Start implementing profiling

build.rs

@@ -23,7 +23,9 @@ fn main() -> Result<(), Box<dyn Error>> {
         .define("VB_LITTLE_ENDIAN", None)
         .define("VB_SIGNED_PROPAGATE", None)
         .define("VB_DIV_GENERIC", None)
+        .define("VB_DIRECT_EXCEPTION", "on_exception")
         .define("VB_DIRECT_EXECUTE", "on_execute")
+        .define("VB_DIRECT_FETCH", "on_fetch")
         .define("VB_DIRECT_FRAME", "on_frame")
         .define("VB_DIRECT_READ", "on_read")
         .define("VB_DIRECT_WRITE", "on_write")

src/emulator.rs

@@ -18,7 +18,7 @@ use tracing::{error, warn};
 
 use crate::{
     audio::Audio,
-    emulator::cart::Cart,
+    emulator::{cart::Cart, profiler::Profiler},
     graphics::TextureSink,
     memory::{MemoryRange, MemoryRegion},
 };
@@ -27,6 +27,7 @@ pub use shrooms_vb_core::{VBKey, VBRegister, VBWatchpointType};
 
 mod address_set;
 mod cart;
+mod profiler;
 mod shrooms_vb_core;
 
 #[derive(PartialEq, Eq, Hash, Clone, Copy, Debug)]
@@ -69,6 +70,7 @@ pub struct EmulatorBuilder {
     audio_on: Arc<[AtomicBool; 2]>,
     linked: Arc<AtomicBool>,
     start_paused: bool,
+    monitor_events: bool,
 }
 
 impl EmulatorBuilder {
@@ -85,6 +87,7 @@ impl EmulatorBuilder {
             audio_on: Arc::new([AtomicBool::new(true), AtomicBool::new(true)]),
             linked: Arc::new(AtomicBool::new(false)),
             start_paused: false,
+            monitor_events: false,
         };
         let client = EmulatorClient {
             queue,
@@ -110,6 +113,13 @@ impl EmulatorBuilder {
         }
     }
 
+    pub fn monitor_events(self, monitor_events: bool) -> Self {
+        Self {
+            monitor_events,
+            ..self
+        }
+    }
+
     pub fn build(self) -> Result<Emulator> {
         let mut emulator = Emulator::new(
             self.commands,
@@ -124,6 +134,9 @@ impl EmulatorBuilder {
         if self.start_paused {
             emulator.pause_sims()?;
         }
+        if self.monitor_events {
+            emulator.monitor_events(SimId::Player1)?;
+        }
         Ok(emulator)
     }
 }
@@ -137,6 +150,7 @@ pub struct Emulator {
     state: Arc<Atomic<EmulatorState>>,
     audio_on: Arc<[AtomicBool; 2]>,
     linked: Arc<AtomicBool>,
+    profilers: [Profiler; 2],
     renderers: HashMap<SimId, TextureSink>,
     messages: HashMap<SimId, mpsc::Sender<Toast>>,
     debuggers: HashMap<SimId, DebugInfo>,
@@ -164,6 +178,7 @@ impl Emulator {
             state,
             audio_on,
             linked,
+            profilers: [Profiler::new(), Profiler::new()],
             renderers: HashMap::new(),
             messages: HashMap::new(),
             debuggers: HashMap::new(),
@@ -213,6 +228,7 @@ impl Emulator {
         }
         let sim = &mut self.sims[index];
         sim.reset();
+        sim.monitor_events(self.profilers[sim_id.to_index()].is_enabled());
         if let Some(cart) = new_cart {
             sim.load_cart(cart.rom.clone(), cart.sram.clone())?;
             self.carts[index] = Some(cart);
@@ -383,6 +399,11 @@ impl Emulator {
         self.watched_regions.insert(range, region);
     }
 
+    fn monitor_events(&mut self, sim_id: SimId) -> Result<()> {
+        self.profilers[sim_id.to_index()].enable(true);
+        self.reset_sim(sim_id, None)
+    }
+
     pub fn run(&mut self) {
         loop {
             let idle = self.tick();
@@ -438,12 +459,20 @@ impl Emulator {
         let p1_running = running && p1_state == SimState::Ready;
         let p2_running = running && p2_state == SimState::Ready;
         let mut idle = !p1_running && !p2_running;
-        if p1_running && p2_running {
+
-            Sim::emulate_many(&mut self.sims);
+        let cycles = self.emulate(p1_running, p2_running);
-        } else if p1_running {
+
-            self.sims[SimId::Player1.to_index()].emulate();
+        // if we're profiling, track events
-        } else if p2_running {
+        for ((sim, profiler), running) in self
-            self.sims[SimId::Player2.to_index()].emulate();
+            .sims
+            .iter_mut()
+            .zip(self.profilers.iter_mut())
+            .zip([p1_running, p2_running])
+        {
+            if !running || !profiler.is_enabled() {
+                continue;
+            }
+            profiler.track(cycles, sim.take_sim_event());
         }
 
