diff --git a/src/pse-sdl/sdl_interface.cpp b/src/pse-sdl/sdl_interface.cpp
index c2a404e6c..daefd8f1e 100644
--- a/src/pse-sdl/sdl_interface.cpp
+++ b/src/pse-sdl/sdl_interface.cpp
@@ -178,6 +178,8 @@ bool SDLInterface::InitializeSystem(const char* filename, s32 save_state_index /
 
   m_system->Reset();
 
+  //m_system->LoadEXE("tests/psxtest_cpu.psxexe");
+
   // Resume execution.
   m_running = true;
   return true;
diff --git a/src/pse/bus.cpp b/src/pse/bus.cpp
index 7a9e1b6c0..fb566d9f6 100644
--- a/src/pse/bus.cpp
+++ b/src/pse/bus.cpp
@@ -2,6 +2,7 @@
 #include "YBaseLib/ByteStream.h"
 #include "YBaseLib/Log.h"
 #include "YBaseLib/String.h"
+#include "cpu_disasm.h"
 #include "dma.h"
 #include "gpu.h"
 #include <cstdio>
@@ -70,6 +71,24 @@ bool Bus::WriteWord(PhysicalMemoryAddress cpu_address, PhysicalMemoryAddress bus
   return DispatchAccess<MemoryAccessType::Write, MemoryAccessSize::Word>(cpu_address, bus_address, value);
 }
 
+void Bus::PatchBIOS(u32 address, u32 value, u32 mask /*= UINT32_C(0xFFFFFFFF)*/)
+{
+  const u32 phys_address = address & UINT32_C(0x1FFFFFFF);
+  const u32 offset = phys_address - BIOS_BASE;
+  Assert(phys_address >= BIOS_BASE && offset < BIOS_SIZE);
+
+  u32 existing_value;
+  std::memcpy(&existing_value, &m_bios[offset], sizeof(existing_value));
+  u32 new_value = (existing_value & ~mask) | value;
+  std::memcpy(&m_bios[offset], &new_value, sizeof(new_value));
+
+  SmallString old_disasm, new_disasm;
+  CPU::DisassembleInstruction(&old_disasm, address, existing_value);
+  CPU::DisassembleInstruction(&new_disasm, address, new_value);
+  Log_InfoPrintf("BIOS-Patch 0x%08X (+0x%X): 0x%08X %s -> %08X %s", address, offset, existing_value,
+                 old_disasm.GetCharArray(), new_value, new_disasm.GetCharArray());
+}
+
 bool Bus::LoadBIOS()
 {
   std::FILE* fp = std::fopen("SCPH1001.BIN", "rb");
@@ -97,9 +116,9 @@ bool Bus::LoadBIOS()
   std::fclose(fp);
 
 #if 1
-  auto Patch = [this](u32 address, u32 value) { std::memcpy(&m_bios[address], &value, sizeof(value)); };
-  Patch(0x6F0C, 0x24010001); // addiu $at, $zero, 1
-  Patch(0x6F14, 0xaf81a9c0); // sw at, -0x5640(gp)
+  // Patch to enable TTY.
+  PatchBIOS(BIOS_BASE + 0x6F0C, 0x24010001);
+  PatchBIOS(BIOS_BASE + 0x6F14, 0xAF81A9C0);
 #endif
 
   return true;
diff --git a/src/pse/bus.h b/src/pse/bus.h
index 084ea7a72..81d7dd86f 100644
--- a/src/pse/bus.h
+++ b/src/pse/bus.h
@@ -29,6 +29,8 @@ public:
   template<MemoryAccessType type, MemoryAccessSize size>
   bool DispatchAccess(PhysicalMemoryAddress cpu_address, PhysicalMemoryAddress bus_address, u32& value);
 
