diff options
| -rw-r--r-- | generic/regc_nfa.c | 1914 | ||||
| -rw-r--r-- | generic/regcomp.c | 41 | ||||
| -rw-r--r-- | generic/regerrs.h | 2 | ||||
| -rw-r--r-- | generic/regex.h | 2 | ||||
| -rw-r--r-- | generic/regguts.h | 24 | ||||
| -rw-r--r-- | tests/reg.test | 59 | ||||
| -rw-r--r-- | tests/regexp.test | 2 |
7 files changed, 1641 insertions, 403 deletions
diff --git a/generic/regc_nfa.c b/generic/regc_nfa.c index 1fad85f..0e0343e 100644 --- a/generic/regc_nfa.c +++ b/generic/regc_nfa.c @@ -34,6 +34,9 @@ #define NISERR() VISERR(nfa->v) #define NERR(e) VERR(nfa->v, (e)) +#define STACK_TOO_DEEP(x) (0) +#define CANCEL_REQUESTED(x) (0) +#define REG_CANCEL 777 /* - newnfa - set up an NFA @@ -59,7 +62,6 @@ newnfa( nfa->nstates = 0; nfa->cm = cm; nfa->v = v; - nfa->size = 0; nfa->bos[0] = nfa->bos[1] = COLORLESS; nfa->eos[0] = nfa->eos[1] = COLORLESS; nfa->parent = parent; /* Precedes newfstate so parent is valid. */ @@ -87,61 +89,6 @@ newnfa( } /* - - TooManyStates - checks if the max states exceeds the compile-time value - ^ static int TooManyStates(struct nfa *); - */ -static int -TooManyStates( - struct nfa *nfa) -{ - struct nfa *parent = nfa->parent; - size_t sz = nfa->size; - - while (parent != NULL) { - sz = parent->size; - parent = parent->parent; - } - if (sz > REG_MAX_STATES) { - return 1; - } - return 0; -} - -/* - - IncrementSize - increases the tracked size of the NFA and its parents. - ^ static void IncrementSize(struct nfa *); - */ -static void -IncrementSize( - struct nfa *nfa) -{ - struct nfa *parent = nfa->parent; - - nfa->size++; - while (parent != NULL) { - parent->size++; - parent = parent->parent; - } -} - -/* - - DecrementSize - increases the tracked size of the NFA and its parents. - ^ static void DecrementSize(struct nfa *); - */ -static void -DecrementSize( - struct nfa *nfa) -{ - struct nfa *parent = nfa->parent; - - nfa->size--; - while (parent != NULL) { - parent->size--; - parent = parent->parent; - } -} - -/* - freenfa - free an entire NFA ^ static void freenfa(struct nfa *); */ @@ -177,20 +124,20 @@ newstate( { struct state *s; - if (TooManyStates(nfa)) { - /* XXX: add specific error for this */ - NERR(REG_ETOOBIG); - return NULL; - } if (nfa->free != NULL) { s = nfa->free; nfa->free = s->next; } else { + if (nfa->v->spaceused >= REG_MAX_COMPILE_SPACE) { + NERR(REG_ETOOBIG); + return NULL; + } s = (struct state *) MALLOC(sizeof(struct state)); if (s == NULL) { NERR(REG_ESPACE); return NULL; } + nfa->v->spaceused += sizeof(struct state); s->oas.next = NULL; s->free = NULL; s->noas = 0; @@ -214,12 +161,6 @@ newstate( } s->prev = nfa->slast; nfa->slast = s; - - /* - * Track the current size and the parent size. - */ - - IncrementSize(nfa); return s; } @@ -290,7 +231,6 @@ freestate( s->prev = NULL; s->next = nfa->free; /* don't delete it, put it on the free list */ nfa->free = s; - DecrementSize(nfa); } /* @@ -309,11 +249,13 @@ destroystate( for (ab=s->oas.next ; ab!=NULL ; ab=abnext) { abnext = ab->next; FREE(ab); + nfa->v->spaceused -= sizeof(struct arcbatch); } s->ins = NULL; s->outs = NULL; s->next = NULL; FREE(s); + nfa->v->spaceused -= sizeof(struct state); } /* @@ -321,6 +263,10 @@ destroystate( ^ static void newarc(struct nfa *, int, pcolor, struct state *, ^ struct state *); */ +/* + * This function checks to make sure that no duplicate arcs are created. + * In general we never want duplicates. + */ static void newarc( struct nfa *nfa, @@ -333,16 +279,42 @@ newarc( assert(from != NULL && to != NULL); - /* - * Check for duplicates. - */ - - for (a=from->outs ; a!=NULL ; a=a->outchain) { - if (a->to == to && a->co == co && a->type == t) { - return; + /* check for duplicate arc, using whichever chain is shorter */ + if (from->nouts <= to->nins) { + for (a = from->outs; a != NULL; a = a->outchain) { + if (a->to == to && a->co == co && a->type == t) { + return; + } + } + } else { + for (a = to->ins; a != NULL; a = a->inchain) { + if (a->from == from && a->co == co && a->type == t) { + return; + } } } + + /* no dup, so create the arc */ + createarc(nfa, t, co, from, to); +} + +/* + * createarc - create a new arc within an NFA + * + * This function must *only* be used after verifying that there is no existing + * identical arc (same type/color/from/to). + */ +static void +createarc( + struct nfa * nfa, + int t, + pcolor co, + struct state * from, + struct state * to) +{ + struct arc *a; + /* the arc is physically allocated within its from-state */ a = allocarc(nfa, from); if (NISERR()) { return; @@ -355,15 +327,21 @@ newarc( a->from = from; /* - * Put the new arc on the beginning, not the end, of the chains. Not only - * is this easier, it has the very useful side effect that deleting the - * most-recently-added arc is the cheapest case rather than the most - * expensive one. + * Put the new arc on the beginning, not the end, of the chains; it's + * simpler here, and freearc() is the same cost either way. See also the + * logic in moveins() and its cohorts, as well as fixempties(). */ - a->inchain = to->ins; + a->inchainRev = NULL; + if (to->ins) { + to->ins->inchainRev = a; + } to->ins = a; a->outchain = from->outs; + a->outchainRev = NULL; + if (from->outs) { + from->outs->outchainRev = a; + } from->outs = a; from->nouts++; @@ -400,14 +378,19 @@ allocarc( */ if (s->free == NULL) { - struct arcbatch *newAb = (struct arcbatch *) - MALLOC(sizeof(struct arcbatch)); + struct arcbatch *newAb; int i; + if (nfa->v->spaceused >= REG_MAX_COMPILE_SPACE) { + NERR(REG_ETOOBIG); + return NULL; + } + newAb = (struct arcbatch *) MALLOC(sizeof(struct arcbatch)); if (newAb == NULL) { NERR(REG_ESPACE); return NULL; } + nfa->v->spaceused += sizeof(struct arcbatch); newAb->next = s->oas.next; s->oas.next = newAb; @@ -436,7 +419,7 @@ freearc( { struct state *from = victim->from; struct state *to = victim->to; - struct arc *a; + struct arc *predecessor; assert(victim->type != 0); @@ -453,16 +436,17 @@ freearc( */ assert(from != NULL); - assert(from->outs != NULL); - a = from->outs; - if (a == victim) { /* simple case: first in chain */ + predecessor = victim->outchainRev; + if (predecessor == NULL) { + assert(from->outs == victim); from->outs = victim->outchain; } else { - for (; a!