diff --git a/core/broadway.cpp b/core/broadway.cpp
index 785e924..d61ff8d 100644
--- a/core/broadway.cpp
+++ b/core/broadway.cpp
@@ -22,7 +22,7 @@ void broadway::run(mem &mem) {
 }
 
 ircolib::u32 broadway::fetch(mem &mem) {
-    ircolib::u32 val = mem.read<ircolib::u32>(pc);
+    ircolib::u32 val = mem.read32(pc);
     pc += 4;
 
     return val;
@@ -48,6 +48,12 @@ void broadway::execute(ircolib::u32 instr, mem &mem) {
     case 31: // mfspr rd, spr
         move_spr(instr);
         break;
+    case 32: // mfspr rd, spr
+        lwz(instr, mem);
+        break;
+    case 36: // stw rs, d(ra)
+        stw(instr, mem);
+        break;
     case 37: // stwu rs, d(ra)
         stwu(instr, mem);
         break;
diff --git a/core/broadway.hpp b/core/broadway.hpp
index c754e82..dafb379 100644
--- a/core/broadway.hpp
+++ b/core/broadway.hpp
@@ -30,7 +30,9 @@ struct broadway {
     void bx(ircolib::u32);
     void bcx(ircolib::u32);
     void move_spr(ircolib::u32);
+    void stw(ircolib::u32, mem &);
     void stwu(ircolib::u32, mem &);
     void sth(ircolib::u32, mem &);
+    void lwz(ircolib::u32, mem &);
 };
 } // namespace weee::core
diff --git a/core/broadway/instructions.cpp b/core/broadway/instructions.cpp
index 1188a01..e9422f7 100644
--- a/core/broadway/instructions.cpp
+++ b/core/broadway/instructions.cpp
@@ -118,12 +118,20 @@ void broadway::stwu(ircolib::u32 instr, mem &mem) {
     if (utils::RA(instr) == 0)
         ircolib::panic("broadway::stwu with ra == 0");
 
-    const ircolib::s32 d = utils::SIMM(instr);
-    const ircolib::u32 EA = gpr[utils::RA(instr)] + d;
-    mem.write(EA, gpr[utils::RS(instr)]);
+    const ircolib::u32 EA = ircolib::s32(gpr[utils::RA(instr)]) + utils::SIMM(instr);
+    mem.write32(EA, gpr[utils::RS(instr)]);
     gpr[utils::RA(instr)] = EA;
 }
 
+void broadway::stw(ircolib::u32 instr, mem &mem) {
+    ircolib::s32 b = gpr[utils::RA(instr)];
+    if (utils::RA(instr) == 0)
+        b = 0;
+
+    const ircolib::u32 EA = b + utils::SIMM(instr);
+    mem.write32(EA, gpr[utils::RS(instr)]);
+}
+
 void broadway::sth(ircolib::u32 instr, mem &mem) {
     ircolib::u32 b = gpr[utils::RA(instr)];
     if (utils::RA(instr) == 0)
@@ -132,6 +140,15 @@ void broadway::sth(ircolib::u32 instr, mem &mem) {
     const ircolib::s32 d = utils::SIMM(instr);
     const ircolib::u32 EA = b + d;
 
-    mem.write(EA, ircolib::u16(gpr[utils::RS(instr)]));
+    mem.write16(EA, ircolib::u16(gpr[utils::RS(instr)]));
+}
+
+void broadway::lwz(ircolib::u32 instr, mem &mem) {
+    ircolib::u32 b = gpr[utils::RA(instr)];
+    if (utils::RA(instr) == 0)
+        b = 0;
+
+    ircolib::u32 ea = ircolib::s32(b) + utils::SIMM(instr);
+    gpr[utils::RD(instr)] = mem.read32(ea);
 }
 } // namespace weee::core
diff --git a/core/broadway/mmio/vi.cpp b/core/broadway/mmio/vi.cpp
index 75e92e4..eaa1283 100644
--- a/core/broadway/mmio/vi.cpp
+++ b/core/broadway/mmio/vi.cpp
@@ -1,15 +1,53 @@
 #include <broadway/mmio/vi.hpp>
-#include "ircolib/log.hpp"
+#include <ircolib/log.hpp>
+#include <mem.hpp>
 
