Title: CSS Anchor Positioning
Shortname: css-anchor-position
Level: 1
Status: ED
Group: csswg
Work Status: exploring
ED: https://drafts.csswg.org/css-anchor-position-1/
Editor: Tab Atkins-Bittner, Google, http://xanthir.com/contact/, w3cid 42199
Editor: Jhey Tompkins, Google, https://twitter.com/jh3yy, w3cid 137616
Editor: Ian Kilpatrick, Google
Abstract: This specification defines 'anchor positioning', where a positioned element can size and position itself relative to one or more "anchor elements" elsewhere on the page.
Introduction {#intro} ===================== While CSS generally determines the position and size of elements according to their parents or other ancestors, [=absolutely positioned=] elements barely participate in their ancestors' layout. Instead, they're sized and positioned explicitly by the [=inset properties=] and [=box alignment properties=], only referencing the final size and position of their [=containing block=]. This provides extreme flexibility, allowing elements to be positioned more or less arbitrarily, including over the top of other elements in ways that the layout methods don't otherwise allow, but in return it's not very expressive-- the element cannot easily express its size and position in terms of other elements on the page. Anchoring {#anchoring} ================================== The anchor functions ''anchor()'' and ''anchor-size()'', defined below, give back some of the expressivity of ordinary layout without compromising on the flexibility and power of [=absolute positioning=]. Using these functions, one can size and position an [=absolutely positioned=] element relative to one or more [=anchor elements=] on the page. The ''@position-set'' rule allows even more flexibility, allowing multiple different sizes/positions to be tried out sequentially until one is found that fits within the [=containing block=]. Anchor-based Positioning: the ''anchor()'' function {#anchor-pos} --------------------------------------------------- An [=absolutely-positioned=] element can use the anchor() function as a value in its [=inset properties=] to refer to the position of one or more [=anchor elements=]. The ''anchor()'' function resolves to a <>.
	<anchor()> = anchor( <>? <>, <>? )
	<> = <> | implicit
	<> = auto | auto-same
	           | top | left | right | bottom
	           | start | end | self-start | self-end
	           | <> | center
The ''anchor()'' function has three arguments: * the <> value specifies how to find the [=anchor element=] it will be drawing positioning information from. If omitted, it behaves as the element's [=default anchor specifier=]. Its possible values are:
: <> :: Specifies the [=anchor name=] it will look for. This name is a [=tree-scoped reference=]. : implicit :: Selects the [=implicit anchor element=] defined for the element, if possible.
See [=target anchor element=] for details. * the <> value refers to the position of the corresponding side of the [=target anchor element=]. The auto and auto-same keywords indicate [=automatic anchor positioning=]. See [[#anchor-auto]] for details. The physical <> keywords (left, right, top, and bottom) are only useful in [=inset properties=] corresponding to their corresponding axis: for example, using ''top'' and ''bottom'' in 'left' or 'right' (or in 'inset-inline-start' if the inline axis is horizontal, etc) results in an [=invalid anchor query=]. The logical <> keywords (start, end, self-start, and self-end) map to one of the physical keywords depending on the property the function is being used in (''top'' or ''bottom'' in the 'top' or 'bottom' properties, etc) and the [=writing mode=] of either the element (for ''self-start'' and ''self-end'') or the [=writing mode=] of the element's [=containing block=] (for ''start'' and ''end''). Issue: Do we need to refer to the anchor element's [=writing mode=]? I think that's too unpredictable to actually do anything useful. If you're specifying your 'inset-inline-start' property, you almost certainly want to refer to an anchor edge relative to your own directions, not an unpredictable edge based on the anchor. A <> value refers to a position a corresponding percentage between the ''start'' and ''end'' sides, with ''0%'' being equivalent to ''start'' and ''100%'' being equivalent to ''end''. The center keyword is equivalent to ''50%''. * the optional <> final argument is a fallback value. If the ''anchor()'' represents an [=invalid anchor query=], it resolves to this value rather that determining its value as detailed below. If omitted, it defaults to ''0px''. Issue: Computed value for anchor() probably needs to be the anchor() function, but with the target anchor element resolved. This allows for transitions to work properly with tree-scoped names, and with changing anchor elements. See Issue 8180. An ''anchor()'' function representing a [=valid anchor query=] resolves at [=used value=] time to the <> that would align the edge of the positioned elements' [=inset-modified containing block=] corresponding to the property the function appears in with the specified border edge of the [=target anchor element=], assuming that all [=scroll containers=] between the [=target anchor element=] and the positioned element's [=containing block=] are scrolled to their initial scroll position (but see 'anchor-scroll'). If the [=target anchor element=] is [=fragmented=], the axis-aligned bounding rectangle of the fragments' border boxes is used instead. Issue: Do we need to control which box we're referring to, so you can align to padding or content edge? If the positioned element has a [=snapshotted scroll offset=], then it is additionally visually shifted by those offsets, as if by an additional ''translate()'' transform.
For example, in ''.bar { top: anchor(--foo top); }'', the ''anchor()'' will resolve to the length that'll line up the .bar element's top edge with the ''--foo'' anchor's top edge. On the other hand, in ''.bar { bottom: anchor(--foo top); }'', it will instead resolve to the length that'll line up the .bar element's bottom edge with the ''--foo'' anchor's top edge. Since 'top' and 'bottom' values specify insets from different edges (the top and bottom of the element's [=containing block=], respectively), the same ''anchor()'' will usually resolve to different lengths in each.
Because the ''anchor()'' function resolves to a <>, it can be used in [=math functions=] like any other length. For example, the following will set up the element so that its [=inset-modified containing block=] is centered on the [=anchor element=] and as wide as possible without overflowing the [=containing block=]:
	.centered-message {
		position: fixed;
		max-width: max-content;
		justify-content: center;

