fix: reimpl handleInterrupt code

passes arm.gba, thumb.gb and armwrestler, fails in actual games
TODO: run FuzzARM debug specific titles

lib/SDL.zig

@@ -1 +1 @@
-Subproject commit f3d7620f85929bca8b7ce2be2fa8a938e164f720
+Subproject commit 401c50ff3d0b83ad4bd89caf580ce1bd90fb5618

lib/known-folders

@@ -1 +1 @@
-Subproject commit 9db1b99219c767d5e24994b1525273fe4031e464
+Subproject commit 24845b0103e611c108d6bc334231c464e699742c

lib/zig-clap

@@ -1 +1 @@
-Subproject commit 996821a3e1f186c9e5cdfd971d742c9815ea590e
+Subproject commit 1c09e0dc31918dd716b1032ad3e1d1c080cbbff1

src/Gui.zig

@@ -8,12 +8,12 @@ const Scheduler = @import("core/scheduler.zig").Scheduler;
 const FpsTracker = @import("core/util.zig").FpsTracker;
 
 const pitch = @import("core/ppu.zig").framebuf_pitch;
+const scale = @import("core/emu.zig").win_scale;
 
 const emu = @import("core/emu.zig");
 const asString = @import("core/util.zig").asString;
 const log = std.log.scoped(.GUI);
 
-const scale = 4;
 const default_title: []const u8 = "ZBA";
 
 window: *SDL.SDL_Window,

src/core/Bus.zig

@@ -77,70 +77,29 @@ pub fn attach(self: *Self, cpu: *Arm7tdmi) void {
     self.cpu = cpu;
 }
 
-pub fn debugRead(self: *const Self, comptime T: type, address: u32) T {
-    const cached = self.sched.tick;
-    defer self.sched.tick = cached;
-
-    // FIXME: This is bad but it's a debug read so I don't care that much?
-    const this = @intToPtr(*Self, @ptrToInt(self));
-
-    return this.read(T, address);
-}
-
-fn readOpenBus(self: *const Self, comptime T: type, address: u32) T {
-    const r15 = self.cpu.?.r[15];
-
-    const word = if (self.cpu.?.cpsr.t.read()) blk: {
-        const page = @truncate(u8, r15 >> 24);
-
-        switch (page) {
-            // EWRAM, PALRAM, VRAM, and Game ROM (16-bit)
-            0x02, 0x05, 0x06, 0x08...0x0D => {
-                const halfword = self.debugRead(u16, r15 + 2);
-                break :blk @as(u32, halfword) << 16 | halfword;
-            },
-            // BIOS or OAM (32-bit)
-            0x00, 0x07 => {
-                const offset: u32 = if (address & 3 == 0b00) 2 else 0;
-                break :blk @as(u32, self.debugRead(u16, (r15 + 2) + offset)) << 16 | self.debugRead(u16, r15 + offset);
-            },
-            // IWRAM (16-bit but special)
-            0x03 => {
-                const offset: u32 = if (address & 3 == 0b00) 2 else 0;
-                break :blk @as(u32, self.debugRead(u16, (r15 + 2) - offset)) << 16 | self.debugRead(u16, r15 + offset);
-            },
-            else => unreachable,
-        }
-    } else self.debugRead(u32, r15 + 4);
-
-    return @truncate(T, rotr(u32, word, 8 * (address & 3)));
-}
-
-fn readBios(self: *Self, comptime T: type, address: u32) T {
-    if (address < Bios.size) return self.bios.checkedRead(T, self.cpu.?.r[15], alignAddress(T, address));
-
-    return self.readOpenBus(T, address);
-}
-
-pub fn read(self: *Self, comptime T: type, address: u32) T {
+pub fn dbgRead(self: *const Self, comptime T: type, address: u32) T {
     const page = @truncate(u8, address >> 24);
-    const align_addr = alignAddress(T, address);
-    defer self.sched.tick += timings[@boolToInt(T == u32)][@truncate(u4, page)];
+    const aligned_addr = forceAlign(T, address);
 
     return switch (page) {
         // General Internal Memory
-        0x00 => self.readBios(T, address),
-        0x02 => self.ewram.read(T, align_addr),
-        0x03 => self.iwram.read(T, align_addr),
-        0x04 => io.read(self, T, align_addr),
+        0x00 => blk: {
+            if (address < Bios.size)
+                break :blk self.bios.dbgRead(T, self.cpu.?.r[15], aligned_addr);
+
+            break :blk self.readOpenBus(T, address);
+        },
+        0x02 => self.ewram.read(T, aligned_addr),
+        0x03 => self.iwram.read(T, aligned_addr),
+        0x04 => io.read(self, T, aligned_addr),
 
         // Internal Display Memory
-        0x05 => self.ppu.palette.read(T, align_addr),
-        0x06 => self.ppu.vram.read(T, align_addr),
-        0x07 => self.ppu.oam.read(T, align_addr),
+        0x05 => self.ppu.palette.read(T, aligned_addr),
+        0x06 => self.ppu.vram.read(T, aligned_addr),
+        0x07 => self.ppu.oam.read(T, aligned_addr),
 
         // External Memory (Game Pak)
-        0x08...0x0D => self.pak.read(T, align_addr),
+        0x08...0x0D => self.pak.dbgRead(T, aligned_addr),
         0x0E...0x0F => blk: {
             const value = self.pak.backup.read(address);
 