 namespace weee::core {
-void video_interface::write(ircolib::u32 addr, ircolib::u16 value) {
-    addr -= 0x0c002000;
+video_interface::video_interface(mem &mem) {
+    mem.register_write16_handler(0x0c002000, 0x0c0020ff,
+                                 [&](ircolib::u32 addr, ircolib::u16 value) { write16(addr, value); });
+    mem.register_write32_handler(0x0c002000, 0x0c0020ff,
+                                 [&](ircolib::u32 addr, ircolib::u32 value) { write32(addr, value); });
+}
+
+void video_interface::write16(ircolib::u32 addr, ircolib::u16 value) {
     switch (addr) {
-    case 2:
+    case 0x00:
+        vtr.raw = value;
+        break;
+    case 0x02:
         dcr.raw = value;
         break;
     default:
-        ircolib::panic("video_interface::write to unimplemented addr 0x{:04X} with value 0x{:04X}", addr, value);
+        ircolib::panic("video_interface::write16 to unimplemented addr 0x{:04X} with value 0x{:04X}", addr, value);
+    }
+}
+
+void video_interface::write32(ircolib::u32 addr, ircolib::u32 value) {
+    switch (addr) {
+    case 0x02:
+        dcr.raw = value;
+        break;
+    case 0x04:
+        htr0.raw = value;
+        break;
+    case 0x08:
+        htr1.raw = value;
+        break;
+    case 0x0c:
+        vto.raw = value;
+        break;
+    case 0x10:
+        vte.raw = value;
+        break;
+    case 0x14:
+        bbei.raw = value;
+        break;
+    case 0x18:
+        bboi.raw = value;
+        break;
+    default:
+        ircolib::panic("video_interface::write32 to unimplemented addr 0x{:04X} with value 0x{:08X}", addr, value);
     }
 }
 } // namespace weee::core
diff --git a/core/broadway/mmio/vi.hpp b/core/broadway/mmio/vi.hpp
index 49e11f6..4828af7 100644
--- a/core/broadway/mmio/vi.hpp
+++ b/core/broadway/mmio/vi.hpp
@@ -2,25 +2,82 @@
 #include <ircolib/types.hpp>
 
 namespace weee::core {
-union DCR {
-    struct {
-        unsigned e : 1;
-        unsigned r : 1;
-        unsigned i : 1;
-        unsigned d : 1;
-        unsigned le0 : 2;
-        unsigned le1 : 2;
-        unsigned fmt : 2;
-        unsigned : 6;
-    };
-
-    ircolib::u16 raw;
-};
+struct mem;
 
 struct video_interface {
-    void write(ircolib::u32, ircolib::u16);
+    video_interface(mem &);
+    void write16(ircolib::u32, ircolib::u16);
+    void write32(ircolib::u32, ircolib::u32);
 
-    private:
-    DCR dcr{};
+    union DCR {
+        struct {
+            unsigned e : 1;
+            unsigned r : 1;
+            unsigned i : 1;
+            unsigned d : 1;
+            unsigned le0 : 2;
+            unsigned le1 : 2;
+            unsigned fmt : 2;
+            unsigned : 6;
+        };
+
+        ircolib::u16 raw;
+    } dcr;
+
+    union HTR0 {
+        struct {
+            unsigned hlw : 9;
+            unsigned : 7;
+            unsigned hce : 7;
+            unsigned : 1;
+            unsigned hcs : 7;
+            unsigned : 1;
+        };
+
+        ircolib::u32 raw;
+    } htr0;
+
+    union HTR1 {
+        struct {
+            unsigned hsy : 7;
+            unsigned hbe : 10;
+            unsigned hbs : 10;
+            unsigned : 5;
+        };
+
+        ircolib::u32 raw;
+    } htr1;
+
+    union VTR {
+        struct {
+            unsigned equ : 4;
+            unsigned acv : 10;
+            unsigned : 2;
+        };
+
+        ircolib::u16 raw;
+    } vtr;
+
+    union VTO {
+        struct {
+            unsigned prb : 10;
+            unsigned : 6;
+            unsigned psb : 10;
+            unsigned : 6;
+        };
+
+        ircolib::u32 raw;
+    } vto, vte;
+
+    union BBI {
+        struct {
+            unsigned bs1 : 5;
+            unsigned be1 : 11;
+            unsigned bs3 : 5;
+            unsigned be3 : 11;
+        };
+
+        ircolib::u32 raw;
+    } bbei, bboi;
 };
 } // namespace weee::core
diff --git a/core/loaders/dol.cpp b/core/loaders/dol.cpp
index bd533fc..0535724 100644
--- a/core/loaders/dol.cpp
+++ b/core/loaders/dol.cpp
@@ -4,7 +4,6 @@
 #include <ircolib/file.hpp>
 #include <ircolib/mem_access.hpp>
 #include <array>
-#include <print>
 
 namespace weee::core {
 bool load_dol(const std::string &path, mem &mem, broadway &broadway) {
@@ -22,22 +21,22 @@ bool load_dol(const std::string &path, mem &mem, broadway &broadway) {
     } hdr;
 
     for (size_t bin_index = 0; bin_index < 0xD8; bin_index += 4) {
-        if (ircolib::is_inside_range_ext(bin_index, 0, 0x1c)) // text file offsets
+        if (ircolib::is_inside_range(bin_index, 0, 0x1b)) // text file offsets
             hdr.text[bin_index / 4].offset = std::byteswap(ircolib::read_access<ircolib::u32>(bin, bin_index));
 
