1# regcomp.sym 2# 3# File has two sections, divided by a line of dashes '-'. 4# 5# Lines beginning with # are ignored, except for those that start with #* 6# which are included in pod/perldebguts.pod. # within a line may be part 7# of a description. 8# 9# First section is for regops, second section is for regmatch-states 10# 11# Note that the order in this file is important. 12# 13# Format for first section: 14# NAME \s+ TYPE, arg-description [struct regnode suffix] [flags] [longjump] ; DESCRIPTION 15# arg-description is currently unused 16# suffix is appended to 'struct_regnode_' giving which one to use. If empty, 17# it means plain 'struct regnode'. If the regnode is a string one, this 18# should instead refer to the base regnode, without the char[1] element 19# of the structure 20# flag <S> means is REGNODE_SIMPLE; flag <V> means is REGNODE_VARIES; <.> is 21# a placeholder 22# longjump is 1 if the (first) argument holds the next offset (instead of the 23# usual 'next_offset' field 24# 25# run perl regen.pl after editing this file 26 27# +- suffix of which struct regnode to use e.g., 28# | +- flags (S or V) struct regnode_1 29# un- | | +- longjmp (0, blank, or 1) blank means 0 30# Name Type used | | | ; comment 31# -------------------------------------------------------------------------- 32# IFMATCH BRANCHJ, off 1 . 1 ; Succeeds if the following matches. 33# UNLESSM BRANCHJ, off 1 . 1 ; Fails if the following matches. 34# SUSPEND BRANCHJ, off 1 V 1 ; "Independent" sub-RE. 35# IFTHEN BRANCHJ, off 1 V 1 ; Switch, should be preceded by switcher. 36# GROUPP GROUPP, num 1 ; Whether the group matched. 37# 38# If we were to start running out of regnodes, many of the ones that are 39# complements could be combined with their non-complement mates. For example, 40# POSIXU could have the flags field have the bottom bit mean do we complement 41# or not, and the type be shifted left 1 bit. Then all that would be needed to 42# extract which to do is a mask for the complement bit, and a right shift for 43# the other, an inconsequential increase in instructions. It might actually be 44# clearer and slightly faster given the case statement and assignment are 45# removed. Note that not everything could be collapsed: NPOSIXA, for example, 46# would require special handling for performance. 47 48 49#* Exit points 50 51END END, no ; End of program. 52SUCCEED END, no ; Return from a subroutine, basically. 53 54#* Line Start Anchors: 55#Note flags field for SBOL indicates if it is a /^/ or a /\A/ 56SBOL BOL, no ; Match "" at beginning of line: /^/, /\A/ 57MBOL BOL, no ; Same, assuming multiline: /^/m 58 59#* Line End Anchors: 60SEOL EOL, no ; Match "" at end of line: /$/ 61MEOL EOL, no ; Same, assuming multiline: /$/m 62EOS EOL, no ; Match "" at end of string: /\z/ 63 64#* Match Start Anchors: 65GPOS GPOS, no ; Matches where last m//g left off. 66 67#* Word Boundary Opcodes: 68# The regops that have varieties that vary depending on the character set regex 69# modifiers have to ordered thusly: /d, /l, /u, /a, /aa. This is because code 70# in regcomp.c uses the enum value of the modifier as an offset from the /d 71# version. The complements must come after the non-complements. 72# BOUND, POSIX and their complements are affected, as well as EXACTF. 73BOUND BOUND, no ; Like BOUNDA for non-utf8, otherwise like BOUNDU 74BOUNDL BOUND, no ; Like BOUND/BOUNDU, but \w and \W are defined by current locale 75BOUNDU BOUND, no ; Match "" at any boundary of a given type using /u rules. 