diff --git a/core/hw/sh4/dyna/blockmanager.cpp b/core/hw/sh4/dyna/blockmanager.cpp
index 33e9c8940..9f147d0ba 100644
--- a/core/hw/sh4/dyna/blockmanager.cpp
+++ b/core/hw/sh4/dyna/blockmanager.cpp
@@ -296,8 +296,8 @@ void bm_ResetCache()
 	{
 		RuntimeBlockInfoPtr block = it.second;
 		block->relink_data = 0;
-		block->pNextBlock = 0;
-		block->pBranchBlock = 0;
+		block->pNextBlock = NULL;
+		block->pBranchBlock = NULL;
 		// needed for the transition to full mmu. Could perhaps limit it to the current block.
 		block->Relink();
 		// Avoid circular references
@@ -312,7 +312,6 @@ void bm_ResetCache()
 	for (auto& block_list : blocks_per_page)
 		block_list.clear();
 
-	// FIXME Grandia II doesn't like it. intermittent reset when pressing start (except if disabling SSA?? TBC)
 	memset(unprotected_pages, 0, sizeof(unprotected_pages));
 
 #ifdef DYNA_OPROF
@@ -516,7 +515,7 @@ void RuntimeBlockInfo::AddRef(RuntimeBlockInfoPtr other)
 }
 
 void RuntimeBlockInfo::RemRef(RuntimeBlockInfoPtr other)
-{ 
+{
 	bm_List::iterator it = std::find(pre_refs.begin(), pre_refs.end(), other);
 	if (it != pre_refs.end())
 		pre_refs.erase(it);
diff --git a/core/hw/sh4/sh4_core.h b/core/hw/sh4/sh4_core.h
index f97ae5d89..a2bd09d4e 100644
--- a/core/hw/sh4/sh4_core.h
+++ b/core/hw/sh4/sh4_core.h
@@ -94,11 +94,6 @@ static INLINE void SetXD(u32 n,f64 val)
 	xf[(n<<1) | 1]=t.sgl[0];
 	xf[(n<<1) | 0]=t.sgl[1];
 }
-//needs to be removed
-u32* Sh4_int_GetRegisterPtr(Sh4RegType reg);
-//needs to be made portable
-void SetFloatStatusReg();
-
 
 bool Do_Interrupt(u32 intEvn);
 bool Do_Exception(u32 epc, u32 expEvn, u32 CallVect);
@@ -155,9 +150,10 @@ static INLINE f64 fixNaN64(f64 f)
 
 // Reduces the precision of the argument f by a given number of bits
 // double have 53 bits of precision so the returned result will have a precision of 53 - bits
+// Note: with -ffast-math c -(c - f) is simplified to ... f, which makes this function a nop
 template<int bits>
 static INLINE double reduce_precision(double f)
 {
-	double c = (double)((1 << bits) + 1) * f;
+	double c = (double)((1ull << bits) + 1) * f;
 	return c - (c - f);
 }
diff --git a/core/hw/sh4/sh4_core_regs.cpp b/core/hw/sh4/sh4_core_regs.cpp
index 0e7c230d9..c583fa7e6 100644
--- a/core/hw/sh4/sh4_core_regs.cpp
+++ b/core/hw/sh4/sh4_core_regs.cpp
@@ -61,7 +61,7 @@ bool UpdateSR()
 u32 old_rm=0xFF;
 u32 old_dn=0xFF;
 
-void SetFloatStatusReg()
+static void SetFloatStatusReg()
 {
 	if ((old_rm!=fpscr.RM) || (old_dn!=fpscr.DN))
 	{
@@ -150,7 +150,7 @@ void UpdateFPSCR()
 }
 
 
-u32* Sh4_int_GetRegisterPtr(Sh4RegType reg)
+static u32* Sh4_int_GetRegisterPtr(Sh4RegType reg)
 {
 	if ((reg>=reg_r0) && (reg<=reg_r15))
 	{
@@ -248,19 +248,6 @@ u32* Sh4_int_GetRegisterPtr(Sh4RegType reg)
 	}
 }
 
