ASCII-compatible variable-width encoding of Unicode, using one to four bytes
UTF-8 is a variable-length character encoding used for electronic communication. Defined by the Unicode Standard, the name is derived from Unicode (or Universal Coded Character Set) Transformation Format – 8-bit.
UTF-8 is capable of encoding all 1,112,064[nb 1] valid character code points in Unicode using one to four one-byte (8-bit) code units. Code points with lower numerical values, which tend to occur more frequently, are encoded using fewer bytes. It was designed for backward compatibility with ASCII: the first 128 characters of Unicode, which correspond one-to-one with ASCII, are encoded using a single byte with the same binary value as ASCII, so that valid ASCII text is valid UTF-8-encoded Unicode as well.
UTF-8 was designed as a superior alternative to UTF-1, a proposed variable-length encoding with partial ASCII compatibility which lacked some features including self-synchronization and fully ASCII-compatible handling of characters such as slashes. Ken Thompson and Rob Pike produced the first implementation for the Plan 9 operating system in September 1992. This led to its adoption by X/Open as its specification for FSS-UTF, which would first be officially presented at USENIX in January 1993 and subsequently adopted by the Internet Engineering Task Force (IETF) in RFC 2277 (BCP 18) for future internet standards work, replacing Single Byte Character Sets such as Latin-1 in older RFCs.
UTF-8 is the dominant encoding for the World Wide Web (and internet technologies), accounting for 98.0% of all web pages, and up to 100.0% for many languages, as of 2022.
The official Internet Assigned Numbers Authority (IANA) code for the encoding is “UTF-8”. All letters are upper-case, and the name is hyphenated. This spelling is used in all the Unicode Consortium documents relating to the encoding. However, the name “utf-8” may be used by all standards conforming to the IANA list (which include CSS, HTML, XML, and HTTP headers), as the declaration is case-insensitive.
Other variants, such as those that omit the hyphen or replace it with a space, i.e. “utf8” or “UTF 8”, are not accepted as correct by the governing standards. Despite this, most web browsers can understand them, and so standards intended to describe existing practice (such as HTML5) may effectively require their recognition.
“UTF-8-BOM” and “UTF-8-NOBOM” are sometimes used for text files which contain or don’t contain a byte order mark (BOM), respectively. In Japan especially, UTF-8 encoding without a BOM is sometimes called “UTF-8N”.
In Windows UTF-8 is codepage 65001 (i.e. CP_UTF8 in source code).
In HP PCL, UTF-8 is called Symbol-ID “18N”.
UTF-8 encodes code points in one to four bytes, depending on the value of the code point. The x characters are replaced by the bits of the code point:
Code point ↔ UTF-8 conversionFirst code pointLast code pointByte 1Byte 2Byte 3Byte 4Code pointsU+0000U+007F0xxxxxxx128U+0080U+07FF110xxxxx10xxxxxx1920U+0800U+FFFF1110xxxx10xxxxxx10xxxxxx[nb 2]61440U+10000[nb 3]U+10FFFF11110xxx10xxxxxx10xxxxxx10xxxxxx The first 128 code points (ASCII) need one byte. The next 1,920 code points need two bytes to encode, which covers the remainder of almost all Latin-script alphabets, and also IPA extensions, Greek, Cyrillic, Coptic, Armenian, Hebrew, Arabic, Syriac, Thaana and N’Ko alphabets, as well as Combining Diacritical Marks. Three bytes are needed for the rest of the Basic Multilingual Plane, which contains virtually all code points in common use, including most Chinese, Japanese and Korean characters. Four bytes are needed for code points in the other planes of Unicode, which include less common CJK characters, various historic scripts, mathematical symbols, and emoji (pictographic symbols).
A “character” can take more than 4 bytes because it is made of more than one code point. For instance a national flag character takes 8 bytes since it’s “constructed from a pair of Unicode scalar values” both from outside the BMP.[nb 4]
Consider the encoding of the euro sign, €:
1. The Unicode code point for € is U+20AC.
2. As this code point lies between U+0800 and U+FFFF, this will take three bytes to encode.
3. Hexadecimal 20AC is binary . The two leading zeros are added because a three-byte encoding needs exactly sixteen bits from the code point.