@@ -153,29 +112,100 @@ pub fn read(self: *Self, comptime T: type, address: u32) T {
 
             break :blk @as(T, value) * multiplier;
         },
-        else => readOpenBus(self, T, address),
+        else => self.readOpenBus(T, address),
+    };
+}
+
+fn readOpenBus(self: *const Self, comptime T: type, address: u32) T {
+    const r15 = self.cpu.?.r[15];
+
+    const word = if (self.cpu.?.cpsr.t.read()) blk: {
+        const page = @truncate(u8, r15 >> 24);
+
+        switch (page) {
+            // EWRAM, PALRAM, VRAM, and Game ROM (16-bit)
+            0x02, 0x05, 0x06, 0x08...0x0D => {
+                const halfword = self.dbgRead(u16, r15 + 2);
+                break :blk @as(u32, halfword) << 16 | halfword;
+            },
+            // BIOS or OAM (32-bit)
+            0x00, 0x07 => {
+                const offset: u32 = if (address & 3 == 0b00) 2 else 0;
+                break :blk @as(u32, self.dbgRead(u16, (r15 + 2) + offset)) << 16 | self.dbgRead(u16, r15 + offset);
+            },
+            // IWRAM (16-bit but special)
+            0x03 => {
+                const offset: u32 = if (address & 3 == 0b00) 2 else 0;
+                break :blk @as(u32, self.dbgRead(u16, (r15 + 2) - offset)) << 16 | self.dbgRead(u16, r15 + offset);
+            },
+            else => unreachable,
+        }
+    } else self.dbgRead(u32, r15 + 4);
+
+    return @truncate(T, rotr(u32, word, 8 * (address & 3)));
+}
+
+pub fn read(self: *Self, comptime T: type, address: u32) T {
+    const page = @truncate(u8, address >> 24);
+    const aligned_addr = forceAlign(T, address);
+
+    self.sched.tick += timings[@boolToInt(T == u32)][@truncate(u4, page)];
+
+    return switch (page) {
+        // General Internal Memory
+        0x00 => blk: {
+            if (address < Bios.size)
+                break :blk self.bios.read(T, self.cpu.?.r[15], aligned_addr);
+
+            break :blk self.readOpenBus(T, address);
+        },
+        0x02 => self.ewram.read(T, aligned_addr),
+        0x03 => self.iwram.read(T, aligned_addr),
+        0x04 => io.read(self, T, aligned_addr),
+
+        // Internal Display Memory
+        0x05 => self.ppu.palette.read(T, aligned_addr),
+        0x06 => self.ppu.vram.read(T, aligned_addr),
+        0x07 => self.ppu.oam.read(T, aligned_addr),
+
+        // External Memory (Game Pak)
+        0x08...0x0D => self.pak.read(T, aligned_addr),
+        0x0E...0x0F => blk: {
+            const value = self.pak.backup.read(address);
+
+            const multiplier = switch (T) {
+                u32 => 0x01010101,
+                u16 => 0x0101,
+                u8 => 1,
+                else => @compileError("Backup: Unsupported read width"),
+            };
+
+            break :blk @as(T, value) * multiplier;
+        },
+        else => self.readOpenBus(T, address),
     };
 }
 
 pub fn write(self: *Self, comptime T: type, address: u32, value: T) void {
     const page = @truncate(u8, address >> 24);
-    const align_addr = alignAddress(T, address);
-    defer self.sched.tick += timings[@boolToInt(T == u32)][@truncate(u4, page)];
+    const aligned_addr = forceAlign(T, address);
+
+    self.sched.tick += timings[@boolToInt(T == u32)][@truncate(u4, page)];
 
     switch (page) {
         // General Internal Memory
-        0x00 => self.bios.write(T, align_addr, value),
-        0x02 => self.ewram.write(T, align_addr, value),
-        0x03 => self.iwram.write(T, align_addr, value),
-        0x04 => io.write(self, T, align_addr, value),
+        0x00 => self.bios.write(T, aligned_addr, value),
+        0x02 => self.ewram.write(T, aligned_addr, value),
+        0x03 => self.iwram.write(T, aligned_addr, value),
+        0x04 => io.write(self, T, aligned_addr, value),
 
         // Internal Display Memory
-        0x05 => self.ppu.palette.write(T, align_addr, value),
-        0x06 => self.ppu.vram.write(T, self.ppu.dispcnt, align_addr, value),
-        0x07 => self.ppu.oam.write(T, align_addr, value),
+        0x05 => self.ppu.palette.write(T, aligned_addr, value),
+        0x06 => self.ppu.vram.write(T, self.ppu.dispcnt, aligned_addr, value),
+        0x07 => self.ppu.oam.write(T, aligned_addr, value),
 
         // External Memory (Game Pak)
-        0x08...0x0D => self.pak.write(T, self.dma[3].word_count, align_addr, value),
+        0x08...0x0D => self.pak.write(T, self.dma[3].word_count, aligned_addr, value),
         0x0E...0x0F => {
             const rotate_by = switch (T) {
                 u32 => address & 3,
@@ -190,7 +220,7 @@ pub fn write(self: *Self, comptime T: type, address: u32, value: T) void {
     }
 }
 
-fn alignAddress(comptime T: type, address: u32) u32 {
+fn forceAlign(comptime T: type, address: u32) u32 {
     return switch (T) {
         u32 => address & 0xFFFF_FFFC,
         u16 => address & 0xFFFF_FFFE,

src/core/bus/Bios.zig

@@ -31,17 +31,22 @@ pub fn deinit(self: Self) void {
     if (self.buf) |buf| self.alloc.free(buf);
 }
 
-pub fn checkedRead(self: *Self, comptime T: type, r15: u32, addr: u32) T {
+pub fn read(self: *Self, comptime T: type, r15: u32, addr: u32) T {
     if (r15 < Self.size) {
         self.addr_latch = addr;
-        return self.read(T, addr);
+        return self.uncheckedRead(T, addr);
     }
 
     log.debug("Rejected read since r15=0x{X:0>8}", .{r15});
-    return @truncate(T, self.read(T, self.addr_latch + 8));
+    return @truncate(T, self.uncheckedRead(T, self.addr_latch + 8));
 }
 
-fn read(self: *const Self, comptime T: type, addr: u32) T {
+pub fn dbgRead(self: *const Self, comptime T: type, r15: u32, addr: u32) T {
+    if (r15 < Self.size) return self.uncheckedRead(T, addr);
+    return @truncate(T, self.uncheckedRead(T, self.addr_latch + 8));
+}
+
+fn uncheckedRead(self: *const Self, comptime T: type, addr: u32) T {
     if (self.buf) |buf| {
         return switch (T) {
             u32, u16, u8 => std.mem.readIntSliceLittle(T, buf[addr..][0..@sizeOf(T)]),

src/core/bus/GamePak.zig

@@ -90,6 +90,33 @@ pub fn read(self: *Self, comptime T: type, address: u32) T {
     };
 }
 
