diff --git a/emulator/emulator.hpp b/emulator/emulator.hpp
index 977dc494..5e61f4bd 100644
--- a/emulator/emulator.hpp
+++ b/emulator/emulator.hpp
@@ -8,7 +8,7 @@ using namespace nall;
 
 namespace Emulator {
   static const string Name = "higan";
-  static const string Version = "094.34";
+  static const string Version = "094.35";
   static const string Author = "byuu";
   static const string License = "GPLv3";
   static const string Website = "http://byuu.org/";
diff --git a/gba/cartridge/cartridge.cpp b/gba/cartridge/cartridge.cpp
index 68d6d80e..39f699ff 100644
--- a/gba/cartridge/cartridge.cpp
+++ b/gba/cartridge/cartridge.cpp
@@ -96,66 +96,56 @@ unsigned Cartridge::ram_size() {
   if(has_sram) return ram.size;
   if(has_eeprom) return eeprom.size;
   if(has_flashrom) return flashrom.size;
-  return 0u;
+  return 0;
 }
 
-uint32 Cartridge::read(uint8 *data, uint32 addr, uint32 size) {
-  if(size == Word) addr &= ~3;
-  if(size == Half) addr &= ~1;
+auto Cartridge::read(uint8 *data, unsigned mode, uint32 addr) -> uint32 {
+  if(mode & Word) addr &= ~3;
+  if(mode & Half) addr &= ~1;
   data += addr;
-  if(size == Word) return data[0] << 0 | data[1] << 8 | data[2] << 16 | data[3] << 24;
-  if(size == Half) return data[0] << 0 | data[1] << 8;
-  return data[0];
+  if(mode & Word) return data[0] << 0 | data[1] << 8 | data[2] << 16 | data[3] << 24;
+  if(mode & Half) return data[0] << 0 | data[1] << 8;
+  if(mode & Byte) return data[0];
+  return 0;  //should never occur
 }
 
-void Cartridge::write(uint8 *data, uint32 addr, uint32 size, uint32 word) {
-  if(size == Word) addr &= ~3;
-  if(size == Half) addr &= ~1;
+auto Cartridge::write(uint8 *data, unsigned mode, uint32 addr, uint32 word) -> void {
+  if(mode & Word) addr &= ~3;
+  if(mode & Half) addr &= ~1;
   data += addr;
-  switch(size) {
-  case Word: data[3] = word >> 24;
-             data[2] = word >> 16;
-  case Half: data[1] = word >>  8;
-  case Byte: data[0] = word >>  0;
+  if(mode & Word) {
+    data[0] = word >>  0;
+    data[1] = word >>  8;
+    data[2] = word >> 16;
+    data[3] = word >> 24;
+  } else if(mode & Half) {
+    data[0] = word >>  0;
+    data[1] = word >>  8;
+  } else if(mode & Byte) {
+    data[0] = word >>  0;
   }
 }
 
 #define RAM_ANALYZE
 
-uint32 Cartridge::read(uint32 addr, uint32 size) {
-  #ifdef RAM_ANALYZE
-  if((addr & 0x0e000000) == 0x0e000000) {
-    static bool once = true;
-    if(once) once = false, print("* SRAM/FlashROM read detected\n");
-  }
-  #endif
+auto Cartridge::rom_read(unsigned mode, uint32 addr) -> uint32 {
+  if(has_eeprom && (addr & eeprom.mask) == eeprom.test) return eeprom.read();
+  return read(rom.data, mode, addr & 0x01ff'ffff);
+}
 
-  if(has_sram     && (addr & 0x0e000000 ) == 0x0e000000 ) return read(ram.data, addr & ram.mask, size);
-  if(has_eeprom   && (addr & eeprom.mask) == eeprom.test) return eeprom.read();
-  if(has_flashrom && (addr & 0x0e000000 ) == 0x0e000000 ) return flashrom.read(addr);
-  if(addr < 0x0e000000) return read(rom.data, addr & 0x01ffffff, size);
+auto Cartridge::rom_write(unsigned mode, uint32 addr, uint32 word) -> void {
+  if(has_eeprom && (addr & eeprom.mask) == eeprom.test) return eeprom.write(word & 1);
+}
+
+auto Cartridge::ram_read(unsigned mode, uint32 addr) -> uint32 {
+  if(has_sram) return read(ram.data, mode, addr & ram.mask);
+  if(has_flashrom) return flashrom.read(addr);
   return cpu.pipeline.fetch.instruction;
 }
 
-void Cartridge::write(uint32 addr, uint32 size, uint32 word) {
-  #ifdef RAM_ANALYZE
-  if((addr & 0x0e000000) == 0x0e000000) {
-    static bool once = true;
-    if(once) once = false, print("* SRAM/FlashROM write detected\n");
-  }
-  if((addr & 0x0f000000) == 0x0d000000) {
-    static bool once = true;
-    if(once) once = false, print("* EEPROM write detected\n");
-  }
-  if((addr & 0x0e00ffff) == 0x0e005555 && (word & 0xff) == 0xaa) {
-    static bool once = true;
-    if(once) once = false, print("* FlashROM write detected\n");
-  }
-  #endif
-
-  if(has_sram     && (addr & 0x0e000000 ) == 0x0e000000 ) return write(ram.data, addr & ram.mask, size, word);
-  if(has_eeprom   && (addr & eeprom.mask) == eeprom.test) return eeprom.write(word & 1);
-  if(has_flashrom && (addr & 0x0e000000 ) == 0x0e000000 ) return flashrom.write(addr, word);
+auto Cartridge::ram_write(unsigned mode, uint32 addr, uint32 word) -> void {
+  if(has_sram) return write(ram.data, mode, addr & ram.mask, word);
+  if(has_flashrom) return flashrom.write(addr, word);
 }
 
 Cartridge::Cartridge() {
diff --git a/gba/cartridge/cartridge.hpp b/gba/cartridge/cartridge.hpp
index 3e480c39..f0e697c3 100644
--- a/gba/cartridge/cartridge.hpp
+++ b/gba/cartridge/cartridge.hpp
@@ -28,13 +28,17 @@ struct Cartridge : property<Cartridge> {
   uint8* ram_data();
   unsigned ram_size();
 
-  uint32 read(uint8* data, uint32 addr, uint32 size);
-  void write(uint8* data, uint32 addr, uint32 size, uint32 word);
+  auto read(uint8* data, unsigned mode, uint32 addr) -> uint32;
+  auto write(uint8* data, unsigned mode, uint32 addr, uint32 word) -> void;
 
-  uint32 read(uint32 addr, uint32 size);
-  void write(uint32 addr, uint32 size, uint32 word);
+  auto rom_read(unsigned mode, uint32 addr) -> uint32;
+  auto rom_write(unsigned mode, uint32 addr, uint32 word) -> void;
+
+  auto ram_read(unsigned mode, uint32 addr) -> uint32;
+  auto ram_write(unsigned mode, uint32 addr, uint32 word) -> void;
 
   void serialize(serializer&);
+
   Cartridge();
   ~Cartridge();
 };
diff --git a/gba/cpu/bus.cpp b/gba/cpu/bus.cpp
new file mode 100644
index 00000000..b6421c6a
--- /dev/null
+++ b/gba/cpu/bus.cpp
@@ -0,0 +1,82 @@
+auto CPU::bus_idle() -> void {
+  prefetch_step(1);
+}
+
+auto CPU::bus_read(unsigned mode, uint32 addr) -> uint32 {
+  addr &= 0x0fff'ffff;
+  unsigned wait = bus_wait(mode, addr);
+  unsigned word = pipeline.fetch.instruction;
+
+  if(addr & 0x0800'0000) {
+    if(mode & Prefetch && regs.wait.control.prefetch) {
+      prefetch_sync(addr);
+      word = prefetch_read();
+      if(mode & Word) word |= prefetch_read() << 16;
+    } else {
+      if(!active.dma) prefetch_wait();
+      step(wait - 1);
+      word = addr < 0x0e00'0000 ? cartridge.rom_read(mode, addr) : cartridge.ram_read(mode, addr);
+      step(1);
+    }
+  } else {
+    prefetch_step(wait - 1);
+         if(addr <  0x0200'0000) word = bios.read(mode, addr);
+    else if(addr <  0x0300'0000) word = ewram_read(mode, addr);
+    else if(addr <  0x0400'0000) word = iwram_read(mode, addr);
+    else if(addr >= 0x0700'0000) word = ppu.oam_read(mode, addr);
+    else if(addr >= 0x0600'0000) word = ppu.vram_read(mode, addr);
+    else if(addr >= 0x0500'0000) word = ppu.pram_read(mode, addr);
+    else if((addr & 0xffff'fc00) == 0x0400'0000) word = bus.mmio[addr & 0x3ff]->read(mode, addr);
+    else if((addr & 0xff00'ffff) == 0x0400'0800) word = ((MMIO*)this)->read(mode, 0x0400'0800 | (addr & 3));
+    prefetch_step(1);
+  }
+
+  return word;
+}
+
+auto CPU::bus_write(unsigned mode, uint32 addr, uint32 word) -> void {
+  addr &= 0x0fff'ffff;
+  unsigned wait = bus_wait(mode, addr);
+
+  if(addr & 0x0800'0000) {
+    if(!active.dma) prefetch_wait();
+    step(wait);
+    addr < 0x0e00'0000 ? cartridge.rom_write(mode, addr, word) : cartridge.ram_write(mode, addr, word);
+  } else {
+    prefetch_step(wait);
+         if(addr  < 0x0200'0000);
+    else if(addr  < 0x0300'0000) ewram_write(mode, addr, word);
+    else if(addr  < 0x0400'0000) iwram_write(mode, addr, word);
+    else if(addr >= 0x0700'0000) ppu.oam_write(mode, addr, word);
+    else if(addr >= 0x0600'0000) ppu.vram_write(mode, addr, word);
+    else if(addr >= 0x0500'0000) ppu.pram_write(mode, addr, word);
+    else if((addr & 0xffff'fc00) == 0x0400'0000) bus.mmio[addr & 0x3ff]->write(mode, addr, word);
+    else if((addr & 0xff00'ffff) == 0x0400'0800) ((MMIO*)this)->write(mode, 0x0400'0800 | (addr & 3), word);
+  }
+}
+
+auto CPU::bus_wait(unsigned mode, uint32 addr) -> unsigned {
+  if(addr < 0x0200'0000) return 1;
+  if(addr < 0x0300'0000) return (16 - regs.memory.control.ewramwait) * (mode & Word ? 2 : 1);
+  if(addr < 0x0500'0000) return 1;
+  if(addr < 0x0700'0000) return mode & Word ? 2 : 1;
+  if(addr < 0x0800'0000) return 1;
+
+  static unsigned timings[] = {5, 4, 3, 9};
+  unsigned n = timings[regs.wait.control.nwait[addr >> 25 & 3]];
+  unsigned s = regs.wait.control.swait[addr >> 25 & 3];
+
+  switch(addr & 0x0e00'0000) {
+  case 0x0800'0000: s = s ? 2 : 3; break;
+  case 0x0a00'0000: s = s ? 2 : 5; break;
+  case 0x0c00'0000: s = s ? 2 : 9; break;
+  case 0x0e00'0000: s = n; break;
+  }
+
+  bool sequential = (mode & Sequential);
+  if((addr & 0x1fffe) == 0) sequential = false;  //N cycle on 16-bit ROM crossing 128KB page boundary (RAM S==N)
+
+  unsigned clocks = sequential ? s : n;
+  if(mode & Word) clocks += s;  //16-bit bus requires two transfers for words
+  return clocks;
+}
diff --git a/gba/cpu/cpu.cpp b/gba/cpu/cpu.cpp
index 1eda3368..2352aa80 100644
--- a/gba/cpu/cpu.cpp
+++ b/gba/cpu/cpu.cpp
@@ -4,6 +4,7 @@ namespace GameBoyAdvance {
 
