| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316 |
- /**
- * @license Copyright (c) 2003-2015, CKSource - Frederico Knabben. All rights reserved.
- * For licensing, see LICENSE.md.
- */
- import Node from './node.js';
- import Text from './text.js';
- import TextProxy from './textproxy.js';
- import Range from './range.js';
- import DocumentFragment from './documentfragment.js';
- import NodeList from './nodelist.js';
- import CKEditorError from '../../utils/ckeditorerror.js';
- /**
- * Contains functions used for composing model tree, grouped together under "model writer" name. Those functions
- * are built on top of {@link engine.model.Node node}, and it's child classes', APIs.
- *
- * Model writer API has multiple advantages and it is highly recommended to use it when changing model tree and nodes:
- * * model writer API {@link engine.model.writer.normalizeNodes normalizes inserted nodes}, which means that you can insert
- * not only {@link engine.model.Node nodes}, but also `String`s, {@link engine.model.TextProxy text proxies} and
- * {@link engine.model.DocumentFragment document fragments},
- * * model writer API operates on {@link engine.model.Position positions}, which means that you have
- * better control over manipulating model tree as positions operate on offsets rather than indexes,
- * * model writer API automatically merges {@link engine.model.TextNode text nodes} with same attributes, which means
- * lower memory usage and better efficiency.
- *
- * @namespace engine.model.writer
- */
- export default {
- insert,
- remove,
- move,
- setAttribute,
- removeAttribute,
- normalizeNodes
- };
- /**
- * Inserts given nodes at given position.
- *
- * @function engine.model.writer.insert
- * @param {engine.model.Position} position Position at which nodes should be inserted.
- * @param {engine.model.NodeSet} nodes Nodes to insert.
- * @returns {engine.model.Range} Range spanning over inserted elements.
- */
- export function insert( position, nodes ) {
- nodes = normalizeNodes( nodes );
- // We have to count offset before inserting nodes because they can get merged and we would get wrong offsets.
- const offset = nodes.reduce( ( sum, node ) => sum + node.offsetSize, 0 );
- const parent = position.parent;
- // Insertion might be in a text node, we should split it if that's the case.
- let index = _splitNodeAtPosition( position );
- // Insert nodes at given index. After splitting we have a proper index and insertion is between nodes,
- // using basic `Element` API.
- parent.insertChildren( index, nodes );
- // Merge text nodes, if possible. Merging is needed only at points where inserted nodes "touch" "old" nodes.
- _mergeNodesAtIndex( parent, index + nodes.length );
- _mergeNodesAtIndex( parent, index );
- return new Range( position, position.getShiftedBy( offset ) );
- }
- /**
- * Removed nodes in given range. Only {@link engine.model.Range#isFlat flat} ranges are accepted.
- *
- * @function engine.model.writer.remove
- * @param {engine.model.Range} range Range containing nodes to remove.
- * @returns {Array.<engine.model.Node>}
- */
- export function remove( range ) {
- if ( !range.isFlat ) {
- /**
- * Trying to remove a range that starts and ends in different element.
- *
- * @error model-writer-remove-range-not-flat
- */
- throw new CKEditorError( 'model-writer-remove-range-not-flat: ' +
- 'Trying to remove a range that starts and ends in different element.' );
- }
- const parent = range.start.parent;
- // Range may be inside text nodes, we have to split them if that's the case.
- const indexStart = _splitNodeAtPosition( range.start );
- const indexEnd = _splitNodeAtPosition( range.end );
- // Remove the text nodes using basic `Element` API.
- const removed = parent.removeChildren( indexStart, indexEnd - indexStart );
- // Merge text nodes, if possible. After some nodes were removed, node before and after removed range will be
- // touching at the position equal to the removed range beginning. We check merging possibility there.
- _mergeNodesAtIndex( parent, indexStart );
- return removed;
- }
- /**
- * Moves nodes in given range to given target position. Only {@link engine.model.Range#isFlat flat} ranges are accepted.
