PARTIAL MATCHING IN PCRE

In normal use of PCRE, if the subject string that is passed to pcre_exec() or pcre_dfa_exec() matches as far as it goes, but is too short to match the entire pattern, PCRE_ERROR_NOMATCH is returned. There are circumstances where it might be helpful to distinguish this case from other cases in which there is no match.

Consider, for example, an application where a human is required to type in data for a field with specific formatting requirements. An example might be a date in the form ddmmmyy, defined by this pattern:

^\d?\d(jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\d\d$

If the application sees the user's keystrokes one by one, and can check that what has been typed so far is potentially valid, it is able to raise an error as soon as a mistake is made, by beeping and not reflecting the character that has been typed, for example. This immediate feedback is likely to be a better user interface than a check that is delayed until the entire string has been entered. Partial matching can also sometimes be useful when the subject string is very long and is not all available at once.

PCRE supports partial matching by means of the PCRE_PARTIAL_SOFT and PCRE_PARTIAL_HARD options, which can be set when calling pcre_exec() or pcre_dfa_exec(). For backwards compatibility, PCRE_PARTIAL is a synonym for PCRE_PARTIAL_SOFT. The essential difference between the two options is whether or not a partial match is preferred to an alternative complete match, though the details differ between the two matching functions. If both options are set, PCRE_PARTIAL_HARD takes precedence.

Setting a partial matching option disables two of PCRE's optimizations. PCRE remembers the last literal byte in a pattern, and abandons matching immediately if such a byte is not present in the subject string. This optimization cannot be used for a subject string that might match only partially. If the pattern was studied, PCRE knows the minimum length of a matching string, and does not bother to run the matching function on shorter strings. This optimization is also disabled for partial matching.

PARTIAL MATCHING USING pcre_exec()

A partial match occurs during a call to pcre_exec() whenever the end of the subject string is reached successfully, but matching cannot continue because more characters are needed. However, at least one character must have been matched. (In other words, a partial match can never be an empty string.)

If PCRE_PARTIAL_SOFT is set, the partial match is remembered, but matching continues as normal, and other alternatives in the pattern are tried. If no complete match can be found, pcre_exec() returns PCRE_ERROR_PARTIAL instead of PCRE_ERROR_NOMATCH. If there are at least two slots in the offsets vector, the first of them is set to the offset of the earliest character that was inspected when the partial match was found. For convenience, the second offset points to the end of the string so that a substring can easily be identified.

For the majority of patterns, the first offset identifies the start of the partially matched string. However, for patterns that contain lookbehind assertions, or \K, or begin with \b or \B, earlier characters have been inspected while carrying out the match. For example:

/(?<=abc)123/

This pattern matches "123", but only if it is preceded by "abc". If the subject string is "xyzabc12", the offsets after a partial match are for the substring "abc12", because all these characters are needed if another match is tried with extra characters added.

If there is more than one partial match, the first one that was found provides the data that is returned. Consider this pattern:

/123\w+X|dogY/

If this is matched against the subject string "abc123dog", both alternatives fail to match, but the end of the subject is reached during matching, so PCRE_ERROR_PARTIAL is returned instead of PCRE_ERROR_NOMATCH. The offsets are set to 3 and 9, identifying "123dog" as the first partial match that was found. (In this example, there are two partial matches, because "dog" on its own partially matches the second alternative.)

If PCRE_PARTIAL_HARD is set for pcre_exec(), it returns PCRE_ERROR_PARTIAL as soon as a partial match is found, without continuing to search for possible complete matches. The difference between the two options can be illustrated by a pattern such as:

/dog(sbody)?/

This matches either "dog" or "dogsbody", greedily (that is, it prefers the longer string if possible). If it is matched against the string "dog" with PCRE_PARTIAL_SOFT, it yields a complete match for "dog". However, if PCRE_PARTIAL_HARD is set, the result is PCRE_ERROR_PARTIAL. On the other hand, if the pattern is made ungreedy the result is different:

/dog(sbody)??/

In this case the result is always a complete match because pcre_exec() finds that first, and it never continues after finding a match. It might be easier to follow this explanation by thinking of the two patterns like this:

 /dog(sbody)?/    is the same as  /dogsbody|dog/
 /dog(sbody)??/   is the same as  /dog|dogsbody/

The second pattern will never match "dogsbody" when pcre_exec() is used, because it will always find the shorter match first.

PARTIAL MATCHING USING pcre_dfa_exec()

The pcre_dfa_exec() function moves along the subject string character by character, without backtracking, searching for all possible matches simultaneously. If the end of the subject is reached before the end of the pattern, there is the possibility of a partial match, again provided that at least one character has matched.