 #include "registers.cpp"
 #include "prefetch.cpp"
+#include "bus.cpp"
 #include "mmio.cpp"
 #include "memory.cpp"
 #include "dma.cpp"
@@ -63,72 +64,6 @@ auto CPU::step(unsigned clocks) -> void {
   sync_step(clocks);
 }
 
-auto CPU::bus_idle(uint32 addr) -> void {
-  step(1);
-  prefetch_step(1);
-}
-
-auto CPU::bus_read(uint32 addr, uint32 size, bool mode) -> uint32 {
-  unsigned wait = bus.wait(addr, size, mode);
-
-  if(addr < 0x0800'0000) {
-    unsigned word = bus.read(addr, size);
-    step(wait);
-    prefetch_step(wait);
-    return word;
-  }
-
-  if(addr < 0x0e00'0000) {
-    if(regs.wait.control.prefetch) {
-      if(mode == Nonsequential) prefetch_start(addr);
-      unsigned word = prefetch_take();
-      if(size == Byte) word = (addr & 1) ? (word >> 8) : (word & 0xff);
-      if(size == Word) word |= prefetch_take() << 16;
-      return word;
-    }
-
-    unsigned word = cartridge.read(addr, size);
-    step(wait);
-    return word;
-  }
-
-  if(addr < 0x1000'0000) {
-    prefetch_stall();
-    unsigned word = bus.read(addr, size);
-    step(wait);
-    return word;
-  }
-
-  step(wait);
-  prefetch_step(wait);
-  return 0x0000'0000;  //open bus?
-}
-
-auto CPU::bus_write(uint32 addr, uint32 size, bool mode, uint32 word) -> void {
-  unsigned wait = bus.wait(addr, size, mode);
-
-  if(addr < 0x0800'0000) {
-    step(wait);
-    prefetch_step(wait);
-    return bus.write(addr, size, word);
-  }
-
-  if(addr < 0x0e00'0000) {
-    prefetch_stall();
-    step(wait);
-    return bus.write(addr, size, word);
-  }
-
-  if(addr < 0x1000'0000) {
-    prefetch_stall();
-    step(wait);
-    return bus.write(addr, size, word);
-  }
-
-  step(wait);
-  prefetch_step(wait);
-}
-
 auto CPU::sync_step(unsigned clocks) -> void {
   ppu.clock -= clocks;
   if(ppu.clock < 0) co_switch(ppu.thread);
@@ -187,6 +122,8 @@ auto CPU::power() -> void {
   pending.dma.hblank = 0;
   pending.dma.hdma = 0;
 
+  active.dma = false;
+
   for(unsigned n = 0x0b0; n <= 0x0df; n++) bus.mmio[n] = this;  //DMA
   for(unsigned n = 0x100; n <= 0x10f; n++) bus.mmio[n] = this;  //Timers
   for(unsigned n = 0x120; n <= 0x12b; n++) bus.mmio[n] = this;  //Serial
diff --git a/gba/cpu/cpu.hpp b/gba/cpu/cpu.hpp
index 33818a0c..734d9503 100644
--- a/gba/cpu/cpu.hpp
+++ b/gba/cpu/cpu.hpp
@@ -14,9 +14,6 @@ struct CPU : Processor::ARM, Thread, MMIO {
   auto main() -> void;
 
   auto step(unsigned clocks) -> void override;
-  auto bus_idle(uint32 addr) -> void override;
-  auto bus_read(uint32 addr, uint32 size, bool mode) -> uint32 override;
-  auto bus_write(uint32 addr, uint32 size, bool mode, uint32 word) -> void override;
 
   auto sync_step(unsigned clocks) -> void;
   auto keypad_run() -> void;
@@ -25,15 +22,21 @@ struct CPU : Processor::ARM, Thread, MMIO {
   CPU();
   ~CPU();
 
+  //bus.cpp
+  auto bus_idle() -> void override;
+  auto bus_read(unsigned mode, uint32 addr) -> uint32 override;
+  auto bus_write(unsigned mode, uint32 addr, uint32 word) -> void override;
+  auto bus_wait(unsigned mode, uint32 addr) -> unsigned;
+
   //mmio.cpp
   auto read(uint32 addr) -> uint8;
   auto write(uint32 addr, uint8 byte) -> void;
 
-  auto iwram_read(uint32 addr, uint32 size) -> uint32;
-  auto iwram_write(uint32 addr, uint32 size, uint32 word) -> void;
+  auto iwram_read(unsigned mode, uint32 addr) -> uint32;
+  auto iwram_write(unsigned mode, uint32 addr, uint32 word) -> void;
 
-  auto ewram_read(uint32 addr, uint32 size) -> uint32;
-  auto ewram_write(uint32 addr, uint32 size, uint32 word) -> void;
+  auto ewram_read(unsigned mode, uint32 addr) -> uint32;
+  auto ewram_write(unsigned mode, uint32 addr, uint32 word) -> void;
 
   //dma.cpp
   auto dma_run() -> void;
diff --git a/gba/cpu/dma.cpp b/gba/cpu/dma.cpp
index c31b30e0..97f18860 100644
--- a/gba/cpu/dma.cpp
+++ b/gba/cpu/dma.cpp
@@ -1,4 +1,6 @@
 auto CPU::dma_run() -> void {
+  active.dma = true;
+
   while(true) {
     bool transferred = false;
     for(auto n : range(4)) {
@@ -13,20 +15,26 @@ auto CPU::dma_run() -> void {
     }
     if(!transferred) break;
   }
+
+  active.dma = false;
 }
 
 auto CPU::dma_exec(Registers::DMA& dma) -> void {
-  unsigned size = dma.control.size ? Word : Half;
-  unsigned mode = dma.run.length == dma.length ? Nonsequential : Sequential;
   unsigned seek = dma.control.size ? 4 : 2;
+  unsigned mode = dma.control.size ? Word : Half;
+  mode |= dma.run.length == dma.length ? Nonsequential : Sequential;
 
-  if(mode == Nonsequential) {
-    idle();
-    idle();
+  if(mode & Nonsequential) {
+    if((dma.source & 0x0800'0000) && (dma.target & 0x0800'0000)) {
+      //ROM -> ROM transfer
+    } else {
+      idle();
+      idle();
+    }
   }
 
-  uint32 word = bus_read(dma.run.source, size, mode);
-  bus_write(dma.run.target, size, mode, word);
+  uint32 word = bus_read(mode, dma.run.source);
+  bus_write(mode, dma.run.target, word);
 
   switch(dma.control.sourcemode) {
   case 0: dma.run.source += seek; break;
diff --git a/gba/cpu/memory.cpp b/gba/cpu/memory.cpp
index 1059fbe1..aecef462 100644
--- a/gba/cpu/memory.cpp
+++ b/gba/cpu/memory.cpp
@@ -1,53 +1,53 @@
-auto CPU::iwram_read(uint32 addr, uint32 size) -> uint32 {
+auto CPU::iwram_read(unsigned mode, uint32 addr) -> uint32 {
   if(regs.memory.control.disable) return cpu.pipeline.fetch.instruction;
 
-  if(size == Word) return iwram_read(addr &~ 2, Half) << 0 | iwram_read(addr | 2, Half) << 16;
-  if(size == Half) return iwram_read(addr &~ 1, Byte) << 0 | iwram_read(addr | 1, Byte) <<  8;
+  if(mode & Word) return iwram_read(Half, addr &~ 2) << 0 | iwram_read(Half, addr | 2) << 16;
+  if(mode & Half) return iwram_read(Byte, addr &~ 1) << 0 | iwram_read(Byte, addr | 1) <<  8;
 
   return iwram[addr & 0x7fff];
 }
 
-auto CPU::iwram_write(uint32 addr, uint32 size, uint32 word) -> void {
+auto CPU::iwram_write(unsigned mode, uint32 addr, uint32 word) -> void {
   if(regs.memory.control.disable) return;
 
-  if(size == Word) {
-    iwram_write(addr &~2, Half, word >>  0);
-    iwram_write(addr | 2, Half, word >> 16);
+  if(mode & Word) {
+    iwram_write(Half, addr &~2, word >>  0);
+    iwram_write(Half, addr | 2, word >> 16);
     return;
   }
 