+  void PatchBIOS(u32 address, u32 value, u32 mask = UINT32_C(0xFFFFFFFF));
+
 private:
   static constexpr u32 DMA_BASE = 0x1F801080;
   static constexpr u32 DMA_SIZE = 0x80;
@@ -42,6 +44,8 @@ private:
   static constexpr u32 EXP2_BASE = 0x1F802000;
   static constexpr u32 EXP2_SIZE = 0x2000;
   static constexpr u32 EXP2_MASK = EXP2_SIZE - 1;
+  static constexpr u32 BIOS_BASE = 0x1FC00000;
+  static constexpr u32 BIOS_SIZE = 0x80000;
 
   bool LoadBIOS();
 
diff --git a/src/pse/bus.inl b/src/pse/bus.inl
index 649290b0b..b2ffc92dc 100644
--- a/src/pse/bus.inl
+++ b/src/pse/bus.inl
@@ -116,13 +116,13 @@ bool Bus::DispatchAccess(PhysicalMemoryAddress cpu_address, PhysicalMemoryAddres
     return (type == MemoryAccessType::Read) ? ReadExpansionRegion2(size, bus_address & EXP2_MASK, value) :
                                               WriteExpansionRegion2(size, bus_address & EXP2_MASK, value);
   }
-  else if (bus_address < 0x1FC00000)
+  else if (bus_address < BIOS_BASE)
   {
     return DoInvalidAccess(type, size, cpu_address, bus_address, value);
   }
-  else if (bus_address < 0x20000000)
+  else if (bus_address < (BIOS_BASE + BIOS_SIZE))
   {
-    return DoBIOSAccess<type, size>(static_cast<u32>(bus_address - UINT32_C(0x1FC00000)), value);
+    return DoBIOSAccess<type, size>(static_cast<u32>(bus_address - BIOS_BASE), value);
   }
   else if (bus_address < 0x1FFE0000)
   {
diff --git a/src/pse/cpu_core.cpp b/src/pse/cpu_core.cpp
index d9717cc68..db46f21e0 100644
--- a/src/pse/cpu_core.cpp
+++ b/src/pse/cpu_core.cpp
@@ -30,10 +30,16 @@ bool Core::DoState(StateWrapper& sw)
   return false;
 }
 
+void Core::SetPC(u32 new_pc)
+{
+  m_regs.npc = new_pc;
+  FlushPipeline();
+}
+
 u8 Core::ReadMemoryByte(VirtualMemoryAddress addr)
 {
   u32 value;
-  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Byte, false>(addr, value);
+  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Byte, false, true>(addr, value);
   return Truncate8(value);
 }
 
@@ -41,7 +47,7 @@ u16 Core::ReadMemoryHalfWord(VirtualMemoryAddress addr)
 {
   Assert(Common::IsAlignedPow2(addr, 2));
   u32 value;
-  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::HalfWord, false>(addr, value);
+  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::HalfWord, false, true>(addr, value);
   return Truncate16(value);
 }
 
@@ -49,27 +55,65 @@ u32 Core::ReadMemoryWord(VirtualMemoryAddress addr)
 {
   Assert(Common::IsAlignedPow2(addr, 4));
   u32 value;
-  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, false>(addr, value);
+  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, false, true>(addr, value);
   return value;
 }
 
 void Core::WriteMemoryByte(VirtualMemoryAddress addr, u8 value)
 {
   u32 value32 = ZeroExtend32(value);
-  DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::Byte, false>(addr, value32);
+  DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::Byte, false, true>(addr, value32);
 }
 
 void Core::WriteMemoryHalfWord(VirtualMemoryAddress addr, u16 value)
 {
   Assert(Common::IsAlignedPow2(addr, 2));
   u32 value32 = ZeroExtend32(value);
-  DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::HalfWord, false>(addr, value32);
+  DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::HalfWord, false, true>(addr, value32);
 }
 
 void Core::WriteMemoryWord(VirtualMemoryAddress addr, u32 value)
 {
   Assert(Common::IsAlignedPow2(addr, 4));
-  DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::Word, false>(addr, value);
+  DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::Word, false, true>(addr, value);
+}
+
+bool Core::SafeReadMemoryByte(VirtualMemoryAddress addr, u8* value)
+{
+  u32 temp = 0;
+  const bool result = DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Byte, false, false>(addr, temp);
+  *value = Truncate8(temp);
+  return result;
+}
+
+bool Core::SafeReadMemoryHalfWord(VirtualMemoryAddress addr, u16* value)
+{
+  u32 temp = 0;
+  const bool result = DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::HalfWord, false, false>(addr, temp);
+  *value = Truncate16(temp);
+  return result;
+}
+
+bool Core::SafeReadMemoryWord(VirtualMemoryAddress addr, u32* value)
+{
+  return DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, false, false>(addr, *value);
+}
+
+bool Core::SafeWriteMemoryByte(VirtualMemoryAddress addr, u8 value)
+{
+  u32 temp = ZeroExtend32(value);
+  return DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::Byte, false, false>(addr, temp);
+}
+
+bool Core::SafeWriteMemoryHalfWord(VirtualMemoryAddress addr, u16 value)
+{
+  u32 temp = ZeroExtend32(value);
+  return DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::HalfWord, false, false>(addr, temp);
+}
+
+bool Core::SafeWriteMemoryWord(VirtualMemoryAddress addr, u32 value)
+{
+  return DoMemoryAccess<MemoryAccessType::Write, MemoryAccessSize::Word, false, false>(addr, value);
 }
 
