I know that / is illegal in Linux, and the following are illegal in Windows
(I think) * . " / \ [ ] : ; | ,
What else am I missing?
I need a comprehensive guide, however, and one that takes into account
double-byte characters. Linking to outside resources is fine with me.
I need to first create a directory on the filesystem using a name that may
contain forbidden characters, so I plan to replace those characters with
underscores. I then need to write this directory and its contents to a zip file
(using Java), so any additional advice concerning the names of zip directories
would be appreciated.
The forbidden printable ASCII characters are:
Linux/Unix:
/ (forward slash)
Windows:
< (less than)
> (greater than)
: (colon - sometimes works, but is actually NTFS Alternate Data Streams)
" (double quote)
/ (forward slash)
\ (backslash)
| (vertical bar or pipe)
? (question mark)
* (asterisk)
Non-printable characters
If your data comes from a source that would permit non-printable characters then there is more to check for.
Linux/Unix:
0 (NULL byte)
Windows:
0-31 (ASCII control characters)
Note: While it is legal under Linux/Unix file systems to create files with control characters in the filename, it might be a nightmare for the users to deal with such files.
Reserved file names
The following filenames are reserved:
Windows:
CON, PRN, AUX, NUL
COM1, COM2, COM3, COM4, COM5, COM6, COM7, COM8, COM9
LPT1, LPT2, LPT3, LPT4, LPT5, LPT6, LPT7, LPT8, LPT9
(both on their own and with arbitrary file extensions, e.g. LPT1.txt).
Other rules
Windows:
Filenames cannot end in a space or dot.
macOS:
You didn't ask for it, but just in case: Colon : and forward slash / depending on context are not permitted (e.g. Finder supports slashes, terminal supports colons). (More details)
A “comprehensive guide” of forbidden filename characters is not going to work on Windows because it reserves filenames as well as characters. Yes, characters like
* " ? and others are forbidden, but there are a infinite number of names composed only of valid characters that are forbidden. For example, spaces and dots are valid filename characters, but names composed only of those characters are forbidden.
Windows does not distinguish between upper-case and lower-case characters, so you cannot create a folder named A if one named a already exists. Worse, seemingly-allowed names like PRN and CON, and many others, are reserved and not allowed. Windows also has several length restrictions; a filename valid in one folder may become invalid if moved to another folder. The rules for
naming files and folders
are on the Microsoft docs.
You cannot, in general, use user-generated text to create Windows directory names. If you want to allow users to name anything they want, you have to create safe names like A, AB, A2 et al., store user-generated names and their path equivalents in an application data file, and perform path mapping in your application.
If you absolutely must allow user-generated folder names, the only way to tell if they are invalid is to catch exceptions and assume the name is invalid. Even that is fraught with peril, as the exceptions thrown for denied access, offline drives, and out of drive space overlap with those that can be thrown for invalid names. You are opening up one huge can of hurt.
Under Linux and other Unix-related systems, there were traditionally only two characters that could not appear in the name of a file or directory, and those are NUL '\0' and slash '/'. The slash, of course, can appear in a pathname, separating directory components.
Rumour1 has it that Steven Bourne (of 'shell' fame) had a directory containing 254 files, one for every single letter (character code) that can appear in a file name (excluding /, '\0'; the name . was the current directory, of course). It was used to test the Bourne shell and routinely wrought havoc on unwary programs such as backup programs.
Other people have covered the rules for Windows filenames, with links to Microsoft and Wikipedia on the topic.
Note that MacOS X has a case-insensitive file system. Current versions of it appear to allow colon : in file names, though historically that was not necessarily always the case:
$ echo a:b > a:b
$ ls -l a:b
-rw-r--r-- 1 jonathanleffler staff 4 Nov 12 07:38 a:b
$
However, at least with macOS Big Sur 11.7, the file system does not allow file names that are not valid UTF-8 strings. That means the file name cannot consist of the bytes that are always invalid in UTF-8 (0xC0, 0xC1, 0xF5-0xFF), and you can't use the continuation bytes 0x80..0xBF as the only byte in a file name. The error given is 92 Illegal byte sequence.
POSIX defines a Portable Filename Character Set consisting of:
A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
a b c d e f g h i j k l m n o p q r s t u v w x y z
0 1 2 3 4 5 6 7 8 9 . _ -
Sticking with names formed solely from those characters avoids most of the problems, though Windows still adds some complications.
1 It was Kernighan & Pike in ['The Practice of Programming'](http://www.cs.princeton.edu/~bwk/tpop.webpage/) who said as much in Chapter 6, Testing, §6.5 Stress Tests:
When Steve Bourne was writing his Unix shell (which came to be known as the Bourne shell), he made a directory of 254 files with one-character names, one for each byte value except '\0' and slash, the two characters that cannot appear in Unix file names. He used that directory for all manner of tests of pattern-matching and tokenization. (The test directory was of course created by a program.) For years afterwards, that directory was the bane of file-tree-walking programs; it tested them to destruction.
Note that the directory must have contained entries . and .., so it was arguably 253 files (and 2 directories), or 255 name entries, rather than 254 files. This doesn't affect the effectiveness of the anecdote, or the careful testing it describes.
TPOP was previously at
http://plan9.bell-labs.com/cm/cs/tpop and
http://cm.bell-labs.com/cm/cs/tpop but both are now (2021-11-12) broken.
See also Wikipedia on TPOP.
Instead of creating a blacklist of characters, you could use a whitelist. All things considered, the range of characters that make sense in a file or directory name context is quite short, and unless you have some very specific naming requirements your users will not hold it against your application if they cannot use the whole ASCII table.