-  if(size == Half) {
-    iwram_write(addr &~1, Byte, word >>  0);
-    iwram_write(addr | 1, Byte, word >>  8);
+  if(mode & Half) {
+    iwram_write(Byte, addr &~1, word >>  0);
+    iwram_write(Byte, addr | 1, word >>  8);
     return;
   }
 
   iwram[addr & 0x7fff] = word;
 }
 
-auto CPU::ewram_read(uint32 addr, uint32 size) -> uint32 {
+auto CPU::ewram_read(unsigned mode, uint32 addr) -> uint32 {
   if(regs.memory.control.disable) return cpu.pipeline.fetch.instruction;
-  if(regs.memory.control.ewram == false) return iwram_read(addr, size);
+  if(!regs.memory.control.ewram) return iwram_read(mode, addr);
 
-  if(size == Word) return ewram_read(addr &~ 2, Half) << 0 | ewram_read(addr | 2, Half) << 16;
-  if(size == Half) return ewram_read(addr &~ 1, Byte) << 0 | ewram_read(addr | 1, Byte) <<  8;
+  if(mode & Word) return ewram_read(Half, addr &~ 2) << 0 | ewram_read(Half, addr | 2) << 16;
+  if(mode & Half) return ewram_read(Byte, addr &~ 1) << 0 | ewram_read(Byte, addr | 1) <<  8;
 
   return ewram[addr & 0x3ffff];
 }
 
-auto CPU::ewram_write(uint32 addr, uint32 size, uint32 word) -> void {
+auto CPU::ewram_write(unsigned mode, uint32 addr, uint32 word) -> void {
   if(regs.memory.control.disable) return;
-  if(regs.memory.control.ewram == false) return iwram_write(addr, size, word);
+  if(!regs.memory.control.ewram) return iwram_write(mode, addr, word);
 
-  if(size == Word) {
-    ewram_write(addr &~2, Half, word >>  0);
-    ewram_write(addr | 2, Half, word >> 16);
+  if(mode & Word) {
+    ewram_write(Half, addr &~2, word >>  0);
+    ewram_write(Half, addr | 2, word >> 16);
     return;
   }
 
-  if(size == Half) {
-    ewram_write(addr &~1, Byte, word >>  0);
-    ewram_write(addr | 1, Byte, word >>  8);
+  if(mode & Half) {
+    ewram_write(Byte, addr &~1, word >>  0);
+    ewram_write(Byte, addr | 1, word >>  8);
     return;
   }
 
diff --git a/gba/cpu/prefetch.cpp b/gba/cpu/prefetch.cpp
index ed3ee42a..56d12e7c 100644
--- a/gba/cpu/prefetch.cpp
+++ b/gba/cpu/prefetch.cpp
@@ -1,43 +1,37 @@
-auto CPU::prefetch_stall() -> void {
-  prefetch.stalled = true;
-}
+auto CPU::prefetch_sync(uint32 addr) -> void {
+  if(addr == prefetch.addr) return;
 
-auto CPU::prefetch_start(uint32 addr) -> void {
-  prefetch.stalled = false;
-  prefetch.slots   = 0;
-  prefetch.input   = 0;
-  prefetch.output  = 0;
-  prefetch.addr    = addr;
-  prefetch.wait    = bus.wait(addr, Half, Nonsequential);
+  prefetch.addr = addr;
+  prefetch.load = addr;
+  prefetch.wait = bus_wait(Half | Nonsequential, prefetch.load);
 }
 
 auto CPU::prefetch_step(unsigned clocks) -> void {
-  if(!regs.wait.control.prefetch || prefetch.stalled) return;
+  step(clocks);
+  if(!regs.wait.control.prefetch || active.dma) return;
 
   prefetch.wait -= clocks;
-  while(prefetch.wait <= 0) {
-    if(prefetch.slots < 8) {
-      prefetch.slot[prefetch.output++] = cartridge.read(prefetch.addr, Half);
-      prefetch.slots++;
-      prefetch.addr += 2;
-    }
-    prefetch.wait += bus.wait(prefetch.addr, Half, Sequential);
+  while(!prefetch.full() && prefetch.wait <= 0) {
+    prefetch.slot[prefetch.load >> 1 & 7] = cartridge.rom_read(Half, prefetch.load);
+    prefetch.load += 2;
+    prefetch.wait += bus_wait(Half | Sequential, prefetch.load);
   }
 }
 
 auto CPU::prefetch_wait() -> void {
-  step(prefetch.wait);
+  if(!regs.wait.control.prefetch || prefetch.full()) return;
+
   prefetch_step(prefetch.wait);
+  prefetch.wait = bus_wait(Half | Nonsequential, prefetch.load);
 }
 
-auto CPU::prefetch_take() -> uint16 {
-  if(prefetch.slots) {
-    step(1);
-    prefetch_step(1);
-  } else {
-    prefetch_wait();
-  }
+auto CPU::prefetch_read() -> uint16 {
+  if(prefetch.empty()) prefetch_step(prefetch.wait);
+  else prefetch_step(1);
 
-  prefetch.slots--;
-  return prefetch.slot[prefetch.input++];
+  if(prefetch.full()) prefetch.wait = bus_wait(Half | Sequential, prefetch.load);
+
+  uint16 half = prefetch.slot[prefetch.addr >> 1 & 7];
+  prefetch.addr += 2;
+  return half;
 }
diff --git a/gba/cpu/prefetch.hpp b/gba/cpu/prefetch.hpp
index cd68acba..9b2e79cc 100644
--- a/gba/cpu/prefetch.hpp
+++ b/gba/cpu/prefetch.hpp
@@ -1,16 +1,14 @@
-struct Prefetch {
+struct {
   uint16 slot[8] = {0};
-  uint3 input = 0;
-  uint3 output = 0;
+  uint32 addr = 0;  //read location of slot buffer
+  uint32 load = 0;  //write location of slot buffer
+  signed wait = 0;  //number of clocks before next slot load
 
-  bool stalled = true;
-  unsigned slots = 0;
-  signed wait = 0;
-  uint32 addr = 0;
+  auto empty() const { return addr == load; }
+  auto full() const { return load - addr == 16; }
 } prefetch;
 
-auto prefetch_stall() -> void;
-auto prefetch_start(uint32 addr) -> void;
+auto prefetch_sync(uint32 addr) -> void;
 auto prefetch_step(unsigned clocks) -> void;
 auto prefetch_wait() -> void;
-auto prefetch_take() -> uint16;
+auto prefetch_read() -> uint16;
diff --git a/gba/cpu/registers.cpp b/gba/cpu/registers.cpp
index aebc7b99..fff279ca 100644
--- a/gba/cpu/registers.cpp
+++ b/gba/cpu/registers.cpp
@@ -221,9 +221,6 @@ auto CPU::Registers::WaitControl::operator=(uint16 source) -> uint16 {
   prefetch = (source >> 14) & 1;
   gametype = (source >> 15) & 1;
   swait[3] = nwait[3];
-
-  cpu.prefetch.stalled = true;
-
   return operator uint16();
 }
 
diff --git a/gba/cpu/state.hpp b/gba/cpu/state.hpp
index 4926a3d8..f43c8340 100644
--- a/gba/cpu/state.hpp
+++ b/gba/cpu/state.hpp
@@ -5,3 +5,7 @@ struct Pending {
     bool hdma;
   } dma;
 } pending;
+
+struct Active {
+  bool dma;
+} active;
diff --git a/gba/gba.hpp b/gba/gba.hpp
index 6336ea5a..19ec42b6 100644
--- a/gba/gba.hpp
+++ b/gba/gba.hpp
@@ -21,8 +21,16 @@ namespace GameBoyAdvance {
 #include <libco/libco.h>
 
 namespace GameBoyAdvance {
-  enum : unsigned { Byte = 8, Half = 16, Word = 32 };
-  enum : bool { Nonsequential = 0, Sequential = 1 };
+  enum : unsigned {       //mode flags for bus_read, bus_write:
+    Nonsequential =   1,  //N cycle
+    Sequential    =   2,  //S cycle
+    Prefetch      =   4,  //instruction fetch (eligible for prefetch)
+    Byte          =   8,  //8-bit access
+    Half          =  16,  //16-bit access
+    Word          =  32,  //32-bit access
+    Load          =  64,  //load operation
+    Store         = 128,  //store operation
+  };
 
   struct Thread {
     ~Thread() {
diff --git a/gba/memory/memory.cpp b/gba/memory/memory.cpp
index 16726c74..d5b1d8e0 100644
--- a/gba/memory/memory.cpp
+++ b/gba/memory/memory.cpp
@@ -3,12 +3,11 @@
 namespace GameBoyAdvance {
 
 #include "mmio.cpp"
-#include "serialization.cpp"
 Bus bus;
 
 struct UnmappedMemory : Memory {
-  auto read(uint32 addr, uint32 size) -> uint32 { return 0; }
-  auto write(uint32 addr, uint32 size, uint32 word) -> void {}
+  auto read(unsigned mode, uint32 addr) -> uint32 override { return 0; }
+  auto write(unsigned mode, uint32 addr, uint32 word) -> void override {}
 };
 
 static UnmappedMemory unmappedMemory;
@@ -31,74 +30,6 @@ auto Bus::mirror(uint32 addr, uint32 size) -> uint32 {
   return base;
 }
 
