class AbstractHeader
An abstract base MIME Header.
@author Chris Corbyn
Hierarchy
- class \Symfony\Component\Mime\Header\AbstractHeader implements \Symfony\Component\Mime\Header\HeaderInterface
Expanded class hierarchy of AbstractHeader
File
-
vendor/
symfony/ mime/ Header/ AbstractHeader.php, line 21
Namespace
Symfony\Component\Mime\HeaderView source
abstract class AbstractHeader implements HeaderInterface {
public const PHRASE_PATTERN = '(?:(?:(?:(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))*(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))|(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])))?[a-zA-Z0-9!#\\$%&\'\\*\\+\\-\\/=\\?\\^_`\\{\\}\\|~]+(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))*(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))|(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])))?)|(?:(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))*(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))|(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])))?"((?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21\\x23-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?"(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))*(?:(?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?(\\((?:(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])|(?:(?:[\\x01-\\x08\\x0B\\x0C\\x0E-\\x19\\x7F]|[\\x21-\\x27\\x2A-\\x5B\\x5D-\\x7E])|(?:\\[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F])|(?1)))*(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])?\\)))|(?:(?:[ \\t]*(?:\\r\\n))?[ \\t])))?))+?)';
private static QpMimeHeaderEncoder $encoder;
private string $name;
private int $lineLength = 76;
private ?string $lang = null;
private string $charset = 'utf-8';
public function __construct(string $name) {
$this->name = $name;
}
public function setCharset(string $charset) : void {
$this->charset = $charset;
}
public function getCharset() : ?string {
return $this->charset;
}
/**
* Set the language used in this Header.
*
* For example, for US English, 'en-us'.
*/
public function setLanguage(string $lang) : void {
$this->lang = $lang;
}
public function getLanguage() : ?string {
return $this->lang;
}
public function getName() : string {
return $this->name;
}
public function setMaxLineLength(int $lineLength) : void {
$this->lineLength = $lineLength;
}
public function getMaxLineLength() : int {
return $this->lineLength;
}
public function toString() : string {
return $this->tokensToString($this->toTokens());
}
/**
* Produces a compliant, formatted RFC 2822 'phrase' based on the string given.
*
* @param string $string as displayed
* @param bool $shorten the first line to make remove for header name
*/
protected function createPhrase(HeaderInterface $header, string $string, string $charset, bool $shorten = false) : string {
// Treat token as exactly what was given
$phraseStr = $string;
// If it's not valid
if (!preg_match('/^' . self::PHRASE_PATTERN . '$/D', $phraseStr)) {
// .. but it is just ascii text, try escaping some characters
// and make it a quoted-string
if (preg_match('/^[\\x00-\\x08\\x0B\\x0C\\x0E-\\x7F]*$/D', $phraseStr)) {
foreach ([
'\\',
'"',
] as $char) {
$phraseStr = str_replace($char, '\\' . $char, $phraseStr);
}
$phraseStr = '"' . $phraseStr . '"';
}
else {
// ... otherwise it needs encoding
// Determine space remaining on line if first line
if ($shorten) {
$usedLength = \strlen($header->getName() . ': ');
}
else {
$usedLength = 0;
}
$phraseStr = $this->encodeWords($header, $string, $usedLength);
}
}
elseif (str_contains($phraseStr, '(')) {
foreach ([
'\\',
'"',
] as $char) {
$phraseStr = str_replace($char, '\\' . $char, $phraseStr);
}
$phraseStr = '"' . $phraseStr . '"';
}
return $phraseStr;
}
/**
* Encode needed word tokens within a string of input.
*/
protected function encodeWords(HeaderInterface $header, string $input, int $usedLength = -1) : string {
$value = '';
$tokens = $this->getEncodableWordTokens($input);
foreach ($tokens as $token) {
// See RFC 2822, Sect 2.2 (really 2.2 ??)
if ($this->tokenNeedsEncoding($token)) {
// Don't encode starting WSP
$firstChar = substr($token, 0, 1);
switch ($firstChar) {
case ' ':
case "\t":
$value .= $firstChar;
$token = substr($token, 1);
}
if (-1 == $usedLength) {
$usedLength = \strlen($header->getName() . ': ') + \strlen($value);
}
$value .= $this->getTokenAsEncodedWord($token, $usedLength);
}
else {
$value .= $token;
}
}
return $value;
}
protected function tokenNeedsEncoding(string $token) : bool {
return (bool) preg_match('~[\\x00-\\x08\\x10-\\x19\\x7F-\\xFF\\r\\n]~', $token);
}
/**
* Splits a string into tokens in blocks of words which can be encoded quickly.
*
* @return string[]
*/
protected function getEncodableWordTokens(string $string) : array {
$tokens = [];
$encodedToken = '';
// Split at all whitespace boundaries
foreach (preg_split('~(?=[\\t ])~', $string) as $token) {
if ($this->tokenNeedsEncoding($token)) {
$encodedToken .= $token;
}
else {
if ('' !== $encodedToken) {
$tokens[] = $encodedToken;
$encodedToken = '';
}
$tokens[] = $token;
}
}
if ('' !== $encodedToken) {
$tokens[] = $encodedToken;
}
foreach ($tokens as $i => $token) {
// whitespace(s) between 2 encoded tokens
if (0 < $i && isset($tokens[$i + 1]) && preg_match('~^[\\t ]+$~', $token) && $this->tokenNeedsEncoding($tokens[$i - 1]) && $this->tokenNeedsEncoding($tokens[$i + 1])) {
$tokens[$i - 1] .= $token . $tokens[$i + 1];
array_splice($tokens, $i, 2);
}
}
return $tokens;
}
/**
* Get a token as an encoded word for safe insertion into headers.