+pub fn dbgRead(self: *const Self, comptime T: type, address: u32) T {
+    const addr = address & 0x1FF_FFFF;
+
+    if (self.backup.kind == .Eeprom) {
+        if (self.isLarge()) {
+            // Addresses 0x1FF_FF00 to 0x1FF_FFFF are reserved from EEPROM accesses if
+            // * Backup type is EEPROM
+            // * Large ROM (Size is greater than 16MB)
+            if (addr > 0x1FF_FEFF)
+                return self.backup.eeprom.dbgRead();
+        } else {
+            // Addresses 0x0D00_0000 to 0x0DFF_FFFF are reserved for EEPROM accesses if
+            // * Backup type is EEPROM
+            // * Small ROM (less than 16MB)
+            if (@truncate(u8, address >> 24) == 0x0D)
+                return self.backup.eeprom.dbgRead();
+        }
+    }
+
+    return switch (T) {
+        u32 => (@as(T, self.get(addr + 3)) << 24) | (@as(T, self.get(addr + 2)) << 16) | (@as(T, self.get(addr + 1)) << 8) | (@as(T, self.get(addr))),
+        u16 => (@as(T, self.get(addr + 1)) << 8) | @as(T, self.get(addr)),
+        u8 => self.get(addr),
+        else => @compileError("GamePak: Unsupported read width"),
+    };
+}
+
 pub fn write(self: *Self, comptime T: type, word_count: u16, address: u32, value: T) void {
     const addr = address & 0x1FF_FFFF;
 

src/core/bus/backup.zig

@@ -340,6 +340,10 @@ const Eeprom = struct {
         return self.reader.read();
     }
 
+    pub fn dbgRead(self: *const Self) u1 {
+        return self.reader.dbgRead();
+    }
+
     pub fn write(self: *Self, word_count: u16, buf: *[]u8, bit: u1) void {
         if (self.guessKind(word_count)) |found| {
             log.info("EEPROM Kind: {}", .{found});
@@ -492,6 +496,19 @@ const Eeprom = struct {
 
             return bit;
         }
+
+        fn dbgRead(self: *const This) u1 {
+            if (!self.enabled) return 1;
+
+            const bit = if (self.i < 4) blk: {
+                break :blk 0;
+            } else blk: {
+                const idx = @intCast(u6, 63 - (self.i - 4));
+                break :blk @truncate(u1, self.data >> idx);
+            };
+
+            return bit;
+        }
     };
 
     const Writer = struct {

src/core/cpu.zig

@@ -6,6 +6,7 @@ const Bit = @import("bitfield").Bit;
 const Bitfield = @import("bitfield").Bitfield;
 const Scheduler = @import("scheduler.zig").Scheduler;
 const FilePaths = @import("util.zig").FilePaths;
+const Logger = @import("util.zig").Logger;
 
 const File = std.fs.File;
 
@@ -225,13 +226,14 @@ pub const thumb = struct {
     }
 };
 
-const enable_logging = false;
+const cpu_logging = @import("emu.zig").cpu_logging;
 const log = std.log.scoped(.Arm7Tdmi);
 
 pub const Arm7tdmi = struct {
     const Self = @This();
 
     r: [16]u32,
+    pipe: Pipline,
     sched: *Scheduler,
     bus: *Bus,
     cpsr: PSR,
@@ -247,13 +249,12 @@ pub const Arm7tdmi = struct {
 
     banked_spsr: [5]PSR,
 
-    log_file: ?*const File,
-    log_buf: [0x100]u8,
-    binary_log: bool,
+    logger: ?Logger,
 
     pub fn init(sched: *Scheduler, bus: *Bus) Self {
         return Self{
             .r = [_]u32{0x00} ** 16,
+            .pipe = Pipline.init(),
             .sched = sched,
             .bus = bus,
             .cpsr = .{ .raw = 0x0000_001F },
@@ -261,15 +262,12 @@ pub const Arm7tdmi = struct {
             .banked_fiq = [_]u32{0x00} ** 10,
             .banked_r = [_]u32{0x00} ** 12,
             .banked_spsr = [_]PSR{.{ .raw = 0x0000_0000 }} ** 5,
-            .log_file = null,
-            .log_buf = undefined,
-            .binary_log = false,
+            .logger = null,
         };
     }
 
-    pub fn useLogger(self: *Self, file: *const File, is_binary: bool) void {
-        self.log_file = file;
-        self.binary_log = is_binary;
+    pub fn attach(self: *Self, log_file: std.fs.File) void {
+        self.logger = Logger.init(log_file);
     }
 
     inline fn bankedIdx(mode: Mode, kind: BankedKind) usize {
@@ -320,8 +318,21 @@ pub const Arm7tdmi = struct {
         return self.bus.io.haltcnt == .Halt;
     }
 
+    pub fn setCpsrNoFlush(self: *Self, value: u32) void {
+        if (value & 0x1F != self.cpsr.raw & 0x1F) self.changeModeFromIdx(@truncate(u5, value & 0x1F));
+        self.cpsr.raw = value;
+    }
+
     pub fn setCpsr(self: *Self, value: u32) void {
         if (value & 0x1F != self.cpsr.raw & 0x1F) self.changeModeFromIdx(@truncate(u5, value & 0x1F));
+
+        const new: PSR = .{ .raw = value };
+        if (self.cpsr.t.read() != new.t.read()) {
+            // If THUMB to ARM or ARM to THUMB, flush pipeline
+            self.r[15] &= if (new.t.read()) ~@as(u32, 1) else ~@as(u32, 3);
+            self.pipe.flush();
+        }
+
         self.cpsr.raw = value;
     }
 
