Tracker: optimisations

[dhardy]

A tracker for notable optimisations (generally low priority so long as the UI appears sufficiently fast).

• Widget identifiers: Widget Identifiers #91
• Fast view filters for large ListData (thousands of items): see View widgets: trait revision, reliable keys #164
• Delaying ListView::update_widgets until event queue is empty: see View widgets: trait revision, reliable keys #164
• Avoid unnecessary redraws when hover target changes (see below)

[dhardy]

Avoid unnecessary redraws when hover target changes

This is done

The Manager::set_hover method (src/event/manager.rs:238) requests a redraw whenever the mouse moves over a different widget, since the "hover status" of a widget may affect how it is drawn (e.g. to highlight a button). Many widgets, however, are not affected by this, thus many of these redraws are unnecessary; unfortunately the Manager doesn't know which widgets have hover highlighting and currently there is no way to communicate this.

Possible ideas:

• Send messages to widgets when mouse-hover status is set/lost, which can request redraws if necessary — this is significant extra event processing but still less work than redrawing
• Extra method in WidgetConfig
• Work towards partial redraws?

[dhardy]

Reduce size of data types

• PressSource is surprisingly large: the value of Touch is 64-bits since on OSX it is a pointer (on all other platforms it is a 32-bit integer); adding the tag makes this a 16-byte type. If the tag could be moved elsewhere (into Event or within the value of Touch) then this and Event would be smaller. Hint: pointers are aligned, thus the smallest 2-3 bits must be zero; unfortunately this requires unsafe platform-specific code.

Current sizes

#[cfg(test)]
#[test]
fn type_sizes() {
    use std::mem::size_of;

    assert_eq!(size_of::<Event>(), 48);
    assert_eq!(size_of::<PressSource>(), 16);
}

[dhardy]

#335 removed usage of the stack_dst crate which was used to avoid some heap allocations. No performance issues were noted as a result and this removed an unsafe and unstable dependency. Eventually Rust may directly support returning unsized trait-impl objects on the stack, but until then it doesn't appear to be a real issue.