SPU: Use bitset to reduce memory for compiler instance

rpcs3/Emu/Cell/SPUCommonRecompiler.cpp

@@ -2482,9 +2482,9 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
 	workload.push_back(entry_point);
 
 	std::memset(m_regmod.data(), 0xff, sizeof(m_regmod));
-	std::memset(m_use_ra.data(), 0xff, sizeof(m_use_ra));
-	std::memset(m_use_rb.data(), 0xff, sizeof(m_use_rb));
-	std::memset(m_use_rc.data(), 0xff, sizeof(m_use_rc));
+	m_use_ra.reset();
+	m_use_rb.reset();
+	m_use_rc.reset();
 	m_targets.clear();
 	m_preds.clear();
 	m_preds[entry_point];
@@ -2579,11 +2579,11 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
 		if (auto iflags = g_spu_iflag.decode(data))
 		{
 			if (+iflags & +spu_iflag::use_ra)
-				m_use_ra[pos / 4] = op.ra;
+				m_use_ra.set(pos / 4);
 			if (+iflags & +spu_iflag::use_rb)
-				m_use_rb[pos / 4] = op.rb;
+				m_use_rb.set(pos / 4);
 			if (+iflags & +spu_iflag::use_rc)
-				m_use_rc[pos / 4] = op.rc;
+				m_use_rc.set(pos / 4);
 		}
 
 		// Analyse instruction
@@ -3010,11 +3010,6 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
 				m_regmod[pos / 4] = s_reg_mfc_size;
 				break;
 			}
-			case MFC_Cmd:
-			{
-				m_use_rb[pos / 4] = s_reg_mfc_eal;
-				break;
-			}
 			default: break;
 			}
 
@@ -3461,10 +3456,13 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
 				reg_save = op.rt;
 			}
 
-			for (auto* _use : {&m_use_ra, &m_use_rb, &m_use_rc})
+			for (auto _use : std::initializer_list<std::pair<u32, bool>>{{op.ra, m_use_ra.test(ia / 4)}
+				, {op.rb, m_use_rb.test(ia / 4)}, {op.rc, m_use_rc.test(ia / 4)}})
 			{
-				if (u8 reg = (*_use)[ia / 4]; reg < s_reg_max)
+				if (_use.second)
 				{
+					const u32 reg = _use.first;
+
 					// Register reg use only if it happens before reg mod
 					if (!block.reg_mod[reg])
 					{
@@ -3479,7 +3477,7 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
 				}
 			}
 
-			if (m_use_rb[ia / 4] == s_reg_mfc_eal)
+			if (type == spu_itype::WRCH && op.ra == MFC_Cmd)
 			{
 				// Expand MFC_Cmd reg use
 				for (u8 reg : {s_reg_mfc_lsa, s_reg_mfc_tag, s_reg_mfc_size})

rpcs3/Emu/Cell/SPURecompiler.h

@@ -202,9 +202,9 @@ class spu_recompiler_base
 	// GPR modified by the instruction (-1 = not set)
 	std::array<u8, 0x10000> m_regmod;
 
-	std::array<u8, 0x10000> m_use_ra;
-	std::array<u8, 0x10000> m_use_rb;
-	std::array<u8, 0x10000> m_use_rc;
+	std::bitset<0x10000> m_use_ra;
+	std::bitset<0x10000> m_use_rb;
+	std::bitset<0x10000> m_use_rc;
 
 	// List of possible targets for the instruction (entry shouldn't exist for simple instructions)
 	std::unordered_map<u32, std::basic_string<u32>, value_hash<u32, 2>> m_targets;

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1839,7 +1839,10 @@ void spu_thread::cpu_work()
 
 	if (has_active_local_bps)
 	{
-		if (local_breakpoints[pc / 4])
+		const u32 pos_at = pc / 4;
+		const u32 pos_bit = 1u << (pos_at % 8);
+
+		if (local_breakpoints[pos_at] & pos_bit)
 		{
 			// Ignore repeatations until a different instruction is issued
 			if (pc != current_bp_pc)

rpcs3/Emu/Cell/SPUThread.h

@@ -816,7 +816,7 @@ class spu_thread : public cpu_thread
 	atomic_t<spu_debugger_mode> debugger_mode{};
 
 	// PC-based breakpoint list
-	std::array<atomic_t<bool>, SPU_LS_SIZE / 4> local_breakpoints{};
+	std::array<atomic_t<u8>, SPU_LS_SIZE / 4 / 8> local_breakpoints{};
 	atomic_t<bool> has_active_local_bps = false;
 	u32 current_bp_pc = umax;
 	bool stop_flag_removal_protection = false;

rpcs3/rpcs3qt/breakpoint_list.cpp

@@ -125,8 +125,10 @@ void breakpoint_list::HandleBreakpointRequest(u32 loc, bool only_add)
 
 		const auto spu = static_cast<spu_thread*>(m_cpu);
 		auto& list = spu->local_breakpoints;
+		const u32 pos_at = loc / 4;
+		const u32 pos_bit = 1u << (pos_at % 8);
 
-		if (list[loc / 4].test_and_invert())
+		if (list[pos_at / 8].fetch_xor(pos_bit) & pos_bit)
 		{
 			if (std::none_of(list.begin(), list.end(), [](auto& val){ return val.load(); }))
 			{

rpcs3/rpcs3qt/debugger_list.cpp

@@ -142,8 +142,17 @@ void debugger_list::ShowAddress(u32 addr, bool select_addr, bool direct)
 	{
 		switch (m_cpu ? m_cpu->id_type() : 0)
 		{
-		case 1: return m_ppu_breakpoint_handler->HasBreakpoint(pc);
-		case 2: return (*spu_bps_list)[pc / 4].load();
+		case 1:
+		{
+			return m_ppu_breakpoint_handler->HasBreakpoint(pc);
+		}
+		case 2:
+		{
+			const u32 pos_at = pc / 4;
+			const u32 pos_bit = 1u << (pos_at % 8);
+
+			return !!((*spu_bps_list)[pos_at] & pos_bit);
+		}
 		default: return false;
 		}
 	};