It does not solve the problem of reserved names in the target file system, but with a whitelist it is easier to mitigate the risks at the source.
In that spirit, this is a range of characters that can be considered safe:
Letters (a-z A-Z) - Unicode characters as well, if needed
Digits (0-9)
Underscore (_)
Hyphen (-)
Space
Dot (.)
And any additional safe characters you wish to allow. Beyond this, you just have to enforce some additional rules regarding spaces and dots. This is usually sufficient:
Name must contain at least one letter or number (to avoid only dots/spaces)
Name must start with a letter or number (to avoid leading dots/spaces)
Name may not end with a dot or space (simply trim those if present, like Explorer does)
This already allows quite complex and nonsensical names. For example, these names would be possible with these rules, and be valid file names in Windows/Linux:
A...........ext
B -.- .ext
In essence, even with so few whitelisted characters you should still decide what actually makes sense, and validate/adjust the name accordingly. In one of my applications, I used the same rules as above but stripped any duplicate dots and spaces.
The easy way to get Windows to tell you the answer is to attempt to rename a file via Explorer and type in a backslash, /, for the new name. Windows will popup a message box telling you the list of illegal characters.
A filename cannot contain any of the following characters:
\ / : * ? " < > |
Microsoft Docs - Naming Files, Paths, and Namespaces - Naming Conventions
Well, if only for research purposes, then your best bet is to look at this Wikipedia entry on Filenames.
If you want to write a portable function to validate user input and create filenames based on that, the short answer is don't. Take a look at a portable module like Perl's File::Spec to have a glimpse to all the hops needed to accomplish such a "simple" task.
Discussing different possible approaches
Difficulties with defining, what's legal and not were already adressed and whitelists were suggested. But not only Windows, but also many unixoid OSes support more-than-8-bit characters such as Unicode. You could here also talk about encodings such as UTF-8. You can consider Jonathan Leffler's comment, where he gives info about modern Linux and describes details for MacOS. Wikipedia states, that (for example) the
modifier letter colon [(See 7. below) is] sometimes used in Windows filenames as it is identical to the colon in the Segoe UI font used for filenames. The [inherited ASCII] colon itself is not permitted.
Therefore, I want to present a much more liberal approach using Unicode Homoglyph characters to replace the "illegal" ones. I found the result in my comparable use-case by far more readable and it's only limited by the used font, which is very broad, 3903 characters for Windows default. Plus you can even restore the original content from the replacements.
Possible choices and research notes
To keep things organized, I will always give the character, it's name and the hexadecimal number representation. The latter is is not case sensitive and leading zeroes can be added or ommitted freely, so for example U+002A and u+2a are equivalent. If available, I'll try to point to more info or alternatives - feel free to show me more or better ones.
Instead of * (U+2A * ASTERISK), you can use one of the many listed, for example U+2217 ∗ (ASTERISK OPERATOR) or the Full Width Asterisk U+FF0A ＊. u+20f0 ⃰ combining asterisk above from combining diacritical marks for symbols might also be a valid choice. You can read 4. for more info about the combining characters.
Instead of . (U+2E . full stop), one of these could be a good option, for example ⋅ U+22C5 dot operator.
Instead of " (U+22 " quotation mark), you can use “ U+201C english leftdoublequotemark, more alternatives see here. I also included some of the good suggestions of Wally Brockway's answer, in this case u+2036 ‶ reversed double prime and u+2033 ″ double prime - I will from now on denote ideas from that source by ¹³.
Instead of / (U+2F / SOLIDUS), you can use ∕ DIVISION SLASH U+2215 (others here), ̸ U+0338 COMBINING LONG SOLIDUS OVERLAY, ̷ COMBINING SHORT SOLIDUS OVERLAY U+0337 or u+2044 ⁄ fraction slash¹³. Be aware about spacing for some characters, including the combining or overlay ones, as they have no width and can produce something like -> ̸th̷is which is ̸th̷is. With added spaces you get -> ̸ th ̷ is, which is ̸ th ̷ is. The second one (COMBINING SHORT SOLIDUS OVERLAY) looks bad in the stackoverflow-font.
Instead of \ (U+5C Reverse solidus), you can use ⧵ U+29F5 Reverse solidus operator (more) or u+20E5 ⃥ combining reverse solidus overlay¹³.
To replace [ (U+5B [ Left square bracket) and ] (U+005D ] Right square bracket), you can use for example U+FF3B［ FULLWIDTH LEFT SQUARE BRACKET and U+FF3D ］FULLWIDTH RIGHT SQUARE BRACKET (from here, more possibilities here).
Instead of : (u+3a : colon), you can use U+2236 ∶ RATIO (for mathematical usage) or U+A789 ꞉ MODIFIER LETTER COLON, (see colon (letter), sometimes used in Windows filenames as it is identical to the colon in the Segoe UI font used for filenames. The colon itself is not permitted ... source and more replacements see here). Another alternative is this one: u+1361 ፡ ethiopic wordspace¹³.
Instead of ; (u+3b ; semicolon), you can use U+037E ; GREEK QUESTION MARK (see here).
For | (u+7c | vertical line), there are some good substitutes such as: U+2223 ∣ DIVIDES, U+0964 । DEVANAGARI DANDA, U+01C0 ǀ LATIN LETTER DENTAL CLICK (the last ones from Wikipedia) or U+2D4F ⵏ Tifinagh Letter Yan. Also the box drawing characters contain various other options.