         if state == EmulatorState::Stepping {
@@ -470,7 +499,7 @@ impl Emulator {
                 let Some(sim) = self.sims.get_mut(sim_id.to_index()) else {
                     continue;
                 };
-                if let Some(reason) = sim.stop_reason() {
+                if let Some(reason) = sim.take_stop_reason() {
                     let stop_reason = match reason {
                         StopReason::Stepped => DebugStopReason::Trace,
                         StopReason::Watchpoint(watch, address) => {
@@ -529,6 +558,19 @@ impl Emulator {
         idle
     }
 
+    fn emulate(&mut self, p1_running: bool, p2_running: bool) -> u32 {
+        const MAX_CYCLES: u32 = 20_000_000;
+        let mut cycles = MAX_CYCLES;
+        if p1_running && p2_running {
+            Sim::emulate_many(&mut self.sims, &mut cycles);
+        } else if p1_running {
+            self.sims[SimId::Player1.to_index()].emulate(&mut cycles);
+        } else if p2_running {
+            self.sims[SimId::Player2.to_index()].emulate(&mut cycles);
+        }
+        MAX_CYCLES - cycles
+    }
+
     fn handle_command(&mut self, command: EmulatorCommand) {
         match command {
             EmulatorCommand::ConnectToSim(sim_id, renderer, messages) => {

src/emulator/profiler.rs

@@ -0,0 +1,23 @@
+use crate::emulator::shrooms_vb_core::SimEvent;
+
+pub struct Profiler {
+    enabled: bool,
+}
+
+impl Profiler {
+    pub fn new() -> Self {
+        Self { enabled: false }
+    }
+
+    pub fn enable(&mut self, enabled: bool) {
+        self.enabled = enabled;
+    }
+
+    pub fn is_enabled(&self) -> bool {
+        self.enabled
+    }
+
+    pub fn track(&mut self, cycles: u32, event: Option<SimEvent>) {
+        println!("profiler {cycles} {event:0x?}");
+    }
+}

src/emulator/shrooms_vb_core.rs

@@ -71,8 +71,16 @@ pub enum VBWatchpointType {
     Access,
 }
 
+type OnException = extern "C" fn(sim: *mut VB, cause: *mut u16) -> c_int;
 type OnExecute =
     extern "C" fn(sim: *mut VB, address: u32, code: *const u16, length: c_int) -> c_int;
+type OnFetch = extern "C" fn(
+    sim: *mut VB,
+    fetch: c_int,
+    address: u32,
+    value: *mut i32,
+    cycles: *mut u32,
+) -> c_int;
 type OnFrame = extern "C" fn(sim: *mut VB) -> c_int;
 type OnRead = extern "C" fn(
     sim: *mut VB,
@@ -135,8 +143,15 @@ unsafe extern "C" {
     fn vb_set_cart_ram(sim: *mut VB, sram: *mut c_void, size: u32) -> c_int;
     #[link_name = "vbSetCartROM"]
     fn vb_set_cart_rom(sim: *mut VB, rom: *mut c_void, size: u32) -> c_int;
+    #[link_name = "vbSetExceptionCallback"]
+    fn vb_set_exception_callback(
+        sim: *mut VB,
+        callback: Option<OnException>,
+    ) -> Option<OnException>;
     #[link_name = "vbSetExecuteCallback"]
     fn vb_set_execute_callback(sim: *mut VB, callback: Option<OnExecute>) -> Option<OnExecute>;
+    #[link_name = "vbSetFetchCallback"]
+    fn vb_set_fetch_callback(sim: *mut VB, callback: Option<OnFetch>) -> Option<OnFetch>;
     #[link_name = "vbSetFrameCallback"]
     fn vb_set_frame_callback(sim: *mut VB, callback: Option<OnFrame>) -> Option<OnFrame>;
     #[link_name = "vbSetKeys"]
@@ -180,11 +195,21 @@ extern "C" fn on_frame(sim: *mut VB) -> c_int {
 }
 
 #[unsafe(no_mangle)]
-extern "C" fn on_execute(sim: *mut VB, address: u32, _code: *const u16, _length: c_int) -> c_int {
+extern "C" fn on_execute(sim: *mut VB, address: u32, code: *const u16, length: c_int) -> c_int {
     // SAFETY: the *mut VB owns its userdata.
     // There is no way for the userdata to be null or otherwise invalid.
     let data: &mut VBState = unsafe { &mut *vb_get_user_data(sim).cast() };
 
