PERLREGUTS(1)
🦋 perlreguts 💅
Perl Programmers Reference Guide — okay bestie, pull up a chair, we're going into the guts of the Perl regex engine together and it's gonna be a whole journey. ✨
🌸 NAME 🌸
perlreguts — literally just: a description of the Perl regular expression engine. That's the whole vibe. 🎀
💖 DESCRIPTION 💖
Okay so, this document is an attempt to shine some light on the guts of the regex engine and how it works. The regex engine represents a significant chunk of the perl codebase, but it's relatively poorly understood — which is honestly kind of iconic of it but also a problem. This document is a meagre attempt at addressing that situation. It's derived from the author's experience, comments in the source code, other papers on the regex engine, feedback on the perl5-porters mail list, and no doubt other places as well. 🫶
NOTICE! It should be clearly understood that the behavior and structures discussed in this represent the state of the engine as the author understood it at the time of writing. It is NOT an API definition. It is purely an internals guide for those who want to hack the regex engine, or understand how the regex engine works. Readers of this document are expected to understand perl's regex syntax and its usage in detail. If you want to learn about the basics of Perl's regular expressions, seeperlre. And if you want to replace the regex engine with your own, seeperlreapi. 💀
👑 OVERVIEW 👑
🧚 A quick note on terms 🧚
There is some debate as to whether to say "regexp" or "regex". In this document we will use the term "regex" unless there is a special reason not to, in which case we will explain why. Periodt. 💅
When speaking about regexes we need to distinguish between their source code form and their internal form. In this document we will use the term "pattern" when we speak of their textual, source code form, and the term "program" when we speak of their internal representation. These correspond to the terms S-regex and B-regex that Mark Jason Dominus employs in his paper on "Rx" ([1] in "REFERENCES"). 🌈
🦋 What is a regular expression engine? 🦋
A regular expression engine is a program that takes a set of constraints specified in a mini-language, and then applies those constraints to a target string, and determines whether or not the string satisfies the constraints. See perlre for a full definition of the language. 📖
In less grandiose terms, the first part of the job is to turn a pattern into something the computer can efficiently use to find the matching point in the string, and the second part is performing the search itself. ✨
To do this we need to produce a program by parsing the text. We then need to execute the program to find the point in the string that matches. And we need to do the whole thing efficiently. 🔥
💜 Structure of a Regexp Program 💜
✨ High Level ✨
Although it is a bit confusing and some people object to the terminology, it is worth taking a look at a comment that has been in regexp.h for years:
This is essentially a linear encoding of a nondeterministic finite-state machine (aka syntax charts or "railroad normal form" in parsing technology).
The term "railroad normal form" is a bit esoteric, with "syntax diagram/charts", or "railroad diagram/charts" being more common terms. Nevertheless it provides a useful mental image of a regex program: each node can be thought of as a unit of track, with a single entry and in most cases a single exit point (there are pieces of track that fork, but statistically not many), and the whole forms a layout with a single entry and single exit point. The matching process can be thought of as a car that moves along the track, with the particular route through the system being determined by the character read at each possible connector point. A car can fall off the track at any point but it may only proceed as long as it matches the track. 🚗💨
Thus the pattern /foo(?:\w+|\d+|\s+)bar/ can be thought of as the following chart:
[start]
|
<foo>
|
+-----+-----+
| | |
<\w+> <\d+> <\s+>
| | |
+-----+-----+
|
<bar>
|
[end]The truth of the matter is that perl's regular expressions these days are much more complex than this kind of structure, but visualising it this way can help when trying to get your bearings, and it matches the current implementation pretty closely. 😌
To be more precise, we will say that a regex program is an encoding of a graph. Each node in the graph corresponds to part of the original regex pattern, such as a literal string or a branch, and has a pointer to the nodes representing the next component to be matched. Since "node" and "opcode" already have other meanings in the perl source, we will call the nodes in a regex program "regops". 👑
The program is represented by an array of regnode structures, one or more of which represent a single regop of the program. Struct regnode is the smallest struct needed, and has a field structure which is shared with all the other larger structures. (Outside this document, the term regnode is sometimes used to mean "regop", which could be confusing.) 🙈
The next pointers of all regops except BRANCH implement concatenation; a next pointer with a BRANCH on both ends of it is connecting two alternatives. [Here we have one of the subtle syntax dependencies: an individual BRANCH (as opposed to a collection of them) is never concatenated with anything because of operator precedence.] 🧠