		--center: anchor(--x 50%);
		--half-distance: min(
			abs(0% - var(--center)),
			abs(100% - var(--center))
		);
		left: calc(var(--center) - var(--half-distance));
		right: calc(var(--center) - var(--half-distance));
		bottom: anchor(--x top);
	}
	
This might be appropriate for an error message on an <{input}> element, for example, as the centering will make it easier to discover which input is being referred to.

Automatic Anchor Positioning

If a positioned element uses the ''anchor()'' function with the ''anchor()/auto'' or ''auto-same'' keywords in one of its [=inset properties=], and the opposing [=inset property=] is ''top/auto'', then the element is using automatic anchor positioning in that property's axis. This will resolve the ''anchor()'' function's <> to the correct side of the anchor element, and automatically create entries in the [=position fallback list=] to flip the positioned element to the opposite side if necessary. [=Automatic anchor positioning=] is only active if the opposite [=inset property=] is ''top/auto''. (For example, if an element had ''top: anchor(auto);'', it would have to also have ''bottom: auto;''.) If this is not the case, the ''anchor()'' function represents an [=invalid anchor query=], and the element does not use [=automatic anchor positioning=] in that axis. When using [=automatic anchor positioning=], the ''anchor()/auto'' <> behaves as the opposite side of the property it's used in. That is, when used in ''top: anchor(auto);'', it's equivalent to ''top: anchor(bottom);''; when used in ''bottom: anchor(auto);'', it's equivalent to ''bottom: anchor(top);''; etc. The ''anchor()/auto-same'' <> behaves as the property it's used in: ''top: anchor(auto-same);'' is equivalent to ''top: anchor(top);'', etc. Additionally, if the element has ''position-fallback: none'', [=automatic anchor positioning=] causes the element to gain a [=position fallback list=] consisting of two entries: * one containing all the base-style properties on the element that are valid to use in ''@try'' rules, with ''anchor()/auto''/''auto-same'' keywords resolved to their appropriate side. * one containing the same, but with the [=inset properties=] in each axis swapped, and the ''anchor()/auto''/''auto-same'' keywords resolved to the opposite sides as well. Note: If the element has a non-none 'position-fallback', these extra entries aren't added. Since the [=position fallback list=] styles override the "base" styles immediately, this will usually mean you wouldn't see a "base" ''anchor(auto)'' show up in the final styles at all, but if that does happen (it's specified in a property that isn't overriden by anything in the [=position fallback list=]), the only effect of the ''anchor()/auto''/''auto-same'' is to resolve to the appropriate side keyword.
For example, the following code using [=automatic anchor positioning=]:
	.foo {
		position: absolute;
		top: calc(.5em + anchor(--foo auto));
	}
	
is equivalent to the following more verbose and explicit code:
	.foo {
		position: absolute;
		position-fallback: --flip;
	}
	@position-fallback --flip {
		@try {
			top: calc(.5em + anchor(--foo bottom));
			bottom: auto;
		}
		@try {
			top: auto;
			bottom: calc(.5em + anchor(--foo top));
		}
	}
	