Instead of , (, U+002C COMMA), you can use for example ‚ U+201A SINGLE LOW-9 QUOTATION MARK (see here).
For ? (U+003F ? QUESTION MARK), these are good candidates: U+FF1F ？ FULLWIDTH QUESTION MARK or U+FE56 ﹖ SMALL QUESTION MARK (from here and here). There are also two more from the Dingbats Block (search for "question") and the u+203d ‽ interrobang¹³.
While my machine seems to accept it unchanged, I still want to include > (u+3e greater-than sign) and < (u+3c less-than sign) for the sake of completeness. The best replacement here is probably also from the quotation block, such as u+203a › single right-pointing angle quotation mark and u+2039 ‹ single left-pointing angle quotation mark respectively. The tifinagh block only contains ⵦ (u+2D66)¹³ to replace <. The last notion is ⋖ less-than with dot u+22D6 and ⋗ greater-than with dot u+22D7.
For additional ideas, you can also look for example into this block. You still want more ideas? You can try to draw your desired character and look at the suggestions here.
How do you type these characters
Say you want to type ⵏ (Tifinagh Letter Yan). To get all of its information, you can always search for this character (ⵏ) on a suited platform such as this Unicode Lookup (please add 0x when you search for hex) or that Unicode Table (that only allows to search for the name, in this case "Tifinagh Letter Yan"). You should obtain its Unicode number U+2D4F and the HTML-code ⵏ (note that 2D4F is hexadecimal for 11599). With this knowledge, you have several options to produce these special characters including the use of
code points to unicode converter or again the Unicode Lookup to reversely convert the numerical representation into the unicode character (remember to set the code point base below to decimal or hexadecimal respectively)
a one-liner makro in Autohotkey: :?*:altpipe::{U+2D4F} to type ⵏ instead of the string altpipe - this is the way I input those special characters, my Autohotkey script can be shared if there is common interest
Alt Characters or alt-codes by pressing and holding alt, followed by the decimal number for the desired character (more info for example here, look at a table here or there). For the example, that would be Alt+11599. Be aware, that many programs do not fully support this windows feature for all of unicode (as of time writing). Microsoft Office is an exception where it usually works, some other OSes provide similar functionality. Typing these chars with Alt-combinations into MS Word is also the way Wally Brockway suggests in his answer¹³ that was already mentionted - if you don't want to transfer all the hexadecimal values to the decimal asc, you can find some of them there¹³.
in MS Office, you can also use ALT + X as described in this MS article to produce the chars
if you rarely need it, you can of course still just copy-paste the special character of your choice instead of typing it
For Windows you can check it using PowerShell
$PathInvalidChars = [System.IO.Path]::GetInvalidPathChars() #36 chars
To display UTF-8 codes you can convert
$enc = [system.Text.Encoding]::UTF8
$PathInvalidChars | foreach { $enc.GetBytes($_) }
$FileNameInvalidChars = [System.IO.Path]::GetInvalidFileNameChars() #41 chars
$FileOnlyInvalidChars = #(':', '*', '?', '\', '/') #5 chars - as a difference
For anyone looking for a regex:
const BLACKLIST = /[<>:"\/\\|?*]/g;
In Windows 10 (2019), the following characters are forbidden by an error when you try to type them:
A file name can't contain any of the following characters:
\ / : * ? " < > |
Here's a c# implementation for windows based on Christopher Oezbek's answer
It was made more complex by the containsFolder boolean, but hopefully covers everything
/// <summary>
/// This will replace invalid chars with underscores, there are also some reserved words that it adds underscore to
/// </summary>
/// <remarks>
/// https://stackoverflow.com/questions/1976007/what-characters-are-forbidden-in-windows-and-linux-directory-names
/// </remarks>
/// <param name="containsFolder">Pass in true if filename represents a folder\file (passing true will allow slash)</param>
public static string EscapeFilename_Windows(string filename, bool containsFolder = false)
{
StringBuilder builder = new StringBuilder(filename.Length + 12);
int index = 0;
// Allow colon if it's part of the drive letter
if (containsFolder)
{
Match match = Regex.Match(filename, #"^\s*[A-Z]:\\", RegexOptions.IgnoreCase);
if (match.Success)
{
builder.Append(match.Value);
index = match.Length;
}
}
// Character substitutions
for (int cntr = index; cntr < filename.Length; cntr++)
{
char c = filename[cntr];
switch (c)
{
case '\u0000':
case '\u0001':
case '\u0002':
case '\u0003':
case '\u0004':
case '\u0005':
case '\u0006':
case '\u0007':
case '\u0008':
case '\u0009':
case '\u000A':
case '\u000B':
case '\u000C':
case '\u000D':
case '\u000E':
case '\u000F':
case '\u0010':
case '\u0011':
case '\u0012':
case '\u0013':
case '\u0014':
case '\u0015':
case '\u0016':
case '\u0017':
case '\u0018':
case '\u0019':
case '\u001A':
case '\u001B':
case '\u001C':
case '\u001D':
case '\u001E':
case '\u001F':
case '<':
case '>':
case ':':
case '"':
case '/':
case '|':
case '?':
case '*':
builder.Append('_');
break;
case '\\':
builder.Append(containsFolder ? c : '_');
break;
default:
builder.Append(c);
break;
}
}
string built = builder.ToString();
if (built == "")
{
return "_";
}
if (built.EndsWith(" ") || built.EndsWith("."))
