Question

I’m trying to find all instances of the keyword “public” in some Java code (with a Python script) that are not in comments or strings, a.k.a. not found following //, in between a /* and a */, and not in between double or single quotes, and which are not part of variable names– i.e. they must be preceded by a space, tab, or newline, and must be followed by the same.

So here’s what I have at the moment–

//.*\spublic\s.*\n
/\*.*\spublic\s.*\*/
".*\spublic\s.*"
'.*\spublic\s.*'

Am I messing this up at all?

But that finds exactly what I’m NOT looking for. How can I turn it around and search the inverse of the sum of those four expressions, as a single regex?

I’ve figured out this probably uses negative look-ahead and look-behind, but I still can’t quite piece it together. Also, for the /**/ regex, I’m concerned that .* doesn’t match newlines, so it would fail to recognize that this public is in a comment:

/*
public
*/

Everything below this point is me thinking on paper and can be disregarded. These thoughts are not fully accurate.

---

Edit:

I daresay (?<!//).*public.* would match anything not in single line comments, so I’m getting the hang of things. I think. But still unsure how to combine everything.

Edit2:

So then– following that idea, I |ed them all to get–

(?<!//).*public.*|(?<!/\*).*public.\*/(?!\*/)|(?<!").*public.*(?!")|(?<!').*public.*(?!')

But I’m not sure about that. //public will not be matched by the first alternate, but it will be matched by the second. I need to AND the look-aheads and look-behinds, not OR the whole thing.

Answer 1

I’m sorry, but I’ll have to break the news to you, that what you are trying to do is impossible. The reason is mostly because Java is not a regular language. As we all know by now, most regex engines provide non-regular features, but Python in particular is lacking something like recursion (PCRE) or balancing groups (.NET) which could do the trick. But let’s look into that in more depth.

First of all, why are your patterns not as good as you think they are? (for the task of matching public inside those literals; similar problems will apply to reversing the logic)

As you have already recognized, you will have problems with line breaks (in the case of /*...*/). This can be solved by either using the modifier/option/flag re.S (which changes the behavior of .) or by using [\s\S] instead of . (because the former matches any character).

But there are other problems. You only want to find surrounding occurrences of the string or comment literals. You are not actually making sure that they are specifically wrapped around the public in question. I’m not sure how much you can put onto a single line in Java, but if you had an arbitrary string, then later a public and then another string on a single line, then your regex would match the public because it can find the " before and after it. Even if that is not possible, if you have two block comments in the same input, then any public between those two block comments would cause a match. So you would need to find a way to assert only that your public is really inside "..." or /*...*/ and not just that these literals can be found anywhere to left of right of it.

Next thing: matches cannot overlap. But your match includes everything from the opening literal until the ending literal. So if you had "public public" that would cause only one match. And capturing cannot help you here. Usually the trick to avoid this is to use lookarounds (which are not included in the match). But (as we will see later) the lookbehind doesn’t work as nicely as you would think, because it cannot be of arbitrary length (only in .NET that is possible).

Now the worst of all. What if you have " inside a comment? That shouldn’t count, right? What if you have // or /* or */ inside a string? That shouldn’t count, right? What about ' inside "-strings and " inside '-strings? Even worse, what about \" inside "-string? So for 100% robustness you would have to do a similar check for your surrounding delimiters as well. And this is usually where regular expressions reach the end of their capabilities and this is why you need a proper parser that walks the input string and builds a whole tree of your code.

But say you never have comment literals inside strings and you never have quotes inside comments (or only matched quotes, because they would constitute a string, and we don’t want public inside strings anyway). So we are basically assuming that every of the literals in question is correctly matched, and they are never nested. In that case you can use a lookahead to check whether you are inside or outside one of the literals (in fact, multiple lookaheads). I’ll get to that shortly.

But there is one more thing left. What does (?<!//).*public.* not work? For this to match it is enough for (?<!//) to match at any single position. e.g. if you just had input // public the engine would try out the negative lookbehind right at the start of the string, (to the left of the start of the string), would find no //, then use .* to consume // and the space and then match public. What you actually want is (?<!//.*)public. This will start the lookbehind from the starting position of public and look all the way to the left through the current line. But… this is a variable-length lookbehind, which is only supported by .NET.

But let’s look into how we can make sure we are really outside of a string. We can use a lookahead to look all the way to the end of the input, and check that there is an even number of quotes on the way.

public(?=[^"]*("[^"]*"[^"]*)*$)

Now if we try really hard we can also ignore escaped quotes when inside of a string:

public(?=[^"]*("(?:[^"\\]|\\.)*"[^"]*)*$)

So once we encounter a " we will accept either non-quote, non-backslash characters, or a backslash character and whatever follows it (that allows escaping of backslash-characters as well, so that in "a string\\" we won’t treat the closing " as being escaped). We can use this with multi-line mode (re.M) to avoid going all the way to the end of the input (because the end of the line is enough):

public(?=[^"\r\n]*("(?:[^"\r\n\\]|\\.)*"[^"\r\n]*)*$)

(re.M is implied for all following patterns)

This is what it looks for single-quoted strings:

public(?=[^'\r\n]*('(?:[^'\r\n\\]|\\.)*'[^'\r\n]*)*$)

For block comments it’s a bit easier, because we only need to look for /* or the end of the string (this time really the end of the entire string), without ever encountering */ on the way. That is done with a negative lookahead at every single position until the end of the search:

public(?=(?:(?![*]/)[\s\S])*(?:/[*]|\Z))

