diff --git a/rpcs3/Emu/ARMv7/ARMv7Context.h b/rpcs3/Emu/ARMv7/ARMv7Context.h
index c474dc6e6d..bdfa8107e2 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Context.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Context.h
@@ -10,17 +10,6 @@ enum ARMv7InstructionSet
 	ThumbEE
 };
 
-union ARMv7Code
-{
-	struct
-	{
-		u16 code0;
-		u16 code1;
-	};
-
-	u32 data;
-};
-
 struct ARMv7Context
 {
 	ARMv7Thread& thread;
diff --git a/rpcs3/Emu/ARMv7/ARMv7Decoder.cpp b/rpcs3/Emu/ARMv7/ARMv7Decoder.cpp
index c28a880722..179e461aed 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Decoder.cpp
+++ b/rpcs3/Emu/ARMv7/ARMv7Decoder.cpp
@@ -1,4 +1,5 @@
 #include "stdafx.h"
+#include <unordered_map>
 #include "Utilities/Log.h"
 #include "ARMv7Thread.h"
 #include "ARMv7Interpreter.h"
@@ -595,31 +596,164 @@ const ARMv7_opcode_t ARMv7_opcode_table[] =
 	ARMv7_OP4(0x0fff, 0xffff, 0x0320, 0xf001, A1, YIELD),
 };
 
-u8 ARMv7Decoder::DecodeMemory(const u32 address)
+struct ARMv7_op2_table_t
 {
-	ARMv7Code code;
-	code.code0 = vm::psv::read16(address & ~1);
-	code.code1 = vm::psv::read16(address + 2 & ~1);
-	u32 arg = address & 0x1 ? code.data : (u32)code.code0 << 16 | code.code1;
+	const ARMv7_opcode_t* data[0x10000];
 
-	//LOG_NOTICE(GENERAL, "code0 = 0x%04x, code1 = 0x%04x, data = 0x%08x, arg = 0x%08x", code.code0, code.code1, code.data, arg);
-
-	// old decoding algorithm
-
-	for (auto& opcode : ARMv7_opcode_table)
+	ARMv7_op2_table_t()
 	{
-		if ((opcode.type < A1) == ((address & 0x1) == 0) && (arg & opcode.mask) == opcode.code)
+		std::vector<const ARMv7_opcode_t*> t2;
+
+		for (auto& opcode : ARMv7_opcode_table)
 		{
-			code.data = opcode.length == 2 ? code.code0 : arg;
-			(*opcode.func)(m_thr.context, code, opcode.type);
-			//LOG_NOTICE(ARMv7, "%s, %s", opcode.name, m_thr.RegsToString());
-			return opcode.length;
+			if (opcode.length == 2)
+			{
+				t2.push_back(&opcode);
+			}
+		}
+
+		for (u32 i = 0; i < 0x10000; i++)
+		{
+			data[i] = nullptr;
+
+			for (auto& opcode : t2)
+			{
+				if (((i << 16) & opcode->mask) == opcode->code)
+				{
+					data[i] = opcode;
+					break;
+				}
+			}
 		}
 	}
 
-	ARMv7_instrs::UNK(m_thr.context, code);
-	return address & 0x1 ? 4 : 2;
+} g_op2t;
 
