Cross-browser issues surrounding the handling of regular expression non-participating capturing groups (which I'll call NPCGs) present several challenges. The standard sucks to begin with, and the three biggest browsers (IE, Firefox, Safari) each disrespect the rules in their own unique ways.
First, I should explain what NPCGs are, as it seems that even some experienced regex users aren't fully aware of or understand the concept. Assuming you're already familiar with the idea of capturing and non-capturing parentheses (see this page if you need a refresher), note that NPCGs are different from groups which capture a zero-length value (i.e., an empty string). This is probably easiest to explain by showing some examples…
The following regexes all potentially contain NPCGs (depending on the data they are run over), because the capturing groups are not required to participate:
/(x)?/
/(x)*/
/(x){0,2}/
/(x)|(y)/
— If this matches, it's guaranteed to contain exactly one NPCG.
/(?!(x))/
— If this matches (which, on its own, it will at least at the end of the string), it's guaranteed to contain an NPCG, because the pattern only succeeds if the match of "x" fails.
/()??/
— This is guaranteed to match within any string and contain an NPCG, because of the use of a lazy ??
quantifier on a capturing group for a zero-length value.
On the other hand, these will never contain an NPCG, because although they are allowed to match a zero-length value, the capturing groups are required to participate:
/(x?)/
/(x*)/
/(x{0,2})/
/((?:xx)?)/
–or– /(xx|)/
— These two are equivalent.
/()?/
–or– /(x?)?/
— These are not required to participate, but their greedy ?
quantifiers ensure they will always succeed at capturing at least an empty string.
So, what's the difference between an NPCG and a group which captures an empty string? I guess that's up to the regex library, but typically, backreferences to NPCGs are assigned a special null or undefined value.
Following are the ECMA-262v3 rules (paraphrased) for how NPCGs should be handled in JavaScript:
- Within a regex, backreferences to NPCGs match an empty string (i.e., the backreferences always succeed). This is unfortunate, since it prevents some fancy patterns which would otherwise be possible (e.g., see my method for mimicking conditionals), and it's atypical compared to many other regular expression engines including Perl 5 (which ECMA-standard regular expressions are supposedly based on), PCRE, .NET, Java, Python, Ruby, JGsoft, and others.
- Within a replacement string, backreferences to NPCGs produce an empty string (i.e., nothing). Unlike the previous point, this is typical elsewhere, and allows you to use a regex like
/a(b)|c(d)/
and replace it with "$1$2"
without having to worry about null pointers or errors about non-participating groups.
- In the result arrays from
RegExp.prototype.exec
, String.prototype.match
(when used with a non-global regex), String.prototype.split
, and the arguments available to callback functions with String.prototype.replace
, NPCGs return undefined
. This is a very logical approach.
References: ECMA-262v3 sectons 15.5.4.11, 15.5.4.14, 15.10.2.1, 15.10.2.3, 15.10.2.8, 15.10.2.9.
Unfortunately, actual browser handling of NPCGs is all over the place, resulting in numerous cross-browser differences which can easily result in subtle (or not so subtle) bugs in your code if you don't know what you're doing. E.g., Firefox incorrectly uses an empty string with the replace()
and split()
methods, but correctly uses undefined
with the exec()
method. Conversely, IE correctly uses undefined
with the replace()
method, incorrectly uses an empty string with the exec()
method, and incorrectly returns neither with the split()
method since it doesn't splice backreferences into the resulting array. As for the handling of backreferences to non-participating groups within regexes (e.g., /(x)?\1y/.test("y")
), Safari uses the more sensible, non-ECMA-compliant approach (returning false
for the previous bit of code), while IE, Firefox, and Opera follow the standard. (If you use /(x?)\1y/.test("y")
instead, all four browsers will correctly return true
.)
Several times I've seen people encounter these differences and diagnose them incorrectly, not having understood the root cause. A recent instance is what prompted this writeup.
Here are cross-browser results from each of the regex and regex-using methods when NPCGs have an impact on the outcome:
Code |
ECMA-262v3 |
IE 5.5 – 7 |
Firefox 2.0.0.6 |
Opera 9.23 |
Safari 3.0.3 |
/(x)?\1y/.test("y") |
true |
true |
true |
true |
false |
/(x)?\1y/.exec("y") |
["y", undefined] |
["y", ""] |
["y", undefined] |
["y", undefined] |
null |
/(x)?y/.exec("y") |
["y", undefined] |
["y", ""] |
["y", undefined] |
["y", undefined] |
["y", undefined] |
"y".match(/(x)?\1y/) |
["y", undefined] |
["y", ""] |
["y", undefined] |
["y", undefined] |
null |
"y".match(/(x)?y/) |
["y", undefined] |
["y", ""] |
["y", undefined] |
["y", undefined] |
["y", undefined] |
"y".match(/(x)?\1y/g) |
["y"] |
["y"] |
["y"] |
["y"] |
null |
"y".split(/(x)?\1y/) |
["", undefined, ""] |
[ ] |
["", "", ""] |
["", undefined, ""] |
["y"] |
"y".split(/(x)?y/) |
["", undefined, ""] |
[ ] |
["", "", ""] |
["", undefined, ""] |
["", ""] |
"y".search(/(x)?\1y/) |
0 |
0 |
0 |
0 |
-1 |
"y".replace(/(x)?\1y/, "z") |
"z" |
"z" |
"z" |
"z" |
"y" |
"y".replace(/(x)?y/, "$1") |
"" |
"" |
"" |
"" |
"" |
"y".replace(/(x)?\1y/,
function($0, $1){
return String($1);
}) |
"undefined" |
"undefined" |
"" |
"undefined" |
"y" |
"y".replace(/(x)?y/,
function($0, $1){
return String($1);
}) |
"undefined" |
"undefined" |
"" |
"undefined" |
"" |
"y".replace(/(x)?y/,
function($0, $1){
return $1;
}) |
"undefined" |
"" |
"" |
"undefined" |
"" |
(Run the tests in your browser.)
The workaround for this mess is to avoid creating any potential for non-participating capturing groups, unless you know exactly what you're doing. Although that shouldn't be necessary, NPCGs are usually easy to avoid anyway. See the examples near the top of this post.
Edit (2007-08-16): I've updated this post with data from the newest versions of the listed browsers. The original data contained a few false negatives for Opera and Safari which resulted from a faulty library used to generate the results.