The operand of some types of regop is a literal string; for others, it is a regop leading into a sub-program. In particular, the operand of a BRANCH node is the first regop of the branch. 🌿
NOTE: As the railroad metaphor suggests, this is not a tree structure: the tail of the branch connects to the thing following the set of BRANCHes. It is a like a single line of railway track that splits as it goes into a station or railway yard and rejoins as it comes out the other side. 🚂
🎀 Regops 🎀
The base structure of a regop is defined in regexp.h as follows:
struct regnode {
U8 flags; /* Various purposes, sometimes overridden */
U8 type; /* Opcode value as specified by regnodes.h */
U16 next_off; /* Offset in size regnode */
};Other larger regnode-like structures are defined in regcomp.h. They are almost like subclasses in that they have the same fields as regnode, with possibly additional fields following in the structure, and in some cases the specific meaning (and name) of some of base fields are overridden. The following is a more complete description. 💅
regnode_1regnode_2
regnode_1 structures have the same header, followed by a single four-byte argument; regnode_2 structures contain two two-byte arguments instead:
regnode_1 U32 arg1;
regnode_2 U16 arg1; U16 arg2;regnode_string
regnode_string structures, used for literal strings, follow the header with a one-byte length and then the string data. Strings are padded on the tail end with zero bytes so that the total length of the node is a multiple of four bytes:
regnode_string char string[1];
U8 str_len; /* overrides flags */regnode_charclass
Bracketed character classes are represented by regnode_charclass structures, which have a four-byte argument and then a 32-byte (256-bit) bitmap indicating which characters in the Latin1 range are included in the class.
regnode_charclass U32 arg1;
char bitmap[ANYOF_BITMAP_SIZE];Various flags whose names begin with ANYOF_ are used for special situations. Above Latin1 matches and things not known until run-time are stored in Perl's pprivate structure. 🔍
regnode_charclass_posixl
There is also a larger form of a char class structure used to represent POSIX char classes under /l matching, called regnode_charclass_posixl which has an additional 32-bit bitmap indicating which POSIX char classes have been included.
regnode_charclass_posixl U32 arg1;
char bitmap[ANYOF_BITMAP_SIZE];
U32 classflags;regnodes.h defines an array called PL_regnode_arg_len[] which gives the size of each opcode in units of size regnode (4-byte). A macro is used to calculate the size of an EXACT node based on its str_len field. 📐
The regops are defined in regnodes.h which is generated from regcomp.sym by regcomp.pl. Currently the maximum possible number of distinct regops is restricted to 256, with about a quarter already used. 😳
A set of macros makes accessing the fields easier and more consistent. These include OP(), which is used to determine the type of a regnode-like structure; NEXT_OFF(), which is the offset to the next node (more on this later); ARG(), ARG1(), ARG2(), ARG_SET(), and equivalents for reading and setting the arguments; and STR_LEN(), STRING() and OPERAND() for manipulating strings and regop bearing types. 🛠️
💀 What regnode is next? 💀
There are two distinct concepts of "next regnode" in the regex engine, and it is important to keep them distinct in your thinking as they overlap conceptually in many places, but where they don't overlap the difference is critical. For the majority of regnode types the two concepts are (nearly) identical in practice. The two types are REGNODE_AFTER which is used heavily during compilation but only occasionally during execution and regnext which is used heavily during execution, and only occasionally during compilation. 🧵
REGNODE_AFTER
This is the "positionally next regnode" in the compiled regex program. For the smaller regnode types it is regnode_ptr+1 under the hood, but as regnode sizes vary and can change over time we offer macros which hide the gory details. 🙈
It is heavily used in the compiler phase but is only used by a few select regnode types in the execution phase. It is also heavily used in the code for dumping the regexp program for debugging. 🐛
There are a selection of macros which can be used to compute this as efficiently as possible depending on the circumstances. The canonical macro is REGNODE_AFTER(), which is the most powerful and should handle any case we have, but is also potentially the slowest. There are two additional macros for the special case that you KNOW the current regnode size is constant, and you know its type or opcode. In which case you can use REGNODE_AFTER_opcode() or REGNODE_AFTER_type(). ⚡
...and that's only the first bite of this whole thing, bestie — keep reading, there's so much more. 💖
In older versions of the regex engine REGNODE_AFTER() was called NEXTOPER but literally everyone found that confusing so it got renamed 💅. There's also a REGNODE_BEFORE(), but bestie, it is unsafe and should not be used in new code. Not me warning you twice. 🙈
🌈 regnext 🌈
