perf: remove an unnecessary buffer copy from WASM asset loading pipeline

crates/bevy_asset/Cargo.toml

@@ -39,7 +39,7 @@ bevy_platform = { path = "../bevy_platform", version = "0.18.0-dev", default-fea
   "std",
 ] }
 
-stackfuture = { version = "0.3", default-features = false }
+stackfuture = { git = "https://github.com/joseph-gio/stackfuture", branch = "local-stack-future", default-features = false }
 atomicow = { version = "1.1", default-features = false, features = ["std"] }
 async-broadcast = { version = "0.7.2", default-features = false }
 async-fs = { version = "2.0", default-features = false }

crates/bevy_asset/src/io/file/sync_file_asset.rs

@@ -3,7 +3,7 @@ use futures_lite::Stream;
 
 use crate::io::{
     get_meta_path, AssetReader, AssetReaderError, AssetWriter, AssetWriterError, AsyncSeek,
-    PathStream, Reader, ReaderRequiredFeatures, Writer,
+    ConditionalSendStackFuture, PathStream, Reader, ReaderRequiredFeatures, Writer,
 };
 
 use alloc::{borrow::ToOwned, boxed::Box, vec::Vec};
@@ -44,9 +44,9 @@ impl Reader for FileReader {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> stackfuture::StackFuture<'a, std::io::Result<usize>, { crate::io::STACK_FUTURE_SIZE }>
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, { crate::io::STACK_FUTURE_SIZE }>
     {
-        stackfuture::StackFuture::from(async { self.0.read_to_end(buf) })
+        ConditionalSendStackFuture::from(async { self.0.read_to_end(buf) })
     }
 }
 

crates/bevy_asset/src/io/memory.rs

@@ -1,6 +1,6 @@
 use crate::io::{
-    AssetReader, AssetReaderError, AssetWriter, AssetWriterError, PathStream, Reader,
-    ReaderRequiredFeatures,
+    AssetReader, AssetReaderError, AssetWriter, AssetWriterError, ConditionalSendStackFuture,
+    PathStream, Reader, ReaderRequiredFeatures,
 };
 use alloc::{borrow::ToOwned, boxed::Box, sync::Arc, vec, vec::Vec};
 use bevy_platform::{
@@ -351,7 +351,7 @@ impl Reader for DataReader {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> stackfuture::StackFuture<'a, std::io::Result<usize>, { super::STACK_FUTURE_SIZE }> {
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, { super::STACK_FUTURE_SIZE }> {
         crate::io::read_to_end(self.data.value(), &mut self.bytes_read, buf)
     }
 }

crates/bevy_asset/src/io/mod.rs

@@ -22,22 +22,19 @@ pub mod gated;
 
 mod source;
 
+pub use futures_io::{AsyncRead, AsyncSeek, AsyncWrite, SeekFrom};
 pub use futures_lite::AsyncWriteExt;
 pub use source::*;
 
 use alloc::{boxed::Box, sync::Arc, vec::Vec};
-use bevy_tasks::{BoxedFuture, ConditionalSendFuture};
+use bevy_tasks::{BoxedFuture, ConditionalSend, ConditionalSendFuture};
 use core::{
     mem::size_of,
     pin::Pin,
     task::{Context, Poll},
 };
-use futures_io::{AsyncRead, AsyncSeek, AsyncWrite};
 use futures_lite::Stream;
-use std::{
-    io::SeekFrom,
-    path::{Path, PathBuf},
-};
+use std::path::{Path, PathBuf};
 use thiserror::Error;
 
 /// Errors that occur while loading assets.
@@ -131,7 +128,14 @@ pub enum SeekKind {
 // a higher maximum necessary.
 pub const STACK_FUTURE_SIZE: usize = 10 * size_of::<&()>();
 
-pub use stackfuture::StackFuture;
+pub use stackfuture::{LocalStackFuture, StackFuture};
+
+#[cfg(target_arch = "wasm32")]
+pub type ConditionalSendStackFuture<'a, T, const STACK_SIZE: usize> =
+    LocalStackFuture<'a, T, STACK_SIZE>;
+#[cfg(not(target_arch = "wasm32"))]
+pub type ConditionalSendStackFuture<'a, T, const STACK_SIZE: usize> =
+    StackFuture<'a, T, STACK_SIZE>;
 