If the element uses [=automatic anchor positioning=] in both axises, it instead adds three entries to the [=position fallback list=]: one reversing just the block axis, one reversing just the inline axis, and finally one reversing both axises at once. ''anchor()/auto'' and ''anchor()/auto-same'' used in a ''@try'' rule cause the rule to insert multiple (2 or 4) sets of entries into the [=position fallback list=], as specified above, if they would validly trigger [=automatic anchor positioning=]. Anchor-based Sizing: the ''anchor-size()'' function {#anchor-size} --------------------------------------------------- An [=absolutely-positioned=] element can use the anchor-size() function in its [=sizing properties=] to refer to the size of one or more [=anchor elements=]. The ''anchor-size()'' function resolves to a <>.
anchor-size() = anchor( <>? <>, <>? )
<> = width | height | block | inline | self-block | self-inline
The ''anchor-size()'' function is similar to ''anchor()'', and takes the same arguments, save that the <> keywords are replaced with <>, referring to the distance between two opposing sides. The physical <> keywords (width and height) refer to the width and height, respectively, of the [=target anchor element=]. Unlike ''anchor()'', there is no restriction on having to match axises; for example, ''width: anchor-size(--foo height);'' is valid. The logical <> keywords (block, inline, self-block, and self-inline) map to one of the physical keywords according to either the [=writing mode=] of the element (for ''self-block'' and ''self-inline'') or the [=writing mode=] of the element's [=containing block=] (for ''anchor-size()/block'' and ''anchor-size()/inline''). An ''anchor-size()'' function representing a [=valid anchor query=] resolves to the <> separating the relevant border edges (either left and right, or top and bottom, whichever is in the specified axis) of the [=target anchor element=]. Taking Scroll Into Account: the 'anchor-scroll' property {#scroll} ------------------------------------------------------------------
Name: anchor-scroll
Value: none | default | <>
Initial: none
Inherited: no
Applies to: [=absolutely-positioned=] elements
Animation Type: discrete
Because scrolling is often done in a separate thread from layout in implementations for performance reasons, but ''anchor()'' can result in both positioning changes (which can be handled in the scrolling thread) and layout changes (which cannot), ''anchor()'' is defined to assume all the [=scroll containers=] between the anchor element and the positioned element's containing block are at their initial scroll position. This means a positioned element will not be aligned with its anchor if any of the scrollers are not at their initial positions. The 'anchor-scroll' property allows an author to compensate for this, without losing the performance benefits of the separate scrolling thread, so long as the positioned element is only anchoring to a single anchor element. Its values are:
: none :: No effect. : default :: Behaves identically to <>, but draws its value from 'anchor-default' on the element. : <> :: Selects a [=target anchor element=] the same as ''anchor()'', which will be compensated for in positioning and fallback.
If 'anchor-scroll' is not ''anchor-scroll/none'' on an [=absolutely-positioned=] element |query el|, and there is a [=target anchor element=] for |query el| given the 'anchor-scroll' value, and at least one ''anchor()'' function on |query el| refers to the same [=target anchor element=], then |query el| has a snapshotted scroll offset, which is a pair of lengths representing a vertical and horizontal offset. The [=snapshotted scroll offset=] is the sum of the offsets from the [=initial scroll position=] of all [=scroll container=] ancestors of the [=target anchor element=], up to but not including |query el|'s [=containing block=]. Issue: Define the precise timing of the snapshot: updated each frame, before style recalc.
Determining The Anchor: the 'anchor-name' property {#determining} -----------------------------------------------------------------
Name: anchor-name
Value: none | <>
Initial: none
Inherited: no
Applies to: all elements that generate a [=principal box=]
Animation Type: discrete
Values are defined as follows:
: none :: The property has no effect. : <> :: If the element generates a [=principal box=], the element is an anchor element, with an anchor name equal to the <>. The [=anchor name=] is a [=tree-scoped name=]. Otherwise, the property has no effect.
The [=anchor functions=] refer to an [=anchor element=] by name. That name is not necessarily unique on the page, however; even if it is, the [=anchor element=] in question might not be capable of anchoring the positioned element.
To determine the target [=anchor element=] given a querying element |query el| and an anchor specifier |anchor spec|: 1. If |anchor spec| is ''implicit'', and the Popover API defines an [=implicit anchor element=] for |query el| which is an [=acceptable anchor element=] for |query el|, return that element. Otherwise, return nothing. 3. Otherwise, |anchor spec| is a <>. Return the first element |el| in tree order that satisfies the following conditions: * |el| is an [=anchor element=] with an [=anchor name=] of |name|. * |el|'s [=anchor name=] and |name| are both associated with the same [=tree=] [=tree/root=]. Note: The [=anchor name=] is a [=tree-scoped name=], while |name| is a [=tree-scoped reference=]. * |el| is an [=acceptable anchor element=] for |query el|. If no element satisfies these conditions, return nothing. Note: The general rule captured by these conditions is that |el| must be fully laid out before |query el| is laid out. CSS's rules about the layout order of stacking contexts give us assurances about this, and the list of conditions above exactly rephrases the stacking context rules into just what's relevant for this purpose, ensuring there is no possibly circularity in anchor positioning. Note: An 'anchor-name' defined by styles in one [=shadow tree=] won't be seen by [=anchor functions=] in styles in a different [=shadow tree=], preserving encapsulation. However, elements in different [=shadow trees=] can still anchor to each other, so long as both the 'anchor-name' and [=anchor function=] come from styles in the same tree, such as by using ''::part()'' to style an element inside a shadow. ([=Implicit anchor elements=] also aren't intrinsically limited to a single tree, but the details of that will depend on the API assigning them.)
An element |el| is a acceptable anchor element for an [=absolutely positioned=] element |query el| if any of the following are true: * |query el| is in a higher [=root layer=] than |el|. * |query el| and |el| are in the same [=root layer=], and all of the following are true: * Either |el| is a descendant of |query el|'s [=containing block=], or |query el|'s [=containing block=] is the [=initial containing block=]. * If |el| has the same [=containing block=] as |query el|, |el| is not [=absolutely positioned=]. * If |el| has a different [=containing block=] from |query el|, the last [=containing block=] in |el|'s [=containing block chain=] before reaching |query el|'s [=containing block=] is not [=absolutely positioned=]. For the purposes of this algorithm, an element is in a particular root layer corresponding to the closest [=inclusive ancestor=] that is in the [=top layer=], or the document if there isn't one. [=Root layers=] are "higher" if their corresponding element is later in the [=top layer=] list; the layer corresponding to the document is lower than all other layers. Note: This wording around "root layer" needs to live in a stacking-context spec, after pulling the [=top layer=] stuff out of [[fullscreen]].
An element can also have an implicit anchor element, used when an [=anchor function=] doesn't specify an explicit [=anchor name=]. Note: The Popover API, for example, defines an [=implicit anchor element=] for a popover-- the element that the popover is attached to.

Default Anchors: the 'anchor-default' property

Name: anchor-default
Value: <>
Initial: implicit
Applies to: [=absolutely positioned=] elements
Inherited: no
Animation type: discrete
The 'anchor-default' property defines the default anchor specifier for all [=anchor functions=] on the element, allowing multiple elements to use the same set of [=anchor functions=] (and [=position fallback lists=]!) while changing which [=anchor element=] each is referring to. Its values are identical to the <> term in ''anchor()'' and ''anchor-size()''.
For example, in the following code both ''.foo'' and ''.bar'' elements can use the same positioning properties and fallback, just changing the anchor element they're referring to:
	.anchored {
		position: absolute;
		position-fallback: --under-then-over;
	}

	@position-fallback --under-then-over {
		@try {
			// No  specified,
			// so it takes from 'anchor-default'.
			top: calc(.5em + anchor(auto));
			bottom: auto;
		}
	}

	.foo.anchored {
		anchor-default: --foo;
	}
	.bar.anchored {
		anchor-default: --bar;
	}
	
Anchor Queries {#queries} -------------- The ''anchor()'' and ''anchor-size()'' functions represent an anchor query: a request for the position of one or more sides of one or more [=anchor elements=]. [=Anchor queries=] are valid only if all of the following conditions are true: * Their function is used on an element that is [=absolutely-positioned=]. * If representing an ''anchor()'' function, the function is being used in an [=inset property=]. * If representing an ''anchor()'' function and the <> keyword is a physical keyword, it's used in an [=inset property=] in the corresponding axis. * If representing an ''anchor-size()'' function, the function is being used in a [=sizing property=]. * There is a [=target anchor element=] for the element and the [=anchor name=] specified in the function. Note: As specified in the definition of ''anchor()'', an [=invalid anchor query=] causes the function to resolve to its fallback value instead. Fallback Sizing/Positioning {#fallback} =========================== Anchor positioning, while powerful, can also be unpredictable. The [=anchor element=] might be anywhere on the page, so positioning an element in any particular fashion (such as above the anchor, or the right of the anchor) might result in the positioned element overflowing its [=containing block=] or being positioned partially off screen. To ameliorate this, an [=absolutely positioned=] element can use the 'position-fallback' property to refer to a ''@position-fallback'' block, giving a list of possible style rules to try out. Each is applied to the element, one by one, and the first that doesn't cause the element to overflow its [=containing block=] is taken as the winner. The 'position-fallback' Property {#fallback-property} --------------------------------
Name: position-fallback
Value: none | <>
Initial: none
Inherited: no
Applies to: [=absolutely-positioned=] elements
Animation type: discrete
Values have the following meanings:
: none :: The property has no effect; the element does not use a [=position fallback list=]. : <> :: If there is a ''@position-fallback'' rule with a name matching the specified ident, then the element uses that [=position fallback list=]. Otherwise, this value has no effect.
The ''@position-fallback'' Rule {#fallback-rule} ------------------------------- The @position-fallback rule defines a [=position fallback list=] with a given name, specifying one or more sets of positioning properties inside of @try blocks that will be applied to an element, with each successive one serving as fallback if the previous would cause the element to partially or fully overflow its [=containing block=]. The grammar of the ''@position-fallback'' rule is:
@position-fallback <> {
	<>
}

@try { <> }
The ''@position-fallback'' rule only accepts ''@try'' rules. The <> specified in the prelude is the rule's name. If multiple ''@position-fallback'' rules are declared with the same name, the last one in document order "wins". The ''@try'' rule only accepts the following [=properties=]: * [=inset properties=] * [=sizing properties=] * [=box alignment properties=] Issue: What exactly are the constraints that determine what's allowed here? Current list is based off of what's reasonable from Chrome's experimental impl. We can make a CQ that keys off of which fallback was used to allow more general styling, at least for descendants. The ''@try'' rules inside a ''@position-fallback'' specify a position fallback list, where each entry consists of the properties specified by each ''@try'', in order. Issue: Would be useful to be able to detect when your anchor(s) are fully off-screen and suppress your display entirely. For example, tooltips living outside the scroller holding the text they're anchored to don't want to just hover over arbitrary parts of the page because their anchor happens to have that position relative to the scrollport. Note: If multiple elements using different anchors want to use the same fallback positioning, just relative to their own anchor elements, omit the <> in ''anchor()'' and specify each element's anchor in 'anchor-default' instead. Note: The most common types of fallback positioning (putting the positioned element on one side of the anchor normally, but flipping to the opposite side if needed) can be done automatically, without using ''@position-fallback'' at all, by using ''anchor()/auto'' or ''anchor()/auto-side'' side values in the ''anchor()'' function. Applying Position Fallback {#fallback-apply} -------------------------- When an element uses a [=position fallback list=], it selects one entry from the list as defined below, and applies those properties to itself as [=used values=]. Note: These have to be applied as used values because we're in the middle of layout right now; defining how they'd interact with the cascade would be extremely confusing *at a minimum*, and perhaps actually circular. In any case, not worth the cost in spec or impl. Issue: This implies that the values can't be transitioned in the usual fashion, since transitions key off of computed values and we're past that point. However, popovers sliding between positions is a common effect in UI libs. Probably should introduce a smooth keyword to 'position-fallback' to trigger automatic "animation" of the fallback'd properties.
To determine the position fallback styles of an element |el|: 1. Let |base styles| be the current used styles of |el|. 2. [=list/For each=] |fallback styles| in the [=position fallback list=]: 1. Apply the styles in |fallback styles| to |el|, overriding the corresponding properties in |base styles|. Perform any specified/computed/used-value time normalizations that are required to make the overridden styles into [=used values=] (such as resolving [=math functions=], etc). Let |adjusted styles| be |el|'s styles after these adjustments. 2. If |el| has a [=snapshotted scroll offset=], then subtract the offsets from |el|'s margin box's position. Also, if any of |el|'s [=inset properties=] are non-auto, subtract the [=snapshotted scroll offset=] for the appropriate axis from their values. Recalculate |el|'s [=inset-modified containing block=] using these shifted values to obtain the |scroll-adjusted IMCB|. 3. If |el|'s margin box is fully contained within the |scroll-adjusted IMCB|, or |fallback styles| is the final entry in the [=position fallback list=], then use |adjusted styles| for |el| and exit this algorithm. Note: Descendants overflowing |el| don't affect this calculation, only |el|'s own [=margin box=].
The styles returned by [=determining the position fallback styles=] are taken as the final values for the specified properties. Implementations may choose to impose an implementation-defined limit on the length of [=position fallback lists=], to limit the amount of excess layout work that may be required. This limit must be at least five.
Nested anchors (an anchored element inside of another anchored element) present the potential for exponential blow-up of layouts when doing fallback, since the grandchild anchored element can cause scrollbars on an ancestor, changing the IMCB for the child anchored element, thus possibly causing the fallback choice to change for it. There are strategies to avoid this, but they're not without costs of their own. We should probably impose a maximum limit as well, to avoid this. However, since *most* usages won't be problematic in the first place, we don't want to restrict them unduly just to prevent weird situations from exploding. Perhaps a complexity budget based on the branching factor at each level? Like, accumulate the product of the fallback list lengths from ancestors, and your fallback list gets limited to not exceed a total product of, say, 1k. Get too deep and you're stuck with your first choice only! But this would allow large, complex fallback lists for top-level stuff, and even some reasonable nesting. (Length-five lists could be nested to depth of 4, for example, if we did go with 1k.) More thought is needed.
For example, the following CSS will first attempt to position a "popover" below the [=element=], but if it doesn't fit on-screen will switch to being above; it defaults to left-aligning, but will switch to right-aligning if that doesn't fit.
	#myPopover {
		position: fixed;
		position-fallback: --button-popover;
		overflow: auto;

		/* The popover is at least as wide as the button */
		min-width: anchor-size(--button width);

		/* The popover is at least as tall as 2 menu items */
		min-height: 6em;
	}

	@position-fallback --button-popover {
		/* First try to align the top, left edge of the popover
		with the bottom, left edge of the button. */
		@try {
			top: anchor(--button bottom);
			left: anchor(--button left);
		}

		/* Next try to align the bottom, left edge of the popover
		with the top, left edge of the button. */
		@try {
			bottom: anchor(--button top);
			left: anchor(--button left);
		}

		/* Next try to align the top, right edge of the popover
		with the bottom, right edge of the button. */
		@try {
			top: anchor(--button bottom);
			right: anchor(--button right);
		}

		/* Finally, try to align the bottom, right edge of the popover
		with the top, right edge of the button. Other positions are possible,
		but this is the final option the author would like the rendering
		engine to try. */
		@try {
			bottom: anchor(--button top);
			right: anchor(--button right);
		}
	}
	
Security Considerations {#sec} ======================= No Security issues have been raised against this document. Privacy Considerations {#priv} ====================== No Privacy issues have been raised against this document.