76BOUNDA BOUND, no ; Match "" at any boundary between \w\W or \W\w, where \w is [_a-zA-Z0-9] 77# All NBOUND nodes are required by code in regexec.c to be greater than all BOUND ones 78NBOUND NBOUND, no ; Like NBOUNDA for non-utf8, otherwise like BOUNDU 79NBOUNDL NBOUND, no ; Like NBOUND/NBOUNDU, but \w and \W are defined by current locale 80NBOUNDU NBOUND, no ; Match "" at any non-boundary of a given type using using /u rules. 81NBOUNDA NBOUND, no ; Match "" betweeen any \w\w or \W\W, where \w is [_a-zA-Z0-9] 82 83#* [Special] alternatives: 84REG_ANY REG_ANY, no 0 S ; Match any one character (except newline). 85SANY REG_ANY, no 0 S ; Match any one character. 86ANYOF ANYOF, sv charclass S ; Match character in (or not in) this class, single char match only 87ANYOFD ANYOF, sv charclass S ; Like ANYOF, but /d is in effect 88ANYOFL ANYOF, sv charclass S ; Like ANYOF, but /l is in effect 89ANYOFPOSIXL ANYOF, sv charclass_posixl S ; Like ANYOFL, but matches [[:posix:]] classes 90 91# Must be sequential 92ANYOFH ANYOFH, sv 1 S ; Like ANYOF, but only has "High" matches, none in the bitmap; the flags field contains the lowest matchable UTF-8 start byte 93ANYOFHb ANYOFH, sv 1 S ; Like ANYOFH, but all matches share the same UTF-8 start byte, given in the flags field 94ANYOFHr ANYOFH, sv 1 S ; Like ANYOFH, but the flags field contains packed bounds for all matchable UTF-8 start bytes. 95ANYOFHs ANYOFH, sv:str 1 S ; Like ANYOFHb, but has a string field that gives the leading matchable UTF-8 bytes; flags field is len 96ANYOFR ANYOFR, packed 1 S ; Matches any character in the range given by its packed args: upper 12 bits is the max delta from the base lower 20; the flags field contains the lowest matchable UTF-8 start byte 97ANYOFRb ANYOFR, packed 1 S ; Like ANYOFR, but all matches share the same UTF-8 start byte, given in the flags field 98# There is no ANYOFRr because khw doesn't think there are likely to be 99# real-world cases where such a large range is used. 100# 101# And khw doesn't believe an ANYOFRs (which would behave like ANYOFHs) is 102# actually worth it. On two-byte UTF-8, the first byte alone is all we need, 103# and ANYOFR already does that. And we don't consider non-Unicode code points 104# or EBCDIC for performance decisions. If we had it, we would be comparing the 105# strings, and if they are equal convert to UV and then test to see if it is in 106# the range. The fast DFA we now use to do the conversion is slower than 107# comparing the strings, but not by much, and negligible in 2 or 3 byte 108# operations. (We don't have to compare the final byte as it has to be 109# different or else this wouldn't be a range.) So we might as well displense 110# with the comparisons that ANYOFRs would do, and go directly to do the 111# conversion . 112 113ANYOFHbbm ANYOFHbbm none bbm S ; Like ANYOFHb, but only for 2-byte UTF-8 characters; uses a bitmap to match the continuation byte 114 115ANYOFM ANYOFM, byte 1 S ; Like ANYOF, but matches an invariant byte as determined by the mask and arg 116NANYOFM ANYOFM, byte 1 S ; complement of ANYOFM 117 118#* POSIX Character Classes: 119# Order of the below is important. See ordering comment above. 