-u32 sh4context_offset_u32(u32 sh4_reg)
-{
-	void* addr=Sh4_int_GetRegisterPtr((Sh4RegType)sh4_reg);
-	u32 offs=(u8*)addr-(u8*)&Sh4cntx;
-	verify(offs<sizeof(Sh4cntx));
-
-	return offs;
-}
-u32 sh4context_offset_regtype(Sh4RegType sh4_reg)
-{
-	return sh4context_offset_u32(sh4_reg);
-}
-
 u32* GetRegPtr(u32 reg)
 {
 	return Sh4_int_GetRegisterPtr((Sh4RegType)reg);
diff --git a/core/hw/sh4/sh4_if.h b/core/hw/sh4/sh4_if.h
index e6b739810..9487df52e 100644
--- a/core/hw/sh4/sh4_if.h
+++ b/core/hw/sh4/sh4_if.h
@@ -280,7 +280,7 @@ struct Sh4Context
 			sr_status_t old_sr;
 			fpscr_t old_fpscr;
 
-			volatile u32 CpuRunning;
+			u32 CpuRunning;
 
 			int sh4_sched_next;
 			u32 interrupt_pend;
@@ -292,9 +292,6 @@ struct Sh4Context
 
 };
 
-u32 sh4context_offset_u32(u32 sh4_reg);
-u32 sh4context_offset_regtype(Sh4RegType sh4_reg);
-
 void DYNACALL do_sqw_mmu(u32 dst);
 extern "C" void DYNACALL do_sqw_nommu_area_3(u32 dst, u8* sqb);
 extern "C" void DYNACALL do_sqw_nommu_area_3_nonvmem(u32 dst, u8* sqb);
diff --git a/core/hw/sh4/sh4_interpreter.h b/core/hw/sh4/sh4_interpreter.h
index 640345790..75096ce93 100644
--- a/core/hw/sh4/sh4_interpreter.h
+++ b/core/hw/sh4/sh4_interpreter.h
@@ -54,7 +54,7 @@ void Sh4_int_SetRegister(Sh4RegType reg,u32 regdata);
 void ExecuteDelayslot();
 void ExecuteDelayslot_RTE();
 
-#define SH4_TIMESLICE (448)	// at 112 Bangai-O doesn't start. 224 is ok
+#define SH4_TIMESLICE 448	// at 112 Bangai-O doesn't start. 224 is ok
 							// at 448 Gundam Side Story hangs on Sega copyright screen, 224 ok, 672 ok(!)
 
 extern "C" {
diff --git a/core/rec-ARM64/rec_arm64.cpp b/core/rec-ARM64/rec_arm64.cpp
index ce18a9f8a..10fd715e0 100644
--- a/core/rec-ARM64/rec_arm64.cpp
+++ b/core/rec-ARM64/rec_arm64.cpp
@@ -71,7 +71,7 @@ static jmp_buf jmp_env;
 static u32 cycle_counter;
 
 static void (*mainloop)(void *context);
-static void (*arm64_intc_sched)();
+static int (*arm64_intc_sched)();
 static void (*arm64_no_update)();
 
 #ifdef PROFILING
@@ -114,8 +114,7 @@ __asm__
 		".hidden ngen_LinkBlock_Shared_stub	\n\t"
 		".globl ngen_LinkBlock_Shared_stub	\n\t"
 	"ngen_LinkBlock_Shared_stub:			\n\t"
-		"mov x0, lr							\n\t"
-		"sub x0, x0, #4						\n\t"	// go before the call
+		"sub x0, lr, #4						\n\t"	// go before the call
 		"bl rdv_LinkBlock					\n\t"   // returns an RX addr
 		"br x0								\n"
 
@@ -403,6 +402,12 @@ public:
 		Label cycles_remaining;
 		B(&cycles_remaining, pl);
 		GenCall(*arm64_intc_sched);
+		Label cpu_running;
+		Cbnz(w0, &cpu_running);
+		Mov(w29, block->vaddr);
+		Str(w29, sh4_context_mem_operand(&next_pc));
+		GenBranch(*arm64_no_update);
+		Bind(&cpu_running);
 		Bind(&cycles_remaining);
 
 #ifdef PROFILING
@@ -1430,8 +1435,8 @@ public:
 		Label intc_sched;
 		Label end_mainloop;
 
-		// void intc_sched()
-		arm64_intc_sched = GetCursorAddress<void (*)()>();
+		// int intc_sched()
+		arm64_intc_sched = GetCursorAddress<int (*)()>();
 		B(&intc_sched);
 
 		// void no_update()
@@ -1518,6 +1523,7 @@ public:
 		GenCallRuntime(UpdateSystem);
 		Mov(lr, x29);
 		Cbnz(w0, &do_interrupts);
+		Ldr(w0, MemOperand(x28, offsetof(Sh4Context, CpuRunning)));
 		Ret();
 
 		Bind(&do_interrupts);