+struct ARMv7_op4t_table_t
+{
+	std::vector<const ARMv7_opcode_t*> table;
+
+	ARMv7_op4t_table_t()
+	{
+		for (auto& opcode : ARMv7_opcode_table)
+		{
+			if (opcode.length == 4 && opcode.type < A1)
+			{
+				table.push_back(&opcode);
+			}
+		}
+	}
+	
+} g_op4t;
+
+std::unordered_map<u32, const ARMv7_opcode_t*> g_opct;
+
+void armv7_decoder_initialize(u32 addr, u32 end_addr, bool dump)
+{
+	// 1. Find every 4-byte thumb instruction and cache it
+	// 2. If some instruction is not recognized, print the error
+	// 3. Possibly print disasm
+	g_opct.clear();
+
+	while (addr < end_addr)
+	{
+		ARMv7Code code = {};
+		code.code0 = vm::psv::read16(addr);
+
+		auto found = g_op2t.data[code.code0];
+
+		if (!found)
+		{
+			code.code1 = code.code0;
+			code.code0 = vm::psv::read16(addr + 2);
+
+			auto op = g_opct.find(code.data);
+			if (op != g_opct.end())
+			{
+				found = op->second;
+			}
+		}
+
+		if (!found)
+		{
+			for (auto opcode : g_op4t.table)
+			{
+				if ((code.data & opcode->mask) == opcode->code)
+				{
+					g_opct[code.data] = (found = opcode);
+					break;
+				}
+			}
+		}
+		
+		if (!found)
+		{
+			LOG_ERROR(ARMv7, "Unknown instruction found at address 0x%08x: %04x %04x", addr, code.code1, code.code0);
+			addr += 4;
+		}
+		else
+		{
+			if (dump) if (found->length == 2)
+			{
+				LOG_NOTICE(ARMv7, "0x%08x: %04x       %s", addr, code.code0, found->name);
+			}
+			else
+			{
+				LOG_NOTICE(ARMv7, "0x%08x: %04x %04x  %s", addr, code.code1, code.code0, found->name);
+			}
+
+			addr += found->length;
+		}
+	}
+
+	while (vm::psv::read16(addr) == 0xf870)
+	{
+		g_opct[0xf8700000 | vm::psv::read16(addr + 2)] = g_op4t.table[0];
+		addr += 16;
+	}
+
+	LOG_NOTICE(ARMv7, "armv7_decoder_initialize() finished, g_opct.size() = %lld", g_opct.size());
+}
+
+u32 ARMv7Decoder::DecodeMemory(const u32 address)
+{
+	if (address & 0x1)
+	{
+		throw "ARMv7Decoder::DecodeMemory() failed (something is wrong with instruction set)";
+	}
+
+	ARMv7Code code = {};
+	code.code0 = vm::psv::read16(address);
+
+	if (auto opcode = g_op2t.data[code.code0])
+	{
+		(*opcode->func)(m_ctx, code, opcode->type);
+		return 2;
+	}
+
+	code.code1 = code.code0;
+	code.code0 = vm::psv::read16(address + 2);
+
+	auto op = g_opct.find(code.data);
+	if (op != g_opct.end())
+	{
+		(*op->second->func)(m_ctx, code, op->second->type);
+		return 4;
+	}
+
+	//for (auto opcode : g_op4t.table)
+	//{
+	//	if ((code.data & opcode->mask) == opcode->code)
+	//	{
+	//		(*opcode->func)(m_ctx, code, opcode->type);
+	//		return 4;
+	//	}
+	//}
+
+	ARMv7_instrs::UNK(m_ctx, code);
+	return 2;
+
+	// "group" decoding algorithm (temporarily disabled)
 
 	//execute_main_group(&m_thr);
 	//// LOG_NOTICE(GENERAL, "%s, %d \n\n", m_thr.m_last_instr_name, m_thr.m_last_instr_size);
diff --git a/rpcs3/Emu/ARMv7/ARMv7Decoder.h b/rpcs3/Emu/ARMv7/ARMv7Decoder.h
index 341ed6a276..229f74014b 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Decoder.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Decoder.h
@@ -1,16 +1,18 @@
 #pragma once
 #include "Emu/CPU/CPUDecoder.h"
 
-class ARMv7Thread;
+struct ARMv7Context;
 
 class ARMv7Decoder : public CPUDecoder
 {
-	ARMv7Thread& m_thr;
+	ARMv7Context& m_ctx;
 
 public:
-	ARMv7Decoder(ARMv7Thread& thr) : m_thr(thr)
+	ARMv7Decoder(ARMv7Context& context) : m_ctx(context)
 	{
 	}
 
