1=head1 NAME 2 3perlreguts - Description of the Perl regular expression engine. 4 5=head1 DESCRIPTION 6 7This document is an attempt to shine some light on the guts of the regex 8engine and how it works. The regex engine represents a significant chunk 9of the perl codebase, but is relatively poorly understood. This document 10is a meagre attempt at addressing this situation. It is derived from the 11author's experience, comments in the source code, other papers on the 12regex engine, feedback on the perl5-porters mail list, and no doubt other 13places as well. 14 15B<NOTICE!> It should be clearly understood that the behavior and 16structures discussed in this represents the state of the engine as the 17author understood it at the time of writing. It is B<NOT> an API 18definition, it is purely an internals guide for those who want to hack 19the regex engine, or understand how the regex engine works. Readers of 20this document are expected to understand perl's regex syntax and its 21usage in detail. If you want to learn about the basics of Perl's 22regular expressions, see L<perlre>. And if you want to replace the 23regex engine with your own, see L<perlreapi>. 24 25=head1 OVERVIEW 26 27=head2 A quick note on terms 28 29There is some debate as to whether to say "regexp" or "regex". In this 30document we will use the term "regex" unless there is a special reason 31not to, in which case we will explain why. 32 33When speaking about regexes we need to distinguish between their source 34code form and their internal form. In this document we will use the term 35"pattern" when we speak of their textual, source code form, and the term 36"program" when we speak of their internal representation. These 37correspond to the terms I<S-regex> and I<B-regex> that Mark Jason 38Dominus employs in his paper on "Rx" ([1] in L</REFERENCES>). 39 40=head2 What is a regular expression engine? 41 42A regular expression engine is a program that takes a set of constraints 43specified in a mini-language, and then applies those constraints to a 44target string, and determines whether or not the string satisfies the 45constraints. See L<perlre> for a full definition of the language. 46 47In less grandiose terms, the first part of the job is to turn a pattern into 48something the computer can efficiently use to find the matching point in 49the string, and the second part is performing the search itself. 50 51To do this we need to produce a program by parsing the text. We then 52need to execute the program to find the point in the string that 53matches. And we need to do the whole thing efficiently. 54 55=head2 Structure of a Regexp Program 56 57=head3 High Level 58 59Although it is a bit confusing and some people object to the terminology, it 60is worth taking a look at a comment that has 61been in F<regexp.h> for years: 62 63I<This is essentially a linear encoding of a nondeterministic 64finite-state machine (aka syntax charts or "railroad normal form" in 65parsing technology).> 66 67The term "railroad normal form" is a bit esoteric, with "syntax 68diagram/charts", or "railroad diagram/charts" being more common terms. 69Nevertheless it provides a useful mental image of a regex program: each 70node can be thought of as a unit of track, with a single entry and in 71most cases a single exit point (there are pieces of track that fork, but 72statistically not many), and the whole forms a layout with a 73single entry and single exit point. The matching process can be thought 74of as a car that moves along the track, with the particular route through 75the system being determined by the character read at each possible 76connector point. A car can fall off the track at any point but it may 77only proceed as long as it matches the track. 78 79Thus the pattern C</foo(?:\w+|\d+|\s+)bar/> can be thought of as the 80following chart: 81 82 [start] 83 | 84 <foo> 85 | 86 +-----+-----+ 87 | | | 88 <\w+> <\d+> <\s+> 89 | | | 90 +-----+-----+ 91 | 92 <bar> 93 | 94 [end] 95 96The truth of the matter is that perl's regular expressions these days are 97much more complex than this kind of structure, but visualising it this way 98can help when trying to get your bearings, and it matches the 99current implementation pretty closely. 100 101To be more precise, we will say that a regex program is an encoding 102of a graph. Each node in the graph corresponds to part of 103the original regex pattern, such as a literal string or a branch, 104and has a pointer to the nodes representing the next component 105to be matched. Since "node" and "opcode" already have other meanings in the 106perl source, we will call the nodes in a regex program "regops". 107 108The program is represented by an array of C<regnode> structures, one or 109more of which represent a single regop of the program. Struct 110C<regnode> is the smallest struct needed, and has a field structure which is 111shared with all the other larger structures. (Outside this document, the term 112"regnode" is sometimes used to mean "regop", which could be confusing.) 113 114=for apidoc Cyh||regnode 115 116The "next" pointers of all regops except C<BRANCH> implement concatenation; 117a "next" pointer with a C<BRANCH> on both ends of it is connecting two 118alternatives. [Here we have one of the subtle syntax dependencies: an 119individual C<BRANCH> (as opposed to a collection of them) is never 120concatenated with anything because of operator precedence.] 