When PCRE_PARTIAL_SOFT is set, PCRE_ERROR_PARTIAL is returned only if there have been no complete matches. Otherwise, the complete matches are returned. However, if PCRE_PARTIAL_HARD is set, a partial match takes precedence over any complete matches. The portion of the string that was inspected when the longest partial match was found is set as the first matching string, provided there are at least two slots in the offsets vector.

Because pcre_dfa_exec() always searches for all possible matches, and there is no difference between greedy and ungreedy repetition, its behaviour is different from pcre_exec when PCRE_PARTIAL_HARD is set. Consider the string "dog" matched against the ungreedy pattern shown above:

/dog(sbody)??/

Whereas pcre_exec() stops as soon as it finds the complete match for "dog", pcre_dfa_exec() also finds the partial match for "dogsbody", and so returns that when PCRE_PARTIAL_HARD is set.

PARTIAL MATCHING AND WORD BOUNDARIES

If a pattern ends with one of sequences \b or \B, which test for word boundaries, partial matching with PCRE_PARTIAL_SOFT can give counter-intuitive results. Consider this pattern:

/\bcat\b/

This matches "cat", provided there is a word boundary at either end. If the subject string is "the cat", the comparison of the final "t" with a following character cannot take place, so a partial match is found. However, pcre_exec() carries on with normal matching, which matches \b at the end of the subject when the last character is a letter, thus finding a complete match. The result, therefore, is not PCRE_ERROR_PARTIAL. The same thing happens with pcre_dfa_exec(), because it also finds the complete match.

Using PCRE_PARTIAL_HARD in this case does yield PCRE_ERROR_PARTIAL, because then the partial match takes precedence.

FORMERLY RESTRICTED PATTERNS

For releases of PCRE prior to 8.00, because of the way certain internal optimizations were implemented in the pcre_exec() function, the PCRE_PARTIAL option (predecessor of PCRE_PARTIAL_SOFT) could not be used with all patterns. From release 8.00 onwards, the restrictions no longer apply, and partial matching with pcre_exec() can be requested for any pattern.

Items that were formerly restricted were repeated single characters and repeated metasequences. If PCRE_PARTIAL was set for a pattern that did not conform to the restrictions, pcre_exec() returned the error code PCRE_ERROR_BADPARTIAL (-13). This error code is no longer in use. The PCRE_INFO_OKPARTIAL call to pcre_fullinfo() to find out if a compiled pattern can be used for partial matching now always returns 1.

EXAMPLE OF PARTIAL MATCHING USING PCRETEST

If the escape sequence \P is present in a pcretest data line, the PCRE_PARTIAL_SOFT option is used for the match. Here is a run of pcretest that uses the date example quoted above:

   re> /^\d?\d(jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\d\d$/
 data> 25jun04\P
  0: 25jun04
  1: jun
 data> 25dec3\P
 Partial match: 23dec3
 data> 3ju\P
 Partial match: 3ju
 data> 3juj\P
 No match
 data> j\P
 No match

The first data string is matched completely, so pcretest shows the matched substrings. The remaining four strings do not match the complete pattern, but the first two are partial matches. Similar output is obtained when pcre_dfa_exec() is used.

If the escape sequence \P is present more than once in a pcretest data line, the PCRE_PARTIAL_HARD option is set for the match.

MULTI-SEGMENT MATCHING WITH pcre_dfa_exec()

When a partial match has been found using pcre_dfa_exec(), it is possible to continue the match by providing additional subject data and calling pcre_dfa_exec() again with the same compiled regular expression, this time setting the PCRE_DFA_RESTART option. You must pass the same working space as before, because this is where details of the previous partial match are stored. Here is an example using pcretest, using the \R escape sequence to set the PCRE_DFA_RESTART option (\D specifies the use of pcre_dfa_exec()):

   re> /^\d?\d(jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\d\d$/
 data> 23ja\P\D
 Partial match: 23ja
 data> n05\R\D
  0: n05

The first call has "23ja" as the subject, and requests partial matching; the second call has "n05" as the subject for the continued (restarted) match. Notice that when the match is complete, only the last part is shown; PCRE does not retain the previously partially-matched string. It is up to the calling program to do that if it needs to.

You can set the PCRE_PARTIAL_SOFT or PCRE_PARTIAL_HARD options with PCRE_DFA_RESTART to continue partial matching over multiple segments. This facility can be used to pass very long subject strings to pcre_dfa_exec().