 /// A type returned from [`AssetReader::read`], which is used to read the contents of a file
 /// (or virtual file) corresponding to an asset.
@@ -154,7 +158,7 @@ pub use stackfuture::StackFuture;
 /// [`SeekKind::AnySeek`] to indicate that they may seek backward, or from the start/end. A reader
 /// implementation may choose to support that, or may just detect those kinds of seeks and return an
 /// error.
-pub trait Reader: AsyncRead + AsyncSeek + Unpin + Send + Sync {
+pub trait Reader: AsyncRead + AsyncSeek + Unpin + ConditionalSend {
     /// Reads the entire contents of this reader and appends them to a vec.
     ///
     /// # Note for implementors
@@ -164,17 +168,17 @@ pub trait Reader: AsyncRead + AsyncSeek + Unpin + Send + Sync {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> StackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
         let future = futures_lite::AsyncReadExt::read_to_end(self, buf);
-        StackFuture::from(future)
+        ConditionalSendStackFuture::from(future)
     }
 }
 
 impl Reader for Box<dyn Reader + '_> {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> StackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
         (**self).read_to_end(buf)
     }
 }
@@ -682,7 +686,7 @@ impl Reader for VecReader {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> StackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
         read_to_end(&self.bytes, &mut self.bytes_read, buf)
     }
 }
@@ -727,7 +731,7 @@ impl Reader for SliceReader<'_> {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> StackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
         read_to_end(self.bytes, &mut self.bytes_read, buf)
     }
 }
@@ -781,8 +785,8 @@ pub(crate) fn read_to_end<'a>(
     source: &'a [u8],
     bytes_read: &'a mut usize,
     dest: &'a mut Vec<u8>,
-) -> StackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
-    StackFuture::from(async {
+) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
+    ConditionalSendStackFuture::from(async {
         if *bytes_read >= source.len() {
             Ok(0)
         } else {

crates/bevy_asset/src/io/processor_gated.rs

@@ -1,6 +1,7 @@
 use crate::{
     io::{
-        AssetReader, AssetReaderError, AssetSourceId, PathStream, Reader, ReaderRequiredFeatures,
+        AssetReader, AssetReaderError, AssetSourceId, ConditionalSendStackFuture, PathStream,
+        Reader, ReaderRequiredFeatures,
     },
     processor::{ProcessStatus, ProcessingState},
     AssetPath,
@@ -155,7 +156,7 @@ impl Reader for TransactionLockedReader<'_> {
     fn read_to_end<'a>(
         &'a mut self,
         buf: &'a mut Vec<u8>,
-    ) -> stackfuture::StackFuture<'a, std::io::Result<usize>, { super::STACK_FUTURE_SIZE }> {
+    ) -> ConditionalSendStackFuture<'a, std::io::Result<usize>, { super::STACK_FUTURE_SIZE }> {
         self.reader.read_to_end(buf)
     }
 }

crates/bevy_asset/src/io/wasm.rs

@@ -1,12 +1,14 @@
 use crate::io::{
-    get_meta_path, AssetReader, AssetReaderError, EmptyPathStream, PathStream, Reader,
-    ReaderRequiredFeatures, VecReader,
+    get_meta_path, AssetReader, AssetReaderError, AsyncRead, AsyncSeek, LocalStackFuture,
+    EmptyPathStream, PathStream, Reader, ReaderRequiredFeatures, SeekFrom, STACK_FUTURE_SIZE,
 };
-use alloc::{borrow::ToOwned, boxed::Box, format};
+use alloc::{borrow::ToOwned, boxed::Box, format, vec::Vec};
+use core::pin::Pin;
+use core::task::{Poll, Context};
 use js_sys::{Uint8Array, JSON};
 use std::path::{Path, PathBuf};
 use tracing::error;
-use wasm_bindgen::{prelude::wasm_bindgen, JsCast, JsValue};
+use wasm_bindgen::{prelude::wasm_bindgen, JsCast, JsValue, UnwrapThrowExt};
 use wasm_bindgen_futures::JsFuture;
 use web_sys::Response;
 