121 122The operand of some types of regop is a literal string; for others, 123it is a regop leading into a sub-program. In particular, the operand 124of a C<BRANCH> node is the first regop of the branch. 125 126B<NOTE>: As the railroad metaphor suggests, this is B<not> a tree 127structure: the tail of the branch connects to the thing following the 128set of C<BRANCH>es. It is a like a single line of railway track that 129splits as it goes into a station or railway yard and rejoins as it comes 130out the other side. 131 132=head3 Regops 133 134The base structure of a regop is defined in F<regexp.h> as follows: 135 136 struct regnode { 137 U8 flags; /* Various purposes, sometimes overridden */ 138 U8 type; /* Opcode value as specified by regnodes.h */ 139 U16 next_off; /* Offset in size regnode */ 140 }; 141 142Other larger C<regnode>-like structures are defined in F<regcomp.h>. They 143are almost like subclasses in that they have the same fields as 144C<regnode>, with possibly additional fields following in 145the structure, and in some cases the specific meaning (and name) 146of some of base fields are overridden. The following is a more 147complete description. 148 149=over 4 150 151=item C<regnode_1> 152 153=item C<regnode_2> 154 155C<regnode_1> structures have the same header, followed by a single 156four-byte argument; C<regnode_2> structures contain two two-byte 157arguments instead: 158 159 regnode_1 U32 arg1; 160 regnode_2 U16 arg1; U16 arg2; 161 162=item C<regnode_string> 163 164C<regnode_string> structures, used for literal strings, follow the header 165with a one-byte length and then the string data. Strings are padded on 166the tail end with zero bytes so that the total length of the node is a 167multiple of four bytes: 168 169 regnode_string char string[1]; 170 U8 str_len; /* overrides flags */ 171 172=item C<regnode_charclass> 173 174Bracketed character classes are represented by C<regnode_charclass> 175structures, which have a four-byte argument and then a 32-byte (256-bit) 176bitmap indicating which characters in the Latin1 range are included in 177the class. 178 179 regnode_charclass U32 arg1; 180 char bitmap[ANYOF_BITMAP_SIZE]; 181 182Various flags whose names begin with C<ANYOF_> are used for special 183situations. Above Latin1 matches and things not known until run-time 184are stored in L</Perl's pprivate structure>. 185 186=item C<regnode_charclass_posixl> 187 188There is also a larger form of a char class structure used to represent 189POSIX char classes under C</l> matching, 190called C<regnode_charclass_posixl> which has an 191additional 32-bit bitmap indicating which POSIX char classes 192have been included. 193 194 regnode_charclass_posixl U32 arg1; 195 char bitmap[ANYOF_BITMAP_SIZE]; 196 U32 classflags; 197 198=back 199 200F<regnodes.h> defines an array called C<PL_regnode_arg_len[]> which gives the size 201of each opcode in units of C<size regnode> (4-byte). A macro is used 202to calculate the size of an C<EXACT> node based on its C<str_len> field. 203 204The regops are defined in F<regnodes.h> which is generated from 205F<regcomp.sym> by F<regcomp.pl>. Currently the maximum possible number 206of distinct regops is restricted to 256, with about a quarter already 207used. 208 209A set of macros makes accessing the fields 210easier and more consistent. These include C<OP()>, which is used to determine 211the type of a C<regnode>-like structure; C<NEXT_OFF()>, which is the offset to 212the next node (more on this later); C<ARG()>, C<ARG1()>, C<ARG2()>, C<ARG_SET()>, 213and equivalents for reading and setting the arguments; and C<STR_LEN()>, 214C<STRING()> and C<OPERAND()> for manipulating strings and regop bearing 215types. 216 217=head3 What regnode is next? 218 219There are two distinct concepts of "next regnode" in the regex engine, 220and it is important to keep them distinct in your thinking as they 221overlap conceptually in many places, but where they don't overlap the 222difference is critical. For the majority of regnode types the two 223concepts are (nearly) identical in practice. The two types are 224C<REGNODE_AFTER> which is used heavily during compilation but only 225occasionally during execution and C<regnext> which is used heavily 226during execution, and only occasionally during compilation. 227 228=over 4 229 230=item "REGNODE_AFTER" 231 232This is the "positionally next regnode" in the compiled regex program. 233For the smaller regnode types it is C<regnode_ptr+1> under the hood, but 234as regnode sizes vary and can change over time we offer macros which 235hide the gory details. 236 237It is heavily used in the compiler phase but is only used by a few 238select regnode types in the execution phase. It is also heavily used in 239the code for dumping the regexp program for debugging. 240 241There are a selection of macros which can be used to compute this as 242efficiently as possible depending on the circumstances. The canonical 243macro is C<REGNODE_AFTER()>, which is the most powerful and should handle 244any case we have, but is also potentially the slowest. There are two 245additional macros for the special case that you KNOW the current regnode 246size is constant, and you know its type or opcode. In which case you can 247use C<REGNODE_AFTER_opcode()> or C<REGNODE_AFTER_type()>. 