-        if (ircolib::is_inside_range_ext(bin_index, 0x1c, 0x48)) // data file offsets
+        if (ircolib::is_inside_range(bin_index, 0x1c, 0x47)) // data file offsets
             hdr.data[(bin_index - 0x1c) / 4].offset = std::byteswap(ircolib::read_access<ircolib::u32>(bin, bin_index));
 
-        if (ircolib::is_inside_range_ext(bin_index, 0x48, 0x64)) // text loading addresses
+        if (ircolib::is_inside_range(bin_index, 0x48, 0x63)) // text loading addresses
             hdr.text[(bin_index - 0x48) / 4].addr = std::byteswap(ircolib::read_access<ircolib::u32>(bin, bin_index));
 
-        if (ircolib::is_inside_range_ext(bin_index, 0x64, 0x90)) // data loading addresses
+        if (ircolib::is_inside_range(bin_index, 0x64, 0x8f)) // data loading addresses
             hdr.data[(bin_index - 0x64) / 4].addr = std::byteswap(ircolib::read_access<ircolib::u32>(bin, bin_index));
 
-        if (ircolib::is_inside_range_ext(bin_index, 0x90, 0xac)) // text sizes
+        if (ircolib::is_inside_range(bin_index, 0x90, 0xab)) // text sizes
             hdr.text[(bin_index - 0x90) / 4].size = std::byteswap(ircolib::read_access<ircolib::u32>(bin, bin_index));
 
-        if (ircolib::is_inside_range_ext(bin_index, 0xac, 0xd8)) // data sizes
+        if (ircolib::is_inside_range(bin_index, 0xac, 0xd7)) // data sizes
             hdr.data[(bin_index - 0xac) / 4].size = std::byteswap(ircolib::read_access<ircolib::u32>(bin, bin_index));
     }
 
diff --git a/core/mem.cpp b/core/mem.cpp
index 61253af..6d39b36 100644
--- a/core/mem.cpp
+++ b/core/mem.cpp
@@ -4,9 +4,22 @@
 #include "ircolib/mem_access.hpp"
 
 namespace weee::core {
-mem::mem() {
+mem::mem() : vi(*this) {
     mem1.resize(24_mib);
     std::fill(mem1.begin(), mem1.end(), 0);
+    register_read16_handler(0x00000000, 0x017fffff,
+                            [&](ircolib::u32 addr) { return ircolib::read_access<ircolib::u16>(mem1, addr); });
+
+    register_read32_handler(0x00000000, 0x017fffff,
+                            [&](ircolib::u32 addr) { return ircolib::read_access<ircolib::u32>(mem1, addr); });
+
+    register_write16_handler(0x00000000, 0x017fffff, [&](ircolib::u32 addr, ircolib::u16 value) {
+        ircolib::write_access<ircolib::u16>(mem1, addr, value);
+    });
+
+    register_write32_handler(0x00000000, 0x017fffff, [&](ircolib::u32 addr, ircolib::u32 value) {
+        ircolib::write_access<ircolib::u32>(mem1, addr, value);
+    });
 }
 
 void mem::copy(std::vector<ircolib::u8> &src, const ircolib::u32 offset) {
@@ -14,7 +27,6 @@ void mem::copy(std::vector<ircolib::u8> &src, const ircolib::u32 offset) {
         ircolib::panic("mem::copy outside mem1 range (src @ 0x{:08} for size 0x{:08X})", offset, src.size());
 
     ircolib::swap_buffer<ircolib::u32>(src);
-    std::println("Copying {} bytes to mem1[{}]", src.size(), offset);
     std::copy(src.begin(), src.end(), mem1.begin() + offset);
 }
 
@@ -23,7 +35,6 @@ void mem::copy(ircolib::u8 *src, const ircolib::u32 size, const ircolib::u32 off
         ircolib::panic("mem::copy outside mem1 range (src @ 0x{:08} for size 0x{:08X})", offset, size);
 
     ircolib::swap_buffer<ircolib::u32>(src, size);
-    std::println("Copying {} bytes to mem1[{}]", size, offset);
     memcpy(&mem1[offset], src, size);
 }
 
