Mirrors/dioxus
2
0
Fork 0
mirror of https://github.com/DioxusLabs/dioxus synced 2024-11-23 04:33:06 +00:00
Code Issues Projects Releases Packages Wiki Activity

wip: working on async diff

Browse source
This commit is contained in:
Jonathan Kelley 2021-08-17 22:25:09 -04:00
parent 882d69571d
commit f41cff571f
7 changed files with 198 additions and 109 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
examples/ssr.rs
View file

@ -5,7 +5,7 @@ use dioxus::ssr;
fn main() {
let mut vdom = VirtualDom::new(App);
vdom.rebuild_in_place().expect("Rebuilding failed");
// vdom.rebuild_in_place().expect("Rebuilding failed");
println!("{}", ssr::render_vdom(&vdom, |c| c));
}
@ -17,3 +17,10 @@ static App: FC<()> = |cx| {
}
))
};
struct MyProps<'a> {
text: &'a str,
}
fn App2<'a>(cx: Context<'a, MyProps>) -> DomTree<'a> {
None
}

86
packages/core/src/diff.rs
View file

@ -1,9 +1,12 @@
//! This module contains the stateful DiffMachine and all methods to diff VNodes, their properties, and their children.
//! The DiffMachine calculates the diffs between the old and new frames, updates the new nodes, and modifies the real dom.
//! This module contains the stateful PriorityFiber and all methods to diff VNodes, their properties, and their children.
//!
//! The [`PriorityFiber`] calculates the diffs between the old and new frames, updates the new nodes, and generates a set
//! of mutations for the RealDom to apply.
//!
//! ## Notice:
//! The inspiration and code for this module was originally taken from Dodrio (@fitzgen) and then modified to support
//! Components, Fragments, Suspense, SubTree memoization, and additional batching operations.
//! Components, Fragments, Suspense, SubTree memoization, incremental diffing, cancelation, NodeRefs, and additional
//! batching operations.
//!
//! ## Implementation Details:
//!
@ -11,12 +14,13 @@
//! --------------------
//! All nodes are addressed by their IDs. The RealDom provides an imperative interface for making changes to these nodes.
//! We don't necessarily require that DOM changes happen instnatly during the diffing process, so the implementor may choose
//! to batch nodes if it is more performant for their application. The expectation is that renderers use a Slotmap for nodes
//! whose keys can be converted to u64 on FFI boundaries.
//! to batch nodes if it is more performant for their application. The element IDs are indicies into the internal element
//! array. The expectation is that implemenetors will use the ID as an index into a Vec of real nodes, allowing for passive
//! garbage collection as the VirtualDOM replaces old nodes.
//!
//! When new nodes are created through `render`, they won't know which real node they correspond to. During diffing, we
//! always make sure to copy over the ID. If we don't do this properly, the ElementId will be populated incorrectly and
//! brick the user's page.
//! When new vnodes are created through `cx.render`, they won't know which real node they correspond to. During diffing,
//! we always make sure to copy over the ID. If we don't do this properly, the ElementId will be populated incorrectly
//! and brick the user's page.
//!
//! ### Fragment Support
//!
@ -26,6 +30,9 @@
//! impossible to craft a fragment with 0 elements - they must always have at least a single placeholder element. This is
//! slightly inefficient, but represents a such an uncommon use case that it is not worth optimizing.
//!
//! Other implementations either don't support fragments or use a "child + sibling" pattern to represent them. Our code is
//! vastly simpler and more performant when we can just create a placeholder element while the fragment has no children.
//!
//! ## Subtree Memoization
//! -----------------------
//! We also employ "subtree memoization" which saves us from having to check trees which take no dynamic content. We can
@ -35,13 +42,15 @@
//! rsx!( div { class: "hello world", "this node is entirely static" } )
//! ```
//! Because the subtrees won't be diffed, their "real node" data will be stale (invalid), so its up to the reconciler to
//! track nodes created in a scope and clean up all relevant data. Support for this is currently WIP
//! track nodes created in a scope and clean up all relevant data. Support for this is currently WIP and depends on comp-time
//! hashing of the subtree from the rsx! macro. We do a very limited form of static analysis via static string pointers as
//! a way of short-circuiting the most expensive checks.
//!
//! ## Bloom Filter and Heuristics
//! ------------------------------
//! For all components, we employ some basic heuristics to speed up allocations and pre-size bump arenas. The heuristics are
//! currently very rough, but will get better as time goes on. For FFI, we recommend using a bloom filter to cache strings.
//!
//! currently very rough, but will get better as time goes on. The information currently tracked includes the size of a
//! bump arena after first render, the number of hooks, and the number of nodes in the tree.
//!
//! ## Garbage Collection
//! ---------------------
@ -53,11 +62,6 @@
//! so the client only needs to maintain a simple list of nodes. By default, Dioxus will not manually clean up old nodes
//! for the client. As new nodes are created, old nodes will be over-written.
//!
//! HEADS-UP:
//! For now, deferred garabge collection is disabled. The code-paths are almost wired up, but it's quite complex to
//! get working safely and efficiently. For now, garabge is collected immediately during diffing. This adds extra
//! overhead, but is faster to implement in the short term.
//!
//! Further Reading and Thoughts
//! ----------------------------
//! There are more ways of increasing diff performance here that are currently not implemented.
@ -67,13 +71,16 @@
use crate::{arena::SharedResources, innerlude::*};
use futures_util::Future;
use fxhash::{FxBuildHasher, FxHashMap, FxHashSet};
use indexmap::IndexSet;
use smallvec::{smallvec, SmallVec};
use std::{any::Any, cell::Cell, cmp::Ordering, marker::PhantomData, pin::Pin};
use std::{
any::Any, cell::Cell, cmp::Ordering, collections::HashSet, marker::PhantomData, pin::Pin,
};
use DomEdit::*;
pub struct DiffMachine<'r, 'bump> {
pub vdom: &'bump SharedResources,
pub struct DiffMachine<'bump> {
vdom: &'bump SharedResources,
pub mutations: Mutations<'bump>,
@ -81,14 +88,10 @@ pub struct DiffMachine<'r, 'bump> {
pub diffed: FxHashSet<ScopeId>,
// will be used later for garbage collection
// we check every seen node and then schedule its eventual deletion
pub seen_scopes: FxHashSet<ScopeId>,
_r: PhantomData<&'r ()>,
}
impl<'r, 'bump> DiffMachine<'r, 'bump> {
impl<'bump> DiffMachine<'bump> {
pub(crate) fn new(
edits: Mutations<'bump>,
cur_scope: ScopeId,
@ -100,7 +103,6 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
vdom: shared,
diffed: FxHashSet::default(),
seen_scopes: FxHashSet::default(),
_r: PhantomData,
}
}
@ -115,12 +117,17 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
vdom: shared,
diffed: FxHashSet::default(),
seen_scopes: FxHashSet::default(),
_r: PhantomData,
}
}
// make incremental progress on the current task
pub fn work(&mut self, is_ready: impl FnMut() -> bool) -> Result<FiberResult> {
todo!()
// Ok(FiberResult::D)
}
//
pub fn diff_scope(&mut self, id: ScopeId) -> Result<()> {
pub async fn diff_scope(&mut self, id: ScopeId) -> Result<()> {
let component = self.get_scope_mut(&id).ok_or_else(|| Error::NotMounted)?;
let (old, new) = (component.frames.wip_head(), component.frames.fin_head());
self.diff_node(old, new);
@ -133,7 +140,11 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
// the real stack should be what it is coming in and out of this function (ideally empty)
//
// each function call assumes the stack is fresh (empty).
pub fn diff_node(&mut self, old_node: &'bump VNode<'bump>, new_node: &'bump VNode<'bump>) {
pub async fn diff_node(
&mut self,
old_node: &'bump VNode<'bump>,
new_node: &'bump VNode<'bump>,
) {
match (&old_node.kind, &new_node.kind) {
// Handle the "sane" cases first.
// The rsx and html macros strongly discourage dynamic lists not encapsulated by a "Fragment".
@ -264,7 +275,8 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
false => {
// the props are different...
scope.run_scope().unwrap();
self.diff_node(scope.frames.wip_head(), scope.frames.fin_head());
self.diff_node(scope.frames.wip_head(), scope.frames.fin_head())
.await;
}
}
@ -296,7 +308,7 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
// This is the case where options or direct vnodes might be used.
// In this case, it's faster to just skip ahead to their diff
if old.children.len() == 1 && new.children.len() == 1 {
self.diff_node(&old.children[0], &new.children[0]);
self.diff_node(&old.children[0], &new.children[0]).await;
return;
}
@ -1059,7 +1071,11 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
// [... parent]
//
// the change list stack is in the same state when this function returns.
fn diff_non_keyed_children(&mut self, old: &'bump [VNode<'bump>], new: &'bump [VNode<'bump>]) {
async fn diff_non_keyed_children(
&mut self,
old: &'bump [VNode<'bump>],
new: &'bump [VNode<'bump>],
) {
// Handled these cases in `diff_children` before calling this function.
//
debug_assert!(!new.is_empty());
@ -1099,15 +1115,15 @@ impl<'r, 'bump> DiffMachine<'r, 'bump> {
self.edit_pop();
// diff the rest
new.iter()
.zip(old.iter())
.for_each(|(new_child, old_child)| self.diff_node(old_child, new_child));
for (new_child, old_child) in new.iter().zip(old.iter()) {
self.diff_node(old_child, new_child).await
}
}
// old.len == new.len -> no nodes added/removed, but perhaps changed
Ordering::Equal => {
for (new_child, old_child) in new.iter().zip(old.iter()) {
self.diff_node(old_child, new_child);
self.diff_node(old_child, new_child).await;
}
}
}

2
packages/core/src/lib.rs
View file

@ -44,6 +44,7 @@ pub(crate) mod innerlude {
pub use crate::scope::*;
pub use crate::util::*;
pub use crate::virtual_dom::*;
pub use crate::yield_now::*;
pub type DomTree<'a> = Option<VNode<'a>>;
pub type FC<P> = fn(Context<P>) -> DomTree;
@ -73,3 +74,4 @@ pub mod scope;
pub mod signals;
pub mod util;
pub mod virtual_dom;
pub mod yield_now;

120
packages/core/src/scheduler.rs
View file

@ -1,3 +1,26 @@
//! Provides resumable task scheduling for Dioxus.
//!
//!
//! ## Design
//!
//! The recent React fiber architecture rewrite enabled pauseable and resumable diffing through the development of
//! something called a "Fiber." Fibers were created to provide a way of "saving a stack frame", making it possible to
//! resume said stack frame at a later time, or to drop it altogether. This made it possible to
//!
//!
//!
//!
//!
//!
//!
//!
//!
//!
//!
//!
//!
//!
use std::any::Any;
use std::any::TypeId;
@ -11,6 +34,7 @@ use futures_util::Future;
use futures_util::FutureExt;
use futures_util::StreamExt;
use indexmap::IndexSet;
use smallvec::SmallVec;
use crate::innerlude::*;
@ -74,13 +98,15 @@ pub struct Scheduler {
shared: SharedResources,
high_priorty: PriorityFiber<'static>,
medium_priority: PriorityFiber<'static>,
low_priority: PriorityFiber<'static>,
waypoints: VecDeque<Waypoint>,
high_priorty: PriortySystem,
medium_priority: PriortySystem,
low_priority: PriortySystem,
}
pub enum FiberResult<'a> {
Done(Mutations<'a>),
Done(&'a mut Mutations<'a>),
Interrupted,
}
@ -97,10 +123,11 @@ impl Scheduler {
garbage_scopes: HashSet::new(),
current_priority: EventPriority::Low,
waypoints: VecDeque::new(),
high_priorty: PriorityFiber::new(),
medium_priority: PriorityFiber::new(),
low_priority: PriorityFiber::new(),
high_priorty: PriortySystem::new(),
medium_priority: PriortySystem::new(),
low_priority: PriortySystem::new(),
}
}
@ -223,21 +250,26 @@ impl Scheduler {
}
pub fn has_work(&self) -> bool {
let has_work = self.high_priorty.has_work()
|| self.medium_priority.has_work()
|| self.low_priority.has_work();
!has_work
self.waypoints.len() > 0
}
pub fn has_pending_garbage(&self) -> bool {
!self.garbage_scopes.is_empty()
}
fn get_current_fiber<'a>(&'a mut self) -> &mut DiffMachine<'a> {
let fib = match self.current_priority {
EventPriority::High => &mut self.high_priorty,
EventPriority::Medium => &mut self.medium_priority,
EventPriority::Low => &mut self.low_priority,
};
unsafe { std::mem::transmute(fib) }
}
/// If a the fiber finishes its works (IE needs to be committed) the scheduler will drop the dirty scope
pub fn work_with_deadline(
&mut self,
mut deadline: &mut Pin<Box<impl FusedFuture<Output = ()>>>,
is_deadline_reached: &mut impl FnMut() -> bool,
) -> FiberResult {
// check if we need to elevate priority
self.current_priority = match (
@ -250,13 +282,10 @@ impl Scheduler {
(false, false, _) => EventPriority::Low,
};
let mut current_fiber = match self.current_priority {
EventPriority::High => &mut self.high_priorty,
EventPriority::Medium => &mut self.medium_priority,
EventPriority::Low => &mut self.low_priority,
};
let mut is_ready = || -> bool { (&mut deadline).now_or_never().is_some() };
todo!()
// TODO: remove this unwrap - proprogate errors out
self.get_current_fiber().work(is_ready).unwrap()
}
// waits for a trigger, canceling early if the deadline is reached
@ -374,45 +403,46 @@ pub struct DirtyScope {
start_tick: u32,
}
// fibers in dioxus aren't exactly the same as React's. Our fibers are more like a "saved state" of the diffing algorithm.
pub struct PriorityFiber<'a> {
// scopes that haven't been updated yet
pending_scopes: Vec<ScopeId>,
/*
A "waypoint" represents a frozen unit in time for the DiffingMachine to resume from. Whenever the deadline runs out
while diffing, the diffing algorithm generates a Waypoint in order to easily resume from where it left off. Waypoints are
fairly expensive to create, especially for big trees, so it's a good idea to pre-allocate them.
pending_nodes: Vec<*const VNode<'a>>,
Waypoints are created pessimisticly, and are only generated when an "Error" state is bubbled out of the diffing machine.
This saves us from wasting cycles book-keeping waypoints for 99% of edits where the deadline is not reached.
*/
pub struct Waypoint {
// the progenitor of this waypoint
root: ScopeId,
// WIP edits
edits: Vec<DomEdit<'a>>,
edits: Vec<DomEdit<'static>>,
started: bool,
// a saved position in the tree
// these indicies continue to map through the tree into children nodes.
// A sequence of usizes is all that is needed to represent the path to a node.
tree_position: SmallVec<[usize; 10]>,
// a fiber is finished when no more scopes or nodes are pending
completed: bool,
seen_scopes: HashSet<ScopeId>,
dirty_scopes: IndexSet<ScopeId>,
invalidate_scopes: HashSet<ScopeId>,
wip_edits: Vec<DomEdit<'a>>,
current_batch_scopes: HashSet<ScopeId>,
priority_level: EventPriority,
}
impl PriorityFiber<'_> {
fn new() -> Self {
pub struct PriortySystem {
pub pending_scopes: Vec<ScopeId>,
pub dirty_scopes: IndexSet<ScopeId>,
}
impl PriortySystem {
pub fn new() -> Self {
Self {
pending_scopes: Vec::new(),
pending_nodes: Vec::new(),
edits: Vec::new(),
started: false,
completed: false,
dirty_scopes: IndexSet::new(),
wip_edits: Vec::new(),
current_batch_scopes: HashSet::new(),
pending_scopes: Default::default(),
dirty_scopes: Default::default(),
}
}
fn has_work(&self) -> bool {
self.dirty_scopes.is_empty()
&& self.wip_edits.is_empty()
&& self.current_batch_scopes.is_empty()
self.pending_scopes.len() > 0 || self.dirty_scopes.len() > 0
}
}

32
packages/core/src/virtual_dom.rs
View file

@ -178,11 +178,12 @@ impl VirtualDom {
///
/// This method will not wait for any suspended nodes to complete.
pub fn run_immediate<'s>(&'s mut self) -> Result<Mutations<'s>> {
use futures_util::FutureExt;
let mut is_ready = || false;
self.run_with_deadline_and_is_ready(futures_util::future::ready(()), &mut is_ready)
.now_or_never()
.expect("this future will always resolve immediately")
todo!()
// use futures_util::FutureExt;
// let mut is_ready = || false;
// self.run_with_deadline(futures_util::future::ready(()), &mut is_ready)
// .now_or_never()
// .expect("this future will always resolve immediately")
}
/// Runs the virtualdom with no time limit.
@ -240,25 +241,6 @@ impl VirtualDom {
pub async fn run_with_deadline<'s>(
&'s mut self,
deadline: impl Future<Output = ()>,
) -> Result<Mutations<'s>> {
use futures_util::FutureExt;
let deadline_future = deadline.shared();
let mut is_ready_deadline = deadline_future.clone();
let mut is_ready = || -> bool { (&mut is_ready_deadline).now_or_never().is_some() };
self.run_with_deadline_and_is_ready(deadline_future, &mut is_ready)
.await
}
/// Runs the virtualdom with a deadline and a custom "check" function.
///
/// Designed this way so "run_immediate" can re-use all the same rendering logic as "run_with_deadline" but the work
/// queue is completely drained;
async fn run_with_deadline_and_is_ready<'s>(
&'s mut self,
deadline: impl Future<Output = ()>,
is_ready: &mut impl FnMut() -> bool,
) -> Result<Mutations<'s>> {
let mut committed_mutations = Mutations::new();
let mut deadline = Box::pin(deadline.fuse());
@ -286,7 +268,7 @@ impl VirtualDom {
// Work through the current subtree, and commit the results when it finishes
// When the deadline expires, give back the work
match self.scheduler.work_with_deadline(&mut deadline, is_ready) {
match self.scheduler.work_with_deadline(&mut deadline) {
FiberResult::Done(mut mutations) => {
committed_mutations.extend(&mut mutations);

52
packages/core/src/yield_now.rs Normal file
View file

@ -0,0 +1,52 @@
use std::future::Future;
use std::pin::Pin;
use std::task::{Context, Poll};
// use crate::task::{Context, Poll};
/// Cooperatively gives up a timeslice to the task scheduler.
///
/// Calling this function will move the currently executing future to the back
/// of the execution queue, making room for other futures to execute. This is
/// especially useful after running CPU-intensive operations inside a future.
///
/// See also [`task::spawn_blocking`].
///
/// [`task::spawn_blocking`]: fn.spawn_blocking.html
///
/// # Examples
///
/// Basic usage:
///
/// ```
/// # async_std::task::block_on(async {
/// #
/// use async_std::task;
///
/// task::yield_now().await;
/// #
/// # })
/// ```
#[inline]
pub async fn yield_now() {
YieldNow(false).await
}
struct YieldNow(bool);
impl Future for YieldNow {
type Output = ();
// The futures executor is implemented as a FIFO queue, so all this future
// does is re-schedule the future back to the end of the queue, giving room
// for other futures to progress.
fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
if !self.0 {
self.0 = true;
cx.waker().wake_by_ref();
Poll::Pending
} else {
Poll::Ready(())
}
}
}

6
packages/web/src/lib.rs
View file

@ -108,9 +108,9 @@ pub async fn run_with_props<T: Properties + 'static>(
let tasks = dom.get_event_sender();
// initialize the virtualdom first
if cfg.hydrate {
dom.rebuild_in_place()?;
}
// if cfg.hydrate {
// dom.rebuild_in_place()?;
// }
let mut websys_dom = dom::WebsysDom::new(
root_el,