-auto Bus::wait(uint32 addr, uint32 size, bool mode) -> unsigned {
-  switch(addr & 0x0f000000) {
-  case 0x00000000: return 1;
-  case 0x01000000: return 1;
-  case 0x02000000: return (16 - cpu.regs.memory.control.ewramwait) * (size == Word ? 2 : 1);
-  case 0x03000000: return 1;
-  case 0x04000000: return 1;
-  case 0x05000000: return 1 + (size == Word);
-  case 0x06000000: return 1 + (size == Word);
-  case 0x07000000: return 1;
-  default: {
-    unsigned n = cpu.regs.wait.control.nwait[addr >> 25 & 3];
-    unsigned s = cpu.regs.wait.control.swait[addr >> 25 & 3];
-
-    static unsigned timing[] = {5, 4, 3, 9};
-    n = timing[n];
-
-    switch(addr & 0x0e000000) {
-    case 0x08000000: s = s ? 2 : 3; break;
-    case 0x0a000000: s = s ? 2 : 5; break;
-    case 0x0c000000: s = s ? 2 : 9; break;
-    case 0x0e000000: s = n; break;
-    }
-
-    bool sequential = (mode == Sequential);
-    if((addr & 0x1fffe) == 0) sequential = false;  //N cycle on 16-bit ROM crossing 128KB page boundary (RAM S==N)
-
-    unsigned clocks = sequential ? s : n;
-    if(size == Word) clocks += s;  //16-bit bus requires two transfers for words
-    return clocks;
-  }
-  }
-}
-
-auto Bus::read(uint32 addr, uint32 size) -> uint32 {
-  switch(addr & 0x0f000000) {
-  case 0x00000000: return bios.read(addr, size);
-  case 0x01000000: return bios.read(addr, size);
-  case 0x02000000: return cpu.ewram_read(addr, size);
-  case 0x03000000: return cpu.iwram_read(addr, size);
-  case 0x4000000:
-    if((addr & 0xfffffc00) == 0x04000000) return mmio[addr & 0x3ff]->read(addr, size);
-    if((addr & 0xff00ffff) == 0x04000800) return ((MMIO&)cpu).read(0x04000800 | (addr & 3), size);
-    return cpu.pipeline.fetch.instruction;
-  case 0x05000000: return ppu.pram_read(addr, size);
-  case 0x06000000: return ppu.vram_read(addr, size);
-  case 0x07000000: return ppu.oam_read(addr, size);
-  default: return cartridge.read(addr, size);
-  }
-}
-
-auto Bus::write(uint32 addr, uint32 size, uint32 word) -> void {
-  switch(addr & 0x0f000000) {
-  case 0x00000000: return;
-  case 0x01000000: return;
-  case 0x02000000: return cpu.ewram_write(addr, size, word);
-  case 0x03000000: return cpu.iwram_write(addr, size, word);
-  case 0x04000000:
-    if((addr & 0xfffffc00) == 0x04000000) return mmio[addr & 0x3ff]->write(addr, size, word);
-    if((addr & 0xff00ffff) == 0x04000800) return ((MMIO&)cpu).write(0x04000800 | (addr & 3), size, word);
-    return;
-  case 0x05000000: return ppu.pram_write(addr, size, word);
-  case 0x06000000: return ppu.vram_write(addr, size, word);
-  case 0x07000000: return ppu.oam_write(addr, size, word);
-  default: return cartridge.write(addr, size, word);
-  }
-}
-
 auto Bus::power() -> void {
   for(auto n : range(0x400)) mmio[n] = &unmappedMemory;
 }
diff --git a/gba/memory/memory.hpp b/gba/memory/memory.hpp
index f1852cea..ec9b9434 100644
--- a/gba/memory/memory.hpp
+++ b/gba/memory/memory.hpp
@@ -1,26 +1,21 @@
 struct Memory {
-  virtual auto read(uint32 addr, uint32 size) -> uint32 = 0;
-  virtual auto write(uint32 addr, uint32 size, uint32 word) -> void = 0;
+  virtual auto read(unsigned mode, uint32 addr) -> uint32 = 0;
+  virtual auto write(unsigned mode, uint32 addr, uint32 word) -> void = 0;
 };
 
 struct MMIO : Memory {
   virtual auto read(uint32 addr) -> uint8 = 0;
   virtual auto write(uint32 addr, uint8 data) -> void = 0;
-  auto read(uint32 addr, uint32 size) -> uint32;
-  auto write(uint32 addr, uint32 size, uint32 word) -> void;
+  auto read(unsigned mode, uint32 addr) -> uint32 final;
+  auto write(unsigned mode, uint32 addr, uint32 word) -> void final;
 };
 
-struct Bus : Memory {
+struct Bus {
   static auto mirror(uint32 addr, uint32 size) -> uint32;
 
-  auto wait(uint32 addr, uint32 size, bool mode) -> unsigned;
-  auto read(uint32 addr, uint32 size) -> uint32;
-  auto write(uint32 addr, uint32 size, uint32 word) -> void;
   auto power() -> void;
 
-  auto serialize(serializer&) -> void;
-
-  Memory* mmio[0x400]{nullptr};
+  Memory* mmio[0x400] = {nullptr};
 };
 
 extern Bus bus;
diff --git a/gba/memory/mmio.cpp b/gba/memory/mmio.cpp
index 8ad6e897..29b8cbe7 100644
--- a/gba/memory/mmio.cpp
+++ b/gba/memory/mmio.cpp
@@ -1,43 +1,35 @@
-auto MMIO::read(uint32 addr, uint32 size) -> uint32 {
+auto MMIO::read(unsigned mode, uint32 addr) -> uint32 {
   uint32 word = 0;
 
-  switch(size) {
-  case Word:
+  if(mode & Word) {
     addr &= ~3;
     word |= read(addr + 0) <<  0;
     word |= read(addr + 1) <<  8;
     word |= read(addr + 2) << 16;
     word |= read(addr + 3) << 24;
-    break;
-  case Half:
+  } else if(mode & Half) {
     addr &= ~1;
     word |= read(addr + 0) <<  0;
     word |= read(addr + 1) <<  8;
-    break;
-  case Byte:
+  } else if(mode & Byte) {
     word |= read(addr + 0) <<  0;
-    break;
   }
 
   return word;
 }
 
-auto MMIO::write(uint32 addr, uint32 size, uint32 word) -> void {
-  switch(size) {
-  case Word:
+auto MMIO::write(unsigned mode, uint32 addr, uint32 word) -> void {
+  if(mode & Word) {
     addr &= ~3;
     write(addr + 0, word >>  0);
     write(addr + 1, word >>  8);
     write(addr + 2, word >> 16);
     write(addr + 3, word >> 24);
-    break;
-  case Half:
+  } else if(mode & Half) {
     addr &= ~1;
     write(addr + 0, word >>  0);
     write(addr + 1, word >>  8);
-    break;
-  case Byte:
+  } else if(mode & Byte) {
     write(addr + 0, word >>  0);
-    break;
   }
 }
diff --git a/gba/memory/serialization.cpp b/gba/memory/serialization.cpp
deleted file mode 100644
index 1987ee2c..00000000
--- a/gba/memory/serialization.cpp
+++ /dev/null
@@ -1,2 +0,0 @@
-auto Bus::serialize(serializer& s) -> void {
-}
diff --git a/gba/ppu/background.cpp b/gba/ppu/background.cpp
index 79515179..46178473 100644
--- a/gba/ppu/background.cpp
+++ b/gba/ppu/background.cpp
@@ -48,7 +48,7 @@ void PPU::render_background_linear(Registers::Background& bg) {
       if(bg.control.screensize & 1) if(tx & 32) offset += 32 * 32;
       if(bg.control.screensize & 2) if(ty & 32) offset += 32 * 32 * (1 + (bg.control.screensize & 1));
       offset = basemap + offset * 2;
-      uint16 mapdata = vram_read(offset, Half);
+      uint16 mapdata = vram_read(Half, offset);
 
       tile.character = mapdata >>  0;
       tile.hflip     = mapdata >> 10;
@@ -57,12 +57,12 @@ void PPU::render_background_linear(Registers::Background& bg) {
 
       if(bg.control.colormode == 0) {
         offset = basechr + tile.character * 32 + (py ^ (tile.vflip ? 7 : 0)) * 4;
-        uint32 word = vram_read(offset, Word);
+        uint32 word = vram_read(Word, offset);
         for(unsigned n = 0; n < 8; n++) data[n] = (word >> (n * 4)) & 15;
       } else {
         offset = basechr + tile.character * 64 + (py ^ (tile.vflip ? 7 : 0)) * 8;
-        uint32 wordlo = vram_read(offset + 0, Word);
-        uint32 wordhi = vram_read(offset + 4, Word);
+        uint32 wordlo = vram_read(Word, offset + 0);
+        uint32 wordhi = vram_read(Word, offset + 4);
         for(unsigned n = 0; n < 4; n++) data[0 + n] = (wordlo >> (n * 8)) & 255;
         for(unsigned n = 0; n < 4; n++) data[4 + n] = (wordhi >> (n * 8)) & 255;
       }
@@ -122,7 +122,7 @@ void PPU::render_background_bitmap(Registers::Background& bg) {
 
   unsigned width  = regs.control.bgmode == 5 ? 160 : 240;
   unsigned height = regs.control.bgmode == 5 ? 128 : 160;
-  unsigned size   = depth ? Half : Byte;
+  unsigned mode   = depth ? Half : Byte;
 
   if(bg.control.mosaic == false || (regs.vcounter % (1 + regs.mosaic.bgvsize)) == 0) {
     bg.hmosaic = bg.lx;
@@ -138,7 +138,7 @@ void PPU::render_background_bitmap(Registers::Background& bg) {
 
     if(px < width && py < height) {
       unsigned offset = py * width + px;
-      unsigned color  = vram_read(basemap + (offset << depth), size);
+      unsigned color  = vram_read(mode, basemap + (offset << depth));
 
       if(depth || color) {  //8bpp color 0 is transparent; 15bpp color is always opaque
         if(depth == 0) color = pram[color];
diff --git a/gba/ppu/memory.cpp b/gba/ppu/memory.cpp
index de319fcf..d87e5638 100644
--- a/gba/ppu/memory.cpp
+++ b/gba/ppu/memory.cpp
@@ -1,65 +1,62 @@
-uint32 PPU::vram_read(uint32 addr, uint32 size) {
+auto PPU::vram_read(unsigned mode, uint32 addr) -> uint32 {
   addr &= (addr & 0x10000) ? 0x17fff : 0x0ffff;
 
-  switch(size) {
-  case Word:
+  if(mode & Word) {
     addr &= ~3;
     return vram[addr + 0] << 0 | vram[addr + 1] << 8 | vram[addr + 2] << 16 | vram[addr + 3] << 24;
-  case Half:
+  } else if(mode & Half) {
     addr &= ~1;
     return vram[addr + 0] << 0 | vram[addr + 1] << 8;
-  case Byte:
+  } else if(mode & Byte) {
     return vram[addr];
+  } else {
+    throw;
   }
 }
 
-void PPU::vram_write(uint32 addr, uint32 size, uint32 word) {
+auto PPU::vram_write(unsigned mode, uint32 addr, uint32 word) -> void {
   addr &= (addr & 0x10000) ? 0x17fff : 0x0ffff;
 