Okay so this one is the regnode you reach by jumping forward by the value of the NEXT_OFF() member of the regnode — or, in a few cases for longer jumps, by the arg1 field of the regnode_1 structure. The subroutine regnext() handles all of that transparently for you, no cap. ✨
In the majority of cases the regnext for a regnode is the regnode that should be executed after the current one has successfully matched, but in some cases that may not be true. In loop control and branch control regnode types the regnext may signify something special: for BRANCH nodes, regnext is the next BRANCH that should be executed if the current one fails execution, and some loop control regnodes set the regnext to be the end of the loop so they can jump to their cleanup if the current iteration fails to match. It's giving "it depends". 💀
Most regnode types do not create a branch in the execution flow, and leaving aside optimizations the two concepts of "next" are the same. For instance the regnext and REGNODE_AFTER of a SBOL opcode are the same during compilation phase. The main place this is not true is BRANCH regnodes where the REGNODE_AFTER represents the start of the pattern in the branch and the regnext represents the linkage to the next BRANCH should this one fail to match, or 0 if it is the last branch. The looping logic for quantifiers also makes similar use of the distinction between the two types, with REGNODE_AFTER being the inside of the loop construct, and the regnext pointing at the end of the loop. 🌀
During compilation the engine may not know what the regnext is for a given node, so during compilation regnext is only used where it must be used and is known to be correct. At the very end of the compilation phase we walk the regex program and correct the regnext data as appropriate, and also perform various optimizations which may result in regnodes that were required during construction becoming redundant, or we may replace a large regnode with a much smaller one and filling in the gap with OPTIMIZED regnodes. So like, we might start with something like this:
BRANCH
EXACT "foo"
BRANCH
EXACT "bar"
EXACT "!"and replace it with something like:
TRIE foo|bar
OPTIMIZED
OPTIMIZED
OPTIMIZED
EXACT "!"the REGNODE_AFTER for the TRIE node would be an OPTIMIZED regnode, and in theory the regnext would be the same as the REGNODE_AFTER. But bestie, it would be so inefficient to execute the OPTIMIZED regnode as a noop three times, so the optimizer fixes the regnext so such nodes are skipped during execution phase. glow up 🫶
During execution phases we use the regnext() almost exclusively, and only use REGNODE_AFTER in special cases where it has a well defined meaning for a given regnode type. For instance /x+/ results in
PLUS
EXACT "x"
ENDthe regnext of the PLUS regnode is the END regnode, and the REGNODE_AFTER of the PLUS regnode is the EXACT regnode. The regnext and REGNODE_AFTER of the EXACT regnode is the END regnode. Obsessed with this tbh. 💖
✨ Process Overview ✨
Broadly speaking, performing a match of a string against a pattern involves the following steps:
- A. Compilation
- Parsing
- Peep-hole optimisation and analysis
- B. Execution
- Start position and no-match optimisations
- Program execution
Where these steps occur in the actual execution of a perl program is determined by whether the pattern involves interpolating any string variables. If interpolation occurs, then compilation happens at run time. If it does not, then compilation is performed at compile time. (The /o modifier changes this, as does qr// to a certain extent.) The engine doesn't really care that much. 🤷♀️
🎀 Compilation 🎀
This code resides primarily in regcomp.c, along with the header files regcomp.h, regexp.h and regnodes.h.
Compilation starts with pregcomp(), which is mostly an initialisation wrapper which farms work out to two other routines for the heavy lifting: the first is reg(), which is the start point for parsing; the second, study_chunk(), is responsible for optimisation. 💪
Initialisation in pregcomp() mostly involves the creation and data-filling of a special structure, RExC_state_t (defined in regcomp.c). Almost all internally-used routines in regcomp.h take a pointer to one of these structures as their first argument, with the name pRExC_state. This structure is used to store the compilation state and contains many fields. Likewise there are many macros which operate on this variable: anything that looks like RExC_xxxx is a macro that operates on this pointer/structure. iykyk 🌸
reg() is the start of the parse process. It is responsible for parsing an arbitrary chunk of pattern up to either the end of the string, or the first closing parenthesis it encounters in the pattern. This means it can be used to parse the top-level regex, or any section inside of a grouping parenthesis. It also handles the "special parens" that perl's regexes have. For instance when parsing /x(?:foo)y/, reg() will at one point be called to parse from the ? symbol up to and including the ). 🦋
Additionally, reg() is responsible for parsing the one or more branches from the pattern, and for "finishing them off" by correctly setting their next pointers. In order to do the parsing, it repeatedly calls out to regbranch(), which is responsible for handling up to the first | symbol it sees.