@@ -424,19 +435,22 @@ pub const Arm7tdmi = struct {
     }
 
     pub fn step(self: *Self) void {
-        if (self.cpsr.t.read()) {
-            const opcode = self.fetch(u16);
-            if (enable_logging) if (self.log_file) |file| self.debug_log(file, opcode);
+        if (self.cpsr.t.read()) blk: {
+            const opcode = @truncate(u16, self.pipe.step(self, u16) orelse break :blk);
+            if (cpu_logging) self.logger.?.mgbaLog(self, opcode);
 
             thumb.lut[thumbIdx(opcode)](self, self.bus, opcode);
-        } else {
-            const opcode = self.fetch(u32);
-            if (enable_logging) if (self.log_file) |file| self.debug_log(file, opcode);
+        } else blk: {
+            const opcode = self.pipe.step(self, u32) orelse break :blk;
+            if (cpu_logging) self.logger.?.mgbaLog(self, opcode);
 
             if (checkCond(self.cpsr, @truncate(u4, opcode >> 28))) {
                 arm.lut[armIdx(opcode)](self, self.bus, opcode);
             }
         }
+
+        if (!self.pipe.flushed) self.r[15] += if (self.cpsr.t.read()) 2 else @as(u32, 4);
+        self.pipe.flushed = false;
     }
 
     pub fn stepDmaTransfer(self: *Self) bool {
@@ -471,27 +485,26 @@ pub const Arm7tdmi = struct {
     pub fn handleInterrupt(self: *Self) void {
         const should_handle = self.bus.io.ie.raw & self.bus.io.irq.raw;
 
-        if (should_handle != 0) {
-            self.bus.io.haltcnt = .Execute;
-            // log.debug("An Interrupt was Fired!", .{});
+        // Return if IME is disabled, CPSR I is set or there is nothing to handle
+        if (!self.bus.io.ime or self.cpsr.i.read() or should_handle == 0) return;
 
-            // Either IME is not true or I in CPSR is true
-            // Don't handle interrupts
-            if (!self.bus.io.ime or self.cpsr.i.read()) return;
-            // log.debug("An interrupt was Handled!", .{});
+        // If pipeline isn't full, return but reschedule the handling of the event
+        if (!self.pipe.isFull()) return;
 
-            // retAddr.gba says r15 on it's own is off by -04h in both ARM and THUMB mode
-            const r15 = self.r[15] + 4;
-            const cpsr = self.cpsr.raw;
+        // log.debug("Handling Interrupt!", .{});
+        self.bus.io.haltcnt = .Execute;
 
-            self.changeMode(.Irq);
-            self.cpsr.t.write(false);
-            self.cpsr.i.write(true);
+        const ret_addr = self.r[15] - if (self.cpsr.t.read()) 2 else @as(u32, 4);
+        const new_spsr = self.cpsr.raw;
 
-            self.r[14] = r15;
-            self.spsr.raw = cpsr;
-            self.r[15] = 0x000_0018;
-        }
+        self.changeMode(.Irq);
+        self.cpsr.t.write(false);
+        self.cpsr.i.write(true);
+
+        self.r[14] = ret_addr;
+        self.spsr.raw = new_spsr;
+        self.r[15] = 0x0000_0018;
+        self.pipe.flush();
     }
 
     inline fn fetch(self: *Self, comptime T: type) T {
@@ -505,10 +518,6 @@ pub const Arm7tdmi = struct {
         return self.bus.read(T, self.r[15]);
     }
 
-    pub fn fakePC(self: *const Self) u32 {
-        return self.r[15] + 4;
-    }
-
     fn debug_log(self: *const Self, file: *const File, opcode: u32) void {
         if (self.binary_log) {
             self.skyLog(file) catch unreachable;
@@ -532,11 +541,11 @@ pub const Arm7tdmi = struct {
         prettyPrintPsr(&self.spsr);
 
         if (self.cpsr.t.read()) {
-            const opcode = self.bus.debugRead(u16, self.r[15] - 4);
+            const opcode = self.bus.dbgRead(u16, self.r[15] - 4);
             const id = thumbIdx(opcode);
             std.debug.print("opcode: ID: 0x{b:0>10} 0x{X:0>4}\n", .{ id, opcode });
         } else {
-            const opcode = self.bus.debugRead(u32, self.r[15] - 4);
+            const opcode = self.bus.dbgRead(u32, self.r[15] - 4);
             const id = armIdx(opcode);
             std.debug.print("opcode: ID: 0x{X:0>3} 0x{X:0>8}\n", .{ id, opcode });
         }
@@ -574,25 +583,6 @@ pub const Arm7tdmi = struct {
         };
     }
 
-    fn skyLog(self: *const Self, file: *const File) !void {
-        var buf: [18 * @sizeOf(u32)]u8 = undefined;
-
-        // Write Registers
-        var i: usize = 0;
-        while (i < 0x10) : (i += 1) {
-            skyWrite(&buf, i, self.r[i]);
-        }
-
-        skyWrite(&buf, 0x10, self.cpsr.raw);
-        skyWrite(&buf, 0x11, if (self.hasSPSR()) self.spsr.raw else self.cpsr.raw);
-        _ = try file.writeAll(&buf);
-    }
-
-    fn skyWrite(buf: []u8, i: usize, num: u32) void {
-        const j = @sizeOf(u32) * i;
-        std.mem.writeIntSliceNative(u32, buf[j..(j + @sizeOf(u32))], num);
-    }
-
     fn mgbaLog(self: *const Self, file: *const File, opcode: u32) !void {
         const thumb_fmt = "{X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} cpsr: {X:0>8} | {X:0>4}:\n";
         const arm_fmt = "{X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} cpsr: {X:0>8} | {X:0>8}:\n";
@@ -613,7 +603,7 @@ pub const Arm7tdmi = struct {
         const r12 = self.r[12];
         const r13 = self.r[13];
         const r14 = self.r[14];
-        const r15 = self.r[15];
+        const r15 = self.r[15] -| if (self.cpsr.t.read()) 2 else @as(u32, 4);
 
         const c_psr = self.cpsr.raw;
 
@@ -621,7 +611,7 @@ pub const Arm7tdmi = struct {
         if (self.cpsr.t.read()) {
             if (opcode >> 11 == 0x1E) {
                 // Instruction 1 of a BL Opcode, print in ARM mode
-                const other_half = self.bus.debugRead(u16, self.r[15]);
+                const other_half = self.bus.debugRead(u16, self.r[15] - 2);
                 const bl_opcode = @as(u32, opcode) << 16 | other_half;
 
                 log_str = try std.fmt.bufPrint(&buf, arm_fmt, .{ r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15, c_psr, bl_opcode });
@@ -665,6 +655,48 @@ pub fn checkCond(cpsr: PSR, cond: u4) bool {
     };
 }
 