- *
- * @param {engine.model.Range} sourceRange Range containing nodes to move.
- * @param {engine.model.Position} targetPosition Position to which nodes should be moved.
- * @returns {engine.model.Range} Range containing moved nodes.
- */
- export function move( sourceRange, targetPosition ) {
- /* jshint validthis:true */
- if ( !sourceRange.isFlat ) {
- /**
- * Trying to move a range that starts and ends in different element.
- *
- * @error model-writer-move-range-not-flat
- */
- throw new CKEditorError( 'model-writer-move-range-not-flat: ' +
- 'Trying to move a range that starts and ends in different element.' );
- }
- const nodes = this.remove( sourceRange );
- // We have to fix `targetPosition` because model changed after nodes from `sourceRange` got removed and
- // that change might have an impact on `targetPosition`.
- targetPosition = targetPosition._getTransformedByDeletion( sourceRange.start, sourceRange.end.offset - sourceRange.start.offset );
- return this.insert( targetPosition, nodes );
- }
- /**
- * Sets given attribute on nodes in given range.
- *
- * @param {engine.model.Range} range Range containing nodes that should have the attribute set.
- * @param {String} key Key of attribute to set.
- * @param {*} value Attribute value.
- */
- export function setAttribute( range, key, value ) {
- // Range might start or end in text nodes, so we have to split them.
- _splitNodeAtPosition( range.start );
- const indexEnd = _splitNodeAtPosition( range.end );
- // Iterate over all items in the range.
- for ( let item of range.getItems() ) {
- // Iterator will return `TextProxy` instances but we know that those text proxies will
- // always represent full text nodes (this is guaranteed thanks to splitting we did before).
- // So, we can operate on those text proxies' text nodes.
- let node = item instanceof TextProxy ? item.textNode : item;
- if ( value !== null ) {
- node.setAttribute( key, value );
- } else {
- node.removeAttribute( key );
- }
- // After attributes changing it may happen that some text nodes can be merged. Try to merge with previous node.
- _mergeNodesAtIndex( node.parent, node.index );
- }
- // Try to merge last changed node with it's previous sibling (not covered by the loop above).
- _mergeNodesAtIndex( range.end.parent, indexEnd );
- }
- /**
- * Removes given attribute from nodes in given range.
- *
- * @param {engine.model.Range} range Range containing nodes that should have the attribute removed.
- * @param {String} key Key of attribute to remove.
- */
- export function removeAttribute( range, key ) {
- /* jshint validthis:true */
- this.setAttribute( range, key, null );
- }
- /**
- * Normalizes given object or an array of objects to an array of {@link engine.model.Node nodes}. See
- * {@link engine.model.NodeSet NodeSet} for details on how normalization is performed.
- *
- * @param {engine.model.NodeSet} nodes Objects to normalize.
- * @returns {Array.<engine.model.Node>} Normalized nodes.
- */
- export function normalizeNodes( nodes ) {
- const normalized = [];
- if ( !( nodes instanceof Array ) ) {
- nodes = [ nodes ];
- }
- // Convert instances of classes other than Node.
- for ( let i = 0; i < nodes.length; i++ ) {
- if ( typeof nodes[ i ] == 'string' ) {
- normalized.push( new Text( nodes[ i ] ) );
- } else if ( nodes[ i ] instanceof TextProxy ) {
- normalized.push( new Text( nodes[ i ].data, nodes[ i ].getAttributes() ) );
- } else if ( nodes[ i ] instanceof DocumentFragment || nodes[ i ] instanceof NodeList ) {
- for ( let child of nodes[ i ] ) {
- normalized.push( child );
- }
- } else if ( nodes[ i ] instanceof Node ) {
- normalized.push( nodes[ i ] );
- }
- // Skip unrecognized type.
- }
- // Merge text nodes.
- for ( let i = 1; i < normalized.length; i++ ) {
- const node = normalized[ i ];
- const prev = normalized[ i - 1 ];
- if ( node instanceof Text && prev instanceof Text && _haveSameAttributes( node, prev ) ) {
- // Doing this instead changing prev.data because .data is readonly.