regbranch() in turn calls regpiece() which handles "things" followed by a quantifier. In order to parse the "things", regatom() is called. This is the lowest level routine, which parses out constant strings, character classes, and the various special symbols like $. If regatom() encounters a ( character it in turn calls reg(). 🔁 (we love a recursion moment)
There used to be two main passes involved in parsing, the first to calculate the size of the compiled program, and the second to actually compile it. But now there is only one main pass, with an initial crude guess based on the length of the input pattern, which is increased if necessary as parsing proceeds, and afterwards, trimmed to the actual amount used. glow up, algorithm ✨
However, it may happen that parsing must be restarted at the beginning when various circumstances occur along the way. An example is if the program turns out to be so large that there are jumps in it that won't fit in the normal 16 bits available. There are two special regops that can hold bigger jump destinations, BRANCHJ and LONGBRANCH. The parse is restarted, and these are used instead of the normal shorter ones. Whenever restarting the parse is required, the function returns failure and sets a flag as to what needs to be done. This is passed up to the top level routine which takes the appropriate action and restarts from scratch. In the case of needing longer jumps, the RExC_use_BRANCHJ flag is set in the RExC_state_t structure, which the functions know to inspect before deciding how to do branches. 😤
In most instances, the function that discovers the issue sets the causal flag and returns failure immediately. "Parsing complications" contains an explicit example of how this works. In other cases, such as a forward reference to a numbered parenthetical grouping, we need to finish the parse to know if that numbered grouping actually appears in the pattern. In those cases, the parse is just redone at the end, with the knowledge of how many groupings occur in it. rerun it, bestie 💅
The routine regtail() is called by both reg() and regbranch() in order to "set the tail pointer" correctly. When executing and we get to the end of a branch, we need to go to the node following the grouping parens. When parsing, however, we don't know where the end will be until we get there, so when we do we must go back and update the offsets as appropriate. regtail is used to make this easier. 🧵
A subtlety of the parsing process means that a regex like /foo/ is originally parsed into an alternation with a single branch. It is only afterwards that the optimiser converts single branch alternations into the simpler form. iconic behaviour honestly 👑
🌟 Parse Call Graph and a Grammar 🌟
The call graph looks like this:
reg() # parse a top level regex, or inside of
# parens
regbranch() # parse a single branch of an alternation
regpiece() # parse a pattern followed by a quantifier
regatom() # parse a simple pattern
regclass() # used to handle a class
reg() # used to handle a parenthesised
# subpattern
....
...
regtail() # finish off the branch
...
regtail() # finish off the branch sequence. Tie each
# branch's tail to the tail of the
# sequence
# (NEW) In Debug mode this is
# regtail_study().A grammar form might be something like this:
atom : constant | class
quant : '*' | '+' | '?' | '{min,max}'
_branch: piece
| piece _branch
| nothing
branch: _branch
| _branch '|' branch
group : '(' branch ')'
_piece: atom | group
piece : _piece
| _piece quant💜 Parsing complications 💜
The implication of the above description is that a pattern containing nested parentheses will result in a call graph which cycles through reg(), regbranch(), regpiece(), regatom(), reg(), regbranch() etc multiple times, until the deepest level of nesting is reached. All the above routines return a pointer to a regnode, which is usually the last regnode added to the program. However, one complication is that reg() returns NULL for parsing (?:) syntax for embedded modifiers, setting the flag TRYAGAIN. The TRYAGAIN propagates upwards until it is captured, in some cases by regatom(), but otherwise unconditionally by regbranch(). Hence it will never be returned by regbranch() to reg(). This flag permits patterns such as (?i)+ to be detected as errors (Quantifier follows nothing in regex; marked by <-- HERE in m/(?i)+ <-- HERE /). 🚨
yeah bestie, that error message is not a vibe. but at least it tells you exactly where the problem is 💀
Another lil plot twist: the way your program gets stored is different depending on whether it needs to hold Unicode, and honestly? You don't always know that until you're already halfway through parsing. 😭 The Unicode version of the program is bigger, and it can't be matched as efficiently. (Peep "Unicode and Localisation Support" below if you're curious why.)
So if the pattern has literal Unicode in it, okay slay, obviously the program needs to store Unicode. But otherwise the parser is optimistic and assumes it can use the more efficient representation, and it starts sizing things on that basis. The drama: if it then bumps into something that must be stored as Unicode — like an "\x{...}" escape sequence representing a character literal — then every single size it already calculated has to be redone, this time with values appropriate for the Unicode representation. 🙈
That's another moment where parsing has to be restarted, and it happens immediately. The function returns failure and sets the flag "RESTART_UTF8" (wrapped up by using the macro "REQUIRE_UTF8"). That restart request gets passed up the call chain in basically the same way, until it's "caught" in Perl_re_op_compile(), which marks the pattern as containing Unicode and restarts the sizing pass. Also possible: constructions inside run-time code blocks turn out to need the Unicode representation — that's signalled by S_compile_runtime_code() returning false to Perl_re_op_compile().