+const Pipline = struct {
+    const Self = @This();
+    stage: [2]?u32,
+    flushed: bool,
+
+    fn init() Self {
+        return .{
+            .stage = [_]?u32{null} ** 2,
+            .flushed = false,
+        };
+    }
+
+    pub fn flush(self: *Self) void {
+        for (self.stage) |*opcode| opcode.* = null;
+        self.flushed = true;
+    }
+
+    pub fn isFull(self: *const Self) bool {
+        return self.stage[0] != null and self.stage[1] != null;
+    }
+
+    pub fn step(self: *Self, cpu: *Arm7tdmi, comptime T: type) ?u32 {
+        comptime std.debug.assert(T == u32 or T == u16);
+
+        const opcode = self.stage[0];
+
+        self.stage[0] = self.stage[1];
+        self.stage[1] = cpu.bus.read(T, cpu.r[15]);
+
+        return opcode;
+    }
+
+    fn reload(self: *Self, cpu: *Arm7tdmi, comptime T: type) void {
+        comptime std.debug.assert(T == u32 or T == u16);
+        const inc = if (T == u32) 4 else 2;
+
+        self.stage[0] = cpu.bus.read(T, cpu.r[15]);
+        self.stage[1] = cpu.bus.read(T, cpu.r[15] + inc);
+        cpu.r[15] += inc * 2;
+    }
+};
+
 pub const PSR = extern union {
     mode: Bitfield(u32, 0, 5),
     t: Bit(u32, 5),

src/core/cpu/arm/block_data_transfer.zig

@@ -55,8 +55,10 @@ pub fn blockDataTransfer(comptime P: bool, comptime U: bool, comptime S: bool, c
 
                 if (L) {
                     cpu.r[15] = bus.read(u32, und_addr);
+                    cpu.pipe.flush();
                 } else {
-                    bus.write(u32, und_addr, cpu.r[15] + 8);
+                    // FIXME: Should r15 on write be +12 ahead?
+                    bus.write(u32, und_addr, cpu.r[15] + 4);
                 }
 
                 cpu.r[rn] = if (U) cpu.r[rn] + 0x40 else cpu.r[rn] - 0x40;
@@ -86,17 +88,23 @@ pub fn blockDataTransfer(comptime P: bool, comptime U: bool, comptime S: bool, c
                     cpu.setUserModeRegister(i, bus.read(u32, address));
                 } else {
                     const value = bus.read(u32, address);
-                    cpu.r[i] = if (i == 0xF) value & 0xFFFF_FFFC else value;
-                    if (S and i == 0xF) cpu.setCpsr(cpu.spsr.raw);
+
+                    cpu.r[i] = value;
+                    if (i == 0xF) {
+                        cpu.r[i] &= ~@as(u32, 3); // Align r15
+                        cpu.pipe.flush();
+
+                        if (S) cpu.setCpsr(cpu.spsr.raw);
+                    }
                 }
             } else {
                 if (S) {
                     // Always Transfer User mode Registers
                     // This happens regardless if r15 is in the list
                     const value = cpu.getUserModeRegister(i);
-                    bus.write(u32, address, value + if (i == 0xF) 8 else @as(u32, 0)); // PC is already 4 ahead to make 12
+                    bus.write(u32, address, value + if (i == 0xF) 4 else @as(u32, 0)); // PC is already 8 ahead to make 12
                 } else {
-                    bus.write(u32, address, cpu.r[i] + if (i == 0xF) 8 else @as(u32, 0));
+                    bus.write(u32, address, cpu.r[i] + if (i == 0xF) 4 else @as(u32, 0));
                 }
             }
         }

src/core/cpu/arm/branch.zig

@@ -9,14 +9,19 @@ const sext = @import("../../util.zig").sext;
 pub fn branch(comptime L: bool) InstrFn {
     return struct {
         fn inner(cpu: *Arm7tdmi, _: *Bus, opcode: u32) void {
-            if (L) cpu.r[14] = cpu.r[15];
-            cpu.r[15] = cpu.fakePC() +% (sext(u32, u24, opcode) << 2);
+            if (L) cpu.r[14] = cpu.r[15] - 4;
+
+            cpu.r[15] +%= sext(u32, u24, opcode) << 2;
+            cpu.pipe.flush();
         }
     }.inner;
 }
 
 pub fn branchAndExchange(cpu: *Arm7tdmi, _: *Bus, opcode: u32) void {
     const rn = opcode & 0xF;
-    cpu.cpsr.t.write(cpu.r[rn] & 1 == 1);
-    cpu.r[15] = cpu.r[rn] & 0xFFFF_FFFE;
+
+    const thumb = cpu.r[rn] & 1 == 1;
+    cpu.r[15] = cpu.r[rn] & if (thumb) ~@as(u32, 1) else ~@as(u32, 3);
+    cpu.cpsr.t.write(thumb);
+    cpu.pipe.flush();
 }

src/core/cpu/arm/data_processing.zig

@@ -13,17 +13,12 @@ pub fn dataProcessing(comptime I: bool, comptime S: bool, comptime instrKind: u4
             const old_carry = @boolToInt(cpu.cpsr.c.read());
 
             // If certain conditions are met, PC is 12 ahead instead of 8
+            // TODO: What are these conditions? I can't remember
             if (!I and opcode >> 4 & 1 == 1) cpu.r[15] += 4;
+            const op1 = cpu.r[rn];
 
-            const op1 = if (rn == 0xF) cpu.fakePC() else cpu.r[rn];
-
-            var op2: u32 = undefined;
-            if (I) {
-                const amount = @truncate(u8, (opcode >> 8 & 0xF) << 1);
-                op2 = rotateRight(S, &cpu.cpsr, opcode & 0xFF, amount);
-            } else {
-                op2 = execute(S, cpu, opcode);
-            }
+            const amount = @truncate(u8, (opcode >> 8 & 0xF) << 1);
+            const op2 = if (I) rotateRight(S, &cpu.cpsr, opcode & 0xFF, amount) else execute(S, cpu, opcode);
 