MULTI-SEGMENT MATCHING WITH pcre_exec()

From release 8.00, pcre_exec() can also be used to do multi-segment matching. Unlike pcre_dfa_exec(), it is not possible to restart the previous match with a new segment of data. Instead, new data must be added to the previous subject string, and the entire match re-run, starting from the point where the partial match occurred. Earlier data can be discarded. Consider an unanchored pattern that matches dates:

   re> /\d?\d(jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\d\d/
 data> The date is 23ja\P
 Partial match: 23ja

At this stage, an application could discard the text preceding "23ja", add on text from the next segment, and call pcre_exec() again. Unlike pcre_dfa_exec(), the entire matching string must always be available, and the complete matching process occurs for each call, so more memory and more processing time is needed.

	Note
	If the pattern contains lookbehind assertions, or \K, or starts with \b or \B, the string that is returned for a partial match will include characters that precede the partially matched string itself, because these must be retained when adding on more characters for a subsequent matching attempt.

ISSUES WITH MULTI-SEGMENT MATCHING

Certain types of pattern may give problems with multi-segment matching, whichever matching function is used.

1. If the pattern contains tests for the beginning or end of a line, you need to pass the PCRE_NOTBOL or PCRE_NOTEOL options, as appropriate, when the subject string for any call does not contain the beginning or end of a line.

2. Lookbehind assertions at the start of a pattern are catered for in the offsets that are returned for a partial match. However, in theory, a lookbehind assertion later in the pattern could require even earlier characters to be inspected, and it might not have been reached when a partial match occurs. This is probably an extremely unlikely case; you could guard against it to a certain extent by always including extra characters at the start.

3. Matching a subject string that is split into multiple segments may not always produce exactly the same result as matching over one single long string, especially when PCRE_PARTIAL_SOFT is used. The section "Partial Matching and Word Boundaries" above describes an issue that arises if the pattern ends with \b or \B. Another kind of difference may occur when there are multiple matching possibilities, because a partial match result is given only when there are no completed matches. This means that as soon as the shortest match has been found, continuation to a new subject segment is no longer possible. Consider again this pcretest example:

   re> /dog(sbody)?/
 data> dogsb\P
  0: dog
 data> do\P\D
 Partial match: do
 data> gsb\R\P\D
  0: g
 data> dogsbody\D
  0: dogsbody
  1: dog

The first data line passes the string "dogsb" to pcre_exec(), setting the PCRE_PARTIAL_SOFT option. Although the string is a partial match for "dogsbody", the result is not PCRE_ERROR_PARTIAL, because the shorter string "dog" is a complete match. Similarly, when the subject is presented to pcre_dfa_exec() in several parts ("do" and "gsb" being the first two) the match stops when "dog" has been found, and it is not possible to continue. On the other hand, if "dogsbody" is presented as a single string, pcre_dfa_exec() finds both matches.

Because of these problems, it is probably best to use PCRE_PARTIAL_HARD when matching multi-segment data. The example above then behaves differently:

   re> /dog(sbody)?/
 data> dogsb\P\P
 Partial match: dogsb
 data> do\P\D
 Partial match: do
 data> gsb\R\P\P\D
 Partial match: gsb

4. Patterns that contain alternatives at the top level which do not all start with the same pattern item may not work as expected when PCRE_DFA_RESTART is used with pcre_dfa_exec(). For example, consider this pattern:

1234|3789

If the first part of the subject is "ABC123", a partial match of the first alternative is found at offset 3. There is no partial match for the second alternative, because such a match does not start at the same point in the subject string. Attempting to continue with the string "7890" does not yield a match because only those alternatives that match at one point in the subject are remembered. The problem arises because the start of the second alternative matches within the first alternative. There is no problem with anchored patterns or patterns such as:

1234|ABCD

where no string can be a partial match for both alternatives. This is not a problem if pcre_exec() is used, because the entire match has to be rerun each time:

   re> /1234|3789/
 data> ABC123\P
 Partial match: 123
 data> 1237890
  0: 3789

Of course, instead of using PCRE_DFA_PARTIAL, the same technique of re-running the entire match can also be used with pcre_dfa_exec(). Another possibility is to work with two buffers. If a partial match at offset n in the first buffer is followed by "no match" when PCRE_DFA_RESTART is used on the second buffer, you can then try a new match starting at offset n+1 in the first buffer.

AUTHOR

Philip Hazel
University Computing Service
Cambridge CB2 3QH, England.

REVISION

Last updated: 19 October 2009
Copyright (c) 1997-2009 University of Cambridge.