4. Because the encoding will be three bytes long, its leading byte starts with three 1s, then a 0 (1110…)
5. The four most significant bits of the code point are stored in the remaining low order four bits of this byte ( ), leaving 12 bits of the code point yet to be encoded (… ).
6. All continuation bytes contain exactly six bits from the code point. So the next six bits of the code point are stored in the low order six bits of the next byte, and 10 is stored in the high order two bits to mark it as a continuation byte (so ).
7. Finally the last six bits of the code point are stored in the low order six bits of the final byte, and again 10 is stored in the high order two bits ( ).
The three bytes can be more concisely written in hexadecimal, as E2 82 AC.
The following table summarizes this conversion, as well as others with different lengths in UTF-8. The colors indicate how bits from the code point are distributed among the UTF-8 bytes. Additional bits added by the UTF-8 encoding process are shown in black.
Examples of UTF-8 encodingCharacterBinary code pointBinary UTF-8Hex UTF-8$U+ £U+00A C2 A3हU+ E0 A4 B9€U+20AC E2 82 AC한U+D55C ED 95 9C𐍈U+ F0 90 8D 88UTF-8’s use of six bits per byte to represent the actual characters being encoded means that octal notation (which uses 3-bit groups) can aid in the comparison of UTF-8 sequences with one another and in manual conversion.
Octal code point ↔ Octal UTF-8 conversionFirst code pointLast code pointCode pointByte 1Byte 2Byte 3Byte xxxxxx xxyy3xx2yy xyyzz34x2yy2zz xyyzz35x2yy2zz xyyzzww36x2yy2zz2wwWith octal notation, the arbitrary octal digits, marked with x, y, z or w in the table, will remain unchanged when converting to or from UTF-8.
Example: Á = U+00C1 = 0301 (in octal) is encoded as in UTF-8 (C3 81 in hex).Example: € = U+20AC = is encoded as in UTF-8 (E2 82 AC in hex).Codepage layout
The following table summarizes usage of UTF-8 code units (individual bytes or octets) in a code page format. The upper half is for bytes used only in single-byte codes, so it looks like a normal code page; the lower half is for continuation bytes and leading bytes and is explained further in the legend below.
7-bit (single-byte) code points. They must not be followed by a continuation byte.Continuation bytes. The tooltip shows in hex the value of the 6 bits they add.[nb 5]Leading bytes for a sequence of multiple bytes, must be followed by exactly N−1 continuation bytes. The tooltip shows the code point range and the Unicode blocks encoded by sequences starting with this byte.Leading bytes where not all arrangements of continuation bytes are valid. E0 and F0 could start overlong encodings. F4 can start code points greater than U+10FFFF. ED can start code points in the range U+D800–U+DFFF, which are invalid UTF-16 surrogate halves.Do not appear in a valid UTF-8 sequence. C0 and C1 could be used only for an “overlong” encoding of a 1-byte character. F5 to FD are leading bytes of 4-byte or longer sequences that can only encode code points larger than U+10FFFF. FE and FF were never assigned any meaning.Overlong encodings
In principle, it would be possible to inflate the number of bytes in an encoding by padding the code point with leading 0s. To encode the euro sign € from the above example in four bytes instead of three, it could be padded with leading 0s until it was 21 bits long – , and encoded as (or F AC in hexadecimal). This is called an overlong encoding.
The standard specifies that the correct encoding of a code point uses only the minimum number of bytes required to hold the significant bits of the code point. Longer encodings are called overlong and are not valid UTF-8 representations of the code point. This rule maintains a one-to-one correspondence between code points and their valid encodings, so that there is a unique valid encoding for each code point. This ensures that string comparisons and searches are well-defined.