 void Core::Branch(u32 target)
@@ -153,7 +197,7 @@ static constexpr bool AddOverflow(u32 old_value, u32 add_value, u32 new_value)
 void Core::DisassembleAndPrint(u32 addr)
 {
   u32 bits;
-  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, true>(addr, bits);
+  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, true, false>(addr, bits);
   PrintInstruction(bits, addr);
 }
 
@@ -182,22 +226,25 @@ void Core::Execute()
 
 void Core::FetchInstruction()
 {
-  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, true>(static_cast<VirtualMemoryAddress>(m_regs.npc),
-                                                                       m_next_instruction.bits);
+  DoMemoryAccess<MemoryAccessType::Read, MemoryAccessSize::Word, true, true>(
+    static_cast<VirtualMemoryAddress>(m_regs.npc), m_next_instruction.bits);
   m_regs.pc = m_regs.npc;
   m_regs.npc += sizeof(m_next_instruction.bits);
 }
 
 void Core::ExecuteInstruction(Instruction inst, u32 inst_pc)
 {
-  if (TRACE_EXECUTION)
-    PrintInstruction(inst.bits, inst_pc);
-
 #if 0
-  if (inst_pc == 0x8005ab80)
+  if (inst_pc == 0xBFC06FF0)
+  {
+    TRACE_EXECUTION = true;
     __debugbreak();
+  }
 #endif
 
+  if (TRACE_EXECUTION)
+    PrintInstruction(inst.bits, inst_pc);
+
   switch (inst.op)
   {
     case InstructionOp::funct:
diff --git a/src/pse/cpu_core.h b/src/pse/cpu_core.h
index 79735f9e1..a5255b171 100644
--- a/src/pse/cpu_core.h
+++ b/src/pse/cpu_core.h
@@ -23,9 +23,22 @@ public:
 
   void Execute();
 
+  const Registers& GetRegs() const { return m_regs; }
+  Registers& GetRegs() { return m_regs; }
+
+  // Sets the PC and flushes the pipeline.
+  void SetPC(u32 new_pc);
+
+  bool SafeReadMemoryByte(VirtualMemoryAddress addr, u8* value);
+  bool SafeReadMemoryHalfWord(VirtualMemoryAddress addr, u16* value);
+  bool SafeReadMemoryWord(VirtualMemoryAddress addr, u32* value);
+  bool SafeWriteMemoryByte(VirtualMemoryAddress addr, u8 value);
+  bool SafeWriteMemoryHalfWord(VirtualMemoryAddress addr, u16 value);
+  bool SafeWriteMemoryWord(VirtualMemoryAddress addr, u32 value);
+
 private:
-  template<MemoryAccessType type, MemoryAccessSize size, bool is_instruction_fetch>
-  void DoMemoryAccess(VirtualMemoryAddress address, u32& value);
+  template<MemoryAccessType type, MemoryAccessSize size, bool is_instruction_fetch, bool raise_exceptions>
+  bool DoMemoryAccess(VirtualMemoryAddress address, u32& value);
 
   u8 ReadMemoryByte(VirtualMemoryAddress addr);
   u16 ReadMemoryHalfWord(VirtualMemoryAddress addr);
diff --git a/src/pse/cpu_core.inl b/src/pse/cpu_core.inl
index 9d4fbc22d..68a1caecf 100644
--- a/src/pse/cpu_core.inl
+++ b/src/pse/cpu_core.inl
@@ -5,8 +5,8 @@
 
 namespace CPU {
 
-template<MemoryAccessType type, MemoryAccessSize size, bool is_instruction_fetch>
-void Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
+template<MemoryAccessType type, MemoryAccessSize size, bool is_instruction_fetch, bool raise_exceptions>
+bool Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
 {
   switch (address >> 29)
   {
@@ -15,13 +15,17 @@ void Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
       if constexpr (type == MemoryAccessType::Write)
       {
         if (m_cop0_regs.sr.Isc)
-          return;
+          return true;
       }
 