-  switch(size) {
-  case Word:
+  if(mode & Word) {
     addr &= ~3;
     vram[addr + 0] = word >>  0;
     vram[addr + 1] = word >>  8;
     vram[addr + 2] = word >> 16;
     vram[addr + 3] = word >> 24;
-    break;
-  case Half:
+  } else if(mode & Half) {
     addr &= ~1;
     vram[addr + 0] = word >>  0;
     vram[addr + 1] = word >>  8;
-    break;
-  case Byte:
+  } else if(mode & Byte) {
     addr &= ~1;
     vram[addr + 0] = word;
     vram[addr + 1] = word;
-    break;
   }
 }
 
-uint32 PPU::pram_read(uint32 addr, uint32 size) {
-  if(size == Word) return pram_read(addr & ~2, Half) << 0 | pram_read(addr | 2, Half) << 16;
-  if(size == Byte) return pram_read(addr, Half) >> ((addr & 1) * 8);
+auto PPU::pram_read(unsigned mode, uint32 addr) -> uint32 {
+  if(mode & Word) return pram_read(Half, addr & ~2) << 0 | pram_read(Half, addr | 2) << 16;
+  if(mode & Byte) return pram_read(Half, addr) >> ((addr & 1) * 8);
   return pram[addr >> 1 & 511];
 }
 
-void PPU::pram_write(uint32 addr, uint32 size, uint32 word) {
-  if(size == Word) {
-    pram_write(addr & ~2, Half, word >>  0);
-    pram_write(addr |  2, Half, word >> 16);
+auto PPU::pram_write(unsigned mode, uint32 addr, uint32 word) -> void {
+  if(mode & Word) {
+    pram_write(Half, addr & ~2, word >>  0);
+    pram_write(Half, addr |  2, word >> 16);
     return;
   }
 
-  if(size == Byte) {
-    return pram_write(addr, Half, word << 8 | word << 0);
+  if(mode & Byte) {
+    return pram_write(Half, addr, word << 8 | word << 0);
   }
 
   pram[addr >> 1 & 511] = word & 0x7fff;
 }
 
-uint32 PPU::oam_read(uint32 addr, uint32 size) {
-  if(size == Word) return oam_read(addr & ~2, Half) << 0 | oam_read(addr | 2, Half) << 16;
-  if(size == Byte) return oam_read(addr, Half) >> ((addr & 1) * 8);
+auto PPU::oam_read(unsigned mode, uint32 addr) -> uint32 {
+  if(mode & Word) return oam_read(Half, addr & ~2) << 0 | oam_read(Half, addr | 2) << 16;
+  if(mode & Byte) return oam_read(Half, addr) >> ((addr & 1) * 8);
 
   auto& obj = object[addr >> 3 & 127];
   auto& par = objectparam[addr >> 5 & 31];
@@ -101,15 +98,15 @@ uint32 PPU::oam_read(uint32 addr, uint32 size) {
   }
 }
 
-void PPU::oam_write(uint32 addr, uint32 size, uint32 word) {
-  if(size == Word) {
-    oam_write(addr & ~2, Half, word >>  0);
-    oam_write(addr |  2, Half, word >> 16);
+auto PPU::oam_write(unsigned mode, uint32 addr, uint32 word) -> void {
+  if(mode & Word) {
+    oam_write(Half, addr & ~2, word >>  0);
+    oam_write(Half, addr |  2, word >> 16);
     return;
   }
 
-  if(size == Byte) {
-    return oam_write(addr, Half, word << 8 | word << 0);
+  if(mode & Byte) {
+    return oam_write(Half, addr, word << 8 | word << 0);
   }
 
   auto& obj = object[addr >> 3 & 127];
diff --git a/gba/ppu/ppu.hpp b/gba/ppu/ppu.hpp
index 7d801ebd..bc017fb1 100644
--- a/gba/ppu/ppu.hpp
+++ b/gba/ppu/ppu.hpp
@@ -16,14 +16,14 @@ struct PPU : Thread, MMIO {
   uint8 read(uint32 addr);
   void write(uint32 addr, uint8 byte);
 
-  uint32 vram_read(uint32 addr, uint32 size);
-  void vram_write(uint32 addr, uint32 size, uint32 word);
+  auto vram_read(unsigned mode, uint32 addr) -> uint32;
+  auto vram_write(unsigned mode, uint32 addr, uint32 word) -> void;
 
-  uint32 pram_read(uint32 addr, uint32 size);
-  void pram_write(uint32 addr, uint32 size, uint32 word);
+  auto pram_read(unsigned mode, uint32 addr) -> uint32;
+  auto pram_write(unsigned mode, uint32 addr, uint32 word) -> void;
 
-  uint32 oam_read(uint32 addr, uint32 size);
-  void oam_write(uint32 addr, uint32 size, uint32 word);
+  auto oam_read(unsigned mode, uint32 addr) -> uint32;
+  auto oam_write(unsigned mode, uint32 addr, uint32 word) -> void;
 
   void render_backgrounds();
   void render_background_linear(Registers::Background&);
diff --git a/gba/system/bios.cpp b/gba/system/bios.cpp
index 9f5514dc..3aebf873 100644
--- a/gba/system/bios.cpp
+++ b/gba/system/bios.cpp
@@ -1,20 +1,22 @@
-uint32 BIOS::read(uint32 addr, uint32 size) {
-  //GBA BIOS is read-protected; only the BIOS itself can read its own memory
-  //when accessed elsewhere; this returns the last value read by the BIOS program
-  if(cpu.r(15) >= 0x02000000) return mdr;
-
-  if(size == Word) return mdr = read(addr &~ 2, Half) << 0 | read(addr | 2, Half) << 16;
-  if(size == Half) return mdr = read(addr &~ 1, Byte) << 0 | read(addr | 1, Byte) <<  8;
-  return mdr = data[addr & 0x3fff];
-}
-
-void BIOS::write(uint32 addr, uint32 size, uint32 word) {
-}
-
 BIOS::BIOS() {
-  data = new uint8[size = 16384]();
+  size = 16384;
+  data = new uint8[size]();
 }
 
 BIOS::~BIOS() {
   delete[] data;
 }
+
+auto BIOS::read(unsigned mode, uint32 addr) -> uint32 {
+  //GBA BIOS is read-protected; only the BIOS itself can read its own memory
+  //when accessed elsewhere; this returns the last value read by the BIOS program
+  if(cpu.r(15) >= 0x02000000) return mdr;
+
+  if(mode & Word) return mdr = read(Half, addr &~ 2) << 0 | read(Half, addr | 2) << 16;
+  if(mode & Half) return mdr = read(Byte, addr &~ 1) << 0 | read(Byte, addr | 1) <<  8;
+
+  return mdr = data[addr & 0x3fff];
+}
+
+auto BIOS::write(unsigned mode, uint32 addr, uint32 word) -> void {
+}
diff --git a/gba/system/serialization.cpp b/gba/system/serialization.cpp
index a8d648e8..f3f044b8 100644
--- a/gba/system/serialization.cpp
+++ b/gba/system/serialization.cpp
@@ -43,7 +43,6 @@ void System::serialize_all(serializer& s) {
   cpu.serialize(s);
   ppu.serialize(s);
   apu.serialize(s);
-  bus.serialize(s);
   player.serialize(s);
 }
 
diff --git a/gba/system/system.hpp b/gba/system/system.hpp
index 7e0686d2..6e82244d 100644
--- a/gba/system/system.hpp
+++ b/gba/system/system.hpp
@@ -3,15 +3,15 @@ enum class Input : unsigned {
 };
 
 struct BIOS : Memory {
-  uint8* data;
-  unsigned size;
-  uint32 mdr;
-
-  uint32 read(uint32 addr, uint32 size);
-  void write(uint32 addr, uint32 size, uint32 word);
-
   BIOS();
   ~BIOS();
+
+  auto read(unsigned mode, uint32 addr) -> uint32 override;
+  auto write(unsigned mode, uint32 addr, uint32 word) -> void override;
+
+  uint8* data = nullptr;
+  unsigned size = 0;
+  uint32 mdr = 0;
 };
 
 struct System {
diff --git a/processor/arm/arm.cpp b/processor/arm/arm.cpp
index 59656fd7..afce5e75 100644
--- a/processor/arm/arm.cpp
+++ b/processor/arm/arm.cpp
@@ -15,6 +15,7 @@ auto ARM::power() -> void {
   processor.power();
   vector(0x00000000, Processor::Mode::SVC);
   pipeline.reload = true;
+  pipeline.nonsequential = true;
   crash = false;
   r(15).modify = [&] {
     pipeline.reload = true;
@@ -29,42 +30,41 @@ auto ARM::exec() -> void {
 }
 
 auto ARM::idle() -> void {
-  processor.nonsequential = true;
-  bus_idle(r(15));
+  pipeline.nonsequential = true;
+  return bus_idle();
 }
 
-auto ARM::read(uint32 addr, uint32 size, bool mode) -> uint32 {
-  if(processor.nonsequential) processor.nonsequential = false, mode = Nonsequential;
-  return bus_read(addr, size, mode);
+auto ARM::read(unsigned mode, uint32 addr) -> uint32 {
+  return bus_read(mode, addr);
 }
 
-auto ARM::load(uint32 addr, uint32 size, bool mode) -> uint32 {
-  if(processor.nonsequential) processor.nonsequential = false, mode = Nonsequential;
-  uint32 word = bus_read(addr, size, mode);
+auto ARM::load(unsigned mode, uint32 addr) -> uint32 {
+  pipeline.nonsequential = true;
+  uint32 word = bus_read(Load | mode, addr);
 
-  if(size == Half) { word &= 0xffff; word |= word << 16; }
-  if(size == Byte) { word &= 0xff; word |= word << 8; word |= word << 16; }
+  if(mode & Half) { word &= 0xffff; word |= word << 16; }
+  if(mode & Byte) { word &= 0xff; word |= word << 8; word |= word << 16; }
 
   word = ror(word, 8 * (addr & 3));
   idle();
 
-  if(size == Half) word &= 0xffff;
-  if(size == Byte) word &= 0xff;
+  if(mode & Half) word &= 0xffff;
+  if(mode & Byte) word &= 0xff;
   return word;
 }
 