248 249In older versions of the regex engine C<REGNODE_AFTER()> was called 250C<NEXTOPER> but this was found to be confusing and it was renamed. There 251is also a C<REGNODE_BEFORE()>, but it is unsafe and should not be used 252in new code. 253 254=item "regnext" 255 256This is the regnode which can be reached by jumping forward by the value 257of the C<NEXT_OFF()> member of the regnode, or in a few cases for longer 258jumps by the C<arg1> field of the C<regnode_1> structure. The subroutine 259C<regnext()> handles this transparently. In the majority of cases the 260C<regnext> for a regnode is the regnode which should be executed after the 261current one has successfully matched, but in some cases this may not be 262true. In loop control and branch control regnode types the regnext may 263signify something special, for BRANCH nodes C<regnext> is the 264next BRANCH that should be executed if the current one fails execution, 265and some loop control regnodes set the regnext to be the end of the loop 266so they can jump to their cleanup if the current iteration fails to match. 267 268=back 269 270Most regnode types do not create a branch in the execution flow, and 271leaving aside optimizations the two concepts of "next" are the same. 272For instance the C<regnext> and C<REGNODE_AFTER> of a SBOL opcode are 273the same during compilation phase. The main place this is not true is 274C<BRANCH> regnodes where the C<REGNODE_AFTER> represents the start of 275the pattern in the branch and the C<regnext> represents the linkage to 276the next BRANCH should this one fail to match, or 0 if it is the last 277branch. The looping logic for quantifiers also makes similar use of 278the distinction between the two types, with C<REGNODE_AFTER> being the 279inside of the loop construct, and the C<regnext> pointing at the end 280of the loop. 281 282During compilation the engine may not know what the regnext is for a 283given node, so during compilation C<regnext> is only used where it must 284be used and is known to be correct. At the very end of the compilation 285phase we walk the regex program and correct the regnext data as 286appropriate, and also perform various optimizations which may result in 287regnodes that were required during construction becoming redundant, or 288we may replace a large regnode with a much smaller one and filling in the 289gap with OPTIMIZED regnodes. Thus we might start with something like 290this: 291 292 BRANCH 293 EXACT "foo" 294 BRANCH 295 EXACT "bar" 296 EXACT "!" 297 298and replace it with something like: 299 300 TRIE foo|bar 301 OPTIMIZED 302 OPTIMIZED 303 OPTIMIZED 304 EXACT "!" 305 306the C<REGNODE_AFTER> for the C<TRIE> node would be an C<OPTIMIZED> 307regnode, and in theory the C<regnext> would be the same as the 308C<REGNODE_AFTER>. But it would be inefficient to execute the OPTIMIZED 309regnode as a noop three times, so the optimizer fixes the C<regnext> so 310such nodes are skipped during execution phase. 311 312During execution phases we use the C<regnext()> almost exclusively, and 313only use C<REGNODE_AFTER> in special cases where it has a well defined 314meaning for a given regnode type. For instance /x+/ results in 315 316 PLUS 317 EXACT "x" 318 END 319 320the C<regnext> of the C<PLUS> regnode is the C<END> regnode, and the 321C<REGNODE_AFTER> of the C<PLUS> regnode is the C<EXACT> regnode. The 322C<regnext> and C<REGNODE_AFTER> of the C<EXACT> regnode is the 323C<END> regnode. 324 325=head1 Process Overview 326 327Broadly speaking, performing a match of a string against a pattern 328involves the following steps: 329 330=over 5 331 332=item A. Compilation 333 334=over 5 335 336=item 1. Parsing 337 338=item 2. Peep-hole optimisation and analysis 339 340=back 341 342=item B. Execution 343 344=over 5 345 346=item 3. Start position and no-match optimisations 347 348=item 4. Program execution 349 350=back 351 352=back 353 354 355Where these steps occur in the actual execution of a perl program is 356determined by whether the pattern involves interpolating any string 357variables. If interpolation occurs, then compilation happens at run time. If it 358does not, then compilation is performed at compile time. (The C</o> modifier changes this, 359as does C<qr//> to a certain extent.) The engine doesn't really care that 360much. 361 362=head2 Compilation 363 364This code resides primarily in F<regcomp.c>, along with the header files 365F<regcomp.h>, F<regexp.h> and F<regnodes.h>. 366 367Compilation starts with C<pregcomp()>, which is mostly an initialisation 368wrapper which farms work out to two other routines for the heavy lifting: the 369first is C<reg()>, which is the start point for parsing; the second, 370C<study_chunk()>, is responsible for optimisation. 371 372Initialisation in C<pregcomp()> mostly involves the creation and data-filling 373of a special structure, C<RExC_state_t> (defined in F<regcomp.c>). 374Almost all internally-used routines in F<regcomp.h> take a pointer to one 375of these structures as their first argument, with the name C<pRExC_state>. 376This structure is used to store the compilation state and contains many 377fields. Likewise there are many macros which operate on this 378variable: anything that looks like C<RExC_xxxx> is a macro that operates on 379this pointer/structure. 