- normalized.splice( i - 1, 2, new Text( prev.data + node.data, prev.getAttributes() ) );
- i--;
- }
- }
- return normalized;
- }
- /**
- * Checks if nodes before and after given index in given element are {@link engine.model.Text text nodes} and
- * merges them into one node if they have same attributes.
- *
- * Merging is done by removing two text nodes and inserting a new text node containing data from both merged text nodes.
- *
- * @ignore
- * @private
- * @param {engine.model.Element} element Parent element of nodes to merge.
- * @param {Number} index Index between nodes to merge.
- */
- function _mergeNodesAtIndex( element, index ) {
- const nodeBefore = element.getChild( index - 1 );
- const nodeAfter = element.getChild( index );
- // Check if both of those nodes are text objects with same attributes.
- if ( nodeBefore instanceof Text && nodeAfter instanceof Text && _haveSameAttributes( nodeBefore, nodeAfter ) ) {
- // Append text of text node after index to the before one.
- const mergedNode = new Text( nodeBefore.data + nodeAfter.data, nodeBefore.getAttributes() );
- // Remove separate text nodes.
- element.removeChildren( index - 1, 2 );
- // Insert merged text node.
- element.insertChildren( index - 1, mergedNode );
- }
- }
- /**
- * Checks if given position is in a text node, and if so, splits the text node in two text nodes, each of them
- * containing a part of original text node.
- *
- * @ignore
- * @private
- * @param {engine.model.Position} position Position at which node should be split.
- * @returns {Number} Index, in position's parent element, between split nodes.
- */
- function _splitNodeAtPosition( position ) {
- const textNode = position.textNode;
- const element = position.parent;
- if ( textNode ) {
- const offsetDiff = position.offset - textNode.startOffset;
- const index = textNode.index;
- element.removeChildren( index, 1 );
- const firstPart = new Text( textNode.data.substr( 0, offsetDiff ), textNode.getAttributes() );
- const secondPart = new Text( textNode.data.substr( offsetDiff ), textNode.getAttributes() );
- element.insertChildren( index, [ firstPart, secondPart ] );
- return index + 1;
- }
- return element.offsetToIndex( position.offset );
- }
- /**
- * Checks whether two given nodes have same attributes.
- *
- * @ignore
- * @private
- * @param {engine.model.Node} nodeA Node to check.
- * @param {engine.model.Node} nodeB Node to check.
- * @returns {Boolean} `true` if nodes have same attributes, `false` otherwise.
- */
- function _haveSameAttributes( nodeA, nodeB ) {
- const iteratorA = nodeA.getAttributes();
- const iteratorB = nodeB.getAttributes();
- for ( let attr of iteratorA ) {
- if ( attr[ 1 ] !== nodeB.getAttribute( attr[ 0 ] ) ) {
- return false;
- }
- iteratorB.next();
- }
- return iteratorB.next().done;
- }
- /**
- * Value that can be normalized to an array of {@link engine.model.Node nodes}.
- *
- * Non-arrays are normalized as follows:
- * * {@link engine.model.Node Node} is left as is,
- * * {@link engine.model.TextProxy TextProxy} and `String` are normalized to {@link engine.model.Text Text},
- * * {@link engine.model.NodeList NodeList} is normalized to an array containing all nodes that are in that node list,
- * * {@link engine.model.DocumentFragment DocumentFragment} is normalized to an array containing all of it's children.
- *
- * Arrays are processed item by item like non-array values and flattened to one array. Normalization always results in
- * a flat array of {@link engine.model.Node nodes}. Consecutive text nodes (or items normalized to text nodes) will be
- * merged if they have same attributes.
- *
- * @typedef {engine.model.Node|engine.model.TextProxy|String|engine.model.NodeList|engine.model.DocumentFragment|Iterable}
- * engine.model.NodeSet
- */
|