Avoid several stack allocations

kelpsyberry · Feb 7, 2024 · aaefca1 · aaefca1
1 parent ab22c0e
commit aaefca1
Show file tree

Hide file tree

Showing 7 changed files with 19 additions and 17 deletions.
diff --git a/core/src/ds_slot/rom/key1.rs b/core/src/ds_slot/rom/key1.rs
@@ -1,19 +1,22 @@
 use crate::{cpu::arm7, utils::Bytes};
 
 #[derive(Clone)]
+#[repr(C)]
 pub struct KeyBuffer<const LEVEL_3: bool> {
     key_buf: [u32; 0x412],
     key_code: [u32; 3],
 }
 
 impl<const LEVEL_3: bool> KeyBuffer<LEVEL_3> {
-    pub fn new<const MODULO: usize>(id_code: u32, arm7_bios: &Bytes<{ arm7::BIOS_SIZE }>) -> Self {
-        let key_code = [id_code, id_code >> 1, id_code << 1];
-        let mut key_buf = [0; 0x412];
-        for (i, word) in key_buf.iter_mut().enumerate() {
+    pub fn new_boxed<const MODULO: usize>(
+        id_code: u32,
+        arm7_bios: &Bytes<{ arm7::BIOS_SIZE }>,
+    ) -> Box<Self> {
+        let mut result = unsafe { Box::<Self>::new_zeroed().assume_init() };
+        result.key_code = [id_code, id_code >> 1, id_code << 1];
+        for (i, word) in result.key_buf.iter_mut().enumerate() {
             *word = arm7_bios.read_le(0x30 + (i << 2));
         }
-        let mut result = KeyBuffer { key_buf, key_code };
         result.apply_key_code::<MODULO>();
         result.apply_key_code::<MODULO>();
         result

diff --git a/core/src/ds_slot/rom/normal.rs b/core/src/ds_slot/rom/normal.rs
@@ -63,7 +63,7 @@ impl Normal {
             contents,
             rom_mask,
             chip_id,
-            key_buf: arm7_bios.map(|bios| Box::new(key1::KeyBuffer::new::<2>(game_code, bios))),
+            key_buf: arm7_bios.map(|bios| key1::KeyBuffer::new_boxed::<2>(game_code, bios)),
             stage: Stage::Initial,
         })
     }

diff --git a/core/src/gpu/engine_3d.rs b/core/src/gpu/engine_3d.rs
@@ -446,7 +446,7 @@ impl Engine3d {
             gx_status: GxStatus(0),
 
             gx_fifo_irq_requested: false,
-            gx_fifo: Box::new(Fifo::new()),
+            gx_fifo: unsafe { Box::new_zeroed().assume_init() },
             gx_pipe: Fifo::new(),
             cur_packed_commands: 0,
             remaining_command_params: 0,
@@ -507,7 +507,7 @@ impl Engine3d {
 
             vert_ram_level: 0,
             poly_ram_level: 0,
-            vert_ram: Box::new([ScreenVertex::new(); 6144]),
+            vert_ram: unsafe { Box::new_zeroed().assume_init() },
             poly_ram: unsafe { Box::new_zeroed().assume_init() },
 
             rendering_state: RenderingState {

diff --git a/render/soft-3d/src/lib.rs b/render/soft-3d/src/lib.rs
@@ -387,7 +387,7 @@ fn depth_test_less_back_facing(a: u32, b: u32, _: PixelAttrs) -> bool {
 impl Renderer {
     pub fn new() -> Self {
         Renderer {
-            color_buffer: Box::new(Scanline([Color::splat(0); 256])),
+            color_buffer: unsafe { Box::new_zeroed().assume_init() },
             depth_buffer: unsafe { Box::new_zeroed().assume_init() },
             attr_buffer: unsafe { Box::new_zeroed().assume_init() },
             polys: Vec::with_capacity(2048),

diff --git a/render/wgpu-2d/src/common/gfx.rs b/render/wgpu-2d/src/common/gfx.rs
@@ -164,7 +164,7 @@ impl GfxThreadChannels {
 struct FrameData {
     output_3d: Box<[Scanline<u32>; SCREEN_HEIGHT]>,
     framebuffer: Box<[[Scanline<BgObjPixel>; SCREEN_HEIGHT]; 2]>,
-    fb_scanline_flags: [[ScanlineFlags; SCREEN_HEIGHT]; 2],
+    fb_scanline_flags: Box<[[ScanlineFlags; SCREEN_HEIGHT]; 2]>,
     engine_3d_enabled: bool,
     frame_index: u64,
 }
@@ -175,7 +175,7 @@ impl Default for FrameData {
             FrameData {
                 output_3d: Box::new_zeroed().assume_init(),
                 framebuffer: Box::new_zeroed().assume_init(),
-                fb_scanline_flags: [[ScanlineFlags::default(); SCREEN_HEIGHT]; 2],
+                fb_scanline_flags: Box::new_zeroed().assume_init(),
                 engine_3d_enabled: false,
                 frame_index: 0,
             }

diff --git a/render/wgpu-2d/src/threaded/lockstep_scanlines.rs b/render/wgpu-2d/src/threaded/lockstep_scanlines.rs
@@ -546,7 +546,7 @@ struct ThreadData {
 
     fns: (FnPtrs<EngineA>, FnPtrs<EngineB>),
     buffers: [Buffers; 2],
-    fb_scanline_flags: [[ScanlineFlags; SCREEN_HEIGHT]; 2],
+    fb_scanline_flags: Box<[[ScanlineFlags; SCREEN_HEIGHT]; 2]>,
     gfx_data: GfxData,
 }
 
@@ -590,7 +590,7 @@ impl ThreadData {
 
                 fns: (FnPtrs::new(), FnPtrs::new()),
                 buffers: [buffers!(), buffers!()],
-                fb_scanline_flags: [[ScanlineFlags::default(); SCREEN_HEIGHT]; 2],
+                fb_scanline_flags: unsafe { Box::new_zeroed().assume_init() },
                 gfx_data,
             },
             color_output_view,

diff --git a/render/wgpu-3d/src/threaded.rs b/render/wgpu-3d/src/threaded.rs
@@ -14,8 +14,7 @@ use emu_utils::triple_buffer;
 use parking_lot::RwLock;
 use std::{
     cell::UnsafeCell,
-    hint,
-    mem::{self, MaybeUninit},
+    hint, mem,
     sync::{
         atomic::{AtomicBool, AtomicU64, AtomicU8, Ordering},
         Arc,
@@ -43,7 +42,7 @@ struct FrameData {
 pub struct Tx {
     shared_data: Arc<SharedData>,
     frame_tx: triple_buffer::Sender<FrameData>,
-    last_gx_data: GxData,
+    last_gx_data: Box<GxData>,
     texture_dirty: [u8; 3],
     tex_pal_dirty: [u8; 3],
     cur_frame_index: u64,
@@ -249,7 +248,7 @@ pub fn init(
         Tx {
             shared_data: Arc::clone(&shared_data),
             frame_tx,
-            last_gx_data: unsafe { MaybeUninit::zeroed().assume_init() },
+            last_gx_data: unsafe { Box::new_zeroed().assume_init() },
             texture_dirty: [0; 3],
             tex_pal_dirty: [0; 3],
             cur_frame_index: 0,