380 381C<reg()> is the start of the parse process. It is responsible for 382parsing an arbitrary chunk of pattern up to either the end of the 383string, or the first closing parenthesis it encounters in the pattern. 384This means it can be used to parse the top-level regex, or any section 385inside of a grouping parenthesis. It also handles the "special parens" 386that perl's regexes have. For instance when parsing C</x(?:foo)y/>, 387C<reg()> will at one point be called to parse from the "?" symbol up to 388and including the ")". 389 390Additionally, C<reg()> is responsible for parsing the one or more 391branches from the pattern, and for "finishing them off" by correctly 392setting their next pointers. In order to do the parsing, it repeatedly 393calls out to C<regbranch()>, which is responsible for handling up to the 394first C<|> symbol it sees. 395 396C<regbranch()> in turn calls C<regpiece()> which 397handles "things" followed by a quantifier. In order to parse the 398"things", C<regatom()> is called. This is the lowest level routine, which 399parses out constant strings, character classes, and the 400various special symbols like C<$>. If C<regatom()> encounters a "(" 401character it in turn calls C<reg()>. 402 403There used to be two main passes involved in parsing, the first to 404calculate the size of the compiled program, and the second to actually 405compile it. But now there is only one main pass, with an initial crude 406guess based on the length of the input pattern, which is increased if 407necessary as parsing proceeds, and afterwards, trimmed to the actual 408amount used. 409 410However, it may happen that parsing must be restarted at the beginning 411when various circumstances occur along the way. An example is if the 412program turns out to be so large that there are jumps in it that won't 413fit in the normal 16 bits available. There are two special regops that 414can hold bigger jump destinations, BRANCHJ and LONGBRANCH. The parse is 415restarted, and these are used instead of the normal shorter ones. 416Whenever restarting the parse is required, the function returns failure 417and sets a flag as to what needs to be done. This is passed up to the 418top level routine which takes the appropriate action and restarts from 419scratch. In the case of needing longer jumps, the C<RExC_use_BRANCHJ> 420flag is set in the C<RExC_state_t> structure, which the functions know 421to inspect before deciding how to do branches. 422 423In most instances, the function that discovers the issue sets the causal 424flag and returns failure immediately. L</Parsing complications> 425contains an explicit example of how this works. In other cases, such as 426a forward reference to a numbered parenthetical grouping, we need to 427finish the parse to know if that numbered grouping actually appears in 428the pattern. In those cases, the parse is just redone at the end, with 429the knowledge of how many groupings occur in it. 430 431The routine C<regtail()> is called by both C<reg()> and C<regbranch()> 432in order to "set the tail pointer" correctly. When executing and 433we get to the end of a branch, we need to go to the node following the 434grouping parens. When parsing, however, we don't know where the end will 435be until we get there, so when we do we must go back and update the 436offsets as appropriate. C<regtail> is used to make this easier. 437 438A subtlety of the parsing process means that a regex like C</foo/> is 439originally parsed into an alternation with a single branch. It is only 440afterwards that the optimiser converts single branch alternations into the 441simpler form. 442 443=head3 Parse Call Graph and a Grammar 444 445The call graph looks like this: 446 447 reg() # parse a top level regex, or inside of 448 # parens 449 regbranch() # parse a single branch of an alternation 450 regpiece() # parse a pattern followed by a quantifier 451 regatom() # parse a simple pattern 452 regclass() # used to handle a class 453 reg() # used to handle a parenthesised 454 # subpattern 455 .... 456 ... 457 regtail() # finish off the branch 458 ... 459 regtail() # finish off the branch sequence. Tie each 460 # branch's tail to the tail of the 461 # sequence 462 # (NEW) In Debug mode this is 463 # regtail_study(). 464 465A grammar form might be something like this: 466 467 atom : constant | class 468 quant : '*' | '+' | '?' | '{min,max}' 469 _branch: piece 470 | piece _branch 471 | nothing 472 branch: _branch 473 | _branch '|' branch 474 group : '(' branch ')' 475 _piece: atom | group 476 piece : _piece 477 | _piece quant 478 479=head3 Parsing complications 480 481The implication of the above description is that a pattern containing nested 482parentheses will result in a call graph which cycles through C<reg()>, 483C<regbranch()>, C<regpiece()>, C<regatom()>, C<reg()>, C<regbranch()> I<etc> 484multiple times, until the deepest level of nesting is reached. All the above 485routines return a pointer to a C<regnode>, which is usually the last regnode 486added to the program. However, one complication is that reg() returns NULL 487for parsing C<(?:)> syntax for embedded modifiers, setting the flag 488C<TRYAGAIN>. The C<TRYAGAIN> propagates upwards until it is captured, in 489some cases by C<regatom()>, but otherwise unconditionally by 490C<regbranch()>. Hence it will never be returned by C<regbranch()> to 491C<reg()>. This flag permits patterns such as C<(?i)+> to be detected as 492errors (I<Quantifier follows nothing in regex; marked by <-- HERE in m/(?i)+ 493<-- HERE />). 