Invalid sequences and error handling
Not all sequences of bytes are valid UTF-8. A UTF-8 decoder should be prepared for:
* invalid bytes
* an unexpected continuation byte
* a non-continuation byte before the end of the character
* the string ending before the end of the character (which can happen in simple string truncation)
* an overlong encoding
* a sequence that decodes to an invalid code point
Many of the first UTF-8 decoders would decode these, ignoring incorrect bits and accepting overlong results. Carefully crafted invalid UTF-8 could make them either skip or create ASCII characters such as NUL, slash, or quotes. Invalid UTF-8 has been used to bypass security validations in high-profile products including Microsoft’s IIS web server and Apache’s Tomcat servlet container. RFC 3629 states “Implementations of the decoding algorithm MUST protect against decoding invalid sequences.” The Unicode Standard requires decoders to “…treat any ill-formed code unit sequence as an error condition. This guarantees that it will neither interpret nor emit an ill-formed code unit sequence.”
Since RFC 3629 (November 2003), the high and low surrogate halves used by UTF-16 (U+D800 through U+DFFF) and code points not encodable by UTF-16 (those after U+10FFFF) are not legal Unicode values, and their UTF-8 encoding must be treated as an invalid byte sequence. Not decoding unpaired surrogate halves makes it impossible to store invalid UTF-16 (such as Windows filenames or UTF-16 that has been split between the surrogates) as UTF-8, while it is possible with WTF-8.
Some implementations of decoders throw exceptions on errors. This has the disadvantage that it can turn what would otherwise be harmless errors (such as a “no such file” error) into a denial of service. For instance early versions of Python 3.0 would exit immediately if the command line or environment variables contained invalid UTF-8. An alternative practice is to replace errors with a replacement character. Since Unicode 6 (October 2010), the standard (chapter 3) has recommended a “best practice” where the error ends as soon as a disallowed byte is encountered. In these decoders E1,A0,C0 is two errors (2 bytes in the first one). This means an error is no more than three bytes long and never contains the start of a valid character, and there are 21,952 different possible errors. The standard also recommends replacing each error with the replacement character “�” (U+FFFD).
Byte order mark
If the Unicode byte order mark (BOM, U+FEFF) character is at the start of a UTF-8 file, the first three bytes will be 0xEF, 0xBB, 0xBF.
The Unicode Standard neither requires nor recommends the use of the BOM for UTF-8, but warns that it may be encountered at the start of a file trans-coded from another encoding. While ASCII text encoded using UTF-8 is backward compatible with ASCII, this is not true when Unicode Standard recommendations are ignored and a BOM is added. A BOM can confuse software that isn’t prepared for it but can otherwise accept UTF-8, e.g. programming languages that permit non-ASCII bytes in string literals but not at the start of the file. Nevertheless, there was and still is software that always inserts a BOM when writing UTF-8, and refuses to correctly interpret UTF-8 unless the first character is a BOM (or the file only contains ASCII).
Declared character set for the 10 million most popular websites since Use of the main encodings on the web from 2001 to 2012 as recorded by Google, with UTF-8 overtaking all others in 2008 and over 60% of the web in 2012 (since then approaching 100%). The ASCII-only figure includes all web pages that only contain ASCII characters, regardless of the declared header. Other encodings of Unicode such as GB2312 are added to “others”.Many standards only support UTF-8, e.g. open JSON exchange requires it (without a byte order mark (BOM)). UTF-8 is also the recommendation from the WHATWG for HTML and DOM specifications, and the Internet Mail Consortium recommends that all e-mail programs be able to display and create mail using UTF-8. The World Wide Web Consortium recommends UTF-8 as the default encoding in XML and HTML (and not just using UTF-8, also declaring it in metadata), “even when all characters are in the ASCII range .. Using non-UTF-8 encodings can have unexpected results”.
Lots of software has the ability to read/write UTF-8, and for some functions (even in some Microsoft products) UTF-8 is the only option. In some cases (e.g. on Windows) it may though require the user to change options from the normal settings, or may require a BOM (byte order mark) as the first character to read the file. Examples of software supporting UTF-8 include Microsoft Word and Microsoft Excel. Most databases support UTF-8 (sometimes the only option as with some file formats), including Microsoft’s since SQL Server 2019, resulting in 35% speed increase, and “nearly 50% reduction in storage requirements.” Microsoft fully supports and recommends UTF-8 for its products such as Windows and Xbox.