120POSIXD POSIXD, none 0 S ; Some [[:class:]] under /d; the FLAGS field gives which one 121POSIXL POSIXD, none 0 S ; Some [[:class:]] under /l; the FLAGS field gives which one 122POSIXU POSIXD, none 0 S ; Some [[:class:]] under /u; the FLAGS field gives which one 123POSIXA POSIXD, none 0 S ; Some [[:class:]] under /a; the FLAGS field gives which one 124NPOSIXD NPOSIXD, none 0 S ; complement of POSIXD, [[:^class:]] 125NPOSIXL NPOSIXD, none 0 S ; complement of POSIXL, [[:^class:]] 126NPOSIXU NPOSIXD, none 0 S ; complement of POSIXU, [[:^class:]] 127NPOSIXA NPOSIXD, none 0 S ; complement of POSIXA, [[:^class:]] 128# End of order is important 129 130CLUMP CLUMP, no 0 V ; Match any extended grapheme cluster sequence 131 132#* Alternation 133 134#* BRANCH The set of branches constituting a single choice are 135#* hooked together with their "next" pointers, since 136#* precedence prevents anything being concatenated to 137#* any individual branch. The "next" pointer of the last 138#* BRANCH in a choice points to the thing following the 139#* whole choice. This is also where the final "next" 140#* pointer of each individual branch points; each branch 141#* starts with the operand node of a BRANCH node. 142#* 143BRANCH BRANCH, node 1 V ; Match this alternative, or the next... 144 145#*Literals 146# NOTE: the relative ordering of these types is important do not change it 147# By convention, folding nodes begin with EXACTF; A digit 8 is in the name if 148# and only if it it requires a UTF-8 target string in order to successfully 149# match. 150 151EXACT EXACT, str ; Match this string (flags field is the length). 152 153#* In a long string node, the U32 argument is the length, and is 154#* immediately followed by the string. 155LEXACT EXACT, len:str 1; Match this long string (preceded by length; flags unused). 156EXACTL EXACT, str ; Like EXACT, but /l is in effect (used so locale-related warnings can be checked for) 157EXACTF EXACT, str ; Like EXACT, but match using /id rules; (string not UTF-8, ASCII folded; non-ASCII not) 158EXACTFL EXACT, str ; Like EXACT, but match using /il rules; (string not likely to be folded) 159EXACTFU EXACT, str ; Like EXACT, but match using /iu rules; (string folded) 160 161# The reason MICRO and SHARP S aren't folded in non-UTF8 patterns is because 162# they would fold to something that requires UTF-8. SHARP S would normally 163# fold to 'ss', but because of /aa, it instead folds to a pair of LATIN SMALL 164# LETTER LONG S characters (U+017F) 165EXACTFAA EXACT, str ; Like EXACT, but match using /iaa rules; (string folded except MICRO in non-UTF8 patterns; doesn't contain SHARP S unless UTF-8; folded length <= unfolded) 166# must immediately follow EXACTFAA 167EXACTFAA_NO_TRIE EXACT, str ; Like EXACTFAA, (string not UTF-8, folded except: MICRO, SHARP S; folded length <= unfolded, not currently trie-able) 168 169# End of important relative ordering. 170 171EXACTFUP EXACT, str ; Like EXACT, but match using /iu rules; (string not UTF-8, folded except MICRO: hence Problematic) 172# In order for a non-UTF-8 EXACTFAA to think the pattern is pre-folded when 173# matching a UTF-8 target string, there would have to be something like an 174# EXACTFAA_MICRO which would not be considered pre-folded for UTF-8 targets, 175# since the fold of the MICRO SIGN would not be done, and would be 176# representable in the UTF-8 target string. 177 178EXACTFLU8 EXACT, str ; Like EXACTFU, but use /il, UTF-8, (string is folded, and everything in it is above 255 179EXACT_REQ8 EXACT, str ; Like EXACT, but only UTF-8 encoded targets can match 180LEXACT_REQ8 EXACT, len:str 1 ; Like LEXACT, but only UTF-8 encoded targets can match 181EXACTFU_REQ8 EXACT, str ; Like EXACTFU, but only UTF-8 encoded targets can match 182# One could add EXACTFAA8 and something that has the same effect for /l, 183# but these would be extremely uncommon 184 185EXACTFU_S_EDGE EXACT, str ; /di rules, but nothing in it precludes /ui, except begins and/or ends with [Ss]; (string not UTF-8; compile-time only) 186 187#*New charclass like patterns 188LNBREAK LNBREAK, none ; generic newline pattern 189 190#*Trie Related 191 192#* Behave the same as A|LIST|OF|WORDS would. The '..C' variants 193#* have inline charclass data (ascii only), the 'C' store it in the 194#* structure. 