494 495Another complication is that the representation used for the program differs 496if it needs to store Unicode, but it's not always possible to know for sure 497whether it does until midway through parsing. The Unicode representation for 498the program is larger, and cannot be matched as efficiently. (See L</Unicode 499and Localisation Support> below for more details as to why.) If the pattern 500contains literal Unicode, it's obvious that the program needs to store 501Unicode. Otherwise, the parser optimistically assumes that the more 502efficient representation can be used, and starts sizing on this basis. 503However, if it then encounters something in the pattern which must be stored 504as Unicode, such as an C<\x{...}> escape sequence representing a character 505literal, then this means that all previously calculated sizes need to be 506redone, using values appropriate for the Unicode representation. This 507is another instance where the parsing needs to be restarted, and it can 508and is done immediately. The function returns failure, and sets the 509flag C<RESTART_UTF8> (encapsulated by using the macro C<REQUIRE_UTF8>). 510This restart request is propagated up the call chain in a similar 511fashion, until it is "caught" in C<Perl_re_op_compile()>, which marks 512the pattern as containing Unicode, and restarts the sizing pass. It is 513also possible for constructions within run-time code blocks to turn out 514to need Unicode representation., which is signalled by 515C<S_compile_runtime_code()> returning false to C<Perl_re_op_compile()>. 516 517The restart was previously implemented using a C<longjmp> in C<regatom()> 518back to a C<setjmp> in C<Perl_re_op_compile()>, but this proved to be 519problematic as the latter is a large function containing many automatic 520variables, which interact badly with the emergent control flow of C<setjmp>. 521 522=head3 Debug Output 523 524Starting in the 5.9.x development version of perl you can C<< use re 525Debug => 'PARSE' >> to see some trace information about the parse 526process. We will start with some simple patterns and build up to more 527complex patterns. 528 529So when we parse C</foo/> we see something like the following table. The 530left shows what is being parsed, and the number indicates where the next regop 531would go. The stuff on the right is the trace output of the graph. The 532names are chosen to be short to make it less dense on the screen. 'tsdy' 533is a special form of C<regtail()> which does some extra analysis. 534 535 >foo< 1 reg 536 brnc 537 piec 538 atom 539 >< 4 tsdy~ EXACT <foo> (EXACT) (1) 540 ~ attach to END (3) offset to 2 541 542The resulting program then looks like: 543 544 1: EXACT <foo>(3) 545 3: END(0) 546 547As you can see, even though we parsed out a branch and a piece, it was ultimately 548only an atom. The final program shows us how things work. We have an C<EXACT> regop, 549followed by an C<END> regop. The number in parens indicates where the C<regnext> of 550the node goes. The C<regnext> of an C<END> regop is unused, as C<END> regops mean 551we have successfully matched. The number on the left indicates the position of 552the regop in the regnode array. 553 554Now let's try a harder pattern. We will add a quantifier, so now we have the pattern 555C</foo+/>. We will see that C<regbranch()> calls C<regpiece()> twice. 556 557 >foo+< 1 reg 558 brnc 559 piec 560 atom 561 >o+< 3 piec 562 atom 563 >< 6 tail~ EXACT <fo> (1) 564 7 tsdy~ EXACT <fo> (EXACT) (1) 565 ~ PLUS (END) (3) 566 ~ attach to END (6) offset to 3 567 568And we end up with the program: 569 570 1: EXACT <fo>(3) 571 3: PLUS(6) 572 4: EXACT <o>(0) 573 6: END(0) 574 575Now we have a special case. The C<EXACT> regop has a C<regnext> of 0. This is 576because if it matches it should try to match itself again. The C<PLUS> regop 577handles the actual failure of the C<EXACT> regop and acts appropriately (going 578to regnode 6 if the C<EXACT> matched at least once, or failing if it didn't). 579 580Now for something much more complex: C</x(?:foo*|b[a][rR])(foo|bar)$/> 581 582 >x(?:foo*|b... 1 reg 583 brnc 584 piec 585 atom 586 >(?:foo*|b[... 3 piec 587 atom 588 >?:foo*|b[a... reg 589 >foo*|b[a][... brnc 590 piec 591 atom 592 >o*|b[a][rR... 5 piec 593 atom 594 >|b[a][rR])... 8 tail~ EXACT <fo> (3) 595 >b[a][rR])(... 9 brnc 596 10 piec 597 atom 598 >[a][rR])(f... 12 piec 599 atom 600 >a][rR])(fo... clas 601 >[rR])(foo|... 14 tail~ EXACT <b> (10) 602 piec 603 atom 604 >rR])(foo|b... clas 605 >)(foo|bar)... 