UTF-8 has been the most common encoding for the World Wide Web since 2008. As of November 2022, UTF-8 accounts for on average 98.0% of all web pages (and 990 of the top 1,000 highest ranked web pages). Although many pages only use ASCII characters to display content, few websites now declare their encoding to only be ASCII instead of UTF-8. Over a third ( of the languages tracked have 100% UTF-8 use.
For local text files UTF-8 usage is lower, and a few legacy single-byte (and a few CJK multi-byte) encodings also remain in use to some degree. The primary cause is editors that do not display or write UTF-8 unless the first character in a file is a byte order mark (BOM), which requires software supporting UTF-8 to ignore it (or strip out), or simply to assume UTF-8 encoded files without it (or always to be used). To support editors that expect the BOM, they would need to add it (but most do not require it). There has been some improvement, Notepad on Windows defaults (catching up with most other editors) to writing UTF-8 without a BOM by default (a change since Windows 7), and some system files on Windows 11 require UTF-8 and don’t require the BOM and almost all files on macOS and Linux are required to be UTF-8 without a BOM. Java 18 changed to defaulting to reading and writing files as UTF-8, and in older versions (e.g. LTS versions) the NIO API only did so. Many other programming languages default to UTF-8 for I/O, including the current Ruby 3.0 and R 4.2.2. All currently supported versions of Python support UTF-8, even on Windows for I/O (but it’s opt-in there for the open() function), and plans to make UTF-8 I/O the default in 3.15 on Windows as for other platforms, and has already made changes to help programmers prepare for this.
The International Organization for Standardization (ISO) set out to compose a universal multi-byte character set in 1989. The draft ISO standard contained a non-required annex called UTF-1 that provided a byte stream encoding of its 32-bit code points. This encoding was not satisfactory on performance grounds, among other problems, and the biggest problem was probably that it did not have a clear separation between ASCII and non-ASCII: new UTF-1 tools would be backward compatible with ASCII-encoded text, but UTF-1-encoded text could confuse existing code expecting ASCII (or extended ASCII), because it could contain continuation bytes in the range 0x21–0x7E that meant something else in ASCII, e.g., 0x2F for ‘/’, the Unix path directory separator, and this example is reflected in the name and introductory text of its replacement. The table below was derived from a textual description in the annex.
Byte 1Byte 2Byte 3Byte 4Byte 51U+0000U+009F00–9F2U+00A0U+00FFA0A0–FF2U+0100U+4015A1–F521–7E, A0–FF3U+4016U+38E2DF6–FB21–7E, A0–FF21–7E, A0–FF5U+38E2EU+7FFFFFFFFC–FF21–7E, A0–FF21–7E, A0–FF21–7E, A0–FF21–7E, A0–FFIn July 1992, the X/Open committee XoJIG was looking for a better encoding. Dave Prosser of Unix System Laboratories submitted a proposal for one that had faster implementation characteristics and introduced the improvement that 7-bit ASCII characters would only represent themselves; all multi-byte sequences would include only bytes where the high bit was set. The name File System Safe UCS Transformation Format (FSS-UTF) and most of the text of this proposal were later preserved in the final specification.
FSS-UTF proposal (1992)Number
Byte 1Byte 2Byte 3Byte 4Byte 51U+0000U+007F0xxxxxxx2U+0080U+207F10xxxxxx1xxxxxxx3U+2080U+8207F110xxxxx1xxxxxxx1xxxxxxx4U+82080U+208207F1110xxxx1xxxxxxx1xxxxxxx1xxxxxxx5U+ U+7FFFFFFF11110xxx1xxxxxxx1xxxxxxx1xxxxxxx1xxxxxxxIn August 1992, this proposal was circulated by an IBM X/Open representative to interested parties. A modification by Ken Thompson of the Plan 9 operating system group at Bell Labs made it self-synchronizing, letting a reader start anywhere and immediately detect character boundaries, at the cost of being somewhat less bit-efficient than the previous proposal. It also abandoned the use of biases and instead added the rule that only the shortest possible encoding is allowed; the additional loss in compactness is relatively insignificant, but readers now have to look out for invalid encodings to avoid reliability and especially security issues. Thompson’s design was outlined on September 2, 1992, on a placemat in a New Jersey diner with Rob Pike. In the following days, Pike and Thompson implemented it and updated Plan 9 to use it throughout, and then communicated their success back to X/Open, which accepted it as the specification for FSS-UTF.
