Update PPUThread.cpp

Author: elad335

rpcs3/Emu/Cell/PPUAnalyser.cpp

@@ -2096,6 +2096,16 @@ bool ppu_module::analyse(u32 lib_toc, u32 entry, const u32 sec_end, const std::b
 
 					if (itype != ppu_itype::BC && itype != ppu_itype::B)
 					{
+						if (i == fend - 4)
+						{
+							if (!(itype & ppu_itype::branch) && itype != ppu_itype::SC)
+							{
+								// Inserts a branch to following code
+								fail = true;
+								break;
+							}
+						}
+
 						continue;
 					}
 
@@ -2106,6 +2116,16 @@ bool ppu_module::analyse(u32 lib_toc, u32 entry, const u32 sec_end, const std::b
 						fail = true;
 						break;
 					}
+
+					if (itype == ppu_itype::BC && (op.bo & 0x14) != 0x14)
+					{
+						if (i == fend - 4)
+						{
+							// Can branch to next
+							fail = true;
+							break;
+						}
+					}
 				}
 
 				if (fail)
@@ -2141,11 +2161,16 @@ bool ppu_module::analyse(u32 lib_toc, u32 entry, const u32 sec_end, const std::b
 
 		if (data.first > 1)
 		{
-			dups_count++;
+			duplicate_map[func.addr] = data.second;
+
+			for (const auto& [addr, size] : func.blocks)
+			{
+				duplicate_map[addr] = data.second + (addr - func.addr);
+			}
 
 			if (func.addr != data.second)
 			{
-				duplicate_map[func.addr] = data.second;
+				dups_count++;
 			}
 
 			ppu_log.trace("Found PPU function duplicate: func 0x%x (%d times)", data.second, data.first);

rpcs3/Emu/Cell/PPUAnalyser.h

@@ -108,6 +108,7 @@ struct ppu_module
 		secs = info.secs;
 		allocations = info.allocations;
 		addr_to_seg_index = info.addr_to_seg_index;
+		duplicate_map = info.duplicate_map;
 	}
 
 	bool analyse(u32 lib_toc, u32 entry, u32 end, const std::basic_string<u32>& applied, std::function<bool()> check_aborted = {});

rpcs3/Emu/Cell/PPUThread.cpp

@@ -4643,8 +4643,15 @@ bool ppu_initialize(const ppu_module& info, bool check_only, u64 file_size)
 			// Copy block or function entry
 			ppu_function& entry = part.funcs.emplace_back(func);
 
+			u32 og_func = entry.addr;
+
+			if (auto it = info.duplicate_map.find(entry.addr); it != info.duplicate_map.end())
+			{
+				og_func = it->second;
+			}
+
 			// Fixup some information
-			entry.name = fmt::format("__0x%x", entry.addr - reloc);
+			entry.name = fmt::format("__0x%x", og_func - reloc);
 
 			if (has_mfvscr && g_cfg.core.ppu_set_sat_bit)
 			{
@@ -5037,6 +5044,8 @@ bool ppu_initialize(const ppu_module& info, bool check_only, u64 file_size)
 
 	bool early_exit = false;
 
+	std::map<std::string, ppu_intrp_func_t> func_ptr_map;
+
 	// Get and install function addresses
 	for (const auto& func : info.funcs)
 	{
@@ -5063,10 +5072,14 @@ bool ppu_initialize(const ppu_module& info, bool check_only, u64 file_size)
 
 		const auto name = fmt::format("__0x%x", og_func - reloc);
 
+		ppu_intrp_func_t dummy{};
+		ppu_intrp_func_t& func_ptr = is_first ? func_ptr_map[name] : dummy;
+
 		// Try to locate existing function if it is not the first time
-		const auto addr = is_first ? ensure(reinterpret_cast<ppu_intrp_func_t>(jit->get(name)))
+		const auto addr = is_first ? (func_ptr ? func_ptr : ensure(reinterpret_cast<ppu_intrp_func_t>(jit->get(name))))
 			: reinterpret_cast<ppu_intrp_func_t>(ensure(jit_mod.funcs[index]));
 
+		func_ptr = addr;
 		jit_mod.funcs.emplace_back(addr);
 
 		if (func.size == 4 && !BLR_func && *info.get_ptr<u32>(func.addr) == ppu_instructions::BLR())
@@ -5155,6 +5168,11 @@ static void ppu_initialize2(jit_compiler& jit, const ppu_module& module_part, co
 	{
 		if (func.size)
 		{
+			if (auto it = module_part.duplicate_map.find(func.addr); it != module_part.duplicate_map.end() && it->second != it->first)
+			{
+				continue;
+			}
+
 			const auto f = cast<Function>(_module->getOrInsertFunction(func.name, _func).getCallee());
 			f->setCallingConv(CallingConv::GHC);
 			f->addParamAttr(1, llvm::Attribute::NoAlias);

rpcs3/Emu/Cell/PPUTranslator.cpp

@@ -148,6 +148,7 @@ Function* PPUTranslator::Translate(const ppu_function& info)
 	const u64 base = m_reloc ? m_reloc->addr : 0;
 	m_addr = info.addr - base;
 	m_attr = info.attr;
+	m_func_base = m_addr;
 
 	// Don't emit check in small blocks without terminator
 	bool need_check = info.size >= 16;
@@ -304,13 +305,30 @@ Value* PPUTranslator::VecHandleResult(Value* val)
 
 Value* PPUTranslator::GetAddr(u64 _add)
 {
+	const bool is_duplicate = m_info.duplicate_map.contains(m_func_base);
+	const auto old_cia = std::exchange(m_cia, nullptr);
+	const auto cia_reg = is_duplicate ? ZExt(RegLoad(m_cia)) : nullptr;
+	const u32 inst_diff = is_duplicate ? m_addr - m_func_base : m_addr;
+
+	m_cia = old_cia;
+	Value* cia = nullptr;
+
 	if (m_reloc)
 	{
 		// Load segment address from global variable, compute actual instruction address
-		return m_ir->CreateAdd(m_ir->getInt64(m_addr + _add), m_seg0);
+		cia = m_ir->CreateAdd(m_ir->getInt64(inst_diff + _add), m_seg0);
+	}
+	else
+	{
+		cia = m_ir->getInt64(inst_diff + _add);
+	}
+
+	if (cia_reg)
+	{
+		cia = m_ir->CreateAdd(cia_reg, cia);
 	}
 
-	return m_ir->getInt64(m_addr + _add);
+	return cia;
 }
 
 Type* PPUTranslator::ScaleType(Type* type, s32 pow2)

rpcs3/Emu/Cell/PPUTranslator.h

@@ -28,6 +28,9 @@ class PPUTranslator final : public cpu_translator
 	// Current position-independent address
 	u64 m_addr = 0;
 
+	// Function start
+	u64 m_func_base = 0;
+
 	// Function attributes
 	bs_t<ppu_attr> m_attr{};