25 tail~ EXACT <a> (12) 606 tail~ BRANCH (3) 607 26 tsdy~ BRANCH (END) (9) 608 ~ attach to TAIL (25) offset to 16 609 tsdy~ EXACT <fo> (EXACT) (4) 610 ~ STAR (END) (6) 611 ~ attach to TAIL (25) offset to 19 612 tsdy~ EXACT <b> (EXACT) (10) 613 ~ EXACT <a> (EXACT) (12) 614 ~ ANYOF[Rr] (END) (14) 615 ~ attach to TAIL (25) offset to 11 616 >(foo|bar)$< tail~ EXACT <x> (1) 617 piec 618 atom 619 >foo|bar)$< reg 620 28 brnc 621 piec 622 atom 623 >|bar)$< 31 tail~ OPEN1 (26) 624 >bar)$< brnc 625 32 piec 626 atom 627 >)$< 34 tail~ BRANCH (28) 628 36 tsdy~ BRANCH (END) (31) 629 ~ attach to CLOSE1 (34) offset to 3 630 tsdy~ EXACT <foo> (EXACT) (29) 631 ~ attach to CLOSE1 (34) offset to 5 632 tsdy~ EXACT <bar> (EXACT) (32) 633 ~ attach to CLOSE1 (34) offset to 2 634 >$< tail~ BRANCH (3) 635 ~ BRANCH (9) 636 ~ TAIL (25) 637 piec 638 atom 639 >< 37 tail~ OPEN1 (26) 640 ~ BRANCH (28) 641 ~ BRANCH (31) 642 ~ CLOSE1 (34) 643 38 tsdy~ EXACT <x> (EXACT) (1) 644 ~ BRANCH (END) (3) 645 ~ BRANCH (END) (9) 646 ~ TAIL (END) (25) 647 ~ OPEN1 (END) (26) 648 ~ BRANCH (END) (28) 649 ~ BRANCH (END) (31) 650 ~ CLOSE1 (END) (34) 651 ~ EOL (END) (36) 652 ~ attach to END (37) offset to 1 653 654Resulting in the program 655 656 1: EXACT <x>(3) 657 3: BRANCH(9) 658 4: EXACT <fo>(6) 659 6: STAR(26) 660 7: EXACT <o>(0) 661 9: BRANCH(25) 662 10: EXACT <ba>(14) 663 12: OPTIMIZED (2 nodes) 664 14: ANYOF[Rr](26) 665 25: TAIL(26) 666 26: OPEN1(28) 667 28: TRIE-EXACT(34) 668 [StS:1 Wds:2 Cs:6 Uq:5 #Sts:7 Mn:3 Mx:3 Stcls:bf] 669 <foo> 670 <bar> 671 30: OPTIMIZED (4 nodes) 672 34: CLOSE1(36) 673 36: EOL(37) 674 37: END(0) 675 676Here we can see a much more complex program, with various optimisations in 677play. At regnode 10 we see an example where a character class with only 678one character in it was turned into an C<EXACT> node. We can also see where 679an entire alternation was turned into a C<TRIE-EXACT> node. As a consequence, 680some of the regnodes have been marked as optimised away. We can see that 681the C<$> symbol has been converted into an C<EOL> regop, a special piece of 682code that looks for C<\n> or the end of the string. 683 684The next pointer for C<BRANCH>es is interesting in that it points at where 685execution should go if the branch fails. When executing, if the engine 686tries to traverse from a branch to a C<regnext> that isn't a branch then 687the engine will know that the entire set of branches has failed. 688 689=head3 Peep-hole Optimisation and Analysis 690 691The regular expression engine can be a weighty tool to wield. On long 692strings and complex patterns it can end up having to do a lot of work 693to find a match, and even more to decide that no match is possible. 694Consider a situation like the following pattern. 695 696 'ababababababababababab' =~ /(a|b)*z/ 697 698The C<(a|b)*> part can match at every char in the string, and then fail 699every time because there is no C<z> in the string. So obviously we can 700avoid using the regex engine unless there is a C<z> in the string. 701Likewise in a pattern like: 702 703 /foo(\w+)bar/ 704 705In this case we know that the string must contain a C<foo> which must be 706followed by C<bar>. We can use Fast Boyer-Moore matching as implemented 707in C<fbm_instr()> to find the location of these strings. If they don't exist 708then we don't need to resort to the much more expensive regex engine. 709Even better, if they do exist then we can use their positions to 710reduce the search space that the regex engine needs to cover to determine 711if the entire pattern matches. 712 713There are various aspects of the pattern that can be used to facilitate 714optimisations along these lines: 715 716=over 5 717 718=item * anchored fixed strings 719 720=item * floating fixed strings 721 722=item * minimum and maximum length requirements 723 724=item * start class 725 726=item * Beginning/End of line positions 727 728=back 729 730Another form of optimisation that can occur is the post-parse "peep-hole" 731optimisation, where inefficient constructs are replaced by more efficient 732constructs. The C<TAIL> regops which are used during parsing to mark the end 733of branches and the end of groups are examples of this. These regops are used 734as place-holders during construction and "always match" so they can be 735"optimised away" by making the things that point to the C<TAIL> point to the 736thing that C<TAIL> points to, thus "skipping" the node. 737 738Another optimisation that can occur is that of "C<EXACT> merging" which is 739where two consecutive C<EXACT> nodes are merged into a single 740regop. An even more aggressive form of this is that a branch 741sequence of the form C<EXACT BRANCH ... EXACT> can be converted into a 742C<TRIE-EXACT> regop. 743 744All of this occurs in the routine C<study_chunk()> which uses a special 745structure C<scan_data_t> to store the analysis that it has performed, and 746does the "peep-hole" optimisations as it goes. 747 748The code involved in C<study_chunk()> is extremely cryptic. Be careful. :-) 749 750=head2 Execution 751 752Execution of a regex generally involves two phases, the first being 753finding the start point in the string where we should match from, 754and the second being running the regop interpreter. 755 756If we can tell that there is no valid start point then we don't bother running 757the interpreter at all. Likewise, if we know from the analysis phase that we 758cannot detect a short-cut to the start position, we go straight to the 759interpreter. 