FSS-UTF (1992) / UTF-8 (1993)Number
Byte 1Byte 2Byte 3Byte 4Byte 5Byte 61U+0000U+007F0xxxxxxx2U+0080U+07FF110xxxxx10xxxxxx3U+0800U+FFFF1110xxxx10xxxxxx10xxxxxx4U+10000U+1FFFFF11110xxx10xxxxxx10xxxxxx10xxxxxx5U+200000U+3FFFFFF111110xx10xxxxxx10xxxxxx10xxxxxx10xxxxxx6U+ U+7FFFFFFF x10xxxxxx10xxxxxx10xxxxxx10xxxxxx10xxxxxxUTF-8 was first officially presented at the USENIX conference in San Diego, from January 25 to 29, 1993. The Internet Engineering Task Force adopted UTF-8 in its Policy on Character Sets and Languages in RFC 2277 (BCP 18) for future internet standards work, replacing Single Byte Character Sets such as Latin-1 in older RFCs.
In November 2003, UTF-8 was restricted by RFC3629 to match the constraints of the UTF-16 character encoding: explicitly prohibiting code points corresponding to the high and low surrogate characters removed more than 3% of the three-byte sequences, and ending at U+10FFFF removed more than 48% of the four-byte sequences and all five- and six-byte sequences.
There are several current definitions of UTF-8 in various standards documents:
* RFC3629 / STD 63 (2003), which establishes UTF-8 as a standard internet protocol element
* RFC5198 defines UTF-8 NFC for Network Interchange (2008)
* ISO/IEC 10646:2014 §9.1 (2014)
* The Unicode Standard, Version 14.0.0 (2021)
They supersede the definitions given in the following obsolete works:
* The Unicode Standard, Version 2.0, Appendix A (1996)
* ISO/IEC :1993 Amendment 2 / Annex R (1996)
* RFC2044 (1996)
* RFC2279 (1998)
* The Unicode Standard, Version 3.0, §2.3 (2000) plus Corrigendum #1 : UTF-8 Shortest Form (2000)
* Unicode Standard Annex #27: Unicode 3.1 (2001)
* The Unicode Standard, Version 5.0 (2006)
* The Unicode Standard, Version 6.0 (2010)
They are all the same in their general mechanics, with the main differences being on issues such as allowed range of code point values and safe handling of invalid input.
Comparison with other encodings
Some of the important features of this encoding are as follows:
* Backward compatibility: Backward compatibility with ASCII and the enormous amount of software designed to process ASCII-encoded text was the main driving force behind the design of UTF-8. In UTF-8, single bytes with values in the range of 0 to 127 map directly to Unicode code points in the ASCII range. Single bytes in this range represent characters, as they do in ASCII. Moreover, 7-bit bytes (bytes where the most significant bit is 0) never appear in a multi-byte sequence, and no valid multi-byte sequence decodes to an ASCII code-point. A sequence of 7-bit bytes is both valid ASCII and valid UTF-8, and under either interpretation represents the same sequence of characters. Therefore, the 7-bit bytes in a UTF-8 stream represent all and only the ASCII characters in the stream. Thus, many text processors, parsers, protocols, file formats, text display programs, etc., which use ASCII characters for formatting and control purposes, will continue to work as intended by treating the UTF-8 byte stream as a sequence of single-byte characters, without decoding the multi-byte sequences. ASCII characters on which the processing turns, such as punctuation, whitespace, and control characters will never be encoded as multi-byte sequences. It is therefore safe for such processors to simply ignore or pass-through the multi-byte sequences, without decoding them. For example, ASCII whitespace may be used to tokenize a UTF-8 stream into words; ASCII line-feeds may be used to split a UTF-8 stream into lines; and ASCII NUL characters can be used to split UTF-8-encoded data into null-terminated strings. Similarly, many format strings used by library functions like “printf” will correctly handle UTF-8-encoded input arguments.
