Add profiling feature

To avoid heavy inlining and provide more context on where time is being
spent.

Cargo.toml

@@ -13,3 +13,6 @@ members = [
 exclude = [
     "rusty-date",
 ]
+
+[profile.release]
+debug = true

ppu/Cargo.toml

@@ -11,6 +11,9 @@ categories = ["embedded", "gaming"]
 keywords = ["embedded", "gameboy", "playdate"]
 readme = "../README.md"
 
+[features]
+profile = []
+
 [dependencies]
 sm83 = { path =  "../sm83", version = "0.1.0" }
 tock-registers = "0.9.0"

ppu/src/lib.rs

@@ -112,6 +112,7 @@ const OBJ_OFFSET_X: usize = 8;
 
 static_assertions::const_assert_eq!(70224, LINE_LENGTH * NUM_LINES);
 
+#[cfg_attr(feature = "profile", inline(never))]
 fn mode_for_current_cycle_count(line_cycles: Cycles, line: usize) -> Mode {
     // This is a simplified implementation that considers a fixed time mode 3.
     let line_cycles: usize = line_cycles.into();
@@ -159,7 +160,8 @@ impl Ppu {
         self.mode
     }
 
-    pub fn update_line_and_cycles(&mut self, cycles: Cycles) {
+    #[cfg_attr(feature = "profile", inline(never))]
+    fn update_line_and_cycles(&mut self, cycles: Cycles) {
         self.cycles = self.cycles + cycles;
         if self.cycles >= Cycles::new(LINE_LENGTH) {
             self.cycles = self.cycles - Cycles::new(LINE_LENGTH);
@@ -171,6 +173,7 @@ impl Ppu {
     }
 
     /// Runs the PPU for the given number of cycles and then returns the PPU state
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn step(
         &mut self,
         cycles: Cycles,
@@ -193,6 +196,7 @@ impl Ppu {
         &self.framebuffer
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn update_lcd_irq(&mut self) -> Interrupts {
         let lyc_eq_ly = self.regs.status.read(STAT::LYC_INT_SELECT) != 0
             && self.regs.status.read(STAT::LYC_EQ_LY) != 0;
@@ -217,6 +221,7 @@ impl Ppu {
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn oam_scan(&mut self) {
         self.selected_oam_entries = heapless::Vec::new();
 
@@ -238,6 +243,7 @@ impl Ppu {
             .collect();
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn step_inner(&mut self, new_mode: Mode, render: bool) -> (Interrupts, PpuResult) {
         const NO_IRQ: Interrupts = Interrupts::new();
         if self.mode == new_mode {
@@ -265,6 +271,7 @@ impl Ppu {
         (NO_IRQ, PpuResult::InProgress(self.mode))
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn draw_line_background(&self, line: &mut [PaletteIndex; DISPLAY_WIDTH]) -> Palette {
         let bg_win_enable = self.regs.lcdc.read(regs::LCDC::BG_AND_WINDOW_ENABLE) != 0;
         if !bg_win_enable {
@@ -319,6 +326,7 @@ impl Ppu {
         self.regs.bg_palette
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn draw_line_window(&self, line: &mut [PaletteIndex; DISPLAY_WIDTH]) {
         let bg_win_enable = self.regs.lcdc.read(regs::LCDC::BG_AND_WINDOW_ENABLE) != 0;
         if !bg_win_enable {
@@ -374,6 +382,7 @@ impl Ppu {
             .for_each(|(dest, palette_index)| *dest = palette_index);
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn draw_line_objects(
         &self,
         bg_line: &[PaletteIndex; DISPLAY_WIDTH],
@@ -445,6 +454,7 @@ impl Ppu {
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn draw_line(&mut self) {
         if self.regs.lcdc.read(regs::LCDC::ENABLE) == 0 {
             return;
@@ -478,6 +488,7 @@ impl Ppu {
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     fn update_registers(&mut self) {
         let line = self.line as u8;
         self.regs.ly = line;

ppu/src/oam.rs

@@ -131,11 +131,13 @@ impl Oam {
         &*self.objects
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn read(&self, address: sm83::memory::Address) -> u8 {
         let (object_idx, object_member_offset) = Self::cpu_addr_to_object_addr(address);
         self.objects[object_idx].read(object_member_offset)
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn write(&mut self, address: sm83::memory::Address, value: u8) {
         let (object_idx, object_member_offset) = Self::cpu_addr_to_object_addr(address);
         self.objects[object_idx].write(object_member_offset, value);

ppu/src/regs.rs

@@ -159,6 +159,7 @@ impl Registers {
             .set((self.status.get() & RO_BITS) | (new_val & !RO_BITS));
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn read(&self, address: sm83::memory::Address) -> u8 {
         match address {
             0xFF40 => self.lcdc.get(),
@@ -179,6 +180,7 @@ impl Registers {
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn write(&mut self, address: sm83::memory::Address, value: u8) {
         match address {
             0xFF40 => self.lcdc.set(value),

ppu/src/vram.rs

@@ -211,6 +211,7 @@ impl TileMap {
         })
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn line(&self, line: usize) -> &[TileIndex; TILE_MAP_WIDTH] {
         let line = (line / TILE_HEIGHT) % TILE_MAP_HEIGHT;
         &self.0[line]
@@ -288,13 +289,15 @@ impl Vram {
         (tile_map_idx, tile_map_address)
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub(crate) fn get_bg_tile_map(&self, map: crate::regs::LCDC::BG_TILE_MAP::Value) -> &TileMap {
         match map {
             crate::regs::LCDC::BG_TILE_MAP::Value::HighMap => &self.0.tile_maps[1],
             crate::regs::LCDC::BG_TILE_MAP::Value::LowMap => &self.0.tile_maps[0],
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub(crate) fn get_win_tile_map(
         &self,
         map: crate::regs::LCDC::WINDOW_TILE_MAP::Value,
@@ -305,6 +308,7 @@ impl Vram {
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub(crate) fn get_tile(
         &self,
         index: TileIndex,
@@ -319,6 +323,7 @@ impl Vram {
         self.0.tile_blocks[block].get_tile(index)
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn read(&self, address: sm83::memory::Address) -> u8 {
         if address < 0x9800 {
             let (blk_idx, blk_address) = Self::vram_address_to_block_address(address);
@@ -329,6 +334,7 @@ impl Vram {
         }
     }
 
+    #[cfg_attr(feature = "profile", inline(never))]
     pub fn write(&mut self, address: sm83::memory::Address, value: u8) {
         if address < 0x9800 {
             let (blk_idx, blk_address) = Self::vram_address_to_block_address(address);

sm83/Cargo.toml

@@ -11,6 +11,9 @@ categories = ["embedded", "gaming"]
 keywords = ["embedded", "gameboy", "playdate"]
 readme = "../README.md"
 
+[features]
+profile = []
+
 [dependencies]
 sm83_decoder_macros = { path = "../sm83_decoder_macros" }