CPU/Recompiler: Check downcount in interpret block

Fixes application locking up with CDS=Max in Driver.

src/core/cpu_recompiler_arm32.cpp

@@ -259,6 +259,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
 #endif
 
   Label dispatch;
+  Label run_events_and_dispatch;
 
   g_enter_recompiler = armAsm->GetCursorAddress<decltype(g_enter_recompiler)>();
   {
@@ -276,6 +277,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
     armAsm->b(lt, &skip_event_check);
 
     g_run_events_and_dispatch = armAsm->GetCursorAddress<const void*>();
+    armAsm->bind(&run_events_and_dispatch);
     armEmitCall(armAsm, reinterpret_cast<const void*>(&TimingEvents::RunEvents), true);
 
     armAsm->bind(&skip_event_check);
@@ -315,6 +317,10 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
   g_interpret_block = armAsm->GetCursorAddress<const void*>();
   {
     armEmitCall(armAsm, reinterpret_cast<const void*>(GetInterpretUncachedBlockFunction()), true);
+    armAsm->ldr(RARG1, PTR(&g_state.pending_ticks));
+    armAsm->ldr(RARG2, PTR(&g_state.downcount));
+    armAsm->cmp(RARG1, RARG2);
+    armAsm->b(ge, &run_events_and_dispatch);
     armAsm->b(&dispatch);
   }
 

src/core/cpu_recompiler_arm64.cpp

@@ -453,6 +453,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
 #endif
 
   Label dispatch;
+  Label run_events_and_dispatch;
 
   g_enter_recompiler = armAsm->GetCursorAddress<decltype(g_enter_recompiler)>();
   {
@@ -476,6 +477,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
     armAsm->b(&dispatch, lt);
 
     g_run_events_and_dispatch = armAsm->GetCursorAddress<const void*>();
+    armAsm->bind(&run_events_and_dispatch);
     armEmitCall(armAsm, reinterpret_cast<const void*>(&TimingEvents::RunEvents), true);
   }
 
@@ -516,6 +518,10 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
   g_interpret_block = armAsm->GetCursorAddress<const void*>();
   {
     armEmitCall(armAsm, reinterpret_cast<const void*>(GetInterpretUncachedBlockFunction()), true);
+    armAsm->ldr(RWARG1, PTR(&g_state.pending_ticks));
+    armAsm->ldr(RWARG2, PTR(&g_state.downcount));
+    armAsm->cmp(RWARG1, RWARG2);
+    armAsm->b(&run_events_and_dispatch, ge);
     armAsm->b(&dispatch);
   }
 

src/core/cpu_recompiler_riscv64.cpp

@@ -241,6 +241,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
   Assembler* rvAsm = &actual_asm;
 
   Label dispatch;
+  Label run_events_and_dispatch;
 
   g_enter_recompiler = reinterpret_cast<decltype(g_enter_recompiler)>(rvAsm->GetCursorPointer());
   {
@@ -264,6 +265,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
     rvAsm->LW(RARG2, PTR(&g_state.downcount));
     rvAsm->BLTU(RARG1, RARG2, &skip_event_check);
 
+    rvAsm->Bind(&run_events_and_dispatch);
     g_run_events_and_dispatch = rvAsm->GetCursorPointer();
     rvEmitCall(rvAsm, reinterpret_cast<const void*>(&TimingEvents::RunEvents));
 
@@ -309,6 +311,9 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
   g_interpret_block = rvAsm->GetCursorPointer();
   {
     rvEmitCall(rvAsm, CodeCache::GetInterpretUncachedBlockFunction());
+    rvAsm->LW(RARG1, PTR(&g_state.pending_ticks));
+    rvAsm->LW(RARG2, PTR(&g_state.downcount));
+    rvAsm->BGE(RARG1, RARG2, &run_events_and_dispatch);
     rvAsm->J(&dispatch);
   }
 

src/core/cpu_recompiler_x64.cpp

@@ -118,6 +118,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
 
   Label dispatch;
   Label exit_recompiler;
+  Label run_events_and_dispatch;
 
   g_enter_recompiler = reinterpret_cast<decltype(g_enter_recompiler)>(const_cast<u8*>(cg->getCurr()));
   {
@@ -143,6 +144,7 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
     cg->jl(dispatch);
 
     g_run_events_and_dispatch = cg->getCurr();
+    cg->L(run_events_and_dispatch);
     cg->call(reinterpret_cast<const void*>(&TimingEvents::RunEvents));
   }
 
@@ -183,6 +185,9 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
   g_interpret_block = cg->getCurr();
   {
     cg->call(CodeCache::GetInterpretUncachedBlockFunction());
+    cg->mov(RWARG1, cg->dword[PTR(&g_state.pending_ticks)]);
+    cg->cmp(RWARG1, cg->dword[PTR(&g_state.downcount)]);
+    cg->jge(run_events_and_dispatch);
     cg->jmp(dispatch);
   }