{
built = built.Substring(0, built.Length - 1) + "_";
}
// These are reserved names, in either the folder or file name, but they are fine if following a dot
// CON, PRN, AUX, NUL, COM0 .. COM9, LPT0 .. LPT9
builder = new StringBuilder(built.Length + 12);
index = 0;
foreach (Match match in Regex.Matches(built, #"(^|\\)\s*(?<bad>CON|PRN|AUX|NUL|COM\d|LPT\d)\s*(\.|\\|$)", RegexOptions.IgnoreCase))
{
Group group = match.Groups["bad"];
if (group.Index > index)
{
builder.Append(built.Substring(index, match.Index - index + 1));
}
builder.Append(group.Value);
builder.Append("_"); // putting an underscore after this keyword is enough to make it acceptable
index = group.Index + group.Length;
}
if (index == 0)
{
return built;
}
if (index < built.Length - 1)
{
builder.Append(built.Substring(index));
}
return builder.ToString();
}
Though the only illegal Unix chars might be / and NULL, although some consideration for command line interpretation should be included.
For example, while it might be legal to name a file 1>&2 or 2>&1 in Unix, file names such as this might be misinterpreted when used on a command line.
Similarly it might be possible to name a file $PATH, but when trying to access it from the command line, the shell will translate $PATH to its variable value.
The .NET Framework System.IO provides the following functions for invalid file system characters:
Path.GetInvalidFileNameChars
Path.GetInvalidPathChars
Those functions should return appropriate results depending on the platform the .NET runtime is running in. That said, the Remarks in the documentation pages for those functions say:
The array returned from this method is not guaranteed to contain the
complete set of characters that are invalid in file and directory
names. The full set of invalid characters can vary by file system.
I always assumed that banned characters in Windows filenames meant that all exotic characters would also be outlawed. The inability to use ?, / and : in particular irked me. One day I discovered that it was virtually only those chars which were banned. Other Unicode characters may be used. So the nearest Unicode characters to the banned ones I could find were identified and MS Word macros were made for them as Alt+?, Alt+: etc. Now I form the filename in Word, using the substitute chars, and copy it to the Windows filename. So far I have had no problems.
Here are the substitute chars (Alt + the decimal Unicode) :
⃰ ⇔ Alt8432
⁄ ⇔ Alt8260
⃥ ⇔ Alt8421
∣ ⇔ Alt8739
ⵦ ⇔ Alt11622
⮚ ⇔ Alt11162
‽ ⇔ Alt8253
፡ ⇔ Alt4961
‶ ⇔ Alt8246
″ ⇔ Alt8243
As a test I formed a filename using all of those chars and Windows accepted it.
This is good enough for me in Python:
def fix_filename(name, max_length=255):
"""
Replace invalid characters on Linux/Windows/MacOS with underscores.
List from https://stackoverflow.com/a/31976060/819417
Trailing spaces & periods are ignored on Windows.
>>> fix_filename(" COM1 ")
'_ COM1 _'
>>> fix_filename("COM10")
'COM10'
>>> fix_filename("COM1,")
'COM1,'
>>> fix_filename("COM1.txt")
'_.txt'
>>> all('_' == fix_filename(chr(i)) for i in list(range(32)))
True
"""
return re.sub(r'[/\\:|<>"?*\0-\x1f]|^(AUX|COM[1-9]|CON|LPT[1-9]|NUL|PRN)(?![^.])|^\s|[\s.]$', "_", name[:max_length], flags=re.IGNORECASE)
See also this outdated list for additional legacy stuff like = in FAT32.
As of 18/04/2017, no simple black or white list of characters and filenames is evident among the answers to this topic - and there are many replies.
The best suggestion I could come up with was to let the user name the file however he likes. Using an error handler when the application tries to save the file, catch any exceptions, assume the filename is to blame (obviously after making sure the save path was ok as well), and prompt the user for a new file name. For best results, place this checking procedure within a loop that continues until either the user gets it right or gives up. Worked best for me (at least in VBA).
In Unix shells, you can quote almost every character in single quotes '. Except the single quote itself, and you can't express control characters, because \ is not expanded. Accessing the single quote itself from within a quoted string is possible, because you can concatenate strings with single and double quotes, like 'I'"'"'m' which can be used to access a file called "I'm" (double quote also possible here).
So you should avoid all control characters, because they are too difficult to enter in the shell. The rest still is funny, especially files starting with a dash, because most commands read those as options unless you have two dashes -- before, or you specify them with ./, which also hides the starting -.
If you want to be nice, don't use any of the characters the shell and typical commands use as syntactical elements, sometimes position dependent, so e.g. you can still use -, but not as first character; same with ., you can use it as first character only when you mean it ("hidden file"). When you are mean, your file names are VT100 escape sequences ;-), so that an ls garbles the output.
When creating internet shortcuts in Windows, to create the file name, it skips illegal characters, except for forward slash, which is converted to minus.
I had the same need and was looking for recommendation or standard references and came across this thread. My current blacklist of characters that should be avoided in file and directory names are:
$CharactersInvalidForFileName = {
"pound" -> "#",
"left angle bracket" -> "<",
"dollar sign" -> "$",
"plus sign" -> "+",
"percent" -> "%",
"right angle bracket" -> ">",
"exclamation point" -> "!",
"backtick" -> "`",
"ampersand" -> "&",
"asterisk" -> "*",
"single quotes" -> "“",
"pipe" -> "|",
"left bracket" -> "{",
"question mark" -> "?",
"double quotes" -> "”",
"equal sign" -> "=",
"right bracket" -> "}",
"forward slash" -> "/",
"colon" -> ":",
"back slash" -> "\\",
"lank spaces" -> "b",
"at sign" -> "#"
};
I would like to use a regular expression to validate user input. I want to allow any combination of letters, numbers, spaces, commas, apostrophes, periods, exclamation marks, and question marks, but I also want to limit the input to 4000 characters. I have come up with the following regular expression to achieve this: /^([a-z]|[0-9]| |,|'|\.|!|\?){1,4000}$/i.