-auto ARM::write(uint32 addr, uint32 size, bool mode, uint32 word) -> void {
-  if(processor.nonsequential) processor.nonsequential = false, mode = Nonsequential;
-  return bus_write(addr, size, mode, word);
+auto ARM::write(unsigned mode, uint32 addr, uint32 word) -> void {
+  pipeline.nonsequential = true;
+  return bus_write(mode, addr, word);
 }
 
-auto ARM::store(uint32 addr, uint32 size, bool mode, uint32 word) -> void {
-  if(size == Half) { word &= 0xffff; word |= word << 16; }
-  if(size == Byte) { word &= 0xff; word |= word << 8; word |= word << 16; }
+auto ARM::store(unsigned mode, uint32 addr, uint32 word) -> void {
+  pipeline.nonsequential = true;
 
-  if(processor.nonsequential) processor.nonsequential = false, mode = Nonsequential;
-  bus_write(addr, size, mode, word);
-  processor.nonsequential = true;
+  if(mode & Half) { word &= 0xffff; word |= word << 16; }
+  if(mode & Byte) { word &= 0xff; word |= word << 8; word |= word << 16; }
+
+  return bus_write(Store | mode, addr, word);
 }
 
 auto ARM::vector(uint32 addr, Processor::Mode mode) -> void {
diff --git a/processor/arm/arm.hpp b/processor/arm/arm.hpp
index e9a93d15..3dcfee07 100644
--- a/processor/arm/arm.hpp
+++ b/processor/arm/arm.hpp
@@ -4,12 +4,20 @@
 namespace Processor {
 
 //Supported Models:
-//* ARMv3 (ST018)
-//* ARMv4 (ARM7TDMI)
+//* ARMv3  (ARM60)
+//* ARMv4T (ARM7TDMI)
 
 struct ARM {
-  enum : unsigned { Byte = 8, Half = 16, Word = 32 };
-  enum : bool { Nonsequential = 0, Sequential = 1 };
+  enum : unsigned {       //mode flags for bus_read, bus_write:
+    Nonsequential =   1,  //N cycle
+    Sequential    =   2,  //S cycle
+    Prefetch      =   4,  //instruction fetch (eligible for prefetch)
+    Byte          =   8,  //8-bit access
+    Half          =  16,  //16-bit access
+    Word          =  32,  //32-bit access
+    Load          =  64,  //load operation
+    Store         = 128,  //store operation
+  };
 
   #include "registers.hpp"
   #include "instructions-arm.hpp"
@@ -17,18 +25,18 @@ struct ARM {
   #include "disassembler.hpp"
 
   virtual auto step(unsigned clocks) -> void = 0;
-  virtual auto bus_idle(uint32 addr) -> void = 0;
-  virtual auto bus_read(uint32 addr, uint32 size, bool mode) -> uint32 = 0;
-  virtual auto bus_write(uint32 addr, uint32 size, bool mode, uint32 word) -> void = 0;
+  virtual auto bus_idle() -> void = 0;
+  virtual auto bus_read(unsigned mode, uint32 addr) -> uint32 = 0;
+  virtual auto bus_write(unsigned mode, uint32 addr, uint32 word) -> void = 0;
 
   //arm.cpp
   auto power() -> void;
   auto exec() -> void;
   auto idle() -> void;
-  auto read(uint32 addr, uint32 size, bool mode) -> uint32;
-  auto load(uint32 addr, uint32 size, bool mode) -> uint32;
-  auto write(uint32 addr, uint32 size, bool mode, uint32 word) -> void;
-  auto store(uint32 addr, uint32 size, bool mode, uint32 word) -> void;
+  auto read(unsigned mode, uint32 addr) -> uint32;
+  auto load(unsigned mode, uint32 addr) -> uint32;
+  auto write(unsigned mode, uint32 addr, uint32 word) -> void;
+  auto store(unsigned mode, uint32 addr, uint32 word) -> void;
   auto vector(uint32 addr, Processor::Mode mode) -> void;
 
   //algorithms.cpp
@@ -44,14 +52,15 @@ struct ARM {
   auto rrx(uint32 source) -> uint32;
 
   //step.cpp
+  auto pipeline_step() -> void;
   auto arm_step() -> void;
   auto thumb_step() -> void;
 
   //serialization.cpp
   auto serialize(serializer&) -> void;
 
-  bool trace{false};
-  uintmax_t instructions{0};
+  bool trace = false;
+  uintmax_t instructions = 0;
 };
 
 }
diff --git a/processor/arm/disassembler.cpp b/processor/arm/disassembler.cpp
index a8ac573e..f08d36e1 100644
--- a/processor/arm/disassembler.cpp
+++ b/processor/arm/disassembler.cpp
@@ -32,7 +32,7 @@ auto ARM::disassemble_arm_instruction(uint32 pc) -> string {
 
   string output{hex<8>(pc), "  "};
 
-  uint32 instruction = read(pc & ~3, Word, Nonsequential);
+  uint32 instruction = read(Word | Nonsequential, pc & ~3);
   output.append(hex<8>(instruction), "  ");
 
   //multiply()
@@ -134,7 +134,7 @@ auto ARM::disassemble_arm_instruction(uint32 pc) -> string {
     if(pre == 1) output.append("]");
     if(pre == 0 || writeback == 1) output.append("!");
 
-    if(rn == 15) output.append(" =0x", hex<4>(read(pc + 8 + (up ? +immediate : -immediate), Half, Nonsequential)));
+    if(rn == 15) output.append(" =0x", hex<4>(read(Half | Nonsequential, pc + 8 + (up ? +immediate : -immediate))));
     return output;
   }
 
@@ -184,8 +184,8 @@ auto ARM::disassemble_arm_instruction(uint32 pc) -> string {
     if(pre == 1) output.append("]");
     if(pre == 0 || writeback == 1) output.append("!");
 
-    if(rn == 15 && half == 1) output.append(" =0x", hex<4>(read(pc + 8 + (up ? +immediate : -immediate), Half, Nonsequential)));
-    if(rn == 15 && half == 0) output.append(" =0x", hex<2>(read(pc + 8 + (up ? +immediate : -immediate), Byte, Nonsequential)));
+    if(rn == 15 && half == 1) output.append(" =0x", hex<4>(read(Half | Nonsequential, pc + 8 + (up ? +immediate : -immediate))));
+    if(rn == 15 && half == 0) output.append(" =0x", hex<2>(read(Byte | Nonsequential, pc + 8 + (up ? +immediate : -immediate))));
     return output;
   }
 
@@ -359,7 +359,7 @@ auto ARM::disassemble_arm_instruction(uint32 pc) -> string {
     if(pre == 1) output.append("]");
     if(pre == 0 || writeback == 1) output.append("!");
 
-    if(rn == 15) output.append(" =0x", hex<8>(read(pc + 8 + (up ? +immediate : -immediate), byte ? Byte : Word, Nonsequential)));
+    if(rn == 15) output.append(" =0x", hex<8>(read((byte ? Byte : Word) | Nonsequential, pc + 8 + (up ? +immediate : -immediate))));
     return output;
   }
 
@@ -457,7 +457,7 @@ auto ARM::disassemble_thumb_instruction(uint32 pc) -> string {
 
   string output{hex<8>(pc), "  "};
 
-  uint16 instruction = read(pc & ~1, Half, Nonsequential);
+  uint16 instruction = read(Half | Nonsequential, pc & ~1);
   output.append(hex<4>(instruction), "  ");
 
   //adjust_register()
@@ -571,7 +571,7 @@ auto ARM::disassemble_thumb_instruction(uint32 pc) -> string {
 
     unsigned rm = ((pc + 4) & ~3) + displacement * 4;
     output.append("ldr ", registers[rd], ",[pc,#0x", hex<3>(rm), "]");
-    output.append(" =0x", hex<8>(read(rm, Word, Nonsequential)));
+    output.append(" =0x", hex<8>(read(Word | Nonsequential, rm)));
 
     return output;
   }
@@ -740,7 +740,7 @@ auto ARM::disassemble_thumb_instruction(uint32 pc) -> string {
   //bl address
   if((instruction & 0xf800) == 0xf000) {
     uint11 offsethi = instruction;
-    instruction = read((pc & ~1) + 2, Half, Nonsequential);
+    instruction = read(Half | Nonsequential, (pc & ~1) + 2);
     uint11 offsetlo = instruction;
 
     int22 displacement = (offsethi << 11) | (offsetlo << 0);
diff --git a/processor/arm/instructions-arm.cpp b/processor/arm/instructions-arm.cpp
index 21a5de5a..27d8016d 100644
--- a/processor/arm/instructions-arm.cpp
+++ b/processor/arm/instructions-arm.cpp
@@ -137,8 +137,8 @@ auto ARM::arm_op_memory_swap() {
   uint4 d = instruction() >> 12;
   uint4 m = instruction();
 
-  uint32 word = load(r(n), byte ? Byte : Word, Nonsequential);
-  store(r(n), byte ? Byte : Word, Nonsequential, r(m));
+  uint32 word = load((byte ? Byte : Word) | Nonsequential, r(n));
+  store((byte ? Byte : Word) | Nonsequential, r(n), r(m));
   r(d) = word;
 }
 
@@ -166,8 +166,8 @@ auto ARM::arm_op_move_half_register() {
   uint32 rm = r(m);
 
   if(pre == 1) rn = up ? rn + rm : rn - rm;
-  if(l == 1) r(d) = load(rn, Half, Nonsequential);
-  if(l == 0) store(rn, Half, Nonsequential, r(d));
+  if(l == 1) r(d) = load(Half | Nonsequential, rn);
+  if(l == 0) store(Half | Nonsequential, rn, r(d));
   if(pre == 0) rn = up ? rn + rm : rn - rm;
 
   if(pre == 0 || writeback == 1) r(n) = rn;
@@ -199,8 +199,8 @@ auto ARM::arm_op_move_half_immediate() {
   uint8 immediate = (ih << 4) + (il << 0);
 
   if(pre == 1) rn = up ? rn + immediate : rn - immediate;
-  if(l == 1) r(d) = load(rn, Half, Nonsequential);
-  if(l == 0) store(rn, Half, Nonsequential, r(d));
+  if(l == 1) r(d) = load(Half | Nonsequential, rn);
+  if(l == 0) store(Half | Nonsequential, rn, r(d));
   if(pre == 0) rn = up ? rn + immediate : rn - immediate;
 
   if(pre == 0 || writeback == 1) r(n) = rn;
@@ -230,7 +230,7 @@ auto ARM::arm_op_load_register() {
   uint32 rm = r(m);
 
   if(pre == 1) rn = up ? rn + rm : rn - rm;
-  uint32 word = load(rn, half ? Half : Byte, Nonsequential);
+  uint32 word = load((half ? Half : Byte) | Nonsequential, rn);
   r(d) = half ? (int16)word : (int8)word;
   if(pre == 0) rn = up ? rn + rm : rn - rm;
 