=NULL && a->outchain!=victim ; a=a->outchain) { - continue; - } - assert(a != NULL); - a->outchain = victim->outchain; + assert(predecessor->outchain == victim); + predecessor->outchain = victim->outchain; + } + if (victim->outchain != NULL) { + assert(victim->outchain->outchainRev == victim); + victim->outchain->outchainRev = predecessor; } from->nouts--; @@ -471,86 +455,95 @@ freearc( */ assert(to != NULL); - assert(to->ins != NULL); - a = to->ins; - if (a == victim) { /* simple case: first in chain */ + predecessor = victim->inchainRev; + if (predecessor == NULL) { + assert(to->ins == victim); to->ins = victim->inchain; } else { - for (; a->inchain!=victim ; a=a->inchain) { - assert(a->inchain != NULL); - continue; - } - a->inchain = victim->inchain; + assert(predecessor->inchain == victim); + predecessor->inchain = victim->inchain; + } + if (victim->inchain != NULL) { + assert(victim->inchain->inchainRev == victim); + victim->inchain->inchainRev = predecessor; } to->nins--; /* - * Clean up and place on free list. + * Clean up and place on from-state's free list. */ victim->type = 0; victim->from = NULL; /* precautions... */ victim->to = NULL; victim->inchain = NULL; + victim->inchainRev = NULL; victim->outchain = NULL; + victim->outchainRev = NULL; victim->freechain = from->free; from->free = victim; } - + /* - - hasnonemptyout - Does state have a non-EMPTY out arc? - ^ static int hasnonemptyout(struct state *); + * changearctarget - flip an arc to have a different to state + * + * Caller must have verified that there is no pre-existing duplicate arc. + * + * Note that because we store arcs in their from state, we can't easily have + * a similar changearcsource function. */ -static int -hasnonemptyout( - struct state *s) +static void +changearctarget(struct arc * a, struct state * newto) { - struct arc *a; + struct state *oldto = a->to; + struct arc *predecessor; - for (a = s->outs; a != NULL; a = a->outchain) { - if (a->type != EMPTY) { - return 1; - } + assert(oldto != newto); + + /* take it off old target's in-chain */ + assert(oldto != NULL); + predecessor = a->inchainRev; + if (predecessor == NULL) { + assert(oldto->ins == a); + oldto->ins = a->inchain; + } else { + assert(predecessor->inchain == a); + predecessor->inchain = a->inchain; } - return 0; -} - -/* - - nonemptyouts - count non-EMPTY out arcs of a state - ^ static int nonemptyouts(struct state *); - */ -static int -nonemptyouts( - struct state *s) -{ - int n = 0; - struct arc *a; + if (a->inchain != NULL) { + assert(a->inchain->inchainRev == a); + a->inchain->inchainRev = predecessor; + } + oldto->nins--; - for (a = s->outs; a != NULL; a = a->outchain) { - if (a->type != EMPTY) { - n++; - } + a->to = newto; + + /* prepend it to new target's in-chain */ + a->inchain = newto->ins; + a->inchainRev = NULL; + if (newto->ins) { + newto->ins->inchainRev = a; } - return n; + newto->ins = a; + newto->nins++; } /* - - nonemptyins - count non-EMPTY in arcs of a state - ^ static int nonemptyins(struct state *); + - hasnonemptyout - Does state have a non-EMPTY out arc? + ^ static int hasnonemptyout(struct state *); */ static int -nonemptyins( +hasnonemptyout( struct state *s) { - int n = 0; struct arc *a; - for (a = s->ins; a != NULL; a = a->inchain) { + for (a = s->outs; a != NULL; a = a->outchain) { if (a->type != EMPTY) { - n++; + return 1; } } - return n; + return 0; } /* @@ -588,13 +581,181 @@ cparc( { newarc(nfa, oa->type, oa->co, from, to); } + +/* + * sortins - sort the in arcs of a state by from/color/type + */ +static void +sortins( + struct nfa * nfa, + struct state * s) +{ + struct arc **sortarray; + struct arc *a; + int n = s->nins; + int i; + + if (n <= 1) { + return; /* nothing to do */ + } + /* make an array of arc pointers ... */ + sortarray = (struct arc **) MALLOC(n * sizeof(struct arc *)); + if (sortarray == NULL) { + NERR(REG_ESPACE); + return; + } + i = 0; + for (a = s->ins; a != NULL; a = a->inchain) { + sortarray[i++] = a; + } + assert(i == n); + /* ... sort the array */ + qsort(sortarray, n, sizeof(struct arc *), sortins_cmp); + /* ... and rebuild arc list in order */ + /* it seems worth special-casing first and last items to simplify loop */ + a = sortarray[0]; + s->ins = a; + a->inchain = sortarray[1]; + a->inchainRev = NULL; + for (i = 1; i < n - 1; i++) { + a = sortarray[i]; + a->inchain = sortarray[i + 1]; + a->inchainRev = sortarray[i - 1]; + } + a = sortarray[i]; + a->inchain = NULL; + a->inchainRev = sortarray[i - 1]; + FREE(sortarray); +} + +static int +sortins_cmp( + const void *a, + const void *b) +{ + const struct arc *aa = *((const struct arc * const *) a); + const struct arc *bb = *((const struct arc * const *) b); + + /* we check the fields in the order they are most likely to be different */ + if (aa->from->no < bb->from->no) { + return -1; + } + if (aa->from->no > bb->from->no) { + return 1; + } + if (aa->co < bb->co) { + return -1; + } + if (aa->co > bb->co) { + return 1; + } + if (aa->type < bb->type) { + return -1; + } + if (aa->type > bb->type) { + return 1; + } + return 0; +} + +/* + * sortouts - sort the out arcs of a state by to/color/type + */ +static void +sortouts( + struct nfa * nfa, + struct state * s) +{ + struct arc **sortarray; + struct arc *a; + int n = s->nouts; + int i; + + if (n <= 1) { + return; /* nothing to do */ + } + /* make an array of arc pointers ... */ + sortarray = (struct arc **) MALLOC(n * sizeof(struct arc *)); + if (sortarray == NULL) { + NERR(REG_ESPACE); + return; + } + i = 0; + for (a = s->outs; a != NULL; a = a->outchain) { + sortarray[i++] = a; + } + assert(i == n); + /* ... sort the array */ + qsort(sortarray, n, sizeof(struct arc *), sortouts_cmp); + /* ... and rebuild arc list in order */ + /* it seems worth special-casing first and last items to simplify loop */ + a = sortarray[0]; + s->outs = a; + a->outchain = sortarray[1]; + a->outchainRev = NULL; + for (i = 1; i < n - 1; i++) { + a = sortarray[i]; + a->outchain = sortarray[i + 1]; + a->outchainRev = sortarray[i - 1]; + } + a = sortarray[i]; + a->outchain = NULL; + a->outchainRev = sortarray[i - 1]; + FREE(sortarray); +} + +static int +sortouts_cmp( + const void *a, + const void *b) +{ + const struct arc *aa = *((const struct arc * const *) a); + const struct arc *bb = *((const struct arc * const *) b); + + /* we check the fields in the order they are most likely to be different */ + if (aa->to->no < bb->to->no) { + return -1; + } + if (aa->to->no > bb->to->no) { + return 1; + } + if (aa->co < bb->co) { + return -1; + } + if (aa->co > bb->co) { + return 1; + } + if (aa->type < bb->type) { + return -1; + } + if (aa->type > bb->type) { + return 1; + } + return 0; +} + +/* + * Common decision logic about whether to use arc-by-arc operations or + * sort/merge. If there's just a few source arcs we cannot recoup the + * cost of sorting the destination arc list, no matter how large it is. + * Otherwise, limit the number of arc-by-arc comparisons to about 1000 + * (a somewhat arbitrary choice, but the breakeven point would probably + * be machine dependent anyway). + */ +#define BULK_ARC_OP_USE_SORT(nsrcarcs, ndestarcs) \ + ((nsrcarcs) < 4 ? 0 : ((nsrcarcs) > 32 || (ndestarcs) > 32)) /* - moveins - move all in arcs of a state to another state * You might think this could be done better by just updating the - * existing arcs, and you would be right if it weren't for the desire + * existing arcs, and you would be right if it weren't for the need * for duplicate suppression, which makes it easier to just make new * ones to exploit the suppression built into newarc. + * + * However, if we have a whole lot of arcs to deal with, retail duplicate + * checks become too slow. In that case we proceed by sorting and merging + * the arc lists, and then we can indeed just update the arcs in-place. + * ^ static void moveins(struct nfa *, struct state *, struct state *); */ static void @@ -603,38 +764,253 @@ moveins( struct state *oldState, struct state *newState) { - struct arc *a; - assert(oldState != newState); - while ((a = oldState->ins) != NULL) { - cparc(nfa, a, a->from, newState); - freearc(nfa, a); + if (!BULK_ARC_OP_USE_SORT(oldState->nins, newState->nins)) { + /* With not too many arcs, just do them one at a time */ + struct arc *a; + + while ((a = oldState->ins) != NULL) { + cparc(nfa, a, a->from, newState); + freearc(nfa, a); + } + } else { + /* + * With many arcs, use a sort-merge approach. Note changearctarget() + * will put the arc onto the front of newState's chain, so it does not + * break our walk through the sorted part of the chain. + */ + struct arc *oa; + struct arc *na; + + /* + * Because we bypass newarc() in this code path, we'd better include a + * cancel check. + */ + if (CANCEL_REQUESTED(nfa->v->re)) { + NERR(REG_CANCEL); + return; + } + + sortins(nfa, oldState); + sortins(nfa, newState); + if (NISERR()) { + return; /* might have failed to sort */ + } + oa = oldState->ins; + na = newState->ins; + while (oa != NULL && na != NULL) { + struct arc *a = oa; + + switch (sortins_cmp(&oa, &na)) { + case -1: + /* newState does not have anything matching oa */ + oa = oa->inchain; + + /* + * Rather than doing createarc+freearc, we can just unlink + * and relink the existing arc struct. + */ + changearctarget(a, newState); + break; + case 0: + /* match, advance in both lists */ + oa = oa->inchain; + na = na->inchain; + /* ... and drop duplicate arc from oldState */ + freearc(nfa, a); + break; + case +1: + /* advance only na; oa might have a match later */ + na = na->inchain; + break; + default: + assert(NOTREACHED); + } + } + while (oa != NULL) { + /* newState does not have anything matching oa */ + struct arc *a = oa; + + oa = oa->inchain; + changearctarget(a, newState); + } } + assert(oldState->nins == 0); assert(oldState->ins == NULL); } /* - copyins - copy in arcs of a state to another state - * Either all arcs, or only non-empty ones as determined by all value. ^ static VOID copyins(struct nfa *, struct state *, struct state *, int); */ static void copyins( struct nfa *nfa, struct state *oldState, - struct state *newState, - int all) + struct state *newState) { - struct arc *a; - assert(oldState != newState); - for (a=oldState->ins ; a!