Fun history fact 👑 — this restart used to be implemented with a "longjmp" in regatom() back to a "setjmp" in Perl_re_op_compile(), but that turned out to be problematic because the latter is a big ol' function with tons of automatic variables, and they interact badly with the emergent control flow of "setjmp". Not it. 💀
🔍 Debug Output 🔍
Starting in the 5.9.x development version of perl you can use re Debug => 'PARSE' to see trace info about the parse process. We'll start with some simple patterns and build up to the complex ones, bestie, don't worry.
So when we parse /foo/ we see something like the table below. The left side shows what's being parsed, and the number tells you where the next regop would go. The stuff on the right is the trace output of the graph. The names are chosen to be short so it's less dense on your screen. 'tsdy' is a special form of regtail() that does some extra analysis.
>foo< 1 reg
brnc
piec
atom
>< 4 tsdy~ EXACT <foo> (EXACT) (1)
~ attach to END (3) offset to 2The resulting program then looks like this:
1: EXACT <foo>(3)
3: END(0)So even though we parsed out a branch and a piece, it was ultimately only an atom. The final program shows us how things work: we have an "EXACT" regop followed by an "END" regop. The number in parens tells you where the "regnext" of the node goes. The "regnext" of an "END" regop is unused, because "END" regops mean we matched successfully 🎉. The number on the left is the position of the regop in the regnode array.
Now let's try a harder pattern. Add a quantifier, so now we've got /foo+/. You'll see that regbranch() calls regpiece() twice.
>foo+< 1 reg
brnc
piec
atom
>o+< 3 piec
atom
>< 6 tail~ EXACT <fo> (1)
7 tsdy~ EXACT <fo> (EXACT) (1)
~ PLUS (END) (3)
~ attach to END (6) offset to 3And we end up with the program:
1: EXACT <fo>(3)
3: PLUS(6)
4: EXACT <o>(0)
6: END(0)Now here's a special case 👀. The "EXACT" regop has a "regnext" of 0. That's because if it matches, it should try to match itself again. The "PLUS" regop handles the actual failure of the "EXACT" regop and acts appropriately — going to regnode 6 if the "EXACT" matched at least once, or failing if it didn't.
Now for something much more complex: /x(?:foo*|b[a][rR])(foo|bar)$/
>x(?:foo*|b... 1 reg
brnc
piec
atom
>(?:foo*|b[... 3 piec
atom
>?:foo*|b[a... reg
>foo*|b[a][... brnc
piec
atom
>o*|b[a][rR... 5 piec
atom
>|b[a][rR])... 8 tail~ EXACT <fo> (3)
>b[a][rR])(... 9 brnc
10 piec
atom
>[a][rR])(f... 12 piec
atom
>a][rR])(fo... clas
>[rR])(foo|... 14 tail~ EXACT <b> (10)
piec
atom
>rR])(foo|b... clas
>)(foo|bar)... 25 tail~ EXACT <a> (12)
tail~ BRANCH (3)
26 tsdy~ BRANCH (END) (9)
~ attach to TAIL (25) offset to 16
tsdy~ EXACT <fo> (EXACT) (4)
~ STAR (END) (6)
~ attach to TAIL (25) offset to 19
tsdy~ EXACT <b> (EXACT) (10)
~ EXACT <a> (EXACT) (12)
~ ANYOF[Rr] (END) (14)
~ attach to TAIL (25) offset to 11
>(foo|bar)$< tail~ EXACT <x> (1)
piec
atom
>foo|bar)$< reg
28 brnc
piec
atom
>|bar)$< 31 tail~ OPEN1 (26)
>bar)$< brnc
32 piec
atom
>)$< 34 tail~ BRANCH (28)
36 tsdy~ BRANCH (END) (31)
~ attach to CLOSE1 (34) offset to 3
tsdy~ EXACT <foo> (EXACT) (29)
~ attach to CLOSE1 (34) offset to 5
tsdy~ EXACT <bar> (EXACT) (32)
~ attach to CLOSE1 (34) offset to 2
>$< tail~ BRANCH (3)
~ BRANCH (9)
~ TAIL (25)
piec
atom
>< 37 tail~ OPEN1 (26)
~ BRANCH (28)
~ BRANCH (31)
~ CLOSE1 (34)
38 tsdy~ EXACT <x> (EXACT) (1)
~ BRANCH (END) (3)
~ BRANCH (END) (9)
~ TAIL (END) (25)
~ OPEN1 (END) (26)
~ BRANCH (END) (28)
~ BRANCH (END) (31)
~ CLOSE1 (END) (34)
~ EOL (END) (36)
~ attach to END (37) offset to 1Resulting in the program:
1: EXACT <x>(3)
3: BRANCH(9)
4: EXACT <fo>(6)
6: STAR(26)
7: EXACT <o>(0)
9: BRANCH(25)
10: EXACT <ba>(14)
12: OPTIMIZED (2 nodes)
14: ANYOF[Rr](26)
25: TAIL(26)
26: OPEN1(28)
28: TRIE-EXACT(34)
[StS:1 Wds:2 Cs:6 Uq:5 #Sts:7 Mn:3 Mx:3 Stcls:bf]
<foo>
<bar>
30: OPTIMIZED (4 nodes)