* Fallback and auto-detection: Only a small subset of possible byte strings are a valid UTF-8 string: several bytes cannot appear; a byte with the high bit set cannot be alone; and further requirements mean that it is extremely unlikely that a readable text in any extended ASCII is valid UTF-8. Part of the popularity of UTF-8 is due to it providing a form of backward compatibility for these as well. A UTF-8 processor which erroneously receives extended ASCII as input can thus “auto-detect” this with very high reliability. A UTF-8 stream may simply contain errors, resulting in the auto-detection scheme producing false positives; but auto-detection is successful in the vast majority of cases, especially with longer texts, and is widely used. It also works to “fall back” or replace 8-bit bytes using the appropriate code-point for a legacy encoding when errors in the UTF-8 are detected, allowing recovery even if UTF-8 and legacy encoding is concatenated in the same file.
* Prefix code: The first byte indicates the number of bytes in the sequence. Reading from a stream can instantaneously decode each individual fully received sequence, without first having to wait for either the first byte of a next sequence or an end-of-stream indication. The length of multi-byte sequences is easily determined by humans as it is simply the number of high-order 1s in the leading byte. An incorrect character will not be decoded if a stream ends mid-sequence.
* Self-synchronization: The leading bytes and the continuation bytes do not share values (continuation bytes start with the bits 10 while single bytes start with 0 and longer lead bytes start with 11). This means a search will not accidentally find the sequence for one character starting in the middle of another character. It also means the start of a character can be found from a random position by backing up at most 3 bytes to find the leading byte. An incorrect character will not be decoded if a stream starts mid-sequence, and a shorter sequence will never appear inside a longer one.
* Sorting order: The chosen values of the leading bytes means that a list of UTF-8 strings can be sorted in code point order by sorting the corresponding byte sequences.
* UTF-8 can encode any Unicode character, avoiding the need to figure out and set a “code page” or otherwise indicate what character set is in use, and allowing output in multiple scripts at the same time. For many scripts there have been more than one single-byte encoding in usage, so even knowing the script was insufficient information to display it correctly.
* The bytes 0xFE and 0xFF do not appear, so a valid UTF-8 stream never matches the UTF-16 byte order mark and thus cannot be confused with it. The absence of 0xFF (0377) also eliminates the need to escape this byte in Telnet (and FTP control connection).
* UTF-8 encoded text is larger than specialized single-byte encodings except for plain ASCII characters. In the case of scripts which used 8-bit character sets with non-Latin characters encoded in the upper half (such as most Cyrillic and Greek alphabet code pages), characters in UTF-8 will be double the size. For some scripts, such as Thai and Devanagari (which is used by various South Asian languages), characters will triple in size. There are even examples where a single byte turns into a composite character in Unicode and is thus six times larger in UTF-8. This has caused objections in India and other countries.
* It is possible in UTF-8 (or any other multi-byte encoding) to split or truncate a string in the middle of a character. If the two pieces are not re-appended later before interpretation as characters, this can introduce an invalid sequence at both the end of the previous section and the start of the next, and some decoders will not preserve these bytes and result in data loss. Because UTF-8 is self-synchronizing this will however never introduce a different valid character, and it is also fairly easy to move the truncation point backward to the start of a character.
* If the code points are all the same size, measurements of a fixed number of them is easy. Due to ASCII-era documentation where “character” is used as a synonym for “byte” this is often considered important. However, by measuring string positions using bytes instead of “characters” most algorithms can be easily and efficiently adapted for UTF-8. Searching for a string within a long string can for example be done byte by byte; the self-synchronization property prevents false positives.
* UTF-8 can encode any Unicode character. Files in different scripts can be displayed correctly without having to choose the correct code page or font. For instance, Chinese and Arabic can be written in the same file without specialized markup or manual settings that specify an encoding.