       if (!m_bus->DispatchAccess<type, size>(address, address & UINT32_C(0x1FFFFFFF), value))
+      {
         Panic("Bus error");
+        return false;
+      }
+
+      return true;
     }
-    break;
 
     case 0x01: // KUSEG 512M-1024M
     case 0x02: // KUSEG 1024M-1536M
@@ -29,26 +33,36 @@ void Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
     {
       // Above 512mb raises an exception.
       Panic("Bad user access");
+      return false;
     }
-    break;
 
     case 0x04: // KSEG0 - physical memory cached
     {
       if constexpr (type == MemoryAccessType::Write)
       {
         if (m_cop0_regs.sr.Isc)
-          return;
+          return true;
       }
 
       if (!m_bus->DispatchAccess<type, size>(address, address & UINT32_C(0x1FFFFFFF), value))
+      {
         Panic("Bus error");
+        return false;
+      }
+
+      return true;
     }
     break;
 
     case 0x05: // KSEG1 - physical memory uncached
     {
       if (!m_bus->DispatchAccess<type, size>(address, address & UINT32_C(0x1FFFFFFF), value))
+      {
         Panic("Bus error");
+        return false;
+      }
+
+      return true;
     }
     break;
 
@@ -61,17 +75,19 @@ void Core::DoMemoryAccess(VirtualMemoryAddress address, u32& value)
           value = m_cache_control;
         else
           WriteCacheControl(value);
+
+        return true;
       }
       else
       {
         Panic("KSEG2 access");
+        return false;
       }
     }
-    break;
 
     default:
       UnreachableCode();
-      break;
+      return false;
   }
 }
 
diff --git a/src/pse/cpu_types.h b/src/pse/cpu_types.h
index a550f151f..e52e3bb2e 100644
--- a/src/pse/cpu_types.h
+++ b/src/pse/cpu_types.h
@@ -179,38 +179,38 @@ struct Registers
 
     struct
     {
-      u32 zero;
-      u32 at;
-      u32 v0;
-      u32 v1;
-      u32 a0;
-      u32 a1;
-      u32 a2;
-      u32 a3;
-      u32 t0;
-      u32 t1;
-      u32 t2;
-      u32 t3;
-      u32 t4;
-      u32 t5;
-      u32 t6;
-      u32 t7;
-      u32 s0;
-      u32 s1;
-      u32 s2;
-      u32 s3;
-      u32 s4;
-      u32 s5;
-      u32 s6;
-      u32 s7;
-      u32 t8;
-      u32 t9;
-      u32 k0;
-      u32 k1;
-      u32 gp;
-      u32 sp;
-      u32 fp;
-      u32 ra;
+      u32 zero;   // r0
+      u32 at;     // r1
+      u32 v0;     // r2
+      u32 v1;     // r3
+      u32 a0;     // r4
+      u32 a1;     // r5
+      u32 a2;     // r6
+      u32 a3;     // r7
+      u32 t0;     // r8
+      u32 t1;     // r9
+      u32 t2;     // r10
+      u32 t3;     // r11
+      u32 t4;     // r12
+      u32 t5;     // r13
+      u32 t6;     // r14
+      u32 t7;     // r15
+      u32 s0;     // r16
+      u32 s1;     // r17
+      u32 s2;     // r18
+      u32 s3;     // r19
+      u32 s4;     // r20
+      u32 s5;     // r21
+      u32 s6;     // r22
+      u32 s7;     // r23
+      u32 t8;     // r24
+      u32 t9;     // r25
+      u32 k0;     // r26
+      u32 k1;     // r27
+      u32 gp;     // r28
+      u32 sp;     // r29
+      u32 fp;     // r30
+      u32 ra;     // r31
     };
   };
 