=NULL ; a=a->inchain) { - if (all || a->type != EMPTY) { + if (!BULK_ARC_OP_USE_SORT(oldState->nins, newState->nins)) { + /* With not too many arcs, just do them one at a time */ + struct arc *a; + + for (a = oldState->ins; a != NULL; a = a->inchain) { cparc(nfa, a, a->from, newState); } + } else { + /* + * With many arcs, use a sort-merge approach. Note that createarc() + * will put new arcs onto the front of newState's chain, so it does + * not break our walk through the sorted part of the chain. + */ + struct arc *oa; + struct arc *na; + + /* + * Because we bypass newarc() in this code path, we'd better include a + * cancel check. + */ + if (CANCEL_REQUESTED(nfa->v->re)) { + NERR(REG_CANCEL); + return; + } + + sortins(nfa, oldState); + sortins(nfa, newState); + if (NISERR()) { + return; /* might have failed to sort */ + } + oa = oldState->ins; + na = newState->ins; + while (oa != NULL && na != NULL) { + struct arc *a = oa; + + switch (sortins_cmp(&oa, &na)) { + case -1: + /* newState does not have anything matching oa */ + oa = oa->inchain; + createarc(nfa, a->type, a->co, a->from, newState); + break; + case 0: + /* match, advance in both lists */ + oa = oa->inchain; + na = na->inchain; + break; + case +1: + /* advance only na; oa might have a match later */ + na = na->inchain; + break; + default: + assert(NOTREACHED); + } + } + while (oa != NULL) { + /* newState does not have anything matching oa */ + struct arc *a = oa; + + oa = oa->inchain; + createarc(nfa, a->type, a->co, a->from, newState); + } + } +} + +/* + * mergeins - merge a list of inarcs into a state + * + * This is much like copyins, but the source arcs are listed in an array, + * and are not guaranteed unique. It's okay to clobber the array contents. + */ +static void +mergeins( + struct nfa * nfa, + struct state * s, + struct arc ** arcarray, + int arccount) +{ + struct arc *na; + int i; + int j; + + if (arccount <= 0) { + return; + } + + /* + * Because we bypass newarc() in this code path, we'd better include a + * cancel check. + */ + if (CANCEL_REQUESTED(nfa->v->re)) { + NERR(REG_CANCEL); + return; + } + + /* Sort existing inarcs as well as proposed new ones */ + sortins(nfa, s); + if (NISERR()) { + return; /* might have failed to sort */ + } + + qsort(arcarray, arccount, sizeof(struct arc *), sortins_cmp); + + /* + * arcarray very likely includes dups, so we must eliminate them. (This + * could be folded into the next loop, but it's not worth the trouble.) + */ + j = 0; + for (i = 1; i < arccount; i++) { + switch (sortins_cmp(&arcarray[j], &arcarray[i])) { + case -1: + /* non-dup */ + arcarray[++j] = arcarray[i]; + break; + case 0: + /* dup */ + break; + default: + /* trouble */ + assert(NOTREACHED); + } + } + arccount = j + 1; + + /* + * Now merge into s' inchain. Note that createarc() will put new arcs + * onto the front of s's chain, so it does not break our walk through the + * sorted part of the chain. + */ + i = 0; + na = s->ins; + while (i < arccount && na != NULL) { + struct arc *a = arcarray[i]; + + switch (sortins_cmp(&a, &na)) { + case -1: + /* s does not have anything matching a */ + createarc(nfa, a->type, a->co, a->from, s); + i++; + break; + case 0: + /* match, advance in both lists */ + i++; + na = na->inchain; + break; + case +1: + /* advance only na; array might have a match later */ + na = na->inchain; + break; + default: + assert(NOTREACHED); + } + } + while (i < arccount) { + /* s does not have anything matching a */ + struct arc *a = arcarray[i]; + + createarc(nfa, a->type, a->co, a->from, s); + i++; } } @@ -648,36 +1024,153 @@ moveouts( struct state *oldState, struct state *newState) { - struct arc *a; - assert(oldState != newState); - while ((a = oldState->outs) != NULL) { - cparc(nfa, a, newState, a->to); - freearc(nfa, a); + if (!BULK_ARC_OP_USE_SORT(oldState->nouts, newState->nouts)) { + /* With not too many arcs, just do them one at a time */ + struct arc *a; + + while ((a = oldState->outs) != NULL) { + cparc(nfa, a, newState, a->to); + freearc(nfa, a); + } + } else { + /* + * With many arcs, use a sort-merge approach. Note that createarc() + * will put new arcs onto the front of newState's chain, so it does + * not break our walk through the sorted part of the chain. + */ + struct arc *oa; + struct arc *na; + + /* + * Because we bypass newarc() in this code path, we'd better include a + * cancel check. + */ + if (CANCEL_REQUESTED(nfa->v->re)) { + NERR(REG_CANCEL); + return; + } + + sortouts(nfa, oldState); + sortouts(nfa, newState); + if (NISERR()) { + return; /* might have failed to sort */ + } + oa = oldState->outs; + na = newState->outs; + while (oa != NULL && na != NULL) { + struct arc *a = oa; + + switch (sortouts_cmp(&oa, &na)) { + case -1: + /* newState does not have anything matching oa */ + oa = oa->outchain; + createarc(nfa, a->type, a->co, newState, a->to); + freearc(nfa, a); + break; + case 0: + /* match, advance in both lists */ + oa = oa->outchain; + na = na->outchain; + /* ... and drop duplicate arc from oldState */ + freearc(nfa, a); + break; + case +1: + /* advance only na; oa might have a match later */ + na = na->outchain; + break; + default: + assert(NOTREACHED); + } + } + while (oa != NULL) { + /* newState does not have anything matching oa */ + struct arc *a = oa; + + oa = oa->outchain; + createarc(nfa, a->type, a->co, newState, a->to); + freearc(nfa, a); + } } + + assert(oldState->nouts == 0); + assert(oldState->outs == NULL); } /* - copyouts - copy out arcs of a state to another state - * Either all arcs, or only non-empty ones as determined by all value. ^ static VOID copyouts(struct nfa *, struct state *, struct state *, int); */ static void copyouts( struct nfa *nfa, struct state *oldState, - struct state *newState, - int all) + struct state *newState) { - struct arc *a; - assert(oldState != newState); - for (a=oldState->outs ; a!