crcbl_shell/

caps.rs

//! [`ShellCaps`] — what a backend can actually do, asked as a question about
//! capability rather than about identity.
//!
//! # The regression this exists to prevent
//!
//! ```ignore
//! if cfg!(target_os = "linux") && std::env::var("WAYLAND_DISPLAY").is_ok() {
//!     // …use the viewport path
//! }
//! ```
//!
//! That code is wrong three ways and all three are invisible until someone
//! files a bug: it is wrong on a Wayland compositor that does not implement
//! `wp_viewporter`, wrong under XWayland, and wrong the day the shell grows a
//! backend nobody thought about. The windowing rules in `docs/notes/backends.md`
//! are explicit — "the
//! renderer picks blit-vs-viewport and windowed-aspect behavior from caps, never
//! from *am I on Wayland*" — and this module is the mechanism. It mirrors
//! `crcbl-hal`'s `Features`, where the renderer asks for `DRAW_INDIRECT_COUNT`
//! rather than for Vulkan.
//!
//! # Caps are per-shell, not per-window
//!
//! A capability here is a property of the connection and the protocols the
//! server advertised, so it is fixed for the shell's lifetime and reported by
//! [`Shell::caps`](crate::Shell::caps) rather than per window. The one thing
//! that genuinely varies per window — its current size and scale — is
//! [`WindowState`](crate::WindowState), not a capability.
//!
//! There is a real edge here, and it is stated rather than hidden: a Wayland
//! compositor can bind a global *after* startup, so a protocol could in
//! principle appear mid-session. Backends must not flip caps bits when that
//! happens; a renderer that picked the blit path at init cannot be told
//! halfway through a frame that it should have picked the viewport path.
//! Capabilities are latched at connection time, and a missing protocol stays
//! missing until the next run.

bitflags::bitflags! {
    /// What a shell backend supports.
    ///
    /// Each flag names a *behaviour a consumer would otherwise have to guess
    /// from the platform*. A flag that no consumer would branch on does not
    /// belong here.
    #[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Hash)]
    pub struct ShellCaps: u32 {
        /// The compositor can scale a smaller buffer to the window's size in
        /// display hardware, so borderless render-scale costs no shader work.
        ///
        /// Wayland's `wp_viewporter`; the browser's CSS-sized canvas. Absent on
        /// X11, where the renderer must do an upscale blit instead. This is the
        /// exact decision the windowing rules call out as the reason caps
        /// exist. Nothing asks for it: the engine does its own upscale on every
        /// platform (`docs/backlog.md`, _The engine owns scaling_).
        const HW_UPSCALE = 1 << 0;

        /// The window system will enforce an aspect ratio during an interactive
        /// resize.
        ///
        /// X11 `WM_NORMAL_HINTS`, Win32 `WM_SIZING`, macOS
        /// `setContentAspectRatio`. **Wayland has no aspect hint**, so a
        /// Wayland backend clears this bit and the renderer letterboxes.
        /// Letterboxing must work regardless — a tiling window manager ignores
        /// the hint on every backend that has one — so this flag chooses
        /// between "good" and "correct", never between "works" and "broken".
        const ASPECT_HINT_HONORED = 1 << 1;

        /// [`Shell::warp_pointer`](crate::Shell::warp_pointer) moves the
        /// pointer.
        ///
        /// X11 `XWarpPointer`, Win32 `SetCursorPos`. Wayland deliberately
        /// forbids it (a client that can move the cursor can spoof clicks); the
        /// engine's answer there is pointer lock plus relative motion, which is
        /// what a camera wanted anyway.
        const POINTER_WARP = 1 << 2;

        /// [`PointerMode::Locked`](crate::PointerMode) is available: the
        /// pointer can be hidden and pinned while relative motion keeps
        /// flowing.
        ///
        /// Wayland `pointer-constraints` + `relative-pointer`, X11 pointer
        /// grab, Win32 clip-and-recentre, the browser's Pointer Lock API.
        ///
        /// The browser is the one that cannot take the lock when asked: it
        /// grants `requestPointerLock` only inside a user gesture, so
        /// [`set_pointer_mode`](crate::Shell::set_pointer_mode) there records
        /// the request and the player's next click completes it. The bit still
        /// means what it says — the mode is available — and a consumer that
        /// needs to know whether the pointer is pinned *now* reads the shape of
        /// the motion events, as
        /// [`PointerMode::Locked`](crate::PointerMode::Locked) describes.
        const POINTER_LOCK = 1 << 3;

        /// [`PointerMode::Confined`](crate::PointerMode) is available: the
        /// pointer stays inside the window but keeps moving visibly.
        ///
        /// A separate bit from [`POINTER_LOCK`](Self::POINTER_LOCK) because
        /// Wayland's `pointer-constraints` advertises lock and confine
        /// independently, and a strategy game wants confine specifically.
        const POINTER_CONFINE = 1 << 4;