@@ -31,46 +42,98 @@ void mem::set(const ircolib::u8 val, const ircolib::u32 size, const ircolib::u32
     if (offset + size >= mem1.size())
         ircolib::panic("mem::set outside mem1 range (@ 0x{:08} for size 0x{:08X})", offset, size);
 
-    std::println("Setting {} bytes to {} from mem1[{}]", size, val, offset);
     memset(&mem1[offset], val, size);
 }
 
-template <>
-ircolib::u32 mem::read(ircolib::u32 addr) {
+ircolib::u8 mem::read8(ircolib::u32 addr) {
     addr &= 0x0FFFFFFF;
-    if (addr > 0x017FFFFF)
-        ircolib::panic("mem::read unimplemented outside mem1 (0x{:08X})", addr);
-
-    return ircolib::read_access<ircolib::u32>(mem1, addr);
-}
-
-template <>
-ircolib::u16 mem::read(ircolib::u32 addr) {
-    addr &= 0x0FFFFFFF;
-    if (addr > 0x017FFFFF)
-        ircolib::panic("mem::read unimplemented outside mem1 (0x{:08X})", addr);
-
-    return ircolib::read_access<ircolib::u16>(mem1, addr);
-}
-
-void mem::write(ircolib::u32 addr, ircolib::u32 value) {
-    addr &= 0x0FFFFFFF;
-    if (addr > 0x017FFFFF)
-        ircolib::panic("mem::write unimplemented outside mem1 (0x{:08X} = 0x{:08X})", addr, value);
-
-    ircolib::write_access<ircolib::u32>(mem1, addr, value);
-}
-
-void mem::write(ircolib::u32 addr, ircolib::u16 value) {
-    addr &= 0x0FFFFFFF;
-    if (addr > 0x017FFFFF && !ircolib::is_inside_range(addr, 0x0C002000, 0x0C0020FF))
-        ircolib::panic("mem::write unimplemented outside mem1 and vi (0x{:08X} = 0x{:08X})", addr, value);
-
-    if (addr <= 0x017FFFFF) {
-        ircolib::write_access<ircolib::u16>(mem1, addr, value);
-        return;
+    for (const auto &handler : read8_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start);
+        }
     }
 
-    vi.write(addr, value);
+    ircolib::panic("mem::read8 unimplemented addr 0x{:08X}", addr);
+    return 0;
+}
+
+ircolib::u16 mem::read16(ircolib::u32 addr) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : read16_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start);
+        }
+    }
+
+    ircolib::panic("mem::read16 unimplemented addr 0x{:08X}", addr);
+    return 0;
+}
+
+ircolib::u32 mem::read32(ircolib::u32 addr) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : read32_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start);
+        }
+    }
+
+    ircolib::panic("mem::read32 unimplemented addr 0x{:08X}", addr);
+    return 0;
+}
+
+ircolib::u64 mem::read64(ircolib::u32 addr) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : read64_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start);
+        }
+    }
+
+    ircolib::panic("mem::read64 unimplemented addr 0x{:08X}", addr);
+    return 0;
+}
+
+void mem::write8(ircolib::u32 addr, ircolib::u8 value) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : write8_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start, value);
+        }
+    }
+
+    ircolib::panic("mem::write8 unimplemented addr 0x{:08X} = 0x{:02X}", addr, value);
+}
+
+void mem::write16(ircolib::u32 addr, ircolib::u16 value) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : write16_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start, value);
+        }
+    }
+
+    ircolib::panic("mem::write16 unimplemented addr 0x{:08X} = 0x{:04X}", addr, value);
+}
+
+void mem::write32(ircolib::u32 addr, ircolib::u32 value) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : write32_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start, value);
+        }
+    }
+
+    ircolib::panic("mem::write32 unimplemented addr 0x{:08X} = 0x{:08X}", addr, value);
+}
+
+void mem::write64(ircolib::u32 addr, ircolib::u64 value) {
+    addr &= 0x0FFFFFFF;
+    for (const auto &handler : write64_handlers) {
+        if (ircolib::is_inside_range(addr, handler.start, handler.end)) {
+            return handler.func(addr - handler.start, value);
+        }
+    }
+
+    ircolib::panic("mem::write64 unimplemented addr 0x{:08X} = 0x{:016X}", addr, value);
 }
 } // namespace weee::core
diff --git a/core/mem.hpp b/core/mem.hpp
index 4371348..3b5b315 100644
--- a/core/mem.hpp
+++ b/core/mem.hpp
@@ -1,21 +1,81 @@
 #pragma once
 #include <vector>
 #include <ircolib/types.hpp>
-#include "broadway/mmio/vi.hpp"
+#include <broadway/mmio/vi.hpp>
+#include <functional>
 
 namespace weee::core {
+template <typename T>
+struct read_handler {
+    std::function<T(ircolib::u32)> func;
+    ircolib::u32 start, end;
+};
+
+template <typename T>
+struct write_handler {
+    std::function<void(ircolib::u32, T)> func;
+    ircolib::u32 start, end;
+};
+
 struct mem {
     mem();
 