760 761The two entry points are C<re_intuit_start()> and C<pregexec()>. These routines 762have a somewhat incestuous relationship with overlap between their functions, 763and C<pregexec()> may even call C<re_intuit_start()> on its own. Nevertheless 764other parts of the perl source code may call into either, or both. 765 766Execution of the interpreter itself used to be recursive, but thanks to the 767efforts of Dave Mitchell in the 5.9.x development track, that has changed: now an 768internal stack is maintained on the heap and the routine is fully 769iterative. This can make it tricky as the code is quite conservative 770about what state it stores, with the result that two consecutive lines in the 771code can actually be running in totally different contexts due to the 772simulated recursion. 773 774=for apidoc pregcomp 775=for apidoc pregexec 776 777=head3 Start position and no-match optimisations 778 779C<re_intuit_start()> is responsible for handling start points and no-match 780optimisations as determined by the results of the analysis done by 781C<study_chunk()> (and described in L</Peep-hole Optimisation and Analysis>). 782 783The basic structure of this routine is to try to find the start- and/or 784end-points of where the pattern could match, and to ensure that the string 785is long enough to match the pattern. It tries to use more efficient 786methods over less efficient methods and may involve considerable 787cross-checking of constraints to find the place in the string that matches. 788For instance it may try to determine that a given fixed string must be 789not only present but a certain number of chars before the end of the 790string, or whatever. 791 792It calls several other routines, such as C<fbm_instr()> which does 793Fast Boyer Moore matching and C<find_byclass()> which is responsible for 794finding the start using the first mandatory regop in the program. 795 796When the optimisation criteria have been satisfied, C<reg_try()> is called 797to perform the match. 798 799=head3 Program execution 800 801C<pregexec()> is the main entry point for running a regex. It contains 802support for initialising the regex interpreter's state, running 803C<re_intuit_start()> if needed, and running the interpreter on the string 804from various start positions as needed. When it is necessary to use 805the regex interpreter C<pregexec()> calls C<regtry()>. 806 807C<regtry()> is the entry point into the regex interpreter. It expects 808as arguments a pointer to a C<regmatch_info> structure and a pointer to 809a string. It returns an integer 1 for success and a 0 for failure. 810It is basically a set-up wrapper around C<regmatch()>. 811 812C<regmatch> is the main "recursive loop" of the interpreter. It is 813basically a giant switch statement that implements a state machine, where 814the possible states are the regops themselves, plus a number of additional 815intermediate and failure states. A few of the states are implemented as 816subroutines but the bulk are inline code. 817 818=head1 MISCELLANEOUS 819 820=head2 Unicode and Localisation Support 821 822When dealing with strings containing characters that cannot be represented 823using an eight-bit character set, perl uses an internal representation 824that is a permissive version of Unicode's UTF-8 encoding[2]. This uses single 825bytes to represent characters from the ASCII character set, and sequences 826of two or more bytes for all other characters. (See L<perlunitut> 827for more information about the relationship between UTF-8 and perl's 828encoding, utf8. The difference isn't important for this discussion.) 829 830No matter how you look at it, Unicode support is going to be a pain in a 831regex engine. Tricks that might be fine when you have 256 possible 832characters often won't scale to handle the size of the UTF-8 character 833set. Things you can take for granted with ASCII may not be true with 834Unicode. For instance, in ASCII, it is safe to assume that 835C<sizeof(char1) == sizeof(char2)>, but in UTF-8 it isn't. Unicode case folding is 836vastly more complex than the simple rules of ASCII, and even when not 837using Unicode but only localised single byte encodings, things can get 838tricky (for example, B<LATIN SMALL LETTER SHARP S> (U+00DF, E<szlig>) 839should match 'SS' in localised case-insensitive matching). 840 841Making things worse is that UTF-8 support was a later addition to the 842regex engine (as it was to perl) and this necessarily made things a lot 843more complicated. Obviously it is easier to design a regex engine with 844Unicode support in mind from the beginning than it is to retrofit it to 845one that wasn't. 846 847Nearly all regops that involve looking at the input string have 848two cases, one for UTF-8, and one not. In fact, it's often more complex 849than that, as the pattern may be UTF-8 as well. 850 851Care must be taken when making changes to make sure that you handle 852UTF-8 properly, both at compile time and at execution time, including 853when the string and pattern are mismatched. 854 855=head2 Base Structures 856 857The C<regexp> structure described in L<perlreapi> is common to all 858regex engines. Two of its fields are intended for the private use 859of the regex engine that compiled the pattern. These are the 860C<intflags> and pprivate members. The C<pprivate> is a void pointer to 861an arbitrary structure whose use and management is the responsibility 862of the compiling engine. perl will never modify either of these 863values. In the case of the stock engine the structure pointed to by 864C<pprivate> is called C<regexp_internal>. 