However, when I attempt to use this regular expression test a subject in PHP with preg_match(), I am given a warning: PHP Warning: preg_match(): Compilation failed: regular expression is too large at offset 37 and the subject fails to be tested.
I find this strange because if I use an infinite quantifier, the test passes just fine (I demonstrate this situation below).
Why is limiting the repetition to 4000 a problem, but infinite repetition not?
regex-test.php:
<?php
$infinite = "/^([a-z]|[0-9]| |,|'|\.|!|\?)*$/i"; // Allows infinite repetition
$fourk = "/^([a-z]|[0-9]| |,|'|\.|!|\?){1,4000}$/i"; // Limits repetition to 4000
$string = "I like apples.";
if ( preg_match($infinite, $string) ){
echo "Passed infinite repetition. \n";
}
if ( preg_match($fourk, $string) ){
echo "Passed maximum repetition of 4000. \n";
}
?>
echos:
Passed infinite repetition
PHP Warning: preg_match(): Compilation failed: regular expression is too large at offset 37 in regex-test.php on line 16
The error is due to its LINK_SIZE, with offset values limiting the compiled pattern size to 64K. This is an expected behavior, explained below, and it's not because of a limit in repetition nor how the pattern is interpreted when compiled.
In this case
As Alan Moore pointed out in his answer, all characters should be in the same character class. I'm more drastic, so allow me to say that pattern is so wrong it makes me cringe.
-No offense, most of us tried that once too. It's just an attempt to underline that in no way such constructs should be used.
There are 3 common pitfalls here in (x|y|z){1,4000}:
Capturing subpatterns should only be used when needed (to store a specific part of the matched text, in order to extract that value or to use it in a backreference). For all other use cases, stick to non-capturing groups or atomic groups. They perform better and save memory.
Capturing subpatterns should not be repeated because the last repetition overwrites the captured text.
-OK, it could be used only in very particular cases.
Alternation (with the |s) adds backtracking states. It's a good practice to try to reduce them as much as you can. In this case, the regex ^[ !',.0-9?A-Z]{1,4000}$/i, would match exactly the same, not only avoiding the error, but also proving better performance.
LINK_SIZE
From "Handling Very Large Patterns" in pcrebuild man page:
Within a compiled pattern, offset values are used to point from one
part to another (for example, from an opening parenthesis to an
alternation metacharacter). By default, in the 8-bit and 16-bit
libraries, two-byte values are used for these offsets, leading to a
maximum size for a compiled pattern of around 64K.
That means the compiled pattern stores an offset value for every subpattern in the alternation, for every repetition of the group. In this case the offsets leave no memory for the rest of the compiled pattern.
This is more clearly expressed in a comment in pcre_internal.h from the PHP dist:
PCRE keeps offsets in its compiled code as 2-byte quantities (always
stored in big-endian order) by default. These are used, for example,
to link from the start of a subpattern to its alternatives and its
end. The use of 2 bytes per offset limits the size of the compiled
regex to around 64K, which is big enough for almost everybody.
Using pcretest, I get the following information:
PCRE version 8.37 2015-04-28
/^([a-z]|[0-9]| |,|'|\.|!|\?){1,575}$/i
Failed: regular expression is too large at offset 36
/^([a-z]|[0-9]| |,|'|\.|!|\?){1,574}$/i
Memory allocation (code space): 65432
There's a Windows version you can download from RexEgg.com.
Regarding other size limitations in PCRE, you can check this post of mine.
Overriding the default LINK_SIZE in PHP
If we had a true reason to use a huge pattern, and this pattern could not be simplified any further by all means, the link size could be increased. However, you can only achieve this by recompiling PHP yourself (therefore, your code won't be portable from now on). It should be the last resort, provided there's no other choice.
Also commented in pcre_internal.h:
The macros are controlled by the value of LINK_SIZE.
This defaults to 2 in the config.h file,
but can be overridden by using -D on the command line.
This is automated on Unix systems via the "configure" command.
PCRE link size can be configured to 3 or 4:
./configure -DLINK_SIZE=4
But keep in mind that longer offsets require additional data, and it will slow down all calls to preg_* functions.
In case of compiling PHP on your own, see Installation on Unix systems or Build your own PHP on Windows.
Looking at the 'regex' engine php uses, pcre here: http://pcre.sourceforge.net/pcre.txt at the limitations section it states:
The maximum length of a compiled pattern is 65539 (sic)
bytes
My guess is that some regex like this:
(123){1,3}
is compiled to something like this
(123)(123)?(123)?
Which makes it bigger than the maximum length
While I agree that the regex compiler shouldn't behave that way, you really shouldn't have encountered this problem. Inside the parentheses, your regex matches exactly one character from a specific set--the definition of a character class. The correct way to write your regex is to list all the characters inside one set of square brackets and forego the parentheses:
/^[a-z0-9 ,'.!?]{1,4000}$/i
That works fine, as this demo shows. However, it was the parentheses that were causing the error (even non-capturing parens cause it), and that doesn't seem right to me, even if they were unnecessary.