But as I said, we’re stumped on the single-line comments for now. But anyway, we can combine the last three regular expressions into one, because lookaheads don’t actually advance the position of the regex engine on the target string:

public(?=[^"\r\n]*("(?:[^"\r\n\\]|\\.)*"[^"\r\n]*)*$)(?=[^'\r\n]*('(?:[^'\r\n\\]|\\.)*'[^'\r\n]*)*$)(?=(?:(?![*]/)[\s\S])*(?:/[*]|\Z))

Now what about those single-line comments? The trick to emulate variable-length lookbehinds is usually to reverse the string and the pattern – which makes the lookbehind a lookahead:

cilbup(?!.*//)

Of course, that means we have to reverse all other patterns, too. The good news is, if we don’t care about escaping, they look exactly the same (because both quotes and block comments are symmetrical). So you could run this pattern on a reversed input:

cilbup(?=[^"\r\n]*("[^"\r\n]*"[^"\r\n]*)*$)(?=[^'\r\n]*('[^'\r\n]*'[^'\r\n]*)*$)(?=(?:(?![*]/)[\s\S])*(?:/[*]|\Z))(?!.*//)

You can then find the match positions in your actual input by using inputLength -foundMatchPosition - foundMatchLength.

Now what about escaping? That get’s quite annoying now, because we have to skip quotes, if they are followed by a backslash. Because of some backtracking issues we need to take care of that in five places. Three times, when consuming non-quote characters (because we need to allow "\ as well now. And twice, when consuming quote characters (using a negative lookahead to make sure there is no backslash after them). Let’s look at double quotes:

cilbup(?=(?:[^"\r\n]|"\\)*(?:"(?!\\)(?:[^"\r\n]|"\\)*"(?!\\)(?:[^"\r\n]|"\\)*)*$)

(It looks horrible, but if you compare it with the pattern that disregards escaping, you will notice the few differences.)

So incorporating that into the above pattern:

cilbup(?=(?:[^"\r\n]|"\\)*(?:"(?!\\)(?:[^"\r\n]|"\\)*"(?!\\)(?:[^"\r\n]|"\\)*)*$)(?=(?:[^'\r\n]|'\\)*(?:'(?!\\)(?:[^'\r\n]|'\\)*'(?!\\)(?:[^'\r\n]|'\\)*)*$)(?=(?:(?![*]/)[\s\S])*(?:/[*]|\Z))(?!.*//)

So this might actually do it for many cases. But as you can see it’s horrible, almost impossible to read, and definitely impossible to maintain.

What were the caveats? No comment literals inside strings, no string literals inside strings of the other type, no string literals inside comments. Plus, we have four independent lookaheads, which will probably take some time (at least I think I have a voided most of backtracking).

In any case, I believe this is as close as you can get with regular expressions.

EDIT:

I just realised I forgot the condition that public must not be part of a longer literal. You included spaces, but what if it’s the first thing in the input? The easiest thing would be to use \b. That matches a position (without including surrounding characters) that is between a word character and a non-word character. However, Java identifiers may contain any Unicode letter or digit, and I’m not sure whether Python’s \b is Unicode-aware. Also, Java identifiers may contain $. Which would break that anyway. Lookarounds to the rescue! Instead of asserting that there is a space character on every side, let’s assert that there is no non-space character. Because we need negative lookarounds for that, we will get the advantage of not including those characters in the match for free:

(?<!\S)cilbup(?!\S)(?=(?:[^"\r\n]|"\\)*(?:"(?!\\)(?:[^"\r\n]|"\\)*"(?!\\)(?:[^"\r\n]|"\\)*)*$)(?=(?:[^'\r\n]|'\\)*(?:'(?!\\)(?:[^'\r\n]|'\\)*'(?!\\)(?:[^'\r\n]|'\\)*)*$)(?=(?:(?![*]/)[\s\S])*(?:/[*]|\Z))(?!.*//)

And because just from scrolling this code snippet to the right one cannot quite grasp how ridiculously huge this regex is, here it is in freespacing mode (re.X) with some annotations:

(?<!\S)      # make sure there is no trailing non-whitespace character
cilbup       # public
(?!\S)       # make sure there is no leading non-whitespace character
(?=          # lookahead (effectively lookbehind!) to ensure we are not inside a
             # string
  (?:[^"\r\n]|"\\)*
             # consume everything except for line breaks and quotes, unless the
             # quote is followed by a backslash (preceded in the actual input)
  (?:        # subpattern that matches two (unescaped) quotes
    "(?!\\)  # a quote that is not followed by a backslash
    (?:[^"\r\n]|"\\)*
             # we've seen that before
    "(?!\\)  # a quote that is not followed by a backslash
    (?:[^"\r\n]|"\\)*
             # we've seen that before
  )*         # end of subpattern - repeat 0 or more times (ensures even no. of ")
  $          # end of line (start of line in actual input)
)            # end of double-quote lookahead
(?=(?:[^'\r\n]|'\\)*(?:'(?!\\)(?:[^'\r\n]|'\\)*'(?!\\)(?:[^'\r\n]|'\\)*)*$)
             # the same horrible bastard again for single quotes
(?=          # lookahead (effectively lookbehind) for block comments
  (?:        # subgroup to consume anything except */
    (?![*]/) # make sure there is no */ coming up
    [\s\S]   # consume an arbitrary character
  )*         # repeat
  (?:/[*]|\Z)# require to find either /* or the end of the string
)            # end of lookahead for block comments
(?!.*//)     # make sure there is no // on this line