-	virtual u8 DecodeMemory(const u32 address);
+	virtual u32 DecodeMemory(const u32 address);
 };
+
+void armv7_decoder_initialize(u32 addr, u32 end_addr, bool dump = false);
diff --git a/rpcs3/Emu/ARMv7/ARMv7DisAsm.cpp b/rpcs3/Emu/ARMv7/ARMv7DisAsm.cpp
index ccb2582cad..6a3626522b 100644
--- a/rpcs3/Emu/ARMv7/ARMv7DisAsm.cpp
+++ b/rpcs3/Emu/ARMv7/ARMv7DisAsm.cpp
@@ -1,4 +1,5 @@
 #include "stdafx.h"
+#if 0
 #include "ARMv7DisAsm.h"
 
 void ARMv7DisAsm::UNK(const u32 data)
@@ -1119,4 +1120,4 @@ void ARMv7DisAsm::UXTH(const u32 data, const ARMv7_encoding type)
 {
 	Write(__FUNCTION__);
 }
-
+#endif
diff --git a/rpcs3/Emu/ARMv7/ARMv7DisAsm.h b/rpcs3/Emu/ARMv7/ARMv7DisAsm.h
index 05375dc167..848baaf7cb 100644
--- a/rpcs3/Emu/ARMv7/ARMv7DisAsm.h
+++ b/rpcs3/Emu/ARMv7/ARMv7DisAsm.h
@@ -1,5 +1,4 @@
 #pragma once
-#include "Emu/ARMv7/ARMv7Opcodes.h"
 #include "Emu/CPU/CPUDisAsm.h"
 
 static const char* g_arm_cond_name[16] =
@@ -10,6 +9,14 @@ static const char* g_arm_cond_name[16] =
 	"gt", "le", "al", "al",
 };
 
+static const char* g_arm_reg_name[16] =
+{
+	"r0", "r1", "r2", "r3",
+	"r4", "r5", "r6", "r7",
+	"r8", "r9", "r10", "r11",
+	"r12", "sp", "lr", "pc",
+};
+
 class ARMv7DisAsm
 	: public CPUDisAsm
 {
@@ -24,6 +31,7 @@ protected:
 		return (u32)dump_pc + imm;
 	}
 
+#if 0
 	std::string GetRegsListString(u16 regs_list)
 	{
 		std::string regs_str;
@@ -316,4 +324,5 @@ protected:
 	virtual void UXTB(const u32 data, const ARMv7_encoding type);
 	virtual void UXTB16(const u32 data, const ARMv7_encoding type);
 	virtual void UXTH(const u32 data, const ARMv7_encoding type);
+#endif
 };
diff --git a/rpcs3/Emu/ARMv7/ARMv7Interpreter.h b/rpcs3/Emu/ARMv7/ARMv7Interpreter.h
index b73b46b2f9..4f187e379b 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Interpreter.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Interpreter.h
@@ -1,5 +1,16 @@
 #pragma once
 
+union ARMv7Code
+{
+	struct
+	{
+		u16 code0;
+		u16 code1;
+	};
+
+	u32 data;
+};
+
 enum ARMv7_encoding
 {
 	T1, T2, T3, T4, A1, A2
diff --git a/rpcs3/Emu/ARMv7/ARMv7Opcodes.h b/rpcs3/Emu/ARMv7/ARMv7Opcodes.h
index 7a7f0e7e53..ed415bf2a2 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Opcodes.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Opcodes.h
@@ -1,16 +1,8 @@
 #pragma once
-
+#if 0
 #include "Emu/ARMv7/ARMv7Thread.h"
 #include "Emu/ARMv7/ARMv7Interpreter.h"
 