             // Undo special condition from above
             if (!I and opcode >> 4 & 1 == 1) cpu.r[15] -= 4;
@@ -67,39 +62,31 @@ pub fn dataProcessing(comptime I: bool, comptime S: bool, comptime instrKind: u4
                 },
                 0x8 => {
                     // TST
-                    if (rd == 0xF) {
-                        undefinedTestBehaviour(cpu);
-                        return;
-                    }
+                    if (rd == 0xF)
+                        return undefinedTestBehaviour(cpu);
 
                     const result = op1 & op2;
                     setTestOpFlags(S, cpu, opcode, result);
                 },
                 0x9 => {
                     // TEQ
-                    if (rd == 0xF) {
-                        undefinedTestBehaviour(cpu);
-                        return;
-                    }
+                    if (rd == 0xF)
+                        return undefinedTestBehaviour(cpu);
 
                     const result = op1 ^ op2;
                     setTestOpFlags(S, cpu, opcode, result);
                 },
                 0xA => {
                     // CMP
-                    if (rd == 0xF) {
-                        undefinedTestBehaviour(cpu);
-                        return;
-                    }
+                    if (rd == 0xF)
+                        return undefinedTestBehaviour(cpu);
 
                     cmp(cpu, op1, op2);
                 },
                 0xB => {
                     // CMN
-                    if (rd == 0xF) {
-                        undefinedTestBehaviour(cpu);
-                        return;
-                    }
+                    if (rd == 0xF)
+                        return undefinedTestBehaviour(cpu);
 
                     cmn(cpu, op1, op2);
                 },
@@ -127,6 +114,8 @@ pub fn dataProcessing(comptime I: bool, comptime S: bool, comptime instrKind: u4
                     setArmLogicOpFlags(S, cpu, rd, result);
                 },
             }
+
+            if (rd == 0xF) cpu.pipe.flush();
         }
     }.inner;
 }
@@ -280,5 +269,5 @@ fn setTestOpFlags(comptime S: bool, cpu: *Arm7tdmi, opcode: u32, result: u32) vo
 
 fn undefinedTestBehaviour(cpu: *Arm7tdmi) void {
     @setCold(true);
-    cpu.setCpsr(cpu.spsr.raw);
+    cpu.setCpsrNoFlush(cpu.spsr.raw);
 }

src/core/cpu/arm/half_signed_data_transfer.zig

@@ -15,20 +15,8 @@ pub fn halfAndSignedDataTransfer(comptime P: bool, comptime U: bool, comptime I:
             const rm = opcode & 0xF;
             const imm_offset_high = opcode >> 8 & 0xF;
 
-            var base: u32 = undefined;
-            if (rn == 0xF) {
-                base = cpu.fakePC();
-                if (!L) base += 4;
-            } else {
-                base = cpu.r[rn];
-            }
-
-            var offset: u32 = undefined;
-            if (I) {
-                offset = imm_offset_high << 4 | rm;
-            } else {
-                offset = cpu.r[rm];
-            }
+            const base = cpu.r[rn] + if (!L and rn == 0xF) 4 else @as(u32, 0);
+            const offset = if (I) imm_offset_high << 4 | rm else cpu.r[rm];
 
             const modified_base = if (U) base +% offset else base -% offset;
             var address = if (P) modified_base else base;

src/core/cpu/arm/single_data_transfer.zig

@@ -14,13 +14,8 @@ pub fn singleDataTransfer(comptime I: bool, comptime P: bool, comptime U: bool,
             const rn = opcode >> 16 & 0xF;
             const rd = opcode >> 12 & 0xF;
 
-            var base: u32 = undefined;
-            if (rn == 0xF) {
-                base = cpu.fakePC();
-                if (!L) base += 4; // Offset of 12
-            } else {
-                base = cpu.r[rn];
-            }
+            // rn is r15 and L is not set, the PC is 12 ahead
+            const base = cpu.r[rn] + if (!L and rn == 0xF) 4 else @as(u32, 0);
 
             const offset = if (I) shifter.immShift(false, cpu, opcode) else opcode & 0xFFF;
 
@@ -40,18 +35,26 @@ pub fn singleDataTransfer(comptime I: bool, comptime P: bool, comptime U: bool,
             } else {
                 if (B) {
                     // STRB
-                    const value = if (rd == 0xF) cpu.r[rd] + 8 else cpu.r[rd];
+                    const value = cpu.r[rd] + if (rd == 0xF) 4 else @as(u32, 0); // PC is 12 ahead
                     bus.write(u8, address, @truncate(u8, value));
                 } else {
                     // STR
-                    const value = if (rd == 0xF) cpu.r[rd] + 8 else cpu.r[rd];
+                    const value = cpu.r[rd] + if (rd == 0xF) 4 else @as(u32, 0);
                     bus.write(u32, address, value);
                 }
             }
 
             address = modified_base;
-            if (W and P or !P) cpu.r[rn] = address;
-            if (L) cpu.r[rd] = result; // This emulates the LDR rd == rn behaviour
+            if (W and P or !P) {
+                cpu.r[rn] = address;
+                if (rn == 0xF) cpu.pipe.flush();
+            }
+
+            if (L) {
+                // This emulates the LDR rd == rn behaviour
+                cpu.r[rd] = result;
+                if (rd == 0xF) cpu.pipe.flush();
+            }
         }
     }.inner;
 }

src/core/cpu/arm/software_interrupt.zig

@@ -6,7 +6,7 @@ pub fn armSoftwareInterrupt() InstrFn {
     return struct {
         fn inner(cpu: *Arm7tdmi, _: *Bus, _: u32) void {
             // Copy Values from Current Mode
-            const r15 = cpu.r[15];
+            const ret_addr = cpu.r[15] - 4;
             const cpsr = cpu.cpsr.raw;
 
             // Switch Mode
@@ -14,9 +14,10 @@ pub fn armSoftwareInterrupt() InstrFn {
             cpu.cpsr.t.write(false); // Force ARM Mode
             cpu.cpsr.i.write(true); // Disable normal interrupts
 