34: CLOSE1(36)
36: EOL(37)
37: END(0)Here we've got a much more complex program, with all kinds of optimisations in play. At regnode 10 we see an example where a character class with only one character in it got turned into an "EXACT" node. We can also see where an entire alternation got turned into a "TRIE-EXACT" node. As a consequence, some of the regnodes have been marked as optimised away. And check it — the "$" symbol has been converted into an "EOL" regop, a special piece of code that looks for "\n" or the end of the string. ✨
The next pointer for "BRANCH"es is interesting because it points at where execution should go if the branch fails. When executing, if the engine tries to traverse from a branch to a "regnext" that isn't a branch, then the engine knows the entire set of branches has failed. 💅
🧚 Peep-hole Optimisation and Analysis 🧚
The regex engine can be a weighty tool to wield, ngl. On long strings and complex patterns it can end up doing a lot of work to find a match, and even more to decide that no match is possible. Consider a situation like this pattern:
'ababababababababababab' =~ /(a|b)*z/The "(a|b)*" part can match at every char in the string, and then fail every time, because there is no "z" in the string. So obviously we can avoid using the regex engine unless there's a "z" in the string. Iconic behaviour honestly.
Likewise in a pattern like:
/foo(\w+)bar/In this case we know the string must contain a "foo" which must be followed by "bar". We can use Fast Boyer-Moore matching as implemented in fbm_instr() to find the location of these strings. If they don't exist, then we don't need to resort to the much more expensive regex engine. Even better: if they do exist, then we can use their positions to reduce the search space that the regex engine needs to cover to figure out if the entire pattern matches. 🫶
There are various aspects of the pattern that can be used to facilitate optimisations along these lines:
- anchored fixed strings
- floating fixed strings
- minimum and maximum length requirements
- start class
- Beginning/End of line positions
okay so there's another glow-up the regex engine can pull off, and it's called post-parse "peep-hole" optimisation — basically it peeks through a little hole at the parsed tree and swaps clunky constructs for sleeker ones. 💅
The TAIL regops? Those are the ones that get used during parsing to mark the end of branches and the end of groups, and they are literally the poster children for this. During construction they're just place-holders, and they "always match", which means they can be "optimised away" — you just make everything that points at the TAIL point at whatever TAIL itself points to, and boom, the node is "skipped" entirely. iconic behaviour, no notes. ✨
Another one that can happen is "EXACT merging", where two consecutive EXACT nodes get smooshed together into one single regop. And then there's the even more feral version: a branch sequence shaped like EXACT BRANCH ... EXACT can get converted into a TRIE-EXACT regop. ate. 👑
All of this lives in the routine study_chunk(), which uses a special structure called scan_data_t to stash the analysis it's done, and does the "peep-hole" optimisations as it goes along.
heads up bestie: the code involved in study_chunk() is extremely cryptic. Be careful. :-)
⚡ EXECUTION ⚡
So running a regex is generally a two-phase situation, and it's giving very much get ready then go. 💖
- find the start point in the string — like, where should we even be matching from?
- run the regop interpreter — the actual matching.
If we can tell there's no valid start point at all, then bestie we do not even bother running the interpreter. Not today. 🛑 And same energy in reverse: if the analysis phase already told us we can't detect a short-cut to the start position, we skip straight to the interpreter.