* UTF-8 is self-synchronizing: character boundaries are easily identified by scanning for well-defined bit patterns in either direction. If bytes are lost due to error or corruption, one can always locate the next valid character and resume processing. If there is a need to shorten a string to fit a specified field, the previous valid character can easily be found. Many multi-byte encodings such as Shift JIS are much harder to resynchronize. This also means that byte-oriented string-searching algorithms can be used with UTF-8 (as a character is the same as a “word” made up of that many bytes), optimized versions of byte searches can be much faster due to hardware support and lookup tables that have only 256 entries. Self-synchronization does however require that bits be reserved for these markers in every byte, increasing the size.
* Efficient to encode using simple bitwise operations. UTF-8 does not require slower mathematical operations such as multiplication or division (unlike Shift JIS, GB 2312 and other encodings).
* UTF-8 will take more space than a multi-byte encoding designed for a specific script. East Asian legacy encodings generally used two bytes per character yet take three bytes per character in UTF-8.
* Byte encodings and UTF-8 are represented by byte arrays in programs, and often nothing needs to be done to a function when converting source code from a byte encoding to UTF-8. UTF-16 is represented by 16-bit word arrays, and converting to UTF-16 while maintaining compatibility with existing ASCII-based programs (such as was done with Windows) requires every API and data structure that takes a string to be duplicated, one version accepting byte strings and another version accepting UTF-16. If backward compatibility is not needed, all string handling still must be modified.
* Text encoded in UTF-8 will be smaller than the same text encoded in UTF-16 if there are more code points below U+0080 than in the range U+0800..U+FFFF. This is true for all modern European languages. It is often true even for languages like Chinese, due to the large number of spaces, newlines, digits, and HTML markup in typical files.
* Most communication (e.g. HTML and IP) and storage (e.g. for Unix) was designed for a stream of bytes. A UTF-16 string must use a pair of bytes for each code unit: * The order of those two bytes becomes an issue and must be specified in the UTF-16 protocol, such as with a byte order mark.
* If an odd number of bytes is missing from UTF-16, the whole rest of the string will be meaningless text. Any bytes missing from UTF-8 will still allow the text to be recovered accurately starting with the next character after the missing bytes.
The following implementations show slight differences from the UTF-8 specification. They are incompatible with the UTF-8 specification and may be rejected by conforming UTF-8 applications.
Unicode Technical Report #26 assigns the name CESU-8 to a nonstandard variant of UTF-8, in which Unicode characters in supplementary planes are encoded using six bytes, rather than the four bytes required by UTF-8. CESU-8 encoding treats each half of a four-byte UTF-16 surrogate pair as a two-byte UCS-2 character, yielding two three-byte UTF-8 characters, which together represent the original supplementary character. Unicode characters within the Basic Multilingual Plane appear as they would normally in UTF-8. The Report was written to acknowledge and formalize the existence of data encoded as CESU-8, despite the Unicode Consortium discouraging its use, and notes that a possible intentional reason for CESU-8 encoding is preservation of UTF-16 binary collation.
CESU-8 encoding can result from converting UTF-16 data with supplementary characters to UTF-8, using conversion methods that assume UCS-2 data, meaning they are unaware of four-byte UTF-16 supplementary characters. It is primarily an issue on operating systems which extensively use UTF-16 internally, such as Microsoft Windows.
In Oracle Database, the UTF8 character set uses CESU-8 encoding, and is deprecated. The AL32UTF8 character set uses standards-compliant UTF-8 encoding, and is preferred.
CESU-8 is prohibited for use in HTML5 documents.
In MySQL, the utf8mb3 character set is defined to be UTF-8 encoded data with a maximum of three bytes per character, meaning only Unicode characters in the Basic Multilingual Plane (i.e. from UCS-2) are supported. Unicode characters in supplementary planes are explicitly not supported. utf8mb3 is deprecated in favor of the utf8mb4 character set, which uses standards-compliant UTF-8 encoding. utf8 is an alias for utf8mb3, but is intended to become an alias to utf8mb4 in a future release of MySQL. It is possible, though unsupported, to store CESU-8 encoded data in utf8mb3, by handling UTF-16 data with supplementary characters as though it is UCS-2.