-            cpu.r[14] = r15; // Resume Execution
+            cpu.r[14] = ret_addr; // Resume Execution
             cpu.spsr.raw = cpsr; // Previous mode CPSR
             cpu.r[15] = 0x0000_0008;
+            cpu.pipe.flush();
         }
     }.inner;
 }

src/core/cpu/barrel_shifter.zig

@@ -18,11 +18,9 @@ pub fn execute(comptime S: bool, cpu: *Arm7tdmi, opcode: u32) u32 {
 
 fn registerShift(comptime S: bool, cpu: *Arm7tdmi, opcode: u32) u32 {
     const rs_idx = opcode >> 8 & 0xF;
+    const rm = cpu.r[opcode & 0xF];
     const rs = @truncate(u8, cpu.r[rs_idx]);
 
-    const rm_idx = opcode & 0xF;
-    const rm = if (rm_idx == 0xF) cpu.fakePC() else cpu.r[rm_idx];
-
     return switch (@truncate(u2, opcode >> 5)) {
         0b00 => logicalLeft(S, &cpu.cpsr, rm, rs),
         0b01 => logicalRight(S, &cpu.cpsr, rm, rs),
@@ -33,9 +31,7 @@ fn registerShift(comptime S: bool, cpu: *Arm7tdmi, opcode: u32) u32 {
 
 pub fn immShift(comptime S: bool, cpu: *Arm7tdmi, opcode: u32) u32 {
     const amount = @truncate(u8, opcode >> 7 & 0x1F);
-
-    const rm_idx = opcode & 0xF;
-    const rm = if (rm_idx == 0xF) cpu.fakePC() else cpu.r[rm_idx];
+    const rm = cpu.r[opcode & 0xF];
 
     var result: u32 = undefined;
     if (amount == 0) {

src/core/cpu/thumb/block_data_transfer.zig

@@ -33,7 +33,8 @@ pub fn fmt14(comptime L: bool, comptime R: bool) InstrFn {
             if (R) {
                 if (L) {
                     const value = bus.read(u32, address);
-                    cpu.r[15] = value & 0xFFFF_FFFE;
+                    cpu.r[15] = value & ~@as(u32, 1);
+                    cpu.pipe.flush();
                 } else {
                     bus.write(u32, address, cpu.r[14]);
                 }
@@ -52,7 +53,13 @@ pub fn fmt15(comptime L: bool, comptime rb: u3) InstrFn {
             const end_address = cpu.r[rb] + 4 * countRlist(opcode);
 
             if (opcode & 0xFF == 0) {
-                if (L) cpu.r[15] = bus.read(u32, address) else bus.write(u32, address, cpu.r[15] + 4);
+                if (L) {
+                    cpu.r[15] = bus.read(u32, address);
+                    cpu.pipe.flush();
+                } else {
+                    bus.write(u32, address, cpu.r[15] + 2);
+                }
+
                 cpu.r[rb] += 0x40;
                 return;
             }

src/core/cpu/thumb/branch.zig

@@ -9,16 +9,13 @@ pub fn fmt16(comptime cond: u4) InstrFn {
     return struct {
         fn inner(cpu: *Arm7tdmi, _: *Bus, opcode: u16) void {
             // B
-            const offset = sext(u32, u8, opcode & 0xFF) << 1;
+            if (cond == 0xE or cond == 0xF)
+                cpu.panic("[CPU/THUMB.16] Undefined conditional branch with condition {}", .{cond});
 
-            const should_execute = switch (cond) {
-                0xE, 0xF => cpu.panic("[CPU/THUMB.16] Undefined conditional branch with condition {}", .{cond}),
-                else => checkCond(cpu.cpsr, cond),
-            };
+            if (!checkCond(cpu.cpsr, cond)) return;
 
-            if (should_execute) {
-                cpu.r[15] = (cpu.r[15] + 2) +% offset;
-            }
+            cpu.r[15] +%= sext(u32, u8, opcode & 0xFF) << 1;
+            cpu.pipe.flush();
         }
     }.inner;
 }
@@ -27,8 +24,8 @@ pub fn fmt18() InstrFn {
     return struct {
         // B but conditional
         fn inner(cpu: *Arm7tdmi, _: *Bus, opcode: u16) void {
-            const offset = sext(u32, u11, opcode & 0x7FF) << 1;
-            cpu.r[15] = (cpu.r[15] + 2) +% offset;
+            cpu.r[15] +%= sext(u32, u11, opcode & 0x7FF) << 1;
+            cpu.pipe.flush();
         }
     }.inner;
 }
@@ -41,13 +38,16 @@ pub fn fmt19(comptime is_low: bool) InstrFn {
 
             if (is_low) {
                 // Instruction 2
-                const old_pc = cpu.r[15];
+                const next_opcode = cpu.r[15] - 2;
 
                 cpu.r[15] = cpu.r[14] +% (offset << 1);
-                cpu.r[14] = old_pc | 1;
+                cpu.r[14] = next_opcode | 1;
+
+                cpu.pipe.flush();
             } else {
                 // Instruction 1
-                cpu.r[14] = (cpu.r[15] + 2) +% (sext(u32, u11, offset) << 12);
+                const lr_offset = sext(u32, u11, offset) << 12;
+                cpu.r[14] = (cpu.r[15] +% lr_offset) & ~@as(u32, 1);
             }
         }
     }.inner;

src/core/cpu/thumb/data_processing.zig

@@ -133,10 +133,9 @@ pub fn fmt12(comptime isSP: bool, comptime rd: u3) InstrFn {
     return struct {
         fn inner(cpu: *Arm7tdmi, _: *Bus, opcode: u16) void {
             // ADD
-            const left = if (isSP) cpu.r[13] else (cpu.r[15] + 2) & 0xFFFF_FFFD;
+            const left = if (isSP) cpu.r[13] else cpu.r[15] & ~@as(u32, 2);
             const right = (opcode & 0xFF) << 2;
-            const result = left + right;
-            cpu.r[rd] = result;
+            cpu.r[rd] = left + right;
         }
     }.inner;
 }

src/core/cpu/thumb/data_transfer.zig

@@ -11,7 +11,9 @@ pub fn fmt6(comptime rd: u3) InstrFn {
         fn inner(cpu: *Arm7tdmi, bus: *Bus, opcode: u16) void {
             // LDR
             const offset = (opcode & 0xFF) << 2;
-            cpu.r[rd] = bus.read(u32, (cpu.r[15] + 2 & 0xFFFF_FFFD) + offset);
+
+            // Bit 1 of the PC intentionally ignored
+            cpu.r[rd] = bus.read(u32, (cpu.r[15] & ~@as(u32, 2)) + offset);
         }
     }.inner;
 }

src/core/cpu/thumb/software_interrupt.zig

@@ -6,7 +6,7 @@ pub fn fmt17() InstrFn {
     return struct {
         fn inner(cpu: *Arm7tdmi, _: *Bus, _: u16) void {
             // Copy Values from Current Mode
-            const r15 = cpu.r[15];
+            const ret_addr = cpu.r[15] - 2;
             const cpsr = cpu.cpsr.raw;
 