The two entry points are re_intuit_start() and pregexec(). These two have a somewhat incestuous relationship ngl — their functions overlap, and pregexec() may even call re_intuit_start() on its own. Regardless, other parts of the perl source code may call into either one, or both. messy but we move. 🙈
Okay so execution of the interpreter itself used to be recursive, but thanks to the efforts of Dave Mitchell in the 5.9.x development track, that changed: now an internal stack is maintained on the heap and the routine is fully iterative. slay. This can make it tricky though, because the code is quite conservative about what state it stores — with the result that two consecutive lines in the code can actually be running in totally different contexts due to the simulated recursion. it's giving illusion magic. 🪄
🔍 Start position and no-match optimisations 🔍
re_intuit_start() is the one responsible for handling start points and no-match optimisations, as determined by the results of the analysis that study_chunk() did (the stuff described in "Peep-hole Optimisation and Analysis").
The basic structure of this routine is:
- try to find the start- and/or end-points of where the pattern could match
- make sure the string is long enough to match the pattern
It tries to use more efficient methods over less efficient methods, and may involve considerable cross-checking of constraints to find the place in the string that matches. For instance it may try to determine that a given fixed string must be not only present but a certain number of chars before the end of the string, or whatever. 🎀
It calls several other routines, like:
fbm_instr()— which does Fast Boyer Moore matchingfind_byclass()— which is responsible for finding the start using the first mandatory regop in the program
When the optimisation criteria have been satisfied, reg_try() is called to perform the match. 💅
🏃 Program execution 🏃
pregexec() is the main entry point for running a regex. It contains support for:
- initialising the regex interpreter's state
- running
re_intuit_start()if needed - running the interpreter on the string from various start positions as needed
When it is necessary to use the regex interpreter, pregexec() calls regtry(). ✨
regtry() is the entry point into the regex interpreter. It expects as arguments:
- a pointer to a
regmatch_infostructure - a pointer to a string
It returns an integer 1 for success and a 0 for failure. It's basically a set-up wrapper around regmatch().
regmatch is the main "recursive loop" of the interpreter. It's basically a giant switch statement that implements a state machine, where the possible states are the regops themselves, plus a number of additional intermediate and failure states. A few of the states are implemented as subroutines but the bulk are inline code. 💀
✨ MISCELLANEOUS ✨
🌈 Unicode and Localisation Support
Okay bestie, buckle up, because this is the part where things get spicy. 🥵
When you're dealing with strings that have characters an eight-bit character set literally cannot represent, perl switches to an internal representation that's a permissive version of Unicode's UTF-8 encoding[2]. So: single bytes handle the ASCII characters, and sequences of two or more bytes handle everything else. (If you wanna know more about how UTF-8 and perl's own encoding utf8 relate to each other, go read perlunitut. For our purposes here though, the difference? Not important, don't stress it.) 🫶
And no matter how you slice it, Unicode support is going to be a pain in a regex engine. 💀 Tricks that are totally fine when you only have 256 possible characters often just... don't scale to the size of the UTF-8 character set. Stuff you can take for granted with ASCII may be flat-out false with Unicode. For example: in ASCII it's safe to assume that sizeof(char1) == sizeof(char2), but in UTF-8? It absolutely is not. Unicode case folding is vastly more complex than the simple little rules of ASCII, and even when you're not using Unicode at all, just localised single byte encodings, things get tricky — for example, LATIN SMALL LETTER SHARP S (U+00DF, ß) should match 'SS' in localised case-insensitive matching. Not me side-eyeing ß. 😭
Making it even worse: UTF-8 support was a later addition to the regex engine (same as it was to perl), and that necessarily made everything a lot more complicated. Obviously it's easier to design a regex engine with Unicode in mind from the very beginning than it is to retrofit it onto one that wasn't. Retrofit era, we've all been there. 💅
Nearly all regops that involve looking at the input string have two cases: one for UTF-8, and one not. And honestly it's often more complex than just two, because the pattern may be UTF-8 as well. Layers on layers on layers. 🧚
Care must be taken when making changes to make sure that you handle UTF-8 properly, both at compile time and at execution time, including when the string and pattern are mismatched. No cap, be careful.
🎀 Base Structures
The regexp structure described in perlreapi is common to all regex engines. Two of its fields are meant for the private use of the regex engine that compiled the pattern. Those are the intflags and pprivate members. The pprivate is a void pointer to an arbitrary structure, and its use and management is entirely the responsibility of the compiling engine. perl will never modify either of these values. Never ever. 🙅
In the case of the stock engine, the structure that pprivate points to is called regexp_internal. Its pprivate and intflags fields contain data specific to each engine.
There are two structures used to store a compiled regular expression. One — the regexp structure described in perlreapi — is populated by the engine currently being used, and some of its fields are read by perl to implement things like the stringification of qr//.