        /// Relative pointer motion is delivered separately from absolute
        /// position, unaccelerated and unclamped by the screen edge.
        ///
        /// This is what a first-person camera must read;
        /// the windowing rules make raw motion a P0 feature, not an
        /// afterthought, on the grounds that an esports-grade engine cannot
        /// derive aim from clamped absolute positions. Without this bit,
        /// [`ShellEvent::PointerMotion`](crate::ShellEvent::PointerMotion)
        /// carries `raw_delta: None` and the camera has to difference absolute
        /// positions — which is wrong at the screen edge and wrong under
        /// pointer acceleration.
        ///
        /// # The one platform that can decline the bypass
        ///
        /// On the desktop backends the unaccelerated stream is a separate
        /// device interface and either exists or does not. In a browser it is
        /// an *option* on the lock — `requestPointerLock({ unadjustedMovement:
        /// true })` — and a browser that does not implement it hands back the
        /// OS-adjusted deltas instead, either by ignoring the key or by
        /// rejecting with `NotSupportedError` and being asked again without it.
        /// The web backend sets this bit because it reads a relative stream and
        /// asks for it unadjusted, which is the decision a camera branches on:
        /// differencing absolute positions is wrong there whatever the browser
        /// answered. It is not a promise that every browser granted the bypass;
        /// the `web` backend's module docs name which browsers do.
        const RAW_POINTER_MOTION = 1 << 5;

        /// The backend is wired to the platform's input method, the framework
        /// CJK and other complex-script input runs through, end to end.
        ///
        /// For a window that named a field with
        /// [`Shell::set_text_input`](crate::Shell::set_text_input), the
        /// composition in progress arrives as
        /// [`ShellEvent::TextPreedit`](crate::ShellEvent::TextPreedit), the
        /// conversion the user picks as
        /// [`ShellEvent::TextCommit`](crate::ShellEvent::TextCommit), and
        /// [`Shell::set_text_input_area`](crate::Shell::set_text_input_area)
        /// is passed on so the input method can place its candidate window.
        /// Whether the user runs an input method is not the shell's to know;
        /// the bit says one would reach the engine.
        ///
        /// Wayland sets it when the compositor offers `text-input-v3`, and
        /// Win32 through `imm32`. Clear on X11, which has no XIM, IBus or
        /// Fcitx client; on AppKit, which does not send the pre-edit yet; and
        /// on the web. Dead keys and Compose sequences are
        /// [`TEXT_COMPOSE`](Self::TEXT_COMPOSE), which a backend can have
        /// without this.
        const TEXT_IME = 1 << 6;

        /// Clipboard get/set works. See
        /// [`clipboard`](crate::clipboard).
        ///
        /// Explicitly a capability because the browser gates clipboard reads
        /// behind a permission prompt and a user gesture, which
        /// the windowing rules call out by name: the editor UI has to
        /// degrade gracefully in-browser rather than offering a paste button
        /// that cannot work.
        const CLIPBOARD = 1 << 7;

        /// Files dragged onto a window arrive as
        /// [`ShellEvent::DroppedFile`](crate::ShellEvent::DroppedFile).
        const DRAG_DROP = 1 << 8;

        /// The shell can create more than one window.
        ///
        /// The MVP editor is single-window, and a canvas backend has exactly
        /// one canvas, so this is not universal. A consumer that wants a
        /// tear-off panel checks this rather than assuming.
        const MULTI_WINDOW = 1 << 9;

        /// Window position in desktop coordinates can be read and set.
        ///
        /// **Wayland cannot do either** — a client is not told where it is and
        /// cannot ask to be moved. Anything that wants to restore a window's
        /// position across sessions (the editor, topic 14's settings store)
        /// must check this bit and degrade to "the compositor decides".
        const WINDOW_POSITION = 1 << 10;

        /// The window system draws the title bar and borders.
        ///
        /// Wayland with `xdg-decoration` and a compositor that opts in, and
        /// every other desktop platform. Without it the UI layer has to draw
        /// its own decorations, which is a layout decision, not a rendering
        /// detail — so it belongs in caps rather than being discovered by
        /// noticing the window looks wrong.
        const SERVER_DECORATIONS = 1 << 11;

        /// Scale factors other than whole integers are reported.
        ///
        /// Wayland `fractional-scale-v1` (in 1/120ths), Win32 per-monitor-v2,
        /// browser `devicePixelRatio`. Without it the shell reports integer
        /// scales only, and a 125% desktop shows up as 1.0 or 2.0 — which the
        /// UI needs to know before it decides whether to snap its layout to a
        /// pixel grid.
        const FRACTIONAL_SCALE = 1 << 12;