*/
protected function getTokenAsEncodedWord(string $token, int $firstLineOffset = 0) : string {
self::$encoder ??= new QpMimeHeaderEncoder();
// Adjust $firstLineOffset to account for space needed for syntax
$charsetDecl = $this->charset;
if (null !== $this->lang) {
$charsetDecl .= '*' . $this->lang;
}
$encodingWrapperLength = \strlen('=?' . $charsetDecl . '?' . self::$encoder->getName() . '??=');
if ($firstLineOffset >= 75) {
// Does this logic need to be here?
$firstLineOffset = 0;
}
$encodedTextLines = explode("\r\n", self::$encoder->encodeString($token, $this->charset, $firstLineOffset, 75 - $encodingWrapperLength));
if ('iso-2022-jp' !== strtolower($this->charset)) {
// special encoding for iso-2022-jp using mb_encode_mimeheader
foreach ($encodedTextLines as $lineNum => $line) {
$encodedTextLines[$lineNum] = '=?' . $charsetDecl . '?' . self::$encoder->getName() . '?' . $line . '?=';
}
}
return implode("\r\n ", $encodedTextLines);
}
/**
* Generates tokens from the given string which include CRLF as individual tokens.
*
* @return string[]
*/
protected function generateTokenLines(string $token) : array {
return preg_split('~(\\r\\n)~', $token, -1, \PREG_SPLIT_DELIM_CAPTURE);
}
/**
* Generate a list of all tokens in the final header.
*/
protected function toTokens(?string $string = null) : array {
$string ??= $this->getBodyAsString();
$tokens = [];
// Generate atoms; split at all invisible boundaries followed by WSP
foreach (preg_split('~(?=[ \\t])~', $string) as $token) {
$newTokens = $this->generateTokenLines($token);
foreach ($newTokens as $newToken) {
$tokens[] = $newToken;
}
}
return $tokens;
}
/**
* Takes an array of tokens which appear in the header and turns them into
* an RFC 2822 compliant string, adding FWSP where needed.
*
* @param string[] $tokens
*/
private function tokensToString(array $tokens) : string {
$lineCount = 0;
$headerLines = [];
$headerLines[] = $this->name . ': ';
$currentLine =& $headerLines[$lineCount++];
// Build all tokens back into compliant header
foreach ($tokens as $i => $token) {
// Line longer than specified maximum or token was just a new line
if ("\r\n" === $token || $i > 0 && \strlen($currentLine . $token) > $this->lineLength && '' !== $currentLine) {
$headerLines[] = '';
$currentLine =& $headerLines[$lineCount++];
}
// Append token to the line
if ("\r\n" !== $token) {
$currentLine .= $token;
}
}
// Implode with FWS (RFC 2822, 2.2.3)
return implode("\r\n", $headerLines);
}
}
Members
Title Sort descending | Modifiers | Object type | Summary | Overriden Title | Overrides |
---|---|---|---|---|---|
AbstractHeader::$charset | private | property | |||
AbstractHeader::$encoder | private static | property | 1 | ||
AbstractHeader::$lang | private | property | |||
AbstractHeader::$lineLength | private | property | |||
AbstractHeader::$name | private | property | |||
AbstractHeader::createPhrase | protected | function | Produces a compliant, formatted RFC 2822 'phrase' based on the string given. | ||
AbstractHeader::encodeWords | protected | function | Encode needed word tokens within a string of input. | ||
AbstractHeader::generateTokenLines | protected | function | Generates tokens from the given string which include CRLF as individual tokens. | ||
AbstractHeader::getCharset | public | function | Overrides HeaderInterface::getCharset | ||
AbstractHeader::getEncodableWordTokens | protected | function | Splits a string into tokens in blocks of words which can be encoded quickly. | ||
AbstractHeader::getLanguage | public | function | Overrides HeaderInterface::getLanguage | ||
AbstractHeader::getMaxLineLength | public | function | Overrides HeaderInterface::getMaxLineLength | ||
AbstractHeader::getName | public | function | Overrides HeaderInterface::getName | ||
AbstractHeader::getTokenAsEncodedWord | protected | function | Get a token as an encoded word for safe insertion into headers. | ||
AbstractHeader::PHRASE_PATTERN | public | constant | |||
AbstractHeader::setCharset | public | function | Overrides HeaderInterface::setCharset | ||
AbstractHeader::setLanguage | public | function | Set the language used in this Header. | Overrides HeaderInterface::setLanguage | |
AbstractHeader::setMaxLineLength | public | function | Overrides HeaderInterface::setMaxLineLength | ||
AbstractHeader::tokenNeedsEncoding | protected | function | 2 | ||
AbstractHeader::tokensToString | private | function | Takes an array of tokens which appear in the header and turns them into an RFC 2822 compliant string, adding FWSP where needed. |
||
AbstractHeader::toString | public | function | Gets this Header rendered as a compliant string. | Overrides HeaderInterface::toString | |
AbstractHeader::toTokens | protected | function | Generate a list of all tokens in the final header. | 1 | |
AbstractHeader::__construct | public | function | 6 | ||
HeaderInterface::getBody | public | function | Gets the body. | 6 | |
HeaderInterface::getBodyAsString | public | function | Gets the header's body, prepared for folding into a final header value. | 6 | |
HeaderInterface::setBody | public | function | Sets the body. | 6 |