For me the problem was an un-escaped ? character
You need to escape it with not one, but to forward slashes \\
My regexp went from (?340202) to (\\?340202)
I am using a lot of regular expressions and stumbled over the question what actually can not be described by a regex.
First example that came to my mind was matching a string like XOOXXXOOOOXXXXX.... This would be a string consisting of an alternating sequence of X's and O's where each subpart consisting only of the character X or O is longer than the previsous sequence of the other character.
Can anybody explain what is the formal limit of a regex? I know this might be a rather academic question but I'm a curious person ;-)
Edit
As I am a php guy I am especially interested in regex described by PCRE standard as described here: http://php.net/manual/en/reference.pcre.pattern.syntax.php
I know that PCRE allows a lot of things that are not part of the original regular expressions like back references.
Mathcing of balanced parentheses seems to be one example that can not be matched by regular expressions in general but it can be matched using PCRE (see http://sandbox.onlinephpfunctions.com/code/fd12b580bb9ad7a19e226219d5146322a41c6e47 for live example):
$data = array('()', '(())', ')(', '(((()', '(((((((((())))))))))', '()()');
$regex = '/^((?:[^()]|\((?1)\))*+)$/';
foreach($data as $d) {
echo "$d matched by regex: " . (preg_match($regex, $d) ? 'yes' : 'no') . "\n";
}
First example that came to my mind was matching a string like XOOXXXOOOOXXXXX.... This would be a string consisting of an alternating sequence of X's and O's where each subpart consisting only of the character X or O is longer than the previsous sequence of the other character.
Yes, that can be done.
To match a non-empty sequence of x's, followed by a greater number of o's, we can use an approach similar to that of the balanced-parentheses regex:
(x(?1)?o)o+
To match a string of x's and o's such that any sequence of x's is followed by a longer sequence of o's (except optionally at the very end), we can build on pattern #1:
^o*(?:(x(?1)?o)o+)*x*$
And of course, we'll also need a variant of pattern #2 with x's and o's flipped:
^x*(?:(o(?1)?x)x+)*o*$
To match a string of x's and o's that meet both of the above criteria, we can convert pattern #2 to a positive lookahead assertion, and renumber the capture-group in pattern #3:
^(?=o*(?:(x(?1)?o)o+)*x*$)x*(?:(o(?2)?x)x+)*o*$
As for the main question . . . I'm confident that a PCRE can match any context-free language, since the support for (?n) outside of the nth capture-group means that you can basically create a subroutine for each of your non-terminals. For example, this context-free grammar:
S → aTb
S → ε
T → cSd
T → eTf
can be written as:
capture-group #1 (S) → (a(?2)b|)
capture-group #2 (T) → (c(?1)d|e(?2)f)
To assemble that into a single regex, we can just concatenate them all, but appending {0} after all but the start non-terminal, and then add ^ and $:
^(a(?2)b|)(c(?1)d|e(?2)f){0}$
But as you saw from your first example, PCREs can match some non-context-free languages as well. (Another example is anbncn, which is a classic example of a non-context-free language. You can match it with PCRE by combining a PCRE for anbncm with a PCRE for ambncn using a forward lookahead assertion. Although the intersection of two regular languages is necessarily regular, the intersection of two context-free languages is not necessarily context-free; but the intersection of the languages defined by two PCREs can be defined by a PCRE.)
The set of all languages that can be recognized by a regular expression is called, not surprisingly, "regular languages".
The next most complicated languages are the context-free languages. They cannot be parsed by any regular expression. The standard example is "all balanced parentheses" -- so "()()" and "(())" but not "(()".
Another good example of a context-free language is HTML.
I don't have definitive evidence that any of these are impossible with things like recursion, balancing groups, self-referencing groups, and appending text to the string being tested. I would be glad to be proven wrong on any or all of these, as I would learn something!
It's pretty bad at math.
For example, I do not believe it is possible using PCRE, to detect a sequence of numbers that is ascending: that is, to match "1 2 7 97 315 316..." but not "127 97 315 316..."
I'm not sure it's possible to even match a sequence contiguously increasing from 1, like "1 2 3...", without exhaustively listing all possibilities like /1( 2( 3(...)?)?)?/ up to the max length you wish to check.
Thee are hacks to make it work by adding known text to the string under test (eg http://www.rexegg.com/regex-trick-line-numbers.html works by adding a series of numbers to the end of the file). But as raw regex, simple math is only possible by brute-forcing.
Another example which I believe it will fail at is "match any sequence which sums to N".
So for N=4, it should match 4, 3 1, 1 3, 2 2, 1 1 1 1, 2 1 1, 1 2 1, 1 1 2, 1 1 1 1, which looks like you could brute-force it, until you realize it also has to match 4 -12 11 0 1.
In the same manner, I don't think you could have it analyze an equation using SI units, and verify whether the units balanced on both sides of the equation. For example, "10N=2kg*5ms^-1". Never mind checking the values, just checking the units are correct.
Then there're all the classes of problems that no computer program can currently accomplish, such as "check if a string has been correctly title cased in English" which would require a context-sensitive natural language parser to correctly detect the different senses of "like" in "Time Flies like an Arrow But Fruit Flies Like a Banana".
I would like to use a regular expression to validate user input. I want to allow any combination of letters, numbers, spaces, commas, apostrophes, periods, exclamation marks, and question marks, but I also want to limit the input to 4000 characters. I have come up with the following regular expression to achieve this: /^([a-z]|[0-9]| |,|'|\.|!|\?){1,4000}$/i.