        /// [`Shell::wait_events`](crate::Shell::wait_events) genuinely blocks.
        ///
        /// An editor idling at zero frames per second on battery depends on
        /// it; a game never calls it. Backends that cannot block — the browser
        /// above all, where blocking the main thread is not a thing that
        /// happens — leave this clear and their `wait_events` returns
        /// immediately, which is always a correct implementation.
        const EVENT_WAIT = 1 << 13;

        /// Fingers arrive as [`ShellEvent::Touch`](crate::ShellEvent::Touch),
        /// with a contact id and a phase, so more than one can be tracked at a
        /// time.
        ///
        /// **Win32 is the one desktop backend that sets it** (`WM_POINTER*`),
        /// beside the browser. Clear on the others, and that is a statement
        /// about the backends rather than about the platforms: a Linux tablet
        /// has a touchscreen, and X11 (XInput2 `XI_TouchBegin`), Wayland
        /// (`wl_touch`) and AppKit (`NSTouch`) all have a way to deliver one.
        /// None of those paths is written, so a game on them sees no contacts
        /// at all — an input that is *absent*, which is a different thing from
        /// one that silently fails.
        ///
        /// A backend that does set it owes the emulated pointer stream for the
        /// primary contact as well; [`ShellEvent::Touch`](crate::ShellEvent::Touch)
        /// says why, and that obligation is what stops a backend turning this on
        /// from breaking every game that binds only the pointer.
        const TOUCH = 1 << 14;

        /// [`WindowRole::Panel`](crate::WindowRole::Panel) can be created:
        /// desktop furniture held against a monitor's edges.
        const PANEL = 1 << 15;

        /// A panel's [`Reserve::OwnExtent`](crate::Reserve::OwnExtent) is
        /// honoured: maximised windows stop at the panel.
        ///
        /// Separate from [`PANEL`](Self::PANEL) because a window system can
        /// have panels and still give nothing but its own furniture reserved
        /// space — macOS, where the panel is created and overlaps what is
        /// behind it.
        const PANEL_RESERVE = 1 << 16;

        /// [`WindowRole::Popup`](crate::WindowRole::Popup) can be created:
        /// a window anchored to another's rectangle.
        const POPUP = 1 << 17;

        /// [`WindowDesc::transparent`](crate::WindowDesc::transparent) is
        /// honoured: where the frame's alpha is below one, the desktop shows
        /// through.
        ///
        /// The window system's half. The frame's half is a swapchain that
        /// carries alpha, premultiplied — which the HAL's surface caps say
        /// whether the driver offers, and which `crcbl::engine::GpuContext`
        /// asks for when the window's surface target says it is transparent
        /// — and a renderer that clears to alpha zero. See
        /// [`WindowDesc::transparent`](crate::WindowDesc::transparent) for the
        /// whole recipe.
        const TRANSPARENT = 1 << 18;

        /// [`Shell::reposition_popup`](crate::Shell::reposition_popup) moves
        /// and resizes a live popup.
        ///
        /// Separate from [`POPUP`](Self::POPUP) because Wayland's
        /// `xdg_popup.reposition` arrived in `xdg_wm_base` 3, and a compositor
        /// offering only 1 or 2 has popups that stay where they opened.
        const POPUP_REPOSITION = 1 << 19;

        /// `Shell::update_accessibility` reaches the platform's assistive
        /// technologies, and the `Accessibility*` events are raised.
        ///
        /// Set only by a backend built with the `accesskit` feature, which is
        /// where that method and those events are, and an adapter for its
        /// platform.
        const ACCESSIBILITY = 1 << 20;

        /// Dead keys and Compose sequences compose: a dead accent and the next
        /// key, or a Compose-key sequence, arrive as one
        /// [`ShellEvent::TextCommit`](crate::ShellEvent::TextCommit) carrying
        /// the composed text.
        ///
        /// libxkbcommon's compose module on Wayland and X11, with the user's
        /// locale's table; `TranslateMessage` on Win32; `interpretKeyEvents:`
        /// on AppKit. Without it a backend may still commit plain typing, but
        /// an accent typed through a dead key never lands on its letter. It
        /// says nothing about an input method, which is
        /// [`TEXT_IME`](Self::TEXT_IME): X11 composes and has none.
        const TEXT_COMPOSE = 1 << 21;

        /// A panel with
        /// [`PanelKeyboard::Exclusive`](crate::PanelKeyboard::Exclusive)
        /// keeps the keyboard while it is mapped: the window system will not
        /// give it to another window, whatever the user clicks.
        ///
        /// Wayland's layer-shell enforces it in the compositor, on the
        /// [`Top`](crate::PanelLayer::Top) and
        /// [`Overlay`](crate::PanelLayer::Overlay) layers; the protocol lets a
        /// compositor use ordinary focus on the lower two. Clear on Win32 and
        /// AppKit, where an exclusive panel takes the keyboard when it is
        /// shown (or hears
        /// [`FocusRefused`](crate::ShellEvent::FocusRefused)) and the user
        /// can click it away, which arrives as
        /// [`Focus`](crate::ShellEvent::Focus) with `focused: false`. A
        /// launcher that should close when it loses the keyboard closes on
        /// that event; a lock screen cannot be built without this bit.
        const PANEL_EXCLUSIVE_KEYBOARD = 1 << 22;