@@ -79,8 +81,8 @@ impl HttpWasmAssetReader {
         match resp.status() {
             200 => {
                 let data = JsFuture::from(resp.array_buffer().unwrap()).await.unwrap();
-                let bytes = Uint8Array::new(&data).to_vec();
-                let reader = VecReader::new(bytes);
+                let bytes = Uint8Array::new(&data);
+                let reader = Uint8ArrayReader::new(bytes);
                 Ok(reader)
             }
             // Some web servers, including itch.io's CDN, return 403 when a requested file isn't present.
@@ -121,3 +123,115 @@ impl AssetReader for HttpWasmAssetReader {
         Ok(false)
     }
 }
+
+/// An [`AsyncRead`] implementation capable of reading a [`Uint8Array`].
+pub struct Uint8ArrayReader {
+    array: Uint8Array,
+    initial_offset: u32,
+}
+
+impl Uint8ArrayReader {
+    /// Create a new [`Uint8ArrayReader`] for `bytes`.
+    pub fn new(array: Uint8Array) -> Self {
+        Self {
+            initial_offset: array.byte_offset(),
+            array,
+        }
+    }
+}
+
+impl AsyncRead for Uint8ArrayReader {
+    fn poll_read(
+        mut self: Pin<&mut Self>,
+        _cx: &mut Context,
+        buf: &mut [u8],
+    ) -> Poll<futures_io::Result<usize>> {
+        let array_len = self.array.length();
+        let n = u32::min(buf.len() as u32, array_len);
+        self.array.subarray(0, n).copy_to(&mut buf[..n as usize]); // NOTE: copy_to will panic if the lengths do not exactly match
+        self.array = self.array.subarray(n, array_len);
+        Poll::Ready(Ok(n as usize))
+    }
+}
+
+impl AsyncSeek for Uint8ArrayReader {
+    fn poll_seek(
+        mut self: Pin<&mut Self>,
+        _cx: &mut Context,
+        seek_from: SeekFrom,
+    ) -> Poll<std::io::Result<u64>> {
+        let array_len = self.array.length();
+        let current_array_buffer_offset = self.array.byte_offset();
+        let array_buffer_end = current_array_buffer_offset + array_len;
+        let new_array_buffer_offset = match seek_from {
+            SeekFrom::Start(from_start) => self
+                .initial_offset
+                .saturating_add(u32::try_from(from_start).unwrap_or(u32::MAX))
+                .min(array_buffer_end),
+            SeekFrom::End(from_end) => {
+                if from_end.is_negative() {
+                    array_buffer_end
+                        .saturating_sub(u32::try_from(from_end.abs()).unwrap_or(u32::MAX))
+                        .max(self.initial_offset)
+                } else {
+                    array_buffer_end
+                }
+            }
+            SeekFrom::Current(from_current) => {
+                if from_current.is_negative() {
+                    current_array_buffer_offset
+                        .saturating_sub(u32::try_from(from_current.abs()).unwrap_or(u32::MAX))
+                        .max(self.initial_offset)
+                } else {
+                    current_array_buffer_offset
+                        .saturating_add(u32::try_from(from_current).unwrap_or(u32::MAX))
+                        .min(array_buffer_end)
+                }
+            }
+        };
+        debug_assert!(new_array_buffer_offset >= self.initial_offset);
+        debug_assert!(new_array_buffer_offset <= array_buffer_end);
+        self.array = self
+            .array
+            .constructor()
+            .call3(
+                &JsValue::UNDEFINED,
+                &self.array.buffer(),
+                &new_array_buffer_offset.into(),
+                &array_buffer_end.into(),
+            )
+            .unwrap_throw()
+            .unchecked_into();
+        Poll::Ready(Ok((new_array_buffer_offset - self.initial_offset).into()))
+    }
+}
+
+impl Reader for Uint8ArrayReader {
+    fn read_to_end<'a>(
+        &'a mut self,
+        buf: &'a mut Vec<u8>,
+    ) -> LocalStackFuture<'a, std::io::Result<usize>, STACK_FUTURE_SIZE> {
+        #[expect(unsafe_code)]
+        LocalStackFuture::from(async {
+            let n = self.array.length();
+            let n_usize = n as usize;
+
+            buf.reserve_exact(n_usize);
+            let spare_capacity = buf.spare_capacity_mut();
+            debug_assert!(spare_capacity.len() >= n_usize);
+            // NOTE: `copy_to_uninit` requires the lengths to match exactly,
+            // and `reserve_exact` may reserve more capacity than required.
+            self.array.copy_to_uninit(&mut spare_capacity[..n_usize]);
+            // SAFETY:
+            // * the vector has enough spare capacity for `n` additional bytes due to `reserve_exact` above
+            // * the bytes have been initialized due to `copy_to_uninit` above.
+            unsafe {
+                let new_len = buf.len() + n_usize;
+                buf.set_len(new_len);
+            }
+            self.array = self.array.subarray(n, n);
+
+            Ok(n_usize)
+        })
+    }
+}

examples/README.md

@@ -828,6 +828,12 @@ python3 -m http.server --directory examples/wasm
 ruby -run -ehttpd examples/wasm
 ```
 
+On Unix platforms, you can utilize the `run-example.sh` script in the wasm directory.
+For example, to run the "many-foxes" example, you can run the following command from the root of the workspace:
+```sh
+./examples/wasm/run-example.sh many_foxes
+```
+
 ##### WebGL2 and WebGPU
 
 Bevy support for WebGPU is being worked on, but is currently experimental.

examples/wasm/run-example.sh

@@ -0,0 +1,19 @@
+#!/bin/sh
+set -euxo pipefail
+
+example_name="$1"
+
+wasm_example_dir=$(dirname -- "$(readlink -f -- "${BASH_SOURCE[0]}")")
+workspace_root=$(realpath "${wasm_example_dir}/../..")
+
+cd "${workspace_root}"
+
+example_output_dir="$wasm_example_dir/target/$example_name"
+rm -rf "$example_output_dir"
+mkdir -p "$example_output_dir"
+rsync -a --exclude='target' --exclude='run-example.sh' "$wasm_example_dir/" "$example_output_dir/"
+
+RUSTFLAGS='--cfg=web_sys_unstable_apis --cfg=getrandom_backend="wasm_js"' cargo build --release --example "$example_name" --target wasm32-unknown-unknown
+wasm-bindgen --out-name wasm_example --out-dir "$example_output_dir/target" --target web "target/wasm32-unknown-unknown/release/examples/$example_name.wasm"
+
+basic-http-server "$example_output_dir"