=NULL ; a=a->outchain) { - if (all || a->type != EMPTY) { + if (!BULK_ARC_OP_USE_SORT(oldState->nouts, newState->nouts)) { + /* With not too many arcs, just do them one at a time */ + struct arc *a; + + for (a = oldState->outs; a != NULL; a = a->outchain) { cparc(nfa, a, newState, a->to); } + } else { + /* + * With many arcs, use a sort-merge approach. Note that createarc() + * will put new arcs onto the front of newState's chain, so it does + * not break our walk through the sorted part of the chain. + */ + struct arc *oa; + struct arc *na; + + /* + * Because we bypass newarc() in this code path, we'd better include a + * cancel check. + */ + if (CANCEL_REQUESTED(nfa->v->re)) { + NERR(REG_CANCEL); + return; + } + + sortouts(nfa, oldState); + sortouts(nfa, newState); + if (NISERR()) { + return; /* might have failed to sort */ + } + oa = oldState->outs; + na = newState->outs; + while (oa != NULL && na != NULL) { + struct arc *a = oa; + + switch (sortouts_cmp(&oa, &na)) { + case -1: + /* newState does not have anything matching oa */ + oa = oa->outchain; + createarc(nfa, a->type, a->co, newState, a->to); + break; + case 0: + /* match, advance in both lists */ + oa = oa->outchain; + na = na->outchain; + break; + case +1: + /* advance only na; oa might have a match later */ + na = na->outchain; + break; + default: + assert(NOTREACHED); + } + } + while (oa != NULL) { + /* newState does not have anything matching oa */ + struct arc *a = oa; + + oa = oa->outchain; + createarc(nfa, a->type, a->co, newState, a->to); + } } } @@ -896,6 +1389,20 @@ specialcolors( - optimize - optimize an NFA ^ static long optimize(struct nfa *, FILE *); */ + + /* + * The main goal of this function is not so much "optimization" (though it + * does try to get rid of useless NFA states) as reducing the NFA to a form + * the regex executor can handle. The executor, and indeed the cNFA format + * that is its input, can only handle PLAIN and LACON arcs. The output of + * the regex parser also includes EMPTY (do-nothing) arcs, as well as + * ^, $, AHEAD, and BEHIND constraint arcs, which we must get rid of here. + * We first get rid of EMPTY arcs and then deal with the constraint arcs. + * The hardest part of either job is to get rid of circular loops of the + * target arc type. We would have to do that in any case, though, as such a + * loop would otherwise allow the executor to cycle through the loop endlessly + * without making any progress in the input string. + */ static long /* re_info bits */ optimize( struct nfa *nfa, @@ -917,17 +1424,23 @@ optimize( if (verbose) { fprintf(f, "\nconstraints:\n"); } + fixconstraintloops(nfa, f); /* get rid of constraint loops */ pullback(nfa, f); /* pull back constraints backward */ pushfwd(nfa, f); /* push fwd constraints forward */ if (verbose) { fprintf(f, "\nfinal cleanup:\n"); } cleanup(nfa); /* final tidying */ +#ifdef REG_DEBUG + if (verbose) { + dumpnfa(nfa, f); + } +#endif return analyze(nfa); /* and analysis */ } /* - - pullback - pull back constraints backward to (with luck) eliminate them + - pullback - pull back constraints backward to eliminate them ^ static void pullback(struct nfa *, FILE *); */ static void @@ -939,6 +1452,7 @@ pullback( struct state *nexts; struct arc *a; struct arc *nexta; + struct state *intermediates; int progress; /* @@ -949,15 +1463,27 @@ pullback( progress = 0; for (s=nfa->states ; s!=NULL && !NISERR() ; s=nexts) { nexts = s->next; + intermediates = NULL; for (a=s->outs ; a!=NULL && !NISERR() ; a=nexta) { nexta = a->outchain; if (a->type == '^' || a->type == BEHIND) { - if (pull(nfa, a)) { + if (pull(nfa, a, &intermediates)) { progress = 1; } } assert(nexta == NULL || s->no != FREESTATE); } + /* clear tmp fields of intermediate states created here */ + while (intermediates != NULL) { + struct state *ns = intermediates->tmp; + + intermediates->tmp = NULL; + intermediates = ns; + } + /* if s is now useless, get rid of it */ + if ((s->nins == 0 || s->nouts == 0) && !s->flag) { + dropstate(nfa, s); + } } if (progress && f != NULL) { dumpnfa(nfa, f); @@ -967,6 +1493,12 @@ pullback( return; } + /* + * Any ^ constraints we were able to pull to the start state can now be + * replaced by PLAIN arcs referencing the BOS or BOL colors. There should + * be no other ^ or BEHIND arcs left in the NFA, though we do not check + * that here (compact() will fail if so). + */ for (a=nfa->pre->outs ; a!=NULL ; a=nexta) { nexta = a->outchain; if (a->type == '^') { @@ -979,15 +1511,28 @@ pullback( /* - pull - pull a back constraint backward past its source state - * A significant property of this function is that it deletes at most - * one state -- the constraint's from state -- and only if the constraint - * was that state's last outarc. + * + * Returns 1 if successful (which it always is unless the source is the + * start state or we have an internal error), 0 if nothing happened. + * + * A significant property of this function is that it deletes no pre-existing + * states, and no outarcs of the constraint's from state other than the given + * constraint arc. This makes the loops in pullback() safe, at the cost that + * we may leave useless states behind. Therefore, we leave it to pullback() + * to delete such states. + * + * If the from state has multiple back-constraint outarcs, and/or multiple + * compatible constraint inarcs, we only need to create one new intermediate + * state per combination of predecessor and successor states. *intermediates + * points to a list of such intermediate states for this from state (chained + * through their tmp fields). ^ static int pull(struct nfa *, struct arc *); */ -static int /* 0 couldn't, 1 could */ +static int pull( struct nfa *nfa, - struct arc *con) + struct arc *con, + struct state **intermediates) { struct state *from = con->from; struct state *to = con->to; @@ -995,10 +1540,7 @@ pull( struct arc *nexta; struct state *s; - if (from == to) { /* circular constraint is pointless */ - freearc(nfa, con); - return 1; - } + assert(from != to); /* should have gotten rid of this earlier */ if (from->flag) { /* can't pull back beyond start */ return 0; } @@ -1008,26 +1550,9 @@ pull( } /* - * DGP 2007-11-15: Cloning a state with a circular constraint on its list - * of outs can lead to trouble [Bug 1810038], so get rid of them first. - */ - - for (a = from->outs; a != NULL; a = nexta) { - nexta = a->outchain; - switch (a->type) { - case '^': - case '$': - case BEHIND: - case AHEAD: - if (from == a->to) { - freearc(nfa, a); - } - break; - } - } - - /* - * First, clone from state if necessary to avoid other outarcs. + * First, clone from state if necessary to avoid other outarcs. This may + * seem wasteful, but it simplifies the logic, and we'll get rid of the + * clone state again at the bottom. */ if (from->nouts > 1) { @@ -1035,10 +1560,12 @@ pull( if (NISERR()) { return 0; } - assert(to != from); /* con is not an inarc */ - copyins(nfa, from, s, 1); /* duplicate inarcs */ + copyins(nfa, from, s); /* duplicate inarcs */ cparc(nfa, con, s, to); /* move constraint arc */ freearc(nfa, con); + if (NISERR()) { + return 0; + } from = s; con = from->outs; } @@ -1048,7 +1575,7 @@ pull( * Propagate the constraint into the from state's inarcs. */ - for (a=from->ins ; a!=NULL ; a=nexta) { + for (a=from->ins ; a!=NULL && !NISERR(); a=nexta) { nexta = a->inchain; switch (combine(con, a)) { case INCOMPATIBLE: /* destroy the arc */ @@ -1057,17 +1584,25 @@ pull( case SATISFIED: /* no action needed */ break; case COMPATIBLE: /* swap the two arcs, more or less */ - s = newstate(nfa); - if (NISERR()) { - return 0; + /* need an intermediate state, but might have one already */ + for (s = *intermediates; s != NULL; s = s->tmp) { + assert(s->nins > 0 && s->nouts > 0); + if (s->ins->from == a->from && s->outs->to == to) { + break; + } } - cparc(nfa, a, s, to); /* anticipate move */ - cparc(nfa, con, a->from, s); - if (NISERR()) { - return 0; + if (s == NULL) { + s = newstate(nfa); + if (NISERR()) { + return 0; + } + s->tmp = *intermediates; + *intermediates = s; } - freearc(nfa, a); - break; + cparc(nfa, con, a->from, s); + cparc(nfa, a, s, to); + freearc(nfa, a); + break; default: assert(NOTREACHED); break; @@ -1079,12 +1614,13 @@ pull( */ moveins(nfa, from, to); - dropstate(nfa, from); /* will free the constraint */ + freearc(nfa, con); + /* from state is now useless, but we leave it to pullback() to clean up */ return 1; } /* - - pushfwd - push forward constraints forward to (with luck) eliminate them + - pushfwd - push forward constraints forward to eliminate them ^ static void pushfwd(struct nfa *, FILE *); */ static void @@ -1096,6 +1632,7 @@ pushfwd( struct state *nexts; struct arc *a; struct arc *nexta; + struct state *intermediates; int progress; /* @@ -1106,14 +1643,25 @@ pushfwd( progress = 0; for (s=nfa->states ; s!=NULL && !NISERR() ; s=nexts) { nexts = s->next; + intermediates = NULL; for (a = s->ins; a != NULL && !NISERR(); a = nexta) { nexta = a->inchain; if (a->type == '$' || a->type == AHEAD) { - if (push(nfa, a)) { + if (push(nfa, a, &intermediates)) { progress = 1; } } - assert(nexta == NULL || s->no != FREESTATE); + } + /* clear tmp fields of intermediate states created here */ + while (intermediates != NULL) { + struct state *ns = intermediates->tmp; + + intermediates->tmp = NULL; + intermediates = ns; + } + /* if s is now useless, get rid of it */ + if ((s->nins == 0 || s->nouts == 0) && !s->flag) { + dropstate(nfa, s); } } if (progress && f != NULL) { @@ -1124,6 +1672,12 @@ pushfwd( return; } + /* + * Any $ constraints we were able to push to the post state can now be + * replaced by PLAIN arcs referencing the EOS or EOL colors. There should + * be no other $ or AHEAD arcs left in the NFA, though we do not check + * that here (compact() will fail if so). + */ for (a = nfa->post->ins; a != NULL; a = nexta) { nexta = a->inchain; if (a->type == '$') { @@ -1136,15 +1690,28 @@ pushfwd( /* - push - push a forward constraint forward past its destination state - * A significant property of this function is that it deletes at most - * one state -- the constraint's to state -- and only if the constraint - * was that state's last inarc. + * + * Returns 1 if successful (which it always is unless the destination is the + * post state or we have an internal error), 0 if nothing happened. + * + * A significant property of this function is that it deletes no pre-existing + * states, and no inarcs of the constraint's to state other than the given + * constraint arc. This makes the loops in pushfwd() safe, at the cost that + * we may leave useless states behind. Therefore, we leave it to pushfwd() + * to delete such states. + * + * If the to state has multiple forward-constraint inarcs, and/or multiple + * compatible constraint outarcs, we only need to create one new intermediate + * state per