865 866Its C<pprivate> and C<intflags> fields contain data 867specific to each engine. 868 869There are two structures used to store a compiled regular expression. 870One, the C<regexp> structure described in L<perlreapi> is populated by 871the engine currently being used and some of its fields read by perl to 872implement things such as the stringification of C<qr//>. 873 874The other structure is pointed to by the C<regexp> struct's 875C<pprivate> and is in addition to C<intflags> in the same struct 876considered to be the property of the regex engine which compiled the 877regular expression; 878 879The regexp structure contains all the data that perl needs to be aware of 880to properly work with the regular expression. It includes data about 881optimisations that perl can use to determine if the regex engine should 882really be used, and various other control info that is needed to properly 883execute patterns in various contexts such as is the pattern anchored in 884some way, or what flags were used during the compile, or whether the 885program contains special constructs that perl needs to be aware of. 886 887In addition it contains two fields that are intended for the private use 888of the regex engine that compiled the pattern. These are the C<intflags> 889and pprivate members. The C<pprivate> is a void pointer to an arbitrary 890structure whose use and management is the responsibility of the compiling 891engine. perl will never modify either of these values. 892 893As mentioned earlier, in the case of the default engines, the C<pprivate> 894will be a pointer to a regexp_internal structure which holds the compiled 895program and any additional data that is private to the regex engine 896implementation. 897 898=head3 Perl's C<pprivate> structure 899 900The following structure is used as the C<pprivate> struct by perl's 901regex engine. Since it is specific to perl it is only of curiosity 902value to other engine implementations. 903 904 typedef struct regexp_internal { 905 regnode *regstclass; 906 struct reg_data *data; 907 struct reg_code_blocks *code_blocks; 908 U32 proglen; 909 U32 name_list_idx; 910 regnode program[1]; 911 } regexp_internal; 912 913Description of the attributes is as follows: 914 915=over 5 916 917=item C<regstclass> 918 919Special regop that is used by C<re_intuit_start()> to check if a pattern 920can match at a certain position. For instance if the regex engine knows 921that the pattern must start with a 'Z' then it can scan the string until 922it finds one and then launch the regex engine from there. The routine 923that handles this is called C<find_by_class()>. Sometimes this field 924points at a regop embedded in the program, and sometimes it points at 925an independent synthetic regop that has been constructed by the optimiser. 926 927=item C<data> 928 929This field points at a C<reg_data> structure, which is defined as follows 930 931 struct reg_data { 932 U32 count; 933 U8 *what; 934 void* data[1]; 935 }; 936 937This structure is used for handling data structures that the regex engine 938needs to handle specially during a clone or free operation on the compiled 939product. Each element in the data array has a corresponding element in the 940what array. During compilation regops that need special structures stored 941will add an element to each array using the add_data() routine and then store 942the index in the regop. 943 944In modern perls the 0th element of this structure is reserved and is NEVER 945used to store anything of use. This is to allow things that need to index 946into this array to represent "no value". 947 948=item C<code_blocks> 949 950This optional structure is used to manage C<(?{})> constructs in the 951pattern. It is made up of the following structures. 952 953 /* record the position of a (?{...}) within a pattern */ 954 struct reg_code_block { 955 STRLEN start; 956 STRLEN end; 957 OP *block; 958 REGEXP *src_regex; 959 }; 960 961 /* array of reg_code_block's plus header info */ 962 struct reg_code_blocks { 963 int refcnt; /* we may be pointed to from a regex 964 and from the savestack */ 965 int count; /* how many code blocks */ 966 struct reg_code_block *cb; /* array of reg_code_block's */ 967 }; 968 969=item C<proglen> 970 971Stores the length of the compiled program in units of regops. 972 973=item C<name_list_idx> 974 975This is the index into the data array where an AV is stored that contains 976the names of any named capture buffers in the pattern, should there be 977any. This is only used in the debugging version of the regex engine and 978when RXp_PAREN_NAMES(prog) is true. It will be 0 if there is no such data. 979 980=item C<program> 981 982Compiled program. Inlined into the structure so the entire struct can be 983treated as a single blob. 984 985=back 986 987=head1 SEE ALSO 988 989L<perlreapi> 990 991L<perlre> 992 993L<perlunitut> 994 995=head1 AUTHOR 996 997by Yves Orton, 2006. 998 999With excerpts from Perl, and contributions and suggestions from 1000Ronald J. Kimball, Dave Mitchell, Dominic Dunlop, Mark Jason Dominus, 1001Stephen McCamant, and David Landgren. 1002 1003Now maintained by Perl 5 Porters. 1004 1005=head1 LICENCE 1006 1007Same terms as Perl. 1008 1009=head1 REFERENCES 1010 1011[1] L<https://perl.plover.com/Rx/paper/> 1012 1013[2] L<https://www.unicode.org/> 1014 1015=cut 1016