruction_is_branch_delay_slot,
-                                                g_state.branch_was_taken, g_state.next_instruction.cop.cop_n),
-    g_state.regs.pc);
 }
 
+void DispatchInterrupt();
+
 // icache stuff
 ALWAYS_INLINE bool IsCachedAddress(VirtualMemoryAddress address)
 {

src/core/cpu_recompiler_code_generator.cpp

@@ -1930,21 +1930,8 @@ bool CodeGenerator::Compile_cop0(const CodeBlockInstruction& cbi)
           EmitBranchIfBitClear(sr_value.host_reg, sr_value.size, 0, &no_interrupt);
           EmitAnd(sr_value.host_reg, sr_value.host_reg, cause_value);
           EmitTest(sr_value.host_reg, Value::FromConstantU32(0xFF00));
-          sr_value.ReleaseAndClear();
-          cause_value.ReleaseAndClear();
           EmitConditionalBranch(Condition::Zero, false, &no_interrupt);
-
+          EmitStoreCPUStructField(offsetof(State, downcount), Value::FromConstantU32(0));
-          EmitBranch(GetCurrentFarCodePointer());
-          SwitchToFarCode();
-
-          // we want to flush pc here
-          m_register_cache.PushState();
-          m_register_cache.FlushAllGuestRegisters(false, true);
-          WriteNewPC(CalculatePC(), false);
-          EmitExceptionExit();
-          m_register_cache.PopState();
-
-          SwitchToNearCode();
           EmitBindLabel(&no_interrupt);
         }
 
@@ -1979,6 +1966,16 @@ bool CodeGenerator::Compile_cop0(const CodeBlockInstruction& cbi)
 
         EmitStoreCPUStructField(offsetof(State, cop0_regs.sr.bits), sr);
 
+        Value cause_value = m_register_cache.AllocateScratch(RegSize_32);
+        EmitLoadCPUStructField(cause_value.host_reg, cause_value.size, offsetof(State, cop0_regs.cause.bits));
+
+        LabelType no_interrupt;
+        EmitAnd(sr.host_reg, sr.host_reg, cause_value);
+        EmitTest(sr.host_reg, Value::FromConstantU32(0xFF00));
+        EmitConditionalBranch(Condition::Zero, false, &no_interrupt);
+        EmitStoreCPUStructField(offsetof(State, downcount), Value::FromConstantU32(0));
+        EmitBindLabel(&no_interrupt);
+
         InstructionEpilogue(cbi);
         return true;
       }

src/core/timing_event.cpp

@@ -3,6 +3,7 @@
 #include "common/log.h"
 #include "common/state_wrapper.h"
 #include "cpu_core.h"
+#include "cpu_core_private.h"
 #include "system.h"
 Log_SetChannel(TimingEvents);
 
@@ -49,8 +50,11 @@ std::unique_ptr<TimingEvent> CreateTimingEvent(std::string name, TickCount perio
 
 void UpdateCPUDowncount()
 {
-  if (!CPU::g_state.frame_done)
+  if (!CPU::g_state.frame_done &&
+      (!CPU::HasPendingInterrupt() || g_settings.cpu_execution_mode == CPUExecutionMode::Interpreter))
+  {
     CPU::g_state.downcount = s_active_events_head->GetDowncount();
+  }
 }
 
 static void SortEvent(TimingEvent* event)