However, when I attempt to use this regular expression test a subject in PHP with preg_match(), I am given a warning: PHP Warning: preg_match(): Compilation failed: regular expression is too large at offset 37 and the subject fails to be tested.
I find this strange because if I use an infinite quantifier, the test passes just fine (I demonstrate this situation below).
Why is limiting the repetition to 4000 a problem, but infinite repetition not?
regex-test.php:
<?php
$infinite = "/^([a-z]|[0-9]| |,|'|\.|!|\?)*$/i"; // Allows infinite repetition
$fourk = "/^([a-z]|[0-9]| |,|'|\.|!|\?){1,4000}$/i"; // Limits repetition to 4000
$string = "I like apples.";
if ( preg_match($infinite, $string) ){
echo "Passed infinite repetition. \n";
}
if ( preg_match($fourk, $string) ){
echo "Passed maximum repetition of 4000. \n";
}
?>
echos:
Passed infinite repetition
PHP Warning: preg_match(): Compilation failed: regular expression is too large at offset 37 in regex-test.php on line 16
The error is due to its LINK_SIZE, with offset values limiting the compiled pattern size to 64K. This is an expected behavior, explained below, and it's not because of a limit in repetition nor how the pattern is interpreted when compiled.
In this case
As Alan Moore pointed out in his answer, all characters should be in the same character class. I'm more drastic, so allow me to say that pattern is so wrong it makes me cringe.
-No offense, most of us tried that once too. It's just an attempt to underline that in no way such constructs should be used.
There are 3 common pitfalls here in (x|y|z){1,4000}:
Capturing subpatterns should only be used when needed (to store a specific part of the matched text, in order to extract that value or to use it in a backreference). For all other use cases, stick to non-capturing groups or atomic groups. They perform better and save memory.
Capturing subpatterns should not be repeated because the last repetition overwrites the captured text.
-OK, it could be used only in very particular cases.
Alternation (with the |s) adds backtracking states. It's a good practice to try to reduce them as much as you can. In this case, the regex ^[ !',.0-9?A-Z]{1,4000}$/i, would match exactly the same, not only avoiding the error, but also proving better performance.
LINK_SIZE
From "Handling Very Large Patterns" in pcrebuild man page:
Within a compiled pattern, offset values are used to point from one
part to another (for example, from an opening parenthesis to an
alternation metacharacter). By default, in the 8-bit and 16-bit
libraries, two-byte values are used for these offsets, leading to a
maximum size for a compiled pattern of around 64K.
That means the compiled pattern stores an offset value for every subpattern in the alternation, for every repetition of the group. In this case the offsets leave no memory for the rest of the compiled pattern.
This is more clearly expressed in a comment in pcre_internal.h from the PHP dist:
PCRE keeps offsets in its compiled code as 2-byte quantities (always
stored in big-endian order) by default. These are used, for example,
to link from the start of a subpattern to its alternatives and its
end. The use of 2 bytes per offset limits the size of the compiled
regex to around 64K, which is big enough for almost everybody.
Using pcretest, I get the following information:
PCRE version 8.37 2015-04-28
/^([a-z]|[0-9]| |,|'|\.|!|\?){1,575}$/i
Failed: regular expression is too large at offset 36
/^([a-z]|[0-9]| |,|'|\.|!|\?){1,574}$/i
Memory allocation (code space): 65432
There's a Windows version you can download from RexEgg.com.
Regarding other size limitations in PCRE, you can check this post of mine.
Overriding the default LINK_SIZE in PHP
If we had a true reason to use a huge pattern, and this pattern could not be simplified any further by all means, the link size could be increased. However, you can only achieve this by recompiling PHP yourself (therefore, your code won't be portable from now on). It should be the last resort, provided there's no other choice.
Also commented in pcre_internal.h:
The macros are controlled by the value of LINK_SIZE.
This defaults to 2 in the config.h file,
but can be overridden by using -D on the command line.
This is automated on Unix systems via the "configure" command.
PCRE link size can be configured to 3 or 4:
./configure -DLINK_SIZE=4
But keep in mind that longer offsets require additional data, and it will slow down all calls to preg_* functions.
In case of compiling PHP on your own, see Installation on Unix systems or Build your own PHP on Windows.
Looking at the 'regex' engine php uses, pcre here: http://pcre.sourceforge.net/pcre.txt at the limitations section it states:
The maximum length of a compiled pattern is 65539 (sic)
bytes
My guess is that some regex like this:
(123){1,3}
is compiled to something like this
(123)(123)?(123)?
Which makes it bigger than the maximum length
While I agree that the regex compiler shouldn't behave that way, you really shouldn't have encountered this problem. Inside the parentheses, your regex matches exactly one character from a specific set--the definition of a character class. The correct way to write your regex is to list all the characters inside one set of square brackets and forego the parentheses:
/^[a-z0-9 ,'.!?]{1,4000}$/i
That works fine, as this demo shows. However, it was the parentheses that were causing the error (even non-capturing parens cause it), and that doesn't seem right to me, even if they were unnecessary.
For me the problem was an un-escaped ? character
You need to escape it with not one, but to forward slashes \\
My regexp went from (?340202) to (\\?340202)
I've just installed a website & legacy CMS onto our server and I'm getting a POSIX compilation error. Luckily it's only appearing in the backend however the client's keen to get rid of it.