Modified UTF-8 (MUTF-8) originated in the Java programming language. In Modified UTF-8, the null character (U+0000) uses the two-byte overlong encoding (hexadecimal C0 80), instead of (hexadecimal 00). Modified UTF-8 strings never contain any actual null bytes but can contain all Unicode code points including U+0000, which allows such strings (with a null byte appended) to be processed by traditional null-terminated string functions. All known Modified UTF-8 implementations also treat the surrogate pairs as in CESU-8.
In normal usage, the language supports standard UTF-8 when reading and writing strings through InputStreamReader and OutputStreamWriter (if it is the platform’s default character set or as requested by the program). However it uses Modified UTF-8 for object serialization among other applications of DataInput and DataOutput, for the Java Native Interface, and for embedding constant strings in class files.
The dex format defined by Dalvik also uses the same modified UTF-8 to represent string values. Tcl also uses the same modified UTF-8 as Java for internal representation of Unicode data, but uses strict CESU-8 for external data.
In WTF-8 (Wobbly Transformation Format, 8-bit) unpaired surrogate halves (U+D800 through U+DFFF) are allowed. This is necessary to store possibly-invalid UTF-16, such as Windows filenames. Many systems that deal with UTF-8 work this way without considering it a different encoding, as it is simpler.
(The term “WTF-8” has also been used humorously to refer to erroneously doubly-encoded UTF-8 sometimes with the implication that CP1252 bytes are the only ones encoded.)
Version 3 of the Python programming language treats each byte of an invalid UTF-8 bytestream as an error (see also changes with new UTF-8 mode in Python 3.7); this gives 128 different possible errors. Extensions have been created to allow any byte sequence that is assumed to be UTF-8 to be losslessly transformed to UTF-16 or UTF-32, by translating the 128 possible error bytes to reserved code points, and transforming those code points back to error bytes to output UTF-8. The most common approach is to translate the codes to U+DC80…U+DCFF which are low (trailing) surrogate values and thus “invalid” UTF-16, as used by Python’s PEP 383 (or “surrogateescape”) approach. Another encoding called MirBSD OPTU-8/16 converts them to U+EF80…U+EFFF in a Private Use Area. In either approach, the byte value is encoded in the low eight bits of the output code point.
These encodings are very useful because they avoid the need to deal with “invalid” byte strings until much later, if at all, and allow “text” and “data” byte arrays to be the same object. If a program wants to use UTF-16 internally these are required to preserve and use filenames that can use invalid UTF-8; as the Windows filesystem API uses UTF-16, the need to support invalid UTF-8 is less there.
For the encoding to be reversible, the standard UTF-8 encodings of the code points used for erroneous bytes must be considered invalid. This makes the encoding incompatible with WTF-8 or CESU-8 (though only for 128 code points). When re-encoding it is necessary to be careful of sequences of error code points which convert back to valid UTF-8, which may be used by malicious software to get unexpected characters in the output, though this cannot produce ASCII characters so it is considered comparatively safe, since malicious sequences (such as cross-site scripting) usually rely on ASCII characters.
1. ^ 17 planes times 216 code points per plane, minus 211 technically-invalid surrogates.
2. ^ 17 planes times 216 code points per plane, minus 211 technically-invalid surrogates.
3. ^ There are enough x bits to encode up to 0x1FFFFF, but the current RFC 3629 §3 limits UTF-8 encoding to code point U+10FFFF, to match the limits of UTF-16. The obsolete RFC 2279 allowed UTF-8 encoding up to (then legal) code point U+7FFFFFF.
4. ^ Some complex emoji characters can take even more than this; the transgender flag emoji (🏳️⚧️), which consists of the five-codepoint sequence U+1F3F3 U+FE0F U+200D U+26A7 U+FE0F, requires sixteen bytes to encode, while that for the flag of Scotland (🏴) requires a total of twenty-eight bytes for the seven-codepoint sequence U+1F3F4 U+E0067 U+E0062 U+E0073 U+E0063 U+E0074 U+E007F.
5. ^ For example, cell 9D says +1D. The hexadecimal number 9D in binary is , and since the 2 highest bits (10) are reserved for marking this as a continuation byte, the remaining 6 bits (011101) have a hexadecimal value of 1D.