-static const char* g_arm_reg_name[16] =
-{
-	"r0", "r1", "r2", "r3",
-	"r4", "r5", "r6", "r7",
-	"r8", "r9", "r10", "r11",
-	"r12", "sp", "lr", "pc",
-};
-#if 0
 using namespace ARMv7_instrs;
 
 struct ARMv7_Instruction
diff --git a/rpcs3/Emu/ARMv7/ARMv7Thread.cpp b/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
index abd6b2791f..e4b1d65291 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
+++ b/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
@@ -106,7 +106,7 @@ void ARMv7Thread::DoRun()
 
 	case 1:
 	case 2:
-		m_dec = new ARMv7Decoder(*this);
+		m_dec = new ARMv7Decoder(context);
 	break;
 	}
 }
diff --git a/rpcs3/Emu/CPU/CPUDecoder.h b/rpcs3/Emu/CPU/CPUDecoder.h
index beb0027055..0a0aea1e0c 100644
--- a/rpcs3/Emu/CPU/CPUDecoder.h
+++ b/rpcs3/Emu/CPU/CPUDecoder.h
@@ -5,7 +5,7 @@
 class CPUDecoder
 {
 public:
-	virtual u8 DecodeMemory(const u32 address)=0;
+	virtual u32 DecodeMemory(const u32 address) = 0;
 
 	virtual ~CPUDecoder() = default;
 };
diff --git a/rpcs3/Emu/Cell/PPCDecoder.cpp b/rpcs3/Emu/Cell/PPCDecoder.cpp
index de2b55fdf9..6425cf7c7a 100644
--- a/rpcs3/Emu/Cell/PPCDecoder.cpp
+++ b/rpcs3/Emu/Cell/PPCDecoder.cpp
@@ -2,7 +2,7 @@
 #include "Emu/Memory/Memory.h"
 #include "PPCDecoder.h"
 
-u8 PPCDecoder::DecodeMemory(const u32 address)
+u32 PPCDecoder::DecodeMemory(const u32 address)
 {
 	u32 instr = vm::read32(address);
 	Decode(instr);
diff --git a/rpcs3/Emu/Cell/PPCDecoder.h b/rpcs3/Emu/Cell/PPCDecoder.h
index b13ab08efd..a653f06e81 100644
--- a/rpcs3/Emu/Cell/PPCDecoder.h
+++ b/rpcs3/Emu/Cell/PPCDecoder.h
@@ -5,9 +5,9 @@
 class PPCDecoder : public CPUDecoder
 {
 public:
-	virtual void Decode(const u32 code)=0;
+	virtual void Decode(const u32 code) = 0;
 
-	virtual u8 DecodeMemory(const u32 address);
+	virtual u32 DecodeMemory(const u32 address);
 
 	virtual ~PPCDecoder() = default;
 };
diff --git a/rpcs3/Emu/Cell/SPURecompiler.h b/rpcs3/Emu/Cell/SPURecompiler.h
index 1b977a6e52..1b85a16d21 100644
--- a/rpcs3/Emu/Cell/SPURecompiler.h
+++ b/rpcs3/Emu/Cell/SPURecompiler.h
@@ -45,7 +45,7 @@ public:
 
 	virtual void Decode(const u32 code);
 
-	virtual u8 DecodeMemory(const u32 address);
+	virtual u32 DecodeMemory(const u32 address);
 };
 
 #define c (*compiler)
diff --git a/rpcs3/Emu/Cell/SPURecompilerCore.cpp b/rpcs3/Emu/Cell/SPURecompilerCore.cpp
index 197f42987e..a829019603 100644
--- a/rpcs3/Emu/Cell/SPURecompilerCore.cpp
+++ b/rpcs3/Emu/Cell/SPURecompilerCore.cpp
@@ -180,7 +180,7 @@ void SPURecompilerCore::Compile(u16 pos)
 	first = false;
 }
 