Warning: preg_match_all() [function.preg-match-all]: Compilation failed:
POSIX collating elements are not supported at offset 32 in
/home/kwecars/public_html/webEdition/we/include/we_classes/SEEM/we_SEEM.class.php
on line 621
From what I can tell it's the newer version of PHP causing the issue. Here's the code:
function getAllHrefs($code){
$trenner = "[\040|\n|\t|\r]*";
$pattern = "/<(a".$trenner."[^>]+href".$trenner."[=\"|=\'|=\\\\|=]*".$trenner.")
([^\'\">\040? \\\]*)([^\"\' \040\\\\>]*)(".$trenner."[^>]*)>/sie";
preg_match_all($pattern, $code, $allLinks); // ---- line 621
return $allLinks;
}
How can I tweak this to work on the newer version of php on this server?
Thanks in advance, my voodoo just isn't strong enough ;)
Your error message that “POSIX collating elements are not supported” deserves some explanation. After all, what in the world is a POSIX collating element anyway, and how can I avoid it?
The short answer is that you have an equals sign inside your square brackets in a place where its use is reserved for future use, assuming we ever get around to implementing it, which is anything but certain. You can tickle this in Perl on the command line this way, which gives a much better error message than PHP is providing:
% perl -le 'print "abc" =~ /[=foo=]/ || "Fail"'
POSIX syntax [= =] is reserved for future extensions in regex; marked by <-- HERE in m/[=foo=] <-- HERE / at -e line 1.
That’s the short answer; the longer answer follows.
Fancy POSIX Character Classes
Inside a square bracketed character class, POSIX admits three different nestedbracketed forms, all indicated using an extra symbol inside the brackets in pairs:
Named POSIX character classes, which are basically like Unicode properties, use an extra colon flanking: [:PROPERTY:], as in [:alpha:].
Collating elements intended to be treated as equivalent to each other, use an extra equals sign flanking them: [=ELEMENTS=], as in [=eéèëê=] in English or French, and [=vw=] in Swedish.
Polygraphs (digraphs, trigraphs, tetragraphs, etc), which are multicharacter elements meant to count as a single character, have an extra dot flanking them: [.DIGRAPH.], as in [.ch.] or [.ll.] per the traditional Spanish alphabet. These are sometimes known as contractions because two or more code points count as though that sequence were a single code point.
Perl supports only the first of these, not the second and third.
They are all awkward to use, because they must be nested inside an extra set of brackets, as in [[:punct:] to mean \pP or \p{punct}. You only need extra braces with Unicode properties when you are selecting one of many, as in [\pL\pN\pM\p{Pc}].
The Intent
The other two were an attempt to support locale-specific linguistic elements in a pre‐Unicode enviornment under legacy 8‑bit locales. For example, to express the traditional Spanish alphabet, which counts acute accents over vowels and diaereses over u’s as the same letter yet which counts a tilde over an n as a different letter altogether, and which furthermore has two digraphs each counting as a distinct letter, you would have to write this in POSIX:
[[=aá=]bc[.ch.]d[=eé=]fgh[=ií=]jkl[.ll.]mnñ[=oó=]pqrst[=uúü=]vwxyz]
You can and sometimes much combine these. For example, in German phonebooks where the three i‑mutated vowels can be spelt without diacritics by inserting a following e:
[a[=ä[.ae.]=]bcdefghijklmno[=ö[.oe.]=]pqrs[=ß[.ss.]=]tu[=ü[.ue.]=]vwxyz]
That way, assuming $ES and $DE are those languages’ respective alphabets, you could say something like
[$ES]{4}
and have it match words like guía, niño, llave, and choco in Spanish; or in German have
[$DE]{6}
and have it match words like tschüß or its uppercase undiacriticked equivalent, TSCHUESS.
The Unicode Way
This is awkward for various reasons, and not just those that are obvious from the two alphabets listed above. It does not admit the notion of combining characters, so you have to add those explicitly for non-normalized text, as in [=e\xE9[.e\x{301.]=].
Unicode has taken another path in how to implement linguistic elements like this. Fortunately, Unicode regular expressions per UTS#18 do not need to support language features tailored for specific languages or locales until Level 3. This is something no one yet has yet implemented.
Note that having SS and ß have the same casefold is not considered a locale tailoring. It is the full casefold for that code point no matter the linguistic context. So those are the same when case is ignored. Strange but true. Given that ß is code point U+00DF, we see that these are the same no matter the locale:
$ perl5.14.0 -E 'say "SS" =~ /^\xDF$/i ? "Pass" : "Fail"'
Pass
$ perl5.14.0 -E 'say "\xDF" =~ /^SS$/i ? "Pass" : "Fail"'
Pass
Although locale tailoring for patterns is still beyond us, collation has been implemented, including with locale support, and you can access it from Perl just fine.
However, PHP does not yet support Unicode collation.
References for Unicode collation include:
ICU’s Collation Concepts document
UTS#10: Unicode Collation Algorithm
Perl’s Unicode::Collate module.
Perl’s Unicode::Collate::Locale module.
[...] are character classes, they match any character between the brackets, you don't have to add | between them. See character classes.
So [abcd] will match a or b or c or d.
If you want to match alternations of more than one character, for example red or blue or yellow, use a sub pattern:
"(red|blue|yellow)"
And you guessed, [abcd] is equivalent to (a|b|c|d).
So here is what you could do for your regex:
For
$trenner = "[\040|\n|\t|\r]*";
Write this instead:
$trenner = "[\040\n\t\r]*";
And for
"[=\"|=\'|=\\\\|=]"
You could do
"(=\"|=\'|=\\\\|=)"
Or
"=[\"'\\\\]?"
BTW you could use \s instead of $trenner (see http://www.php.net/manual/en/regexp.reference.escape.php)