195# NOTE: the relative order of the TRIE-like regops is significant 196 197TRIE TRIE, trie 1 ; Match many EXACT(F[ALU]?)? at once. flags==type 198TRIEC TRIE,trie charclass ; Same as TRIE, but with embedded charclass data 199 200# For start classes, contains an added fail table. 201AHOCORASICK TRIE, trie 1 ; Aho Corasick stclass. flags==type 202AHOCORASICKC TRIE,trie charclass ; Same as AHOCORASICK, but with embedded charclass data 203 204#*Do nothing types 205 206NOTHING NOTHING, no ; Match empty string. 207#*A variant of above which delimits a group, thus stops optimizations 208TAIL NOTHING, no ; Match empty string. Can jump here from outside. 209 210#*Loops 211 212#* STAR,PLUS '?', and complex '*' and '+', are implemented as 213#* circular BRANCH structures. Simple cases 214#* (one character per match) are implemented with STAR 215#* and PLUS for speed and to minimize recursive plunges. 216#* 217STAR STAR, node 0 V ; Match this (simple) thing 0 or more times: /A{0,}B/ where A is width 1 char 218PLUS PLUS, node 0 V ; Match this (simple) thing 1 or more times: /A{1,}B/ where A is width 1 char 219 220CURLY CURLY, sv 3 V ; Match this (simple) thing {n,m} times: /A{m,n}B/ where A is width 1 char 221CURLYN CURLY, no 3 V ; Capture next-after-this simple thing: /(A){m,n}B/ where A is width 1 char 222CURLYM CURLY, no 3 V ; Capture this medium-complex thing {n,m} times: /(A){m,n}B/ where A is fixed-length 223CURLYX CURLY, sv 3 V ; Match/Capture this complex thing {n,m} times. 224 225#*This terminator creates a loop structure for CURLYX 226WHILEM WHILEM, no 0 V ; Do curly processing and see if rest matches. 227 228#*Buffer related 229 230#*OPEN,CLOSE,GROUPP ...are numbered at compile time. 231OPEN OPEN, num 1 ; Mark this point in input as start of #n. 232CLOSE CLOSE, num 1 ; Close corresponding OPEN of #n. 233SROPEN SROPEN, none ; Same as OPEN, but for script run 234SRCLOSE SRCLOSE, none ; Close preceding SROPEN 235 236REF REF, num 2 V ; Match some already matched string 237REFF REF, num 2 V ; Match already matched string, using /di rules. 238REFFL REF, num 2 V ; Match already matched string, using /li rules. 239# N?REFF[AU] could have been implemented using the FLAGS field of the 240# regnode, but by having a separate node type, we can use the existing switch 241# statement to avoid some tests 242REFFU REF, num 2 V ; Match already matched string, usng /ui. 243REFFA REF, num 2 V ; Match already matched string, using /aai rules. 244 245#*Named references. Code in regcomp.c assumes that these all are after 246#*the numbered references 247REFN REF, no-sv 2 V ; Match some already matched string 248REFFN REF, no-sv 2 V ; Match already matched string, using /di rules. 249REFFLN REF, no-sv 2 V ; Match already matched string, using /li rules. 250REFFUN REF, num 2 V ; Match already matched string, using /ui rules. 251REFFAN REF, num 2 V ; Match already matched string, using /aai rules. 252 253#*Support for long RE 254LONGJMP LONGJMP, off 1 . 1 ; Jump far away. 255BRANCHJ BRANCHJ, off 2 V 1 ; BRANCH with long offset. 256 257#*Special Case Regops 258IFMATCH BRANCHJ, off 1 . 1 ; Succeeds if the following matches; non-zero flags "f", next_off "o" means lookbehind assertion starting "f..(f-o)" characters before current 259UNLESSM BRANCHJ, off 1 . 1 ; Fails if the following matches; non-zero flags "f", next_off "o" means lookbehind assertion starting "f..(f-o)" characters before current 260SUSPEND BRANCHJ, off 1 V 1 ; "Independent" sub-RE. 261IFTHEN BRANCHJ, off 1 V 1 ; Switch, should be preceded by switcher. 262GROUPP GROUPP, num 1 ; Whether the group matched. 263 264#*The heavy worker 265 266EVAL EVAL, evl/flags 2 ; Execute some Perl code. 267 268#*Modifiers 269 270MINMOD MINMOD, no ; Next operator is not greedy. 271LOGICAL LOGICAL, no ; Next opcode should set the flag only. 272 273#*This is not used yet 274RENUM BRANCHJ, off 1 . 1 ; Group with independently numbered parens. 275 276#*Regex Subroutines 277GOSUB GOSUB, num/ofs 2 ; recurse to paren arg1 at (signed) ofs arg2 278 279#*Special conditionals 280GROUPPN GROUPPN, no-sv 1 ; Whether the group matched. 281INSUBP INSUBP, num 1 ; Whether we are in a specific recurse. 282DEFINEP DEFINEP, none 1 ; Never execute directly. 283 284#*Backtracking Verbs 285ENDLIKE ENDLIKE, none ; Used only for the type field of verbs 286OPFAIL ENDLIKE, no-sv 1 ; Same as (?!), but with verb arg 287ACCEPT ENDLIKE, no-sv/num 2 ; Accepts the current matched string, with verbar 288 289#*Verbs With Arguments 290VERB VERB, no-sv 1 ; Used only for the type field of verbs 291PRUNE VERB, no-sv 1 ; Pattern fails at this startpoint if no-backtracking through this 292MARKPOINT VERB, no-sv 1 ; Push the current location for rollback by cut. 293SKIP VERB, no-sv 1 ; On failure skip forward (to the mark) before retrying 294COMMIT VERB, no-sv 1 ; Pattern fails outright if backtracking through this 295CUTGROUP VERB, no-sv 1 ; On failure go to the next alternation in the group 296 297#*Control what to keep in $&. 298KEEPS KEEPS, no ; $& begins here. 299 300#*Validate that lookbehind IFMATCH and UNLESSM end at the right place 301LOOKBEHIND_END END, no ; Return from lookbehind (IFMATCH/UNLESSM) and validate position 302 303# NEW STUFF SOMEWHERE ABOVE THIS LINE. Stuff that regexec.c: find_byclass() 304# and regrepeat() use should go way above, near LNBREAK to allow a more compact 305# jump table to be generated for their switch() statements 306 307################################################################################ 308 309#*SPECIAL REGOPS 310 311#* This is not really a node, but an optimized away piece of a "long" 312#* node. To simplify debugging output, we mark it as if it were a node 313OPTIMIZED NOTHING, off ; Placeholder for dump. 314 315#* Special opcode with the property that no opcode in a compiled program 316#* will ever be of this type. Thus it can be used as a flag value that 317#* no other opcode has been seen. END is used similarly, in that an END 318#* node cant be optimized. So END implies "unoptimizable" and PSEUDO 319#* mean "not seen anything to optimize yet". 320PSEUDO PSEUDO, off ; Pseudo opcode for internal use. 321 322REGEX_SET REGEX_SET, depth p S ; Regex set, temporary node used in pre-optimization compilation 323 324------------------------------------------------------------------------------- 325# Format for second section: 326# REGOP \t typelist [ \t typelist] 327# typelist= namelist 328# = namelist:FAIL 329# = name:count 330 331# Anything below is a state 332# 333# 334TRIE next:FAIL 335EVAL B,postponed_AB:FAIL 336CURLYX end:FAIL 337WHILEM A_pre,A_min,A_max,B_min,B_max:FAIL 338BRANCH next:FAIL 339CURLYM A,B:FAIL 340IFMATCH A:FAIL 341CURLY B_min,B_max:FAIL 342COMMIT next:FAIL 343MARKPOINT next:FAIL 344SKIP next:FAIL 345CUTGROUP next:FAIL 346KEEPS next:FAIL 347REF next:FAIL 348