combination of predecessor and successor states. *intermediates + * points to a list of such intermediate states for this to state (chained + * through their tmp fields). ^ static int push(struct nfa *, struct arc *); */ -static int /* 0 couldn't, 1 could */ +static int push( struct nfa *nfa, - struct arc *con) + struct arc *con, + struct state **intermediates) { struct state *from = con->from; struct state *to = con->to; @@ -1152,10 +1719,7 @@ push( struct arc *nexta; struct state *s; - if (to == from) { /* circular constraint is pointless */ - freearc(nfa, con); - return 1; - } + assert(to != from); /* should have gotten rid of this earlier */ if (to->flag) { /* can't push forward beyond end */ return 0; } @@ -1165,29 +1729,9 @@ push( } /* - * DGP 2007-11-15: Here we duplicate the same protections as appear - * in pull() above to avoid troubles with cloning a state with a - * circular constraint on its list of ins. It is not clear whether - * this is necessary, or is protecting against a "can't happen". - * Any test case that actually leads to a freearc() call here would - * be a welcome addition to the test suite. - */ - - for (a = to->ins; a != NULL; a = nexta) { - nexta = a->inchain; - switch (a->type) { - case '^': - case '$': - case BEHIND: - case AHEAD: - if (a->from == to) { - freearc(nfa, a); - } - break; - } - } - /* - * First, clone to state if necessary to avoid other inarcs. + * First, clone to state if necessary to avoid other inarcs. This may + * seem wasteful, but it simplifies the logic, and we'll get rid of the + * clone state again at the bottom. */ if (to->nins > 1) { @@ -1195,9 +1739,12 @@ push( if (NISERR()) { return 0; } - copyouts(nfa, to, s, 1); /* duplicate outarcs */ - cparc(nfa, con, from, s); /* move constraint */ + copyouts(nfa, to, s); /* duplicate outarcs */ + cparc(nfa, con, from, s); /* move constraint arc */ freearc(nfa, con); + if (NISERR()) { + return 0; + } to = s; con = to->ins; } @@ -1207,7 +1754,7 @@ push( * Propagate the constraint into the to state's outarcs. */ - for (a = to->outs; a != NULL; a = nexta) { + for (a = to->outs; a != NULL && !NISERR(); a = nexta) { nexta = a->outchain; switch (combine(con, a)) { case INCOMPATIBLE: /* destroy the arc */ @@ -1216,17 +1763,25 @@ push( case SATISFIED: /* no action needed */ break; case COMPATIBLE: /* swap the two arcs, more or less */ - s = newstate(nfa); - if (NISERR()) { - return 0; + /* need an intermediate state, but might have one already */ + for (s = *intermediates; s != NULL; s = s->tmp) { + assert(s->nins > 0 && s->nouts > 0); + if (s->ins->from == from && s->outs->to == a->to) { + break; + } } - cparc(nfa, con, s, a->to); /* anticipate move */ - cparc(nfa, a, from, s); - if (NISERR()) { - return 0; + if (s == NULL) { + s = newstate(nfa); + if (NISERR()) { + return 0; + } + s->tmp = *intermediates; + *intermediates = s; } - freearc(nfa, a); - break; + cparc(nfa, con, s, a->to); + cparc(nfa, a, from, s); + freearc(nfa, a); + break; default: assert(NOTREACHED); break; @@ -1238,7 +1793,8 @@ push( */ moveouts(nfa, to, from); - dropstate(nfa, to); /* will free the constraint */ + freearc(nfa, con); + /* to state is now useless, but we leave it to pushfwd() to clean up */ return 1; } @@ -1316,6 +1872,12 @@ fixempties( struct state *nexts; struct arc *a; struct arc *nexta; + int totalinarcs; + struct arc **inarcsorig; + struct arc **arcarray; + int arccount; + int prevnins; + int nskip; /* * First, get rid of any states whose sole out-arc is an EMPTY, @@ -1361,42 +1923,129 @@ fixempties( dropstate(nfa, s); } + if (NISERR()) { + return; + } + /* - * For each remaining NFA state, find all other states that are - * reachable from it by a chain of one or more EMPTY arcs. Then - * generate new arcs that eliminate the need for each such chain. + * For each remaining NFA state, find all other states from which it is + * reachable by a chain of one or more EMPTY arcs. Then generate new arcs + * that eliminate the need for each such chain. + * + * We could replace a chain of EMPTY arcs that leads from a "from" state + * to a "to" state either by pushing non-EMPTY arcs forward (linking + * directly from "from"'s predecessors to "to") or by pulling them back + * (linking directly from "from" to "to"'s successors). We choose to + * always do the former; this choice is somewhat arbitrary, but the + * approach below requires that we uniformly do one or the other. + * + * Suppose we have a chain of N successive EMPTY arcs (where N can easily + * approach the size of the NFA). All of the intermediate states must + * have additional inarcs and outarcs, else they'd have been removed by + * the steps above. Assuming their inarcs are mostly not empties, we will + * add O(N^2) arcs to the NFA, since a non-EMPTY inarc leading to any one + * state in the chain must be duplicated to lead to all its successor + * states as well. So there is no hope of doing less than O(N^2) work; + * however, we should endeavor to keep the big-O cost from being even + * worse than that, which it can easily become without care. In + * particular, suppose we were to copy all S1's inarcs forward to S2, and + * then also to S3, and then later we consider pushing S2's inarcs forward + * to S3. If we include the arcs already copied from S1 in that, we'd be + * doing O(N^3) work. (The duplicate-arc elimination built into newarc() + * and its cohorts would get rid of the extra arcs, but not without cost.) + * + * We can avoid this cost by treating only arcs that existed at the