<?php

namespace MediaWiki\Extension\DiscussionTools;

use Config;
use DateInterval;
use DateTime;
use DateTimeImmutable;
use DateTimeZone;
use DOMElement;
use DOMNode;
use DOMText;
use IP;
use Language;
use MediaWiki\MediaWikiServices;
use MWException;
use Title;

// TODO clean up static vs non-static
// TODO consider making timestamp parsing not a returned function

class CommentParser {
	private const SIGNATURE_SCAN_LIMIT = 100;

	/** @var DOMElement */
	private $rootNode;

	/** @var ThreadItem[] */
	private $threadItems;
	/** @var CommentItem[] */
	private $commentItems;
	/** @var CommentItem[] */
	private $commentsById;
	/** @var HeadingItem[] */
	private $threads;

	/** @var Config */
	private $config;

	private $dateFormat;
	private $digits;
	/** @var string[] */
	private $contLangMessages;
	private $localTimezone;
	private $timezones;

	/**
	 * @param DomElement $rootNode Root node of content to parse
	 * @param Language $language Content language
	 * @param Config $config
	 * @param array $data
	 */
	public function __construct( DOMElement $rootNode, Language $language, Config $config, array $data = [] ) {
		$this->rootNode = $rootNode;
		$this->config = $config;

		if ( !$data ) {
			// TODO: Instead of passing data used for mocking, mock the methods that fetch the data.
			$data = Data::getLocalData( null, $config, $language );
		}

		$this->dateFormat = $data['dateFormat'];
		$this->digits = $data['digits'];
		$this->contLangMessages = $data['contLangMessages'];
		$this->localTimezone = $data['localTimezone'];
		$this->timezones = $data['timezones'];
	}

	/**
	 * @param DomElement $rootNode Root node
	 * @return CommentParser
	 */
	public static function newFromGlobalState( DOMElement $rootNode ) : CommentParser {
		return new static(
			$rootNode,
			MediaWikiServices::getInstance()->getContentLanguage(),
			MediaWikiServices::getInstance()->getMainConfig()
		);
	}

	/**
	 * Return the next leaf node in the tree order that is not an empty or whitespace-only text node.
	 *
	 * In other words, this returns a text node with content other than whitespace, or an element node
	 * with no children, that follows the given node.
	 *
	 * @param DOMNode $node Node to start searching at. This node's children are ignored.
	 * @return DOMNode
	 */
	private function nextInterestingLeafNode( DOMNode $node ) : DOMNode {
		$rootNode = $this->rootNode;
		$treeWalker = new TreeWalker(
			$rootNode,
			NodeFilter::SHOW_ELEMENT | NodeFilter::SHOW_TEXT,
			function ( $n ) use ( $node, $rootNode ) {
				// Ignore this node and its descendants
				// (unless it's the root node, this is a special case for "fakeHeading" handling)
				if ( $node !== $rootNode && ( $n === $node || $n->parentNode === $node ) ) {
					return NodeFilter::FILTER_REJECT;
				}
				if (
					(
						$n->nodeType === XML_TEXT_NODE &&
						CommentUtils::htmlTrim( $n->nodeValue ) !== ''
					) ||
					(
						$n->nodeType === XML_CDATA_SECTION_NODE &&
						CommentUtils::htmlTrim( $n->nodeValue ) !== ''
					) ||
					( $n->nodeType === XML_ELEMENT_NODE && !$n->firstChild )
				) {
					return NodeFilter::FILTER_ACCEPT;
				}
				return NodeFilter::FILTER_SKIP;
			}
		);
		$treeWalker->currentNode = $node;
		$treeWalker->nextNode();
		if ( !$treeWalker->currentNode ) {
			throw new MWException( 'nextInterestingLeafNode not found' );
		}
		return $treeWalker->currentNode;
	}

	/**
	 * @param string[] $values Values to match
	 * @return string Regular expression
	 */
	private static function regexpAlternateGroup( array $values ) : string {
		return '(' . implode( '|', array_map( function ( string $x ) {
			return preg_quote( $x, '/' );
		}, $values ) ) . ')';
	}

	/**
	 * @param string[] $messageKeys Message keys
	 * @return string[] Message values
	 */
	private function getMessages( array $messageKeys ) : array {
		return array_map( function ( string $key ) {
			return $this->contLangMessages[$key];
		}, $messageKeys );
	}

	/**
	 * Get a regexp that matches timestamps generated using the given date format.
	 *
	 * This only supports format characters that are used by the default date format in any of
	 * MediaWiki's languages, namely: D, d, F, G, H, i, j, l, M, n, Y, xg, xkY (and escape characters),
	 * and only dates when MediaWiki existed, let's say 2000 onwards (Thai dates before 1941 are
	 * complicated).
	 *
	 * @param string $format Date format
	 * @param string $digitsRegexp Regular expression matching a single localised digit, e.g. '[0-9]'
	 * @param array $tzAbbrs Associative array mapping localised timezone abbreviations to
	 *   IANA abbreviations, for the local timezone, e.g. [ 'EDT' => 'EDT', 'EST' => 'EST' ]
	 * @return string Regular expression
	 */
	private function getTimestampRegexp(
		string $format, string $digitsRegexp, array $tzAbbrs
	) : string {
		$formatLength = strlen( $format );
		$s = '';
		// Adapted from Language::sprintfDate()
		for ( $p = 0; $p < $formatLength; $p++ ) {
			$num = false;
			$code = $format[ $p ];
			if ( $code === 'x' && $p < $formatLength - 1 ) {
				$code .= $format[++$p];
			}
			if ( $code === 'xk' && $p < $formatLength - 1 ) {
				$code .= $format[++$p];
			}

			switch ( $code ) {
				case 'xx' :
					$s .= 'x';
					break;
				case 'xg':
					$s .= self::regexpAlternateGroup(
						$this->getMessages( Language::MONTH_GENITIVE_MESSAGES )
					);
					break;
				case 'd':
					$num = '2';
					break;
				case 'D':
					$s .= self::regexpAlternateGroup(
						$this->getMessages( Language::WEEKDAY_ABBREVIATED_MESSAGES )
					);
					break;
				case 'j':
					$num = '1,2';
					break;
				case 'l':
					$s .= self::regexpAlternateGroup(
						$this->getMessages( Language::WEEKDAY_MESSAGES )
					);
					break;
				case 'F':
					$s .= self::regexpAlternateGroup(
						$this->getMessages( Language::MONTH_MESSAGES )
					);
					break;
				case 'M':
					$s .= self::regexpAlternateGroup(
						$this->getMessages( Language::MONTH_ABBREVIATED_MESSAGES )
					);
					break;
				case 'n':
					$num = '1,2';
					break;
				case 'Y':
					$num = '4';
					break;
				case 'xkY':
					$num = '4';
					break;
				case 'G':
					$num = '1,2';
					break;
				case 'H':
					$num = '2';
					break;
				case 'i':
					$num = '2';
					break;
				case '\\':
					// Backslash escaping
					if ( $p < $formatLength - 1 ) {
						$s .= preg_quote( $format[++$p], '/' );
					} else {
						$s .= preg_quote( '\\', '/' );
					}
					break;
				case '"':
					// Quoted literal
					if ( $p < $formatLength - 1 ) {
						$endQuote = strpos( $format, '"', $p + 1 );
						if ( $endQuote === false ) {
							// No terminating quote, assume literal "
							$s .= '"';
						} else {
							$s .= preg_quote( substr( $format, $p + 1, $endQuote - $p - 1 ), '/' );
							$p = $endQuote;
						}
					} else {
						// Quote at end of string, assume literal "
						$s .= '"';
					}
					break;
				default:
					$s .= preg_quote( $format[$p], '/' );
			}
			if ( $num !== false ) {
				$s .= '(' . $digitsRegexp . '{' . $num . '})';
			}
		}

		$tzRegexp = self::regexpAlternateGroup( array_keys( $tzAbbrs ) );

		// Hard-coded parentheses and space like in Parser::pstPass2
		// Ignore some invisible Unicode characters that often sneak into copy-pasted timestamps (T245784)
		// \uNNNN syntax can only be used from PHP 7.3
		return '/' . $s . '[\\x{200E}\\x{200F}]? [\\x{200E}\\x{200F}]?\\(' . $tzRegexp . '\\)/u';
	}

	/**
	 * Get a function that parses timestamps generated using the given date format, based on the result
	 * of matching the regexp returned by getTimestampRegexp()
	 *
	 * @param string $format Date format, as used by MediaWiki
	 * @param string|null $digits Localised digits from 0 to 9, e.g. `0123456789`
	 * @param string $localTimezone Local timezone IANA name, e.g. `America/New_York`
	 * @param array $tzAbbrs Map of localised timezone abbreviations to IANA abbreviations
	 *   for the local timezone, e.g. [ 'EDT' => 'EDT', 'EST' => 'EST' ]
	 * @return callable Parser function
	 */
	private function getTimestampParser(
		string $format, ?string $digits, string $localTimezone, array $tzAbbrs
	) : callable {
		$untransformDigits = function ( string $text ) use ( $digits ) {
			if ( !$digits ) {
				return $text;
			}
			return preg_replace_callback(
				'/[' . $digits . ']/',
				function ( array $m ) use ( $digits ) {
					return (string)strpos( $digits, $m[0] );
				},
				$text
			);
		};

		$formatLength = strlen( $format );
		$matchingGroups = [];
		for ( $p = 0; $p < $formatLength; $p++ ) {
			$code = $format[$p];
			if ( $code === 'x' && $p < $formatLength - 1 ) {
				$code .= $format[++$p];
			}
			if ( $code === 'xk' && $p < $formatLength - 1 ) {
				$code .= $format[++$p];
			}

			switch ( $code ) {
				case 'xx':
					break;
				case 'xg':
				case 'd':
				case 'j':
				case 'D':
				case 'l':
				case 'F':
				case 'M':
				case 'n':
				case 'Y':
				case 'xkY':
				case 'G':
				case 'H':
				case 'i':
					$matchingGroups[] = $code;
					break;
				case '\\':
					// Backslash escaping
					if ( $p < $formatLength - 1 ) {
						$p++;
					}
					break;
				case '"':
					// Quoted literal
					if ( $p < $formatLength - 1 ) {
						$endQuote = strpos( $format, '"', $p + 1 );
						if ( $endQuote !== false ) {
							$p = $endQuote;
						}
					}
					break;
				default:
					break;
			}
		}

		return function ( array $match ) use (
			$matchingGroups, $untransformDigits, $localTimezone, $tzAbbrs
		) {
			if ( is_array( $match[0] ) ) {
				// Strip PREG_OFFSET_CAPTURE data
				$match = array_map( function ( array $tuple ) {
					return $tuple[0];
				}, $match );
			}
			$year = 0;
			$monthIdx = 0;
			$day = 0;
			$hour = 0;
			$minute = 0;
			foreach ( $matchingGroups as $i => $code ) {
				$text = $match[$i + 1];
				switch ( $code ) {
					case 'xg':
						$monthIdx = array_search( $text, $this->getMessages( Language::MONTH_GENITIVE_MESSAGES ) );
						break;
					case 'd':
					case 'j':
						$day = intval( $untransformDigits( $text ) );
						break;
					case 'D':
					case 'l':
						// Day of the week - unused
						break;
					case 'F':
						$monthIdx = array_search( $text, $this->getMessages( Language::MONTH_MESSAGES ) );
						break;
					case 'M':
						$monthIdx = array_search( $text, $this->getMessages( Language::MONTH_ABBREVIATED_MESSAGES ) );
						break;
					case 'n':
						$monthIdx = intval( $untransformDigits( $text ) ) - 1;
						break;
					case 'Y':
						$year = intval( $untransformDigits( $text ) );
						break;
					case 'xkY':
						// Thai year
						$year = intval( $untransformDigits( $text ) ) - 543;
						break;
					case 'G':
					case 'H':
						$hour = intval( $untransformDigits( $text ) );
						break;
					case 'i':
						$minute = intval( $untransformDigits( $text ) );
						break;
					default:
						// TODO throw NotImplementedException or whatever it's called
						throw new MWException( 'Not implemented' );
				}
			}

			// The last matching group is the timezone abbreviation
			$tzAbbr = $tzAbbrs[ end( $match ) ];

			// Most of the time, the timezone abbreviation is not necessary to parse the date, since we
			// can assume all times are in the wiki's local timezone.
			$date = new DateTime();
			// setTimezone must be called before setDate/setTime
			$date->setTimezone( new DateTimeZone( $localTimezone ) );
			$date->setDate( $year, $monthIdx + 1, $day );
			$date->setTime( $hour, $minute, 0 );

			// But during the "fall back" at the end of DST, some times will happen twice.
			// Since the timezone abbreviation disambiguates the DST/non-DST times, we can detect
			// when PHP chose the wrong one, and then try the other one. It appears that PHP always
			// uses the later (non-DST) hour, but that behavior isn't documented, so we account for both.
			if ( $date->format( 'T' ) !== $tzAbbr ) {
				$altDate = clone $date;
				if ( $date->format( 'I' ) ) {
					// Parsed time is DST, try non-DST by advancing one hour
					$altDate->add( new DateInterval( 'PT1H' ) );
				} else {
					// Parsed time is non-DST, try DST by going back one hour
					$altDate->sub( new DateInterval( 'PT1H' ) );
				}
				if ( $altDate->format( 'T' ) === $tzAbbr ) {
					$date = $altDate;
					$discussionToolsWarning = 'Timestamp has timezone abbreviation for the wrong time';
				} else {
					$discussionToolsWarning = 'Ambiguous time at DST switchover was parsed';
				}
			}

			// Now set the timezone back to UTC for formatting
			$date->setTimezone( new DateTimeZone( 'UTC' ) );
			$date = DateTimeImmutable::createFromMutable( $date );
			if ( isset( $discussionToolsWarning ) ) {
				// @phan-suppress-next-line PhanUndeclaredProperty
				$date->discussionToolsWarning = $discussionToolsWarning;
			}

			return $date;
		};
	}

	/**
	 * Get a regular expression that matches timestamps in the local date format.
	 *
	 * This calls getTimestampRegexp() with predefined data for the current wiki.
	 *
	 * @return string Regular expression
	 */
	public function getLocalTimestampRegexp() : string {
		return $this->getTimestampRegexp(
			$this->dateFormat,
			$this->digits ? "[$this->digits]" : '\\d',
			$this->timezones
		);
	}

	/**
	 * Get a function that parses timestamps in the local date format, based on the result
	 * of matching the regexp returned by getLocalTimestampRegexp().
	 *
	 * This calls getTimestampParser() with predefined data for the current wiki.
	 *
	 * @return callable Parser function
	 */
	private function getLocalTimestampParser() : callable {
		return $this->getTimestampParser(
			$this->dateFormat,
			$this->digits,
			$this->localTimezone,
			$this->timezones
		);
	}

	/**
	 * Get the indent level of $node, relative to $rootNode.
	 *
	 * The indent level is the number of lists inside of which it is nested.
	 *
	 * @param DOMNode $node
	 * @return int
	 */
	private function getIndentLevel( DOMNode $node ) : int {
		$indent = 0;
		while ( $node ) {
			if ( $node === $this->rootNode ) {
				break;
			}
			$nodeName = strtolower( $node->nodeName );
			if ( $nodeName === 'li' || $nodeName === 'dd' ) {
				$indent++;
			}
			$node = $node->parentNode;
		}
		return $indent;
	}

	/**
	 * Find a user signature preceding a timestamp.
	 *
	 * The signature includes the timestamp node.
	 *
	 * A signature must contain at least one link to the user's userpage, discussion page or
	 * contributions (and may contain other links). The link may be nested in other elements.
	 *
	 * This function returns a two-element array. The first element is an array of sibling nodes
	 * comprising the signature, in reverse order (with $timestampNode or its parent node as the last
	 * element). The second element is the username (null for unsigned comments).
	 *
	 * @param DOMText $timestampNode Text node
	 * @param DOMNode|null $until Node to stop searching at
	 * @return array [ nodes, username ]
	 */
	private function findSignature( DOMText $timestampNode, ?DOMNode $until = null ) : array {
		// Support timestamps being linked to the diff introducing the comment:
		// if the timestamp node is the only child of a link node, use the link node instead
		if (
			!$timestampNode->previousSibling && !$timestampNode->nextSibling &&
			strtolower( $timestampNode->parentNode->nodeName ) === 'a'
		) {
			$timestampNode = $timestampNode->parentNode;
		}

		$node = $timestampNode;
		$sigNodes = [ $node ];
		$sigUsername = null;
		$length = 0;
		$lastLinkNode = $timestampNode;

		while (
			( $node = $node->previousSibling ) && $length < self::SIGNATURE_SCAN_LIMIT && $node !== $until
		) {
			$sigNodes[] = $node;
			$length += $node->textContent ? strlen( $node->textContent ) : 0;
			if ( !( $node instanceof DOMElement ) ) {
				continue;
			}

			$links = [];
			if ( strtolower( $node->nodeName ) === 'a' ) {
				$links = [ $node ];
			} else {
				// Handle links nested in formatting elements.
				// Helpful accidental feature: users whose signature is not detected in full (due to
				// text formatting) can just wrap it in a <span> to fix that.
				// "Ten Pound Hammer • (What did I screw up now?)"
				// "« Saper // dyskusja »"
				$links = $node->getElementsByTagName( 'a' );
			}
			if ( !count( $links ) ) {
				continue;
			}

			// Find the earliest link that links to the user's user page
			foreach ( $links as $link ) {
				$username = null;
				$title = CommentUtils::getTitleFromUrl( $link->getAttribute( 'href' ) );
				if ( !$title ) {
					continue;
				}
				if ( $title->getNamespace() === NS_USER || $title->getNamespace() === NS_USER_TALK ) {
					$username = $title->getText();
					if ( strpos( $username, '/' ) !== false ) {
						continue;
					}
				} elseif ( $title->isSpecial( 'Contributions' ) ) {
					$parts = explode( '/', $title->getText(), 2 );
					if ( !isset( $parts[1] ) ) {
						continue;
					}
					// Normalize the username: users may link to their contributions with an unnormalized name
					$userpage = Title::makeTitleSafe( NS_USER, $parts[1] );
					if ( !$userpage ) {
						continue;
					}
					$username = $userpage->getText();
				}
				if ( !$username ) {
					continue;
				}
				if ( IP::isIPv6( $username ) ) {
					// Bot-generated links "Preceding unsigned comment added by" have non-standard case
					$username = strtoupper( $username );
				}

				// Accept the first link to the user namespace, then only accept links to that user
				if ( $sigUsername === null ) {
					$sigUsername = $username;
				}
				if ( $username === $sigUsername ) {
					$lastLinkNode = $node;
					break;
				}
			}
			// Keep looking if a node with links wasn't a link to a user page
			// "Doc James (talk · contribs · email)"
		}
		// Pop excess text nodes
		while ( end( $sigNodes ) !== $lastLinkNode ) {
			array_pop( $sigNodes );
		}
		return [ $sigNodes, $sigUsername ];
	}

	/**
	 * Callback for TreeWalker that will skip over nodes where we don't want to detect
	 * comments (or section headings).
	 *
	 * @param DOMNode $node
	 * @return int Appropriate NodeFilter constant
	 */
	public static function acceptOnlyNodesAllowingComments( DOMNode $node ) {
		// The table of contents has a heading that gets erroneously detected as a section
		if ( $node instanceof DOMElement && $node->getAttribute( 'id' ) === 'toc' ) {
			return NodeFilter::FILTER_REJECT;
		}
		return NodeFilter::FILTER_ACCEPT;
	}

	/**
	 * Find a timestamps in a given text node
	 *
	 * @param DOMText $node Text node
	 * @param string $timestampRegex Timestamp regex
	 * @return array|null Match data
	 */
	public function findTimestamp( DOMText $node, string $timestampRegex ) : ?array {
			$nodeText = '';

			while ( $node ) {
				$nodeText .= $node->nodeValue;

				// In Parsoid HTML, entities are represented as a 'mw:Entity' node, rather than normal HTML
				// entities. On Arabic Wikipedia, the "UTC" timezone name contains some non-breaking spaces,
				// which apparently are often turned into &nbsp; entities by buggy editing tools. To handle
				// this, we must piece together the text, so that our regexp can match those timestamps.
				if (
				( $nextSibling = $node->nextSibling ) &&
				$nextSibling instanceof DOMElement &&
				$nextSibling->getAttribute( 'typeof' ) === 'mw:Entity'
				) {
				$nodeText .= $nextSibling->firstChild->nodeValue;

					// If the entity is followed by more text, do this again
					if (
					$nextSibling->nextSibling &&
					$nextSibling->nextSibling instanceof DOMText
					) {
					$node = $nextSibling->nextSibling;
					} else {
						$node = null;
					}
				} else {
					$node = null;
				}
			}

			$matchData = null;
			// Allows us to mimic match.index in #getComments
			if ( preg_match( $timestampRegex, $nodeText, $matchData, PREG_OFFSET_CAPTURE ) ) {
			return $matchData;
			}
		return null;
	}

	/**
	 * Get all discussion comments (and headings) within a DOM subtree.
	 *
	 * This returns a flat list, use getThreads() to get a tree structure starting at section headings.
	 *
	 * For example, for a MediaWiki discussion like this (we're dealing with HTML DOM here,
	 * the wikitext syntax is just for illustration):
	 *
	 *     == A ==
	 *     B. ~~~~
	 *     : C.
	 *     : C. ~~~~
	 *     :: D. ~~~~
	 *     ::: E. ~~~~
	 *     ::: F. ~~~~
	 *     : G. ~~~~
	 *     H. ~~~~
	 *     : I. ~~~~
	 *
	 * This function would return a structure like:
	 *
	 *     [
	 *       HeadingItem( { level: 0, range: (h2: A)        } ),
	 *       CommentItem( { level: 1, range: (p: B)         } ),
	 *       CommentItem( { level: 2, range: (li: C, li: C) } ),
	 *       CommentItem( { level: 3, range: (li: D)        } ),
	 *       CommentItem( { level: 4, range: (li: E)        } ),
	 *       CommentItem( { level: 4, range: (li: F)        } ),
	 *       CommentItem( { level: 2, range: (li: G)        } ),
	 *       CommentItem( { level: 1, range: (p: H)         } ),
	 *       CommentItem( { level: 2, range: (li: I)        } )
	 *     ]
	 *
	 * @return ThreadItem[] Thread items
	 */
	public function getThreadItems() : array {
		if ( !$this->threadItems ) {
			$this->buildThreads();
		}
		return $this->threadItems;
	}

	/**
	 * Same as getFlatThreadItems, but only returns the CommentItems
	 *
	 * @return CommentItem[] Comment items
	 */
	public function getCommentItems() : array {
		if ( !$this->commentItems ) {
			$this->buildThreads();
		}
		return $this->commentItems;
	}

	/**
	 * Find a CommentItem by its ID
	 *
	 * @param string $id Comment ID
	 * @return CommentItem|null Comment item, null if not found
	 */
	public function findCommentById( string $id ) : ?CommentItem {
		if ( !$this->commentsById ) {
			$this->buildThreads();
		}
		return $this->commentsById[$id] ?? null;
	}

	private function buildThreadItems() : void {
		$timestampRegex = $this->getLocalTimestampRegexp();
		$commentItems = [];
		$threadItems = [];
		$dfParser = $this->getLocalTimestampParser();

		// Placeholder heading in case there are comments in the 0th section
		$range = new ImmutableRange( $this->rootNode, 0, $this->rootNode, 0 );
		$fakeHeading = new HeadingItem( $range, true );

		$curComment = $fakeHeading;

		$treeWalker = new TreeWalker(
			$this->rootNode,
			NodeFilter::SHOW_ELEMENT | NodeFilter::SHOW_TEXT,
			[ self::class, 'acceptOnlyNodesAllowingComments' ]
		);
		while ( $node = $treeWalker->nextNode() ) {
			if ( $node instanceof DOMElement && preg_match( '/^h[1-6]$/i', $node->tagName ) ) {
				$range = new ImmutableRange( $node, 0, $node, $node->childNodes->length );
				$curComment = new HeadingItem( $range );
				$threadItems[] = $curComment;
			} elseif ( $node instanceof DOMText && ( $match = $this->findTimestamp( $node, $timestampRegex ) ) ) {
				$warnings = [];
				$foundSignature = $this->findSignature( $node, $curComment->getRange()->endContainer );
				$author = $foundSignature[1];
				$firstSigNode = end( $foundSignature[0] );
				$lastSigNode = $foundSignature[0][0];

				if ( !$author ) {
					// Ignore timestamps for which we couldn't find a signature. It's probably not a real
					// comment, but just a false match due to a copypasted timestamp.
					continue;
				}

				// Everything from the last comment up to here is the next comment
				$startNode = $this->nextInterestingLeafNode( $curComment->getRange()->endContainer );
				$offset = $lastSigNode === $node ?
					$match[0][1] + strlen( $match[0][0] ) :
					CommentUtils::childIndexOf( $lastSigNode ) + 1;
				$range = new ImmutableRange(
					$startNode->parentNode,
					CommentUtils::childIndexOf( $startNode ),
					$lastSigNode === $node ? $node : $lastSigNode->parentNode,
					$offset
				);
				$sigRange = new ImmutableRange(
					$firstSigNode->parentNode,
					CommentUtils::childIndexOf( $firstSigNode ),
					$lastSigNode === $node ? $node : $lastSigNode->parentNode,
					$offset
				);

				$startLevel = $this->getIndentLevel( $startNode ) + 1;
				$endLevel = $this->getIndentLevel( $node ) + 1;
				if ( $startLevel !== $endLevel ) {
					$warnings[] = 'Comment starts and ends with different indentation';
				}

				// Avoid generating multiple comments when there is more than one signature on a single "line".
				// Often this is done when someone edits their comment later and wants to add a note about that.
				// (Or when another person corrects a typo, or strikes out a comment, etc.) Multiple comments
				// within one paragraph/list-item result in a confusing double "Reply" button, and we also have
				// no way to indicate which one you're replying to (this might matter in the future for
				// notifications or something).
				if (
					$curComment instanceof CommentItem &&
					(
						CommentUtils::closestElement(
							$node, [ 'li', 'dd', 'p' ]
						) ?? $node->parentNode
					) ===
					(
						CommentUtils::closestElement(
							$curComment->getRange()->endContainer, [ 'li', 'dd', 'p' ]
						) ?? $curComment->getRange()->endContainer->parentNode
					)
				) {
					// Merge this with the previous comment. Use that comment's author and timestamp.
					$curComment->setRange(
						$curComment->getRange()->setEnd( $range->endContainer, $range->endOffset )
					);
					$curComment->addSignatureRange( $sigRange );
					$curComment->setLevel( min( min( $startLevel, $endLevel ), $curComment->getLevel() ) );

					continue;
				}

				$dateTime = $dfParser( $match );
				if ( isset( $dateTime->discussionToolsWarning ) ) {
					$warnings[] = $dateTime->discussionToolsWarning;
				}

				$curComment = new CommentItem(
					// Should this use the indent level of $startNode or $node?
					min( $startLevel, $endLevel ),
					$range,
					[ $sigRange ],
					// ISO 8601 date. Almost DateTimeInterface::RFC3339_EXTENDED, but ending with 'Z' instead
					// of '+00:00', like Date#toISOString in JavaScript.
					$dateTime->format( 'Y-m-d\TH:i:s.v\Z' ),
					$author
				);
				if ( $warnings ) {
					$curComment->addWarnings( $warnings );
				}
				$commentItems[] = $curComment;
				$threadItems[] = $curComment;
			}
		}

		// Insert the fake placeholder heading if there are any comments in the 0th section
		// (before the first real heading)
		if ( count( $threadItems ) && !( $threadItems[ 0 ] instanceof HeadingItem ) ) {
			array_unshift( $threadItems, $fakeHeading );
		}

		$this->commentItems = $commentItems;
		$this->threadItems = $threadItems;
	}

	/**
	 * Group discussion comments into threads and associate replies to original messages.
	 *
	 * Each thread must begin with a heading. Original messages in the thread are treated as replies to
	 * its heading. Other replies are associated based on the order and indentation level.
	 *
	 * Note that the objects in `comments` are extended in-place with the additional data.
	 *
	 * For example, for a MediaWiki discussion like this (we're dealing with HTML DOM here,
	 * the wikitext syntax is just for illustration):
	 *
	 *     == A ==
	 *     B. ~~~~
	 *     : C.
	 *     : C. ~~~~
	 *     :: D. ~~~~
	 *     ::: E. ~~~~
	 *     ::: F. ~~~~
	 *     : G. ~~~~
	 *     H. ~~~~
	 *     : I. ~~~~
	 *
	 * This function would return a structure like:
	 *
	 *     [
	 *       HeadingItem( { level: 0, range: (h2: A), replies: [
	 *         CommentItem( { level: 1, range: (p: B), replies: [
	 *           CommentItem( { level: 2, range: (li: C, li: C), replies: [
	 *             CommentItem( { level: 3, range: (li: D), replies: [
	 *               CommentItem( { level: 4, range: (li: E), replies: [] },
	 *               CommentItem( { level: 4, range: (li: F), replies: [] },
	 *             ] },
	 *           ] },
	 *           CommentItem( { level: 2, range: (li: G), replies: [] },
	 *         ] },
	 *         CommentItem( { level: 1, range: (p: H), replies: [
	 *           CommentItem( { level: 2, range: (li: I), replies: [] },
	 *         ] },
	 *       ] } )
	 *     ]
	 *
	 * @return HeadingItem[] Tree structure of comments, top-level items are the headings.
	 */
	public function getThreads() : array {
		if ( !$this->threads ) {
			$this->buildThreads();
		}
		return $this->threads;
	}

	private function buildThreads() : void {
		if ( !$this->threadItems ) {
			$this->buildThreadItems();
		}

		$threads = [];
		$replies = [];
		$this->commentsById = [];

		foreach ( $this->threadItems as &$threadItem ) {
			if ( $threadItem instanceof HeadingItem ) {
				// We don't need ids for section headings right now, but we might in the future
				// e.g. if we allow replying directly to sections (adding top-level comments)
				$id = null;
			} elseif ( $threadItem instanceof CommentItem ) {
				$id = ( $threadItem->getAuthor() ?? '' ) . '|' . $threadItem->getTimestamp();

				// If there would be multiple comments with the same ID (i.e. the user left multiple comments
				// in one edit, or within a minute), append sequential numbers
				$number = 0;
				while ( isset( $this->commentsById["$id|$number"] ) ) {
					$number++;
				}
				$id = "$id|$number";
				$this->commentsById[$id] = $threadItem;
				// This modifies the original objects in $comments!
				$threadItem->setId( $id );
			} else {
				throw new MWException( 'Unknown ThreadItem type' );
			}

			if ( count( $replies ) < $threadItem->getLevel() ) {
				// Someone skipped an indentation level (or several). Pretend that the previous reply
				// covers multiple indentation levels, so that following comments get connected to it.
				$threadItem->addWarning( 'Comment skips indentation level' );
				while ( count( $replies ) < $threadItem->getLevel() ) {
					// FIXME this will clone the reply, not just set a reference
					$replies[] = end( $replies );
				}
			}

			if ( $threadItem instanceof HeadingItem ) {
				// New root (thread)
				$threads[] = $threadItem;
			} elseif ( isset( $replies[ $threadItem->getLevel() - 1 ] ) ) {
				// Add as a reply to the closest less-nested comment
				$threadItem->setParent( $replies[ $threadItem->getLevel() - 1 ] );
				$threadItem->getParent()->addReply( $threadItem );
			} else {
				$threadItem->addWarning( 'Comment could not be connected to a thread' );
			}

			$replies[ $threadItem->getLevel() ] = $threadItem;
			// Cut off more deeply nested replies
			array_splice( $replies, $threadItem->getLevel() + 1 );
		}

		$this->threads = $threads;
	}
}