@@ -263,7 +263,7 @@ auto ARM::arm_op_load_immediate() {
   uint8 immediate = (ih << 4) + (il << 0);
 
   if(pre == 1) rn = up ? rn + immediate : rn - immediate;
-  uint32 word = load(rn, half ? Half : Byte, Nonsequential);
+  uint32 word = load((half ? Half : Byte) | Nonsequential, rn);
   r(d) = half ? (int16)word : (int8)word;
   if(pre == 0) rn = up ? rn + immediate : rn - immediate;
 
@@ -437,8 +437,8 @@ auto ARM::arm_op_move_immediate_offset() {
   auto& rd = r(d);
 
   if(pre == 1) rn = up ? rn + rm : rn - rm;
-  if(l == 1) rd = load(rn, byte ? Byte : Word, Nonsequential);
-  if(l == 0) store(rn, byte ? Byte : Word, Nonsequential, rd);
+  if(l == 1) rd = load((byte ? Byte : Word) | Nonsequential, rn);
+  if(l == 0) store((byte ? Byte : Word) | Nonsequential, rn, rd);
   if(pre == 0) rn = up ? rn + rm : rn - rm;
 
   if(pre == 0 || writeback == 1) r(n) = rn;
@@ -482,8 +482,8 @@ auto ARM::arm_op_move_register_offset() {
   if(mode == 3) rm = rs ? ror(rm, rs) : rrx(rm);
 
   if(pre == 1) rn = up ? rn + rm : rn - rm;
-  if(l == 1) rd = load(rn, byte ? Byte : Word, Nonsequential);
-  if(l == 0) store(rn, byte ? Byte : Word, Nonsequential, rd);
+  if(l == 1) rd = load((byte ? Byte : Word) | Nonsequential, rn);
+  if(l == 0) store((byte ? Byte : Word) | Nonsequential, rn, rd);
   if(pre == 0) rn = up ? rn + rm : rn - rm;
 
   if(pre == 0 || writeback == 1) r(n) = rn;
@@ -521,11 +521,13 @@ auto ARM::arm_op_move_multiple() {
 
   if(usr) processor.setMode(Processor::Mode::USR);
 
+  unsigned sequential = Nonsequential;
   for(unsigned m = 0; m < 16; m++) {
     if(list & 1 << m) {
-      if(l == 1) r(m) = read(rn, Word, Nonsequential);
-      if(l == 0) write(rn, Word, Nonsequential, r(m));
+      if(l == 1) r(m) = read(Word | sequential, rn);
+      if(l == 0) write(Word | sequential, rn, r(m));
       rn += 4;
+      sequential = Sequential;
     }
   }
 
@@ -540,7 +542,7 @@ auto ARM::arm_op_move_multiple() {
       }
     }
   } else {
-    processor.nonsequential = true;
+    pipeline.nonsequential = true;
   }
 
   if(writeback) {
diff --git a/processor/arm/instructions-thumb.cpp b/processor/arm/instructions-thumb.cpp
index a8587389..83f27d80 100644
--- a/processor/arm/instructions-thumb.cpp
+++ b/processor/arm/instructions-thumb.cpp
@@ -145,7 +145,7 @@ auto ARM::thumb_op_load_literal() {
   uint8 displacement = instruction();
 
   unsigned rm = (r(15) & ~3) + displacement * 4;
-  r(d) = load(rm, Word, Nonsequential);
+  r(d) = load(Word | Nonsequential, rm);
 }
 
 //(ld(r,s),str){b,h} rd,[rn,rm]
@@ -161,14 +161,14 @@ auto ARM::thumb_op_move_register_offset() {
   uint3 d = instruction() >> 0;
 
   switch(opcode) {
-  case 0: store(r(n) + r(m), Word, Nonsequential, r(d));        break;  //STR
-  case 1: store(r(n) + r(m), Half, Nonsequential, r(d));        break;  //STRH
-  case 2: store(r(n) + r(m), Byte, Nonsequential, r(d));        break;  //STRB
-  case 3: r(d) =  (int8)load(r(n) + r(m), Byte, Nonsequential); break;  //LDSB
-  case 4: r(d) = load(r(n) + r(m), Word, Nonsequential);        break;  //LDR
-  case 5: r(d) = load(r(n) + r(m), Half, Nonsequential);        break;  //LDRH
-  case 6: r(d) = load(r(n) + r(m), Byte, Nonsequential);        break;  //LDRB
-  case 7: r(d) = (int16)load(r(n) + r(m), Half, Nonsequential); break;  //LDSH
+  case 0: store(Word | Nonsequential, r(n) + r(m), r(d));        break;  //STR
+  case 1: store(Half | Nonsequential, r(n) + r(m), r(d));        break;  //STRH
+  case 2: store(Byte | Nonsequential, r(n) + r(m), r(d));        break;  //STRB
+  case 3: r(d) =  (int8)load(Byte | Nonsequential, r(n) + r(m)); break;  //LDSB
+  case 4: r(d) = load(Word | Nonsequential, r(n) + r(m));        break;  //LDR
+  case 5: r(d) = load(Half | Nonsequential, r(n) + r(m));        break;  //LDRH
+  case 6: r(d) = load(Byte | Nonsequential, r(n) + r(m));        break;  //LDRB
+  case 7: r(d) = (int16)load(Half | Nonsequential, r(n) + r(m)); break;  //LDSH
   }
 }
 
@@ -184,8 +184,8 @@ auto ARM::thumb_op_move_word_immediate() {
   uint3 n = instruction() >> 3;
   uint3 d = instruction() >> 0;
 
-  if(l == 1) r(d) = load(r(n) + offset * 4, Word, Nonsequential);
-  if(l == 0) store(r(n) + offset * 4, Word, Nonsequential, r(d));
+  if(l == 1) r(d) = load(Word | Nonsequential, r(n) + offset * 4);
+  if(l == 0) store(Word | Nonsequential, r(n) + offset * 4, r(d));
 }
 
 //(ldr,str)b rd,[rn,#offset]
@@ -200,8 +200,8 @@ auto ARM::thumb_op_move_byte_immediate() {
   uint3 n = instruction() >> 3;
   uint3 d = instruction() >> 0;
 
-  if(l == 1) r(d) = load(r(n) + offset, Byte, Nonsequential);
-  if(l == 0) store(r(n) + offset, Byte, Nonsequential, r(d));
+  if(l == 1) r(d) = load(Byte | Nonsequential, r(n) + offset);
+  if(l == 0) store(Byte | Nonsequential, r(n) + offset, r(d));
 }
 
 //(ldr,str)h rd,[rn,#offset]
@@ -216,8 +216,8 @@ auto ARM::thumb_op_move_half_immediate() {
   uint3 n = instruction() >> 3;
   uint3 d = instruction() >> 0;
 
-  if(l == 1) r(d) = load(r(n) + offset * 2, Half, Nonsequential);
-  if(l == 0) store(r(n) + offset * 2, Half, Nonsequential, r(d));
+  if(l == 1) r(d) = load(Half | Nonsequential, r(n) + offset * 2);
+  if(l == 0) store(Half | Nonsequential, r(n) + offset * 2, r(d));
 }
 
 //(ldr,str) rd,[sp,#immediate]
@@ -230,8 +230,8 @@ auto ARM::thumb_op_move_stack() {
   uint3 d = instruction() >> 8;
   uint8 immediate = instruction();
 
-  if(l == 1) r(d) = load(r(13) + immediate * 4, Word, Nonsequential);
-  if(l == 0) store(r(13) + immediate * 4, Word, Nonsequential, r(d));
+  if(l == 1) r(d) = load(Word | Nonsequential, r(13) + immediate * 4);
+  if(l == 0) store(Word | Nonsequential, r(13) + immediate * 4, r(d));
 }
 
 //add rd,{pc,sp},#immediate
@@ -275,24 +275,30 @@ auto ARM::thumb_op_stack_multiple() {
   if(l == 1) sp = r(13);
   if(l == 0) sp = r(13) - (bit::count(list) + branch) * 4;
 
+  unsigned sequential = Nonsequential;
   for(unsigned m = 0; m < 8; m++) {
     if(list & 1 << m) {
-      if(l == 1) r(m) = read(sp, Word, Nonsequential);  //POP
-      if(l == 0) write(sp, Word, Nonsequential, r(m));  //PUSH
+      if(l == 1) r(m) = read(Word | sequential, sp);  //POP
+      if(l == 0) write(Word | sequential, sp, r(m));  //PUSH
       sp += 4;
+      sequential = Sequential;
     }
   }
 
   if(branch) {
     //note: ARMv5+ POP sets cpsr().t
-    if(l == 1) r(15) = read(sp, Word, Nonsequential);  //POP
-    if(l == 0) write(sp, Word, Nonsequential, r(14));  //PUSH
+    if(l == 1) r(15) = read(Word | Nonsequential, sp);  //POP
+    if(l == 0) write(Word | Nonsequential, sp, r(14));  //PUSH
     sp += 4;
   }
 
-  if(l == 1) idle();
-  if(l == 1) r(13) += (bit::count(list) + branch) * 4;
-  if(l == 0) r(13) -= (bit::count(list) + branch) * 4;
+  if(l == 1) {
+    idle();
+    r(13) += (bit::count(list) + branch) * 4;
+  } else {
+    pipeline.nonsequential = true;
+    r(13) -= (bit::count(list) + branch) * 4;
+  }
 }
 
 //(ldmia,stmia) rn!,{r...}
@@ -308,8 +314,8 @@ auto ARM::thumb_op_move_multiple() {
 
   for(unsigned m = 0; m < 8; m++) {
     if(list & 1 << m) {
-      if(l == 1) r(m) = read(rn, Word, Nonsequential);  //LDMIA
-      if(l == 0) write(rn, Word, Nonsequential, r(m));  //STMIA
+      if(l == 1) r(m) = read(Word | Nonsequential, rn);  //LDMIA
+      if(l == 0) write(Word | Nonsequential, rn, r(m));  //STMIA
       rn += 4;
     }
   }
diff --git a/processor/arm/registers.cpp b/processor/arm/registers.cpp
index efcf1621..da91d6f3 100644
--- a/processor/arm/registers.cpp
+++ b/processor/arm/registers.cpp
@@ -11,7 +11,6 @@ auto ARM::Processor::power() -> void {
   pc = 0;
 
   carryout = false;
-  nonsequential = false;
   irqline = false;
 
   cpsr = 0;
@@ -61,19 +60,4 @@ auto ARM::Processor::setMode(Mode mode) -> void {
   }
 }
 
-auto ARM::pipeline_step() -> void {
-  pipeline.execute = pipeline.decode;
-  pipeline.decode = pipeline.fetch;
-
-  if(cpsr().t == 0) {
-    r(15).data += 4;
-    pipeline.fetch.address = r(15) & ~3;
-    pipeline.fetch.instruction = read(pipeline.fetch.address, Word, Sequential);
-  } else {
-    r(15).data += 2;
-    pipeline.fetch.address = r(15) & ~1;
-    pipeline.fetch.instruction = read(pipeline.fetch.address, Half, Sequential);
-  }
-}
-
 #endif
diff --git a/processor/arm/registers.hpp b/processor/arm/registers.hpp
index 115268d9..ab332ced 100644
--- a/processor/arm/registers.hpp
+++ b/processor/arm/registers.hpp
@@ -50,6 +50,7 @@ struct PSR {
 
 struct Pipeline {
   bool reload = false;
+  bool nonsequential = false;
 
   struct Instruction {
     uint32 address = 0;
@@ -106,7 +107,6 @@ struct Processor {
   GPR pc;
   PSR cpsr;
   bool carryout = false;
-  bool nonsequential = false;
   bool irqline = false;
 
   GPR* r[16] = {nullptr};
@@ -120,8 +120,6 @@ Processor processor;
 Pipeline pipeline;
 bool crash = false;
 
-auto pipeline_step() -> void;
-
 alwaysinline auto r(unsigned n) -> GPR& { return *processor.r[n]; }
 alwaysinline auto cpsr() -> PSR& { return processor.cpsr; }
 alwaysinline auto spsr() -> PSR& { return *processor.spsr; }
diff --git a/processor/arm/serialization.cpp b/processor/arm/serialization.cpp
index cbf4ab6c..59c65a4d 100644
--- a/processor/arm/serialization.cpp
+++ b/processor/arm/serialization.cpp
@@ -55,10 +55,10 @@ auto ARM::serialize(serializer& s) -> void {
   s.integer(processor.pc.data);
   processor.cpsr.serialize(s);
   s.integer(processor.carryout);
-  s.integer(processor.nonsequential);
   s.integer(processor.irqline);
 
   s.integer(pipeline.reload);
+  s.integer(pipeline.nonsequential);
   s.integer(pipeline.execute.address);
   s.integer(pipeline.execute.instruction);
   s.integer(pipeline.decode.address);
diff --git a/processor/arm/step.cpp b/processor/arm/step.cpp
index c070f515..c88bbf11 100644
--- a/processor/arm/step.cpp
+++ b/processor/arm/step.cpp
@@ -1,10 +1,31 @@
+auto ARM::pipeline_step() -> void {
+  pipeline.execute = pipeline.decode;
+  pipeline.decode = pipeline.fetch;
+
+  unsigned sequential = Sequential;
+  if(pipeline.nonsequential) {
+    pipeline.nonsequential = false;
+    sequential = Nonsequential;
+  }
+
+  if(cpsr().t == 0) {
+    r(15).data += 4;
+    pipeline.fetch.address = r(15) & ~3;
+    pipeline.fetch.instruction = read(Prefetch | Word | sequential, pipeline.fetch.address);
+  } else {
+    r(15).data += 2;
+    pipeline.fetch.address = r(15) & ~1;
+    pipeline.fetch.instruction = read(Prefetch | Half | sequential, pipeline.fetch.address);
+  }
+}
+
 auto ARM::arm_step() -> void {
   if(pipeline.reload) {
     pipeline.reload = false;
     r(15).data &= ~3;
 
     pipeline.fetch.address = r(15) & ~3;
-    pipeline.fetch.instruction = read(pipeline.fetch.address, Word, Nonsequential);
+    pipeline.fetch.instruction = read(Prefetch | Word | Nonsequential, pipeline.fetch.address);
 
     pipeline_step();
   }
@@ -61,7 +82,7 @@ auto ARM::thumb_step() -> void {
     r(15).data &= ~1;
 
     pipeline.fetch.address = r(15) & ~1;
-    pipeline.fetch.instruction = read(pipeline.fetch.address, Half, Nonsequential);
+    pipeline.fetch.instruction = read(Prefetch | Half | Nonsequential, pipeline.fetch.address);
 
     pipeline_step();
   }
diff --git a/sfc/chip/armdsp/armdsp.hpp b/sfc/chip/armdsp/armdsp.hpp
index 30ca130f..019c31aa 100644
--- a/sfc/chip/armdsp/armdsp.hpp
+++ b/sfc/chip/armdsp/armdsp.hpp
@@ -1,4 +1,4 @@
-//ARMv3 (ARM6)
+//ARMv3 (ARM60)
 
 struct ArmDSP : Processor::ARM, Coprocessor {
   uint8* programROM;
@@ -11,9 +11,9 @@ struct ArmDSP : Processor::ARM, Coprocessor {
   void enter();
 
   void step(unsigned clocks) override;
-  void bus_idle(uint32 addr) override;
-  uint32 bus_read(uint32 addr, uint32 size, bool mode) override;
-  void bus_write(uint32 addr, uint32 size, bool mode, uint32 word) override;
+  void bus_idle() override;
+  uint32 bus_read(unsigned mode, uint32 addr) override;
+  void bus_write(unsigned mode, uint32 addr, uint32 word) override;
 
   uint8 mmio_read(unsigned addr);
   void mmio_write(unsigned addr, uint8 data);
diff --git a/sfc/chip/armdsp/memory.cpp b/sfc/chip/armdsp/memory.cpp
index 8a57d6ca..cfd8a304 100644
--- a/sfc/chip/armdsp/memory.cpp
+++ b/sfc/chip/armdsp/memory.cpp
@@ -3,32 +3,33 @@
 //note: timings are completely unverified
 //due to the ST018 chip design (on-die ROM), testing is nearly impossible
 
-void ArmDSP::bus_idle(uint32 addr) {
+void ArmDSP::bus_idle() {
   step(1);
 }
 
-uint32 ArmDSP::bus_read(uint32 addr, uint32 size, bool mode) {
+uint32 ArmDSP::bus_read(unsigned mode, uint32 addr) {
   step(1);
 
-  static auto memory = [&](const uint8 *memory, uint32 addr, uint32 size) -> uint32 {
-    switch(size) {
-    case Word:
+  static auto memory = [&](const uint8 *memory, unsigned mode, uint32 addr) -> uint32 {
+    if(mode & Word) {
       memory += addr & ~3;
       return memory[0] << 0 | memory[1] << 8 | memory[2] << 16 | memory[3] << 24;
-    case Byte:
+    } else if(mode & Byte) {
       return memory[addr];
+    } else {
+      return 0;  //should never occur
     }
   };
 
   switch(addr & 0xe0000000) {
-  case 0x00000000: return memory(programROM, addr & 0x1ffff, size);
+  case 0x00000000: return memory(programROM, mode, addr & 0x1ffff);
   case 0x20000000: return pipeline.fetch.instruction;
   case 0x40000000: break;
   case 0x60000000: return 0x40404001;
   case 0x80000000: return pipeline.fetch.instruction;
-  case 0xa0000000: return memory(dataROM, addr & 0x7fff, size);
+  case 0xa0000000: return memory(dataROM, mode, addr & 0x7fff);
   case 0xc0000000: return pipeline.fetch.instruction;
-  case 0xe0000000: return memory(programRAM, addr & 0x3fff, size);
+  case 0xe0000000: return memory(programRAM, mode, addr & 0x3fff);
   }
 
   addr &= 0xe000003f;
@@ -44,25 +45,22 @@ uint32 ArmDSP::bus_read(uint32 addr, uint32 size, bool mode) {
     return bridge.status();
   }
 
-  return 0u;
+  return 0;
 }
 
-void ArmDSP::bus_write(uint32 addr, uint32 size, bool mode, uint32 word) {
+void ArmDSP::bus_write(unsigned mode, uint32 addr, uint32 word) {
   step(1);
 
-  static auto memory = [](uint8 *memory, uint32 addr, uint32 size, uint32 word) {
-    switch(size) {
-    case Word:
+  static auto memory = [](uint8 *memory, unsigned mode, uint32 addr, uint32 word) {
+    if(mode & Word) {
       memory += addr & ~3;
       *memory++ = word >>  0;
       *memory++ = word >>  8;
       *memory++ = word >> 16;
       *memory++ = word >> 24;
-      break;
-    case Byte:
+    } else if(mode & Byte) {
       memory += addr;
       *memory++ = word >>  0;
-      break;
     }
   };
 
@@ -74,7 +72,7 @@ void ArmDSP::bus_write(uint32 addr, uint32 size, bool mode, uint32 word) {
   case 0x80000000: return;
   case 0xa0000000: return;
   case 0xc0000000: return;
-  case 0xe0000000: return memory(programRAM, addr & 0x3fff, size, word);
+  case 0xe0000000: return memory(programRAM, mode, addr & 0x3fff, word);
   }
 
   addr &= 0xe000003f;