diff --git a/src/pse/system.cpp b/src/pse/system.cpp
index 6ff0d4541..35afe7b06 100644
--- a/src/pse/system.cpp
+++ b/src/pse/system.cpp
@@ -47,3 +47,90 @@ void System::RunFrame()
   while (current_frame_number == m_frame_number)
     m_cpu->Execute();
 }
+
+bool System::LoadEXE(const char* filename)
+{
+#pragma pack(push, 1)
+  struct EXEHeader
+  {
+    char id[8];            // 0x000-0x007 PS-X EXE
+    char pad1[8];          // 0x008-0x00F
+    u32 initial_pc;        // 0x010
+    u32 initial_gp;        // 0x014
+    u32 load_address;      // 0x018
+    u32 file_size;         // 0x01C excluding 0x800-byte header
+    u32 unk0;              // 0x020
+    u32 unk1;              // 0x024
+    u32 memfill_start;     // 0x028
+    u32 memfill_size;      // 0x02C
+    u32 initial_sp_base;   // 0x030
+    u32 initial_sp_offset; // 0x034
+    u32 reserved[5];       // 0x038-0x04B
+    char marker[0x7B4];    // 0x04C-0x7FF
+  };
+  static_assert(sizeof(EXEHeader) == 0x800);
+#pragma pack(pop)
+
+  std::FILE* fp = std::fopen(filename, "rb");
+  if (!fp)
+    return false;
+
+  EXEHeader header;
+  if (std::fread(&header, sizeof(header), 1, fp) != 1)
+  {
+    std::fclose(fp);
+    return false;
+  }
+
+  if (header.memfill_size > 0)
+  {
+    const u32 words_to_write = header.memfill_size / 4;
+    u32 address = header.memfill_start & ~UINT32_C(3);
+    for (u32 i = 0; i < words_to_write; i++)
+    {
+      m_cpu->SafeWriteMemoryWord(address, 0);
+      address += sizeof(u32);
+    }
+  }
+
+  if (header.file_size >= 4)
+  {
+    std::vector<u32> data_words(header.file_size / 4);
+    if (std::fread(data_words.data(), header.file_size, 1, fp) != 1)
+    {
+      std::fclose(fp);
+      return false;
+    }
+
+    const u32 num_words = header.file_size / 4;
+    u32 address = header.load_address;
+    for (u32 i = 0; i < num_words; i++)
+    {
+      m_cpu->SafeWriteMemoryWord(address, data_words[i]);
+      address += sizeof(u32);
+    }
+  }
+
+  std::fclose(fp);
+
+  // patch the BIOS to jump to the executable directly
+  {
+    const u32 r_pc = header.load_address;
+    const u32 r_gp = header.initial_gp;
+    const u32 r_sp = header.initial_sp_base;
+    const u32 r_fp = header.initial_sp_base + header.initial_sp_offset;
+
+    // pc has to be done first because we can't load it in the delay slot
+    m_bus->PatchBIOS(0xBFC06FF0, UINT32_C(0x3C080000) | r_pc >> 16);                // lui $t0, (r_pc >> 16)
+    m_bus->PatchBIOS(0xBFC06FF4, UINT32_C(0x35080000) | (r_pc & UINT32_C(0xFFFF))); // ori $t0, $t0, (r_pc & 0xFFFF)
+    m_bus->PatchBIOS(0xBFC06FF8, UINT32_C(0x3C1C0000) | r_gp >> 16);                // lui $gp, (r_gp >> 16)
+    m_bus->PatchBIOS(0xBFC06FFC, UINT32_C(0x379C0000) | (r_gp & UINT32_C(0xFFFF))); // ori $gp, $gp, (r_gp & 0xFFFF)
+    m_bus->PatchBIOS(0xBFC07000, UINT32_C(0x3C1D0000) | r_sp >> 16);                // lui $sp, (r_sp >> 16)
+    m_bus->PatchBIOS(0xBFC07004, UINT32_C(0x37BD0000) | (r_sp & UINT32_C(0xFFFF))); // ori $sp, $sp, (r_sp & 0xFFFF)
+    m_bus->PatchBIOS(0xBFC07008, UINT32_C(0x3C1E0000) | r_fp >> 16);                // lui $fp, (r_fp >> 16)
+    m_bus->PatchBIOS(0xBFC0700C, UINT32_C(0x01000008));                             // jr $t0
+    m_bus->PatchBIOS(0xBFC07010, UINT32_C(0x37DE0000) | (r_fp & UINT32_C(0xFFFF))); // ori $fp, $fp, (r_fp & 0xFFFF)
+  }
+
+  return true;
+}
diff --git a/src/pse/system.h b/src/pse/system.h
index 0338f3e9f..615818885 100644
--- a/src/pse/system.h
+++ b/src/pse/system.h
@@ -28,6 +28,8 @@ public:
 
   void RunFrame();
 
+  bool LoadEXE(const char* filename);
+
 private:
   HostInterface* m_host_interface;
   std::unique_ptr<CPU::Core> m_cpu;