CPU/Recompiler: Don't use far code for mtc0 cache check

Redundant for a few instructions.

src/core/cpu_recompiler_arm32.cpp

@@ -2265,15 +2265,19 @@ void CPU::ARM32Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
-    SwitchToFarCodeIfBitSet(changed_bits, 16);
+    Label caches_unchanged;
+    armAsm->tst(changed_bits, 1u << 16);
+    armAsm->b(eq, &caches_unchanged);
     EmitCall(reinterpret_cast<const void*>(&CPU::UpdateMemoryPointers));
     armAsm->ldr(RARG1, PTR(ptr)); // reload value for interrupt test below
+    armAsm->bind(&caches_unchanged);
+
+    // might need to reload fastmem base too
     if (CodeCache::IsUsingFastmem() && m_block->HasFlag(CodeCache::BlockFlags::ContainsLoadStoreInstructions) &&
         IsHostRegAllocated(RMEMBASE.GetCode()))
     {
       FreeHostReg(RMEMBASE.GetCode());
     }
-    SwitchToNearCode(true);
 
     TestInterrupts(RARG1);
   }

src/core/cpu_recompiler_arm64.cpp

@@ -2428,11 +2428,13 @@ void CPU::ARM64Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
-    SwitchToFarCodeIfBitSet(changed_bits, 16);
+    Label caches_unchanged;
+    armAsm->tbz(changed_bits, 16, &caches_unchanged);
     EmitCall(reinterpret_cast<const void*>(&CPU::UpdateMemoryPointers));
     armAsm->ldr(RWARG1, PTR(ptr)); // reload value for interrupt test below
+    if (CodeCache::IsUsingFastmem())
       armAsm->ldr(RMEMBASE, PTR(&g_state.fastmem_base));
-    SwitchToNearCode(true);
+    armAsm->bind(&caches_unchanged);
 
     TestInterrupts(RWARG1);
   }

src/core/cpu_recompiler_riscv64.cpp

@@ -2283,13 +2283,15 @@ void CPU::RISCV64Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
+    Label caches_unchanged;
     rvAsm->SRLIW(RSCRATCH, changed_bits, 16);
     rvAsm->ANDI(RSCRATCH, RSCRATCH, 1);
-    SwitchToFarCode(true, &Assembler::BEQ, RSCRATCH, zero);
+    rvAsm->BEQ(RSCRATCH, zero, &caches_unchanged);
     EmitCall(reinterpret_cast<const void*>(&CPU::UpdateMemoryPointers));
     rvAsm->LW(new_value, PTR(ptr));
+    if (CodeCache::IsUsingFastmem())
       rvAsm->LD(RMEMBASE, PTR(&g_state.fastmem_base));
-    SwitchToNearCode(true);
+    rvAsm->Bind(&caches_unchanged);
 
     TestInterrupts(RARG1);
   }

src/core/cpu_recompiler_x64.cpp

@@ -2241,12 +2241,15 @@ void CPU::X64Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
+    Label caches_unchanged;
     cg->test(changed_bits, 1u << 16);
-    SwitchToFarCode(true, &CodeGenerator::jnz);
+    cg->jz(caches_unchanged);
     cg->call(&CPU::UpdateMemoryPointers);
     cg->mov(RWARG2, cg->dword[PTR(ptr)]); // reload value for interrupt test below
+    if (CodeCache::IsUsingFastmem())
       cg->mov(RMEMBASE, cg->qword[PTR(&g_state.fastmem_base)]);
-    SwitchToNearCode(true);
+
+    cg->L(caches_unchanged);
 
     TestInterrupts(RWARG2);
   }