-    template <typename T>
-    T read(ircolib::u32);
-    void write(ircolib::u32, ircolib::u32);
-    void write(ircolib::u32, ircolib::u16);
+    void register_read8_handler(ircolib::u32 start, ircolib::u32 end,
+                                const std::function<ircolib::u8(ircolib::u32)> &func) {
+        read8_handlers.push_back({std::move(func), start, end});
+    }
+    void register_read16_handler(ircolib::u32 start, ircolib::u32 end,
+                                 const std::function<ircolib::u16(ircolib::u32)> &func) {
+        read16_handlers.push_back({std::move(func), start, end});
+    }
+    void register_read32_handler(ircolib::u32 start, ircolib::u32 end,
+                                 const std::function<ircolib::u32(ircolib::u32)> &func) {
+        read32_handlers.push_back({std::move(func), start, end});
+    }
+    void register_read64_handler(ircolib::u32 start, ircolib::u32 end,
+                                 const std::function<ircolib::u64(ircolib::u32)> &func) {
+        read64_handlers.push_back({std::move(func), start, end});
+    }
+    void register_write8_handler(ircolib::u32 start, ircolib::u32 end,
+                                 const std::function<void(ircolib::u32, ircolib::u8)> &func) {
+        write8_handlers.push_back({std::move(func), start, end});
+    }
+    void register_write16_handler(ircolib::u32 start, ircolib::u32 end,
+                                  const std::function<void(ircolib::u32, ircolib::u16)> &func) {
+        write16_handlers.push_back({std::move(func), start, end});
+    }
+    void register_write32_handler(ircolib::u32 start, ircolib::u32 end,
+                                  const std::function<void(ircolib::u32, ircolib::u32)> &func) {
+        write32_handlers.push_back({std::move(func), start, end});
+    }
+    void register_write64_handler(ircolib::u32 start, ircolib::u32 end,
+                                  const std::function<void(ircolib::u32, ircolib::u64)> &func) {
+        write64_handlers.push_back({std::move(func), start, end});
+    }
+
+    ircolib::u8 read8(ircolib::u32);
+    ircolib::u16 read16(ircolib::u32);
+    ircolib::u32 read32(ircolib::u32);
+    ircolib::u64 read64(ircolib::u32);
+    void write8(ircolib::u32, ircolib::u8);
+    void write16(ircolib::u32, ircolib::u16);
+    void write32(ircolib::u32, ircolib::u32);
+    void write64(ircolib::u32, ircolib::u64);
+
     void copy(std::vector<ircolib::u8> &src, const ircolib::u32 offset);
     void copy(ircolib::u8 *src, const ircolib::u32 size, const ircolib::u32 offset);
     void set(const ircolib::u8 val, const ircolib::u32 size, const ircolib::u32 offset);
 
     private:
+    std::vector<read_handler<ircolib::u8>> read8_handlers{};
+    std::vector<read_handler<ircolib::u16>> read16_handlers{};
+    std::vector<read_handler<ircolib::u32>> read32_handlers{};
+    std::vector<read_handler<ircolib::u64>> read64_handlers{};
+    std::vector<write_handler<ircolib::u8>> write8_handlers{};
+    std::vector<write_handler<ircolib::u16>> write16_handlers{};
+    std::vector<write_handler<ircolib::u32>> write32_handlers{};
+    std::vector<write_handler<ircolib::u64>> write64_handlers{};
+
     std::vector<ircolib::u8> mem1;
     video_interface vi;
 };
diff --git a/external/ircolib/mem_access.hpp b/external/ircolib/mem_access.hpp
index 05d8d97..bb83b77 100644
--- a/external/ircolib/mem_access.hpp
+++ b/external/ircolib/mem_access.hpp
@@ -30,11 +30,6 @@ static inline constexpr bool is_inside_range(const std::integral auto &addr, con
     return addr >= start && addr <= end;
 }
 
-static inline constexpr bool is_inside_range_ext(const std::integral auto &addr, const std::integral auto &start,
-                                                 const std::integral auto &end) {
-    return addr >= start && addr < end;
-}
-
 template <typename T>
 static constexpr inline T read_access(const u8 *data, const u32 index);
 template <typename T>