             // Switch Mode
@@ -14,9 +14,10 @@ pub fn fmt17() InstrFn {
             cpu.cpsr.t.write(false); // Force ARM Mode
             cpu.cpsr.i.write(true); // Disable normal interrupts
 
-            cpu.r[14] = r15; // Resume Execution
+            cpu.r[14] = ret_addr; // Resume Execution
             cpu.spsr.raw = cpsr; // Previous mode CPSR
             cpu.r[15] = 0x0000_0008;
+            cpu.pipe.flush();
         }
     }.inner;
 }

src/core/emu.zig

@@ -12,8 +12,11 @@ const Thread = std.Thread;
 const Atomic = std.atomic.Atomic;
 const Allocator = std.mem.Allocator;
 
-const sync_audio = false;
-const sync_video: RunKind = .UnlimitedFPS;
+// TODO: Move these to a TOML File
+const sync_audio = true; // Enable Audio Sync
+const sync_video: RunKind = .LimitedFPS; // Configure Video Sync
+pub const win_scale = 3; // 1x, 2x, 3x, etc. Window Scaling
+pub const cpu_logging = false; // Enable detailed CPU logging
 
 // 228 Lines which consist of 308 dots (which are 4 cycles long)
 const cycles_per_frame: u64 = 228 * (308 * 4); //280896

src/core/util.zig

@@ -1,6 +1,7 @@
 const std = @import("std");
 const builtin = @import("builtin");
 const Log2Int = std.math.Log2Int;
+const Arm7tdmi = @import("cpu.zig").Arm7tdmi;
 
 // Sign-Extend value of type `T` to type `U`
 pub fn sext(comptime T: type, comptime U: type, value: T) T {
@@ -112,3 +113,64 @@ pub fn writeUndefined(log: anytype, comptime format: []const u8, args: anytype)
     log.warn(format, args);
     if (builtin.mode == .Debug) std.debug.panic("TODO: Implement I/O Register", .{});
 }
+
+pub const Logger = struct {
+    const Self = @This();
+
+    buf: std.io.BufferedWriter(4096 << 2, std.fs.File.Writer),
+
+    pub fn init(file: std.fs.File) Self {
+        return .{
+            .buf = .{ .unbuffered_writer = file.writer() },
+        };
+    }
+
+    pub fn print(self: *Self, comptime format: []const u8, args: anytype) !void {
+        try self.buf.writer().print(format, args);
+    }
+
+    pub fn mgbaLog(self: *Self, arm7tdmi: *const Arm7tdmi, opcode: u32) void {
+        const fmt_base = "{X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} {X:0>8} cpsr: {X:0>8} | ";
+        const thumb_fmt = fmt_base ++ "{X:0>4}:\n";
+        const arm_fmt = fmt_base ++ "{X:0>8}:\n";
+
+        if (arm7tdmi.cpsr.t.read()) {
+            if (opcode >> 11 == 0x1E) {
+                // Instruction 1 of a BL Opcode, print in ARM mode
+                const low = arm7tdmi.bus.dbgRead(u16, arm7tdmi.r[15]);
+                const bl_opcode = @as(u32, opcode) << 16 | low;
+
+                self.print(arm_fmt, Self.fmtArgs(arm7tdmi, bl_opcode)) catch @panic("failed to write to log file");
+            } else {
+                self.print(thumb_fmt, Self.fmtArgs(arm7tdmi, opcode)) catch @panic("failed to write to log file");
+            }
+        } else {
+            self.print(arm_fmt, Self.fmtArgs(arm7tdmi, opcode)) catch @panic("failed to write to log file");
+        }
+    }
+
+    fn fmtArgs(arm7tdmi: *const Arm7tdmi, opcode: u32) FmtArgTuple {
+        return .{
+            arm7tdmi.r[0],
+            arm7tdmi.r[1],
+            arm7tdmi.r[2],
+            arm7tdmi.r[3],
+            arm7tdmi.r[4],
+            arm7tdmi.r[5],
+            arm7tdmi.r[6],
+            arm7tdmi.r[7],
+            arm7tdmi.r[8],
+            arm7tdmi.r[9],
+            arm7tdmi.r[10],
+            arm7tdmi.r[11],
+            arm7tdmi.r[12],
+            arm7tdmi.r[13],
+            arm7tdmi.r[14],
+            arm7tdmi.r[15] - if (arm7tdmi.cpsr.t.read()) 2 else @as(u32, 4),
+            arm7tdmi.cpsr.raw,
+            opcode,
+        };
+    }
+};
+
+const FmtArgTuple = std.meta.Tuple(&.{ u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32, u32 });

src/main.zig

@@ -14,6 +14,7 @@ const Allocator = std.mem.Allocator;
 const log = std.log.scoped(.CLI);
 const width = @import("core/ppu.zig").width;
 const height = @import("core/ppu.zig").height;
+const arm7tdmi_logging = @import("core/emu.zig").cpu_logging;
 pub const log_level = if (builtin.mode != .Debug) .info else std.log.default_level;
 
 // TODO: Reimpl Logging
@@ -48,6 +49,10 @@ pub fn main() anyerror!void {
 
     var arm7tdmi = Arm7tdmi.init(&scheduler, &bus);
 
+    const log_file: ?std.fs.File = if (arm7tdmi_logging) try std.fs.cwd().createFile("zba.log", .{}) else null;
+    defer if (log_file) |file| file.close();
+
+    if (log_file) |file| arm7tdmi.attach(file);
     bus.attach(&arm7tdmi); // TODO: Shrink Surface (only CPSR and  r15?)
     if (paths.bios == null) arm7tdmi.fastBoot();