-u8 SPURecompilerCore::DecodeMemory(const u32 address)
+u32 SPURecompilerCore::DecodeMemory(const u32 address)
 {
 	assert(CPU.ls_offset == address - CPU.PC);
 	const u32 m_offset = CPU.ls_offset;
diff --git a/rpcs3/Loader/ELF32.cpp b/rpcs3/Loader/ELF32.cpp
index 988d1d779c..80906f88ca 100644
--- a/rpcs3/Loader/ELF32.cpp
+++ b/rpcs3/Loader/ELF32.cpp
@@ -6,6 +6,7 @@
 #include "ELF32.h"
 #include "Emu/Cell/SPUThread.h"
 #include "Emu/ARMv7/ARMv7Thread.h"
+#include "Emu/ARMv7/ARMv7Decoder.h"
 #include "Emu/ARMv7/PSVFuncList.h"
 #include "Emu/System.h"
 
@@ -98,17 +99,8 @@ namespace loader
 
 				u32 entry = 0; // actual entry point (ELFs entry point is ignored)
 				u32 fnid_addr = 0;
-
-				// load section names
-				//assert(m_ehdr.data_le.e_shstrndx < m_shdrs.size());
-				//const u32 sname_off = m_shdrs[m_ehdr.data_le.e_shstrndx].data_le.sh_offset;
-				//const u32 sname_size = m_shdrs[m_ehdr.data_le.e_shstrndx].data_le.sh_size;
-				//const u32 sname_base = sname_size ? Memory.PSV.RAM.AllocAlign(sname_size) : 0;
-				//if (sname_base)
-				//{
-				//	m_stream->Seek(handler::get_stream_offset() + sname_off);
-				//	m_stream->Read(vm::get_ptr<void>(sname_base), sname_size);
-				//}
+				u32 code_start = 0;
+				u32 code_end = 0;
 
 				for (auto& shdr : m_shdrs)
 				{
@@ -125,7 +117,14 @@ namespace loader
 						name.push_back(c);
 					}
 
-					if (!strcmp(name.c_str(), ".sceModuleInfo.rodata"))
+					if (!strcmp(name.c_str(), ".text"))
+					{
+						LOG_NOTICE(LOADER, ".text analysis...");
+
+						code_start = shdr.data_le.sh_addr;
+						code_end = shdr.data_le.sh_size + code_start;
+					}
+					else if (!strcmp(name.c_str(), ".sceModuleInfo.rodata"))
 					{
 						LOG_NOTICE(LOADER, ".sceModuleInfo.rodata analysis...");
 
@@ -190,6 +189,8 @@ namespace loader
 								vm::psv::write16(addr + 2, 0); // index 0 (unimplemented stub)
 								vm::psv::write32(addr + 4, nid); // nid
 							}
+
+							code_end = std::min<u32>(addr, code_end);
 						}
 					}
 					else if (!strcmp(name.c_str(), ".sceRefs.rodata"))
@@ -238,6 +239,8 @@ namespace loader
 					}
 				}
 
+				armv7_decoder_initialize(code_start, code_end);
+
 				arm7_thread(entry & ~1 /* TODO: Thumb/ARM encoding selection */, "main_thread").args({ Emu.GetPath()/*, "-emu"*/ }).run();
 				break;
 			}
diff --git a/rpcs3/Loader/ELF64.cpp b/rpcs3/Loader/ELF64.cpp
index 706aee8000..c943495597 100644
--- a/rpcs3/Loader/ELF64.cpp
+++ b/rpcs3/Loader/ELF64.cpp
@@ -516,7 +516,7 @@ namespace loader
 
 									if (module && !module->Load(nid))
 									{
-										LOG_WARNING(LOADER, "Unimplemented function '%s' in '%s' module (HLE)", SysCalls::GetHLEFuncName(nid).c_str(), module_name.c_str());
+										LOG_ERROR(LOADER, "Unimplemented function '%s' in '%s' module (HLE)", SysCalls::GetHLEFuncName(nid).c_str(), module_name.c_str());
 									}
 									else //if (Ini.HLELogging.GetValue())
 									{