        /// [`Shell::set_cursor`](crate::Shell::set_cursor) with a shape
        /// changes the cursor the user sees, to the platform's own themed
        /// cursor for that [`CursorIcon`](crate::CursorIcon).
        ///
        /// Win32's stock `user32` cursors, AppKit's `NSCursor`, and the
        /// canvas's CSS `cursor` on the web; on Wayland, `cursor-shape-v1`
        /// where the compositor offers it, which hands the compositor the
        /// shape's name and lets it draw the user's theme. Clear on X11, which
        /// has no cursor theme loader yet. Without it, `set_cursor` still
        /// accepts and records a shape — a UI calls it on every hover, and an
        /// error there would be one every caller had to handle — and hiding
        /// with `None` works everywhere; only the shape goes unseen.
        const CURSOR_SHAPE = 1 << 23;
    }
}

impl ShellCaps {
    /// The set a full-featured desktop backend is expected to report.
    ///
    /// Not a promise about any particular platform — it is the reference point
    /// tests and [`HeadlessShell`](crate::HeadlessShell) start from, and the
    /// value a doc example can name without pretending to know what the machine
    /// running it supports.
    pub const DESKTOP: Self = Self::all();

    /// Whether pointer capture in either form is available.
    #[inline]
    #[must_use]
    pub const fn has_pointer_capture(self) -> bool {
        self.intersects(Self::POINTER_LOCK.union(Self::POINTER_CONFINE))
    }

    /// Whether a first-person camera can get trustworthy aim input: the pointer
    /// can be locked *and* relative motion is delivered.
    ///
    /// Both halves are required, and this helper exists because forgetting the
    /// second one produces a camera that works until the pointer reaches the
    /// edge of the screen.
    #[inline]
    #[must_use]
    pub const fn has_mouselook(self) -> bool {
        self.contains(Self::POINTER_LOCK.union(Self::RAW_POINTER_MOTION))
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn mouselook_needs_both_halves() {
        assert!(ShellCaps::DESKTOP.has_mouselook());
        // Pointer lock alone is the trap: aim breaks at the screen edge.
        assert!(!ShellCaps::POINTER_LOCK.has_mouselook());
        assert!(!ShellCaps::RAW_POINTER_MOTION.has_mouselook());
        assert!((ShellCaps::POINTER_LOCK | ShellCaps::RAW_POINTER_MOTION).has_mouselook());
    }

    #[test]
    fn pointer_capture_is_either_form() {
        assert!(ShellCaps::POINTER_LOCK.has_pointer_capture());
        assert!(ShellCaps::POINTER_CONFINE.has_pointer_capture());
        assert!(!ShellCaps::POINTER_WARP.has_pointer_capture());
        assert!(!ShellCaps::empty().has_pointer_capture());
    }

    #[test]
    fn a_wayland_shaped_cap_set_is_expressible() {
        // The point of the type: the exact combination Wayland produces — hardware
        // upscale yes, aspect hint no, pointer warp no — is a value, not a
        // `#[cfg]`.
        let wayland = ShellCaps::HW_UPSCALE
            | ShellCaps::POINTER_LOCK
            | ShellCaps::POINTER_CONFINE
            | ShellCaps::RAW_POINTER_MOTION
            | ShellCaps::TEXT_IME
            | ShellCaps::TEXT_COMPOSE
            | ShellCaps::CLIPBOARD
            | ShellCaps::DRAG_DROP
            | ShellCaps::MULTI_WINDOW
            | ShellCaps::SERVER_DECORATIONS
            | ShellCaps::FRACTIONAL_SCALE
            | ShellCaps::EVENT_WAIT
            | ShellCaps::CURSOR_SHAPE;
        assert!(wayland.has_mouselook());
        assert!(!wayland.contains(ShellCaps::POINTER_WARP));
        assert!(!wayland.contains(ShellCaps::ASPECT_HINT_HONORED));
        assert!(!wayland.contains(ShellCaps::WINDOW_POSITION));
        // `wl_touch` exists and this backend does not bind it, which is the
        // whole difference between a platform and a backend.
        assert!(!wayland.contains(ShellCaps::TOUCH));
        assert!(ShellCaps::DESKTOP.contains(wayland));
        assert_eq!(ShellCaps::default(), ShellCaps::empty());
    }
}