+    if data.monitor.enabled {
+        // SAFETY: length is the length of code, in elements
+        let code = unsafe { slice::from_raw_parts(code, length as usize) };
+        if data.monitor.detect_event(sim, address, code) {
+            // Something interesting will happen after this instruction is run.
+            // The on_fetch callback will fire when it does.
+            unsafe { vb_set_fetch_callback(sim, Some(on_fetch)) };
+        }
+    }
+
     let mut stopped = data.stop_reason.is_some();
     if data.step_from.is_some_and(|s| s != address) {
         data.step_from = None;
@@ -199,6 +224,35 @@ extern "C" fn on_execute(sim: *mut VB, address: u32, _code: *const u16, _length:
     if stopped { 1 } else { 0 }
 }
 
+#[unsafe(no_mangle)]
+extern "C" fn on_fetch(
+    sim: *mut VB,
+    _fetch: c_int,
+    _address: u32,
+    _value: *mut i32,
+    _cycles: *mut u32,
+) -> c_int {
+    // SAFETY: the *mut VB owns its userdata.
+    // There is no way for the userdata to be null or otherwise invalid.
+    let data: &mut VBState = unsafe { &mut *vb_get_user_data(sim).cast() };
+    data.monitor.event = data.monitor.queued_event.take();
+    unsafe { vb_set_exception_callback(sim, Some(on_exception)) };
+    unsafe { vb_set_fetch_callback(sim, None) };
+    1
+}
+
+#[unsafe(no_mangle)]
+extern "C" fn on_exception(sim: *mut VB, cause: *mut u16) -> c_int {
+    // SAFETY: the *mut VB owns its userdata.
+    // There is no way for the userdata to be null or otherwise invalid.
+    let data: &mut VBState = unsafe { &mut *vb_get_user_data(sim).cast() };
+    data.monitor.event = data.monitor.queued_event.take();
+    data.monitor.queued_event = Some(SimEvent::Interrupt(unsafe { *cause }));
+    unsafe { vb_set_exception_callback(sim, None) };
+    unsafe { vb_set_fetch_callback(sim, Some(on_fetch)) };
+    if data.monitor.event.is_some() { 1 } else { 0 }
+}
+
 #[unsafe(no_mangle)]
 extern "C" fn on_read(
     sim: *mut VB,
@@ -260,6 +314,83 @@ extern "C" fn on_write(
     0
 }
 
+#[allow(dead_code)]
+#[derive(Debug)]
+pub enum SimEvent {
+    Call(u32),
+    Return,
+    Interrupt(u16),
+    Reti,
+}
+
+struct EventMonitor {
+    enabled: bool,
+    event: Option<SimEvent>,
+    queued_event: Option<SimEvent>,
+}
+
+impl EventMonitor {
+    fn new() -> Self {
+        Self {
+            enabled: false,
+            event: None,
+            queued_event: None,
+        }
+    }
+
+    fn detect_event(&mut self, sim: *mut VB, address: u32, code: &[u16]) -> bool {
+        self.queued_event = self.do_detect_event(sim, address, code);
+        self.queued_event.is_some()
+    }
+
+    fn do_detect_event(&self, sim: *mut VB, address: u32, code: &[u16]) -> Option<SimEvent> {
+        const JAL_OPCODE: u16 = 0b101011;
+        const JMP_OPCODE: u16 = 0b000110;
+        const RETI_OPCODE: u16 = 0b011001;
+
+        const fn format_i_reg_1(code: &[u16]) -> u8 {
+            (code[0] & 0x1f) as u8
+        }
+
+        const fn format_iv_disp(code: &[u16]) -> i32 {
+            let value = ((code[0] & 0x3ff) as i32) << 16 | (code[1] as i32);
+            value << 6 >> 6
+        }
+
+        let opcode = code[0] >> 10;
+
+        if opcode == JAL_OPCODE {
+            let disp = format_iv_disp(code);
+            if disp != 4 {
+                // JAL .+4 is how programs get r31 to a known value for indirect calls
+                // (which we detect later.)
+                // Any other JAL is a function call.
+                return Some(SimEvent::Call(address.wrapping_add_signed(disp)));
+            }
+        }
+
+        if opcode == JMP_OPCODE {
+            let jmp_reg = format_i_reg_1(code);
+            if jmp_reg == 31 {
+                // JMP[r31] is a return
+                return Some(SimEvent::Return);
+            }
+            let r31 = unsafe { vb_get_program_register(sim, 31) };
+            if r31 as u32 == address.wrapping_add(2) {
+                // JMP anywhere else, if r31 points to after the JMP, is an indirect call
+                let target = unsafe { vb_get_program_register(sim, jmp_reg as u32) };
+                return Some(SimEvent::Call(target as u32));
+            }
+        }
+
+        if opcode == RETI_OPCODE {
+            return Some(SimEvent::Reti);
+        }
+
+        None
+    }
+}
+
 const AUDIO_CAPACITY_SAMPLES: usize = 834 * 4;
 const AUDIO_CAPACITY_FLOATS: usize = AUDIO_CAPACITY_SAMPLES * 2;
 pub const EXPECTED_FRAME_SIZE: usize = 834 * 2;
@@ -267,6 +398,7 @@ pub const EXPECTED_FRAME_SIZE: usize = 834 * 2;
 struct VBState {
     frame_seen: bool,
     stop_reason: Option<StopReason>,
+    monitor: EventMonitor,
     step_from: Option<u32>,
     breakpoints: Vec<u32>,
     read_watchpoints: AddressSet,
@@ -277,6 +409,7 @@ struct VBState {
 impl VBState {
     fn needs_execute_callback(&self) -> bool {
         self.step_from.is_some()
+            || self.monitor.enabled
             || !self.breakpoints.is_empty()
             || !self.read_watchpoints.is_empty()
             || !self.write_watchpoints.is_empty()
@@ -311,6 +444,7 @@ impl Sim {
         let state = VBState {
             frame_seen: false,
             stop_reason: None,
+            monitor: EventMonitor::new(),
             step_from: None,
             breakpoints: vec![],
             read_watchpoints: AddressSet::new(),
@@ -332,6 +466,23 @@ impl Sim {
         unsafe { vb_reset(self.sim) };
     }
 
+    pub fn monitor_events(&mut self, enabled: bool) {
+        let state = self.get_state();
+        state.monitor.enabled = enabled;
+        state.monitor.event = None;
+        state.monitor.queued_event = None;
+        if enabled {
+            unsafe { vb_set_execute_callback(self.sim, Some(on_execute)) };
+            unsafe { vb_set_exception_callback(self.sim, Some(on_exception)) };
+        } else {
+            if !state.needs_execute_callback() {
+                unsafe { vb_set_execute_callback(self.sim, None) };
+            }
+            unsafe { vb_set_exception_callback(self.sim, None) };
+            unsafe { vb_set_fetch_callback(self.sim, None) };
+        }
+    }
+
     pub fn load_cart(&mut self, mut rom: Vec<u8>, mut sram: Vec<u8>) -> Result<()> {
         self.unload_cart();
 
@@ -394,16 +545,14 @@ impl Sim {
         unsafe { vb_set_peer(self.sim, ptr::null_mut()) };
     }
 
-    pub fn emulate(&mut self) {
+    pub fn emulate(&mut self, cycles: &mut u32) {
-        let mut cycles = 20_000_000;
+        unsafe { vb_emulate(self.sim, cycles) };
-        unsafe { vb_emulate(self.sim, &mut cycles) };
     }
 
-    pub fn emulate_many(sims: &mut [Sim]) {
+    pub fn emulate_many(sims: &mut [Sim], cycles: &mut u32) {
-        let mut cycles = 20_000_000;
         let count = sims.len() as c_uint;
         let sims = sims.as_mut_ptr().cast();
-        unsafe { vb_emulate_ex(sims, count, &mut cycles) };
+        unsafe { vb_emulate_ex(sims, count, cycles) };
     }
 
     pub fn read_pixels(&mut self, buffers: &mut [u8]) -> bool {
@@ -632,7 +781,7 @@ impl Sim {
         Some(string)
     }
 
-    pub fn stop_reason(&mut self) -> Option<StopReason> {
+    pub fn take_stop_reason(&mut self) -> Option<StopReason> {
         let data = self.get_state();
         let reason = data.stop_reason.take();
         if !data.needs_execute_callback() {
@@ -641,6 +790,11 @@ impl Sim {
         reason
     }
 
+    pub fn take_sim_event(&mut self) -> Option<SimEvent> {
+        let data = self.get_state();
+        data.monitor.event.take()
+    }
+
     fn get_state(&mut self) -> &mut VBState {
         // SAFETY: the *mut VB owns its userdata.
         // There is no way for the userdata to be null or otherwise invalid.

src/main.rs

@@ -31,6 +31,9 @@ struct Args {
     /// Start a GDB/LLDB debug server on this port.
     #[arg(short, long)]
     debug_port: Option<u16>,
+    /// Enable profiling a game
+    #[arg(short, long)]
+    profile: bool,
 }
 
 fn init_logger() {
@@ -106,6 +109,12 @@ fn main() -> Result<()> {
         }
         builder = builder.start_paused(true);
     }
+    if args.profile {
+        if args.rom.is_none() {
+            bail!("to start profiling, please select a game.");
+        }
+        builder = builder.monitor_events(true)
+    }
 
     ThreadBuilder::default()
         .name("Emulator".to_owned())