The other structure is pointed to by the regexp struct's pprivate, and it is, in addition to intflags in that same struct, considered to be the property of the regex engine which compiled the regular expression. 🫡
The regexp structure contains all the data that perl needs to be aware of to properly work with the regular expression. It includes data about optimisations that perl can use to determine if the regex engine should really be used, plus various other control info needed to properly execute patterns in various contexts — like, is the pattern anchored in some way? What flags were used during the compile? Does the program contain special constructs that perl needs to be aware of? All that.
On top of that, it contains two fields intended for the private use of the regex engine that compiled the pattern: the intflags and pprivate members. The pprivate is a void pointer to an arbitrary structure whose use and management is the responsibility of the compiling engine. perl will never modify either of these values. (Yes, bestie, I know I said this already — the page says it twice, so I say it twice. Consistency icon. 👑)
As mentioned earlier, for the default engines, the pprivate will be a pointer to a regexp_internal structure, which holds the compiled program and any additional data that is private to the regex engine implementation.
💜 Perl's pprivate structure
The following structure is used as the pprivate struct by perl's regex engine. Since it's specific to perl, it's only of curiosity value to other engine implementations. Like, look but don't touch energy.
typedef struct regexp_internal {
regnode *regstclass;
struct reg_data *data;
struct reg_code_blocks *code_blocks;
U32 proglen;
U32 name_list_idx;
regnode program[1];
} regexp_internal;Description of the attributes is as follows, mwah:
regstclass— Special regop that is used byre_intuit_start()to check if a pattern can match at a certain position. For instance, if the regex engine knows that the pattern must start with a'Z', then it can scan the string until it finds one and then launch the regex engine from there. The routine that handles this is calledfind_by_class(). Sometimes this field points at a regop embedded in the program, and sometimes it points at an independent synthetic regop that has been constructed by the optimiser. Multitasking king. ✨data— This field points at areg_datastructure, which is defined as follows:
struct reg_data {
U32 count;
U8 *what;
void* data[1];
};This structure is used for handling data structures that the regex engine needs to handle specially during a clone or free operation on the compiled product. Each element in the data array has a corresponding element in the what array. During compilation, regops that need special structures stored will add an element to each array using the add_data() routine, and then store the index in the regop.
In modern perls the 0th element of this structure is reserved and is NEVER used to store anything of use. This is to allow things that need to index into this array to represent "no value". Respect the reserved slot. 🚫
code_blocks— This optional structure is used to manage(?{})constructs in the pattern. It is made up of the following structures:
/* record the position of a (?{...}) within a pattern */
struct reg_code_block {
STRLEN start;
STRLEN end;
OP *block;
REGEXP *src_regex;
};
/* array of reg_code_block's plus header info */
struct reg_code_blocks {
int refcnt; /* we may be pointed to from a regex
and from the savestack */
int count; /* how many code blocks */
struct reg_code_block *cb; /* array of reg_code_block's */
};proglen— Stores the length of the compiled program in units of regops. Short, sweet, measured. 📏name_list_idx— This is the index into thedataarray where an AV is stored that contains the names of any named capture buffers in the pattern, should there be any. This is only used in the debugging version of the regex engine and whenRXp_PAREN_NAMES(prog)is true. It will be0if there is no such data. Falsy queen. 🌸program— Compiled program. Inlined into the structure so the entire struct can be treated as a single blob. Blob era. 🦋
✨ SEE ALSO ✨
- perlreapi
- perlre
- perlunitut
🌟 AUTHOR 🌟
by Yves Orton, 2006. Iconic behaviour, honestly. 👑
With excerpts from Perl, and contributions and suggestions from Ronald J. Kimball, Dave Mitchell, Dominic Dunlop, Mark Jason Dominus, Stephen McCamant, and David Landgren. The whole group chat showed up for this one. 🫶
Now maintained by Perl 5 Porters.
💖 LICENCE 💖
Same terms as Perl. Simple as that, periodt.
🔗 REFERENCES 🔗
okay bestie that's the whole entire page, we made it to the end and honestly? we ate. 💅✨ this was perl v5.42.3, dated 2026-08-03, and the page itself is PERLREGUTS(1). go forth and regex responsibly, ily, mwah 🎀🦋💖
📦 where this page comes from
- fedora 43: perl-doc 5.42.3-524.fc43 · (GPL-1.0-or-later OR Artistic-1.0-Perl) AND Unicode-3.0 AND LicenseRef-Fedora-UltraPermissive
- fedora 44: perl-doc 5.42.3-525.fc44 · (GPL-1.0-or-later OR Artistic-1.0-Perl) AND Unicode-3.0 AND LicenseRef-Fedora-UltraPermissive
- 💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash on : a language model's rewrite of the original, under its terms