Skip to content

Latest commit

 

History

History
99 lines (72 loc) · 6.03 KB

architecture.md

File metadata and controls

99 lines (72 loc) · 6.03 KB

Architecture

Overview

The LazySignals API provides convenience functions to create an ECS entity for each semantic structure that contains the proper components (bundles). The overall design is inspired by similar structures in functional programming, such as the infamous Haskell monad. The specifics of the developer API are inspired by the TC 39 Signals proposal.

It is lossy by default, which means if multiple signals are sent in the same tick, whichever signal is applied most recently contains the value that will be propagated to the Computeds, Effects, and Actions. The prior signal values within the tick are overwritten before they are sent on as each command calls the merge_next function.

The mappings of valid combinations of components to bundles is enumerated in the next sections.

Primitives

A propagator aggregates data from its dependencies. For a Computed, the data is merged into its LazySignalsState. An Effect may uses its source values to perform some side effects. An Action is an Effect but instead of having exclusive world access, it returns a CommandQueue to be evaluated by a LazySignals system when the Action completes. It uses Bevy async tasks (bevy_tasks) and is the default method to spawn long running operations.

The commands returned by an Action as well as the closures passed to Effects may themselves send new signals, to be evaluated during the standard LazySignals update cycle.

A LazySignalsState component holds the value, bound by the traits defined by LazySignalsData. An ImmutableState component stores the type information required for reflection and is populated with the return value of the init_component call when the LazyImmutableState is created.

To send a signal, merge the next_value and add a SendSignal component.

To form a Computed, add a ComputedImmutable component to the Signal entity.

A LazyEffect component identifies an Effect. A LazyEffect can contain an Action instead, which does not have exclusive world access, but returns a CommandQueue to be applied by the LazySignals update system. An Effect can also return a closure that defines a system function.

Exclusive Systems

Check Tasks

The task checking system checks the status of each Action marked with RunningTask. If a task is completed, RunningTask is removed and any commands in the returned CommandQueue are applied.

Init System

The init system runs every tick. Newly added Computed, Effect, and Action components will have an InitDependencies component to mark them. These systems just run subscribe for each of the sources and triggers so that the relevant LazySignals components are notified at the proper time.

Signal Processing

During processing, a (should be brief) write lock for the world is obtained. If the value of a signal is unchanged, the SendSignal for each sent signal is simply discarded. The next_value is also set with a special default error to disambiguate it from None as an intentional data value. Otherwise, each Signal's data field is replaced with next_value. The Signal is marked with ValueChanged to avoid having to constantly override Bevy's internal change detection during the exclusive systems. Subscribers are added to a "running" set and removed from the LazySignalsState's subscribers, which is cleared. This prevents to need to manage subscriptions manually when LazySignals entities are despawned. Finally, the SendSignal component is removed.

The initial "running" set is iterated. If the item is a Computed, then add a ComputeMemo component to mark it for update. If it is an Effect or Action, add a DeferredEffect component to mark it for scheduling. Effects may be triggered, which possibly means sending a signal with a None value, or triggering upstream effects and tasks with a unit or typed but possibly unchanged value (e g. to represent a button press).

Walk the subscriber tree, adding each item's subscribers to the "next_running" set and removing them from its own subscribers. As each item processes, add it to a "processed" set and do not add any item to the "next_running" set if it exists in the "processed" set. When the current "running" set is exhausted, run the next one. The system exits when each item in the final running set finishes and the next running set is empty.

Memo Processing

The closure in the Computed component of every entity marked with a ComputeMemo component runs and the result is stored in the LazyImmutableState. As each value is read, the Computed is added to the next_subscribers of the source entity. If the value is itself a Computed, it will recompute if it's marked with Dirty. Otherwise it simply returns the value. If the value is different, ValueChanged will be added after the closure is evaluated, which will be used to limit which effects are scheduled next. The Dirty component is removed whether the value changed or not.

A stack is kept of all running operations. If any source is dirty, the Computed will put itself and its dirty sources on the stack (TODO verify sources are filtered by Dirty). This avoids the use of direct recursion and stack overflows. The system exits when each item in the stack finishes.

Effect Processing

The effects system examimes the dependencies of each entity with a DeferredEffect component. If any dependency of an Effect is changed, the Effect closure is called after placing the Effect into the "running" set. Effects will also run if the entity has a Triggered component.

The first 4 systems can be run as needed in between systems that need to have signals processed between them. It is recommended to only run the effects once per tick to avoid running the same effects if triggered more than once. Alternatively, care must be taken to make sure effects can be triggered safely repeatedly or else that the situation is avoided.

The system exits when each item in the "running" set finishes. Actions are processed like Effects, but their closures do not receive a &mut World and instead must return a CommandQueue. Also an Action spawns an async Bevy task instead of evaluating its closure during the system.