start + * of this phase as candidates to be pushed forward. To identify those, + * we remember the first inarc each state had to start with. We rely on + * the fact that newarc() and friends put new arcs on the front of their + * to-states' inchains, and that this phase never deletes arcs, so that + * the original arcs must be the last arcs in their to-states' inchains. * - * If we just do this straightforwardly, the algorithm gets slow in - * complex graphs, because the same arcs get copied to all - * intermediate states of an EMPTY chain, and then uselessly pushed - * repeatedly to the chain's final state; we waste a lot of time in - * newarc's duplicate checking. To improve matters, we decree that - * any state with only EMPTY out-arcs is "doomed" and will not be - * part of the final NFA. That can be ensured by not adding any new - * out-arcs to such a state. Having ensured that, we need not update - * the state's in-arcs list either; all arcs that might have gotten - * pushed forward to it will just get pushed directly to successor - * states. This eliminates most of the useless duplicate arcs. + * So the process here is that, for each state in the NFA, we gather up + * all non-EMPTY inarcs of states that can reach the target state via + * EMPTY arcs. We then sort, de-duplicate, and merge these arcs into the + * target state's inchain. (We can safely use sort-merge for this as long + * as we update each state's original-arcs pointer after we add arcs to + * it; the sort step of mergeins probably changed the order of the old + * arcs.) + * + * Another refinement worth making is that, because we only add non-EMPTY + * arcs during this phase, and all added arcs have the same from-state as + * the non-EMPTY arc they were cloned from, we know ahead of time that any + * states having only EMPTY outarcs will be useless for lack of outarcs + * after we drop the EMPTY arcs. (They cannot gain non-EMPTY outarcs if + * they had none to start with.) So we need not bother to update the + * inchains of such states at all. + */ + + /* Remember the states' first original inarcs */ + /* ... and while at it, count how many old inarcs there are altogether */ + inarcsorig = (struct arc **) MALLOC(nfa->nstates * sizeof(struct arc *)); + if (inarcsorig == NULL) { + NERR(REG_ESPACE); + return; + } + totalinarcs = 0; + for (s = nfa->states; s != NULL; s = s->next) { + inarcsorig[s->no] = s->ins; + totalinarcs += s->nins; + } + + /* + * Create a workspace for accumulating the inarcs to be added to the + * current target state. totalinarcs is probably a considerable + * overestimate of the space needed, but the NFA is unlikely to be large + * enough at this point to make it worth being smarter. */ + arcarray = (struct arc **) MALLOC(totalinarcs * sizeof(struct arc *)); + if (arcarray == NULL) { + NERR(REG_ESPACE); + FREE(inarcsorig); + return; + } + + /* And iterate over the target states */ for (s = nfa->states; s != NULL && !NISERR(); s = s->next) { - for (s2 = emptyreachable(s, s); s2 != s && !NISERR(); - s2 = nexts) { - /* - * If s2 is doomed, we decide that (1) we will always push - * arcs forward to it, not pull them back to s; and (2) we - * can optimize away the push-forward, per comment above. - * So do nothing. - */ - if (s2->flag || hasnonemptyout(s2)) { - replaceempty(nfa, s, s2); - } + /* Ignore target states without non-EMPTY outarcs, per note above */ + if (!s->flag && !hasnonemptyout(s)) { + continue; + } - /* Reset the tmp fields as we walk back */ - nexts = s2->tmp; - s2->tmp = NULL; + /* Find predecessor states and accumulate their original inarcs */ + arccount = 0; + for (s2 = emptyreachable(nfa, s, s, inarcsorig); s2 != s; s2 = nexts) { + /* Add s2's original inarcs to arcarray[], but ignore empties */ + for (a = inarcsorig[s2->no]; a != NULL; a = a->inchain) { + if (a->type != EMPTY) { + arcarray[arccount++] = a; + } + } + + /* Reset the tmp fields as we walk back */ + nexts = s2->tmp; + s2->tmp = NULL; + } + s->tmp = NULL; + assert(arccount <= totalinarcs); + + /* Remember how many original inarcs this state has */ + prevnins = s->nins; + + /* Add non-duplicate inarcs to target state */ + mergeins(nfa, s, arcarray, arccount); + + /* Now we must update the state's inarcsorig pointer */ + nskip = s->nins - prevnins; + a = s->ins; + while (nskip-- > 0) { + a = a->inchain; } - s->tmp = NULL; + inarcsorig[s->no] = a; } + + FREE(arcarray); + FREE(inarcsorig); + if (NISERR()) { return; } @@ -1432,89 +2081,604 @@ fixempties( } /* - - emptyreachable - recursively find all states reachable from s by EMPTY arcs + - emptyreachable - recursively find all states that can reach s by EMPTY arcs * The return value is the last such state found. Its tmp field links back * to the next-to-last such state, and so on back to s, so that all these * states can be located without searching the whole NFA. + * + * Since this is only used in fixempties(), we pass in the inarcsorig[] array + * maintained by that function. This lets us skip over all new inarcs, which + * are certainly not EMPTY arcs. + * * The maximum recursion depth here is equal to the length of the longest * loop-free chain of EMPTY arcs, which is surely no more than the size of - * the NFA, and in practice will be a lot less than that. + * the NFA, and in practice will be less than that. ^ static struct state *emptyreachable(struct state *, struct state *); */ static struct state * emptyreachable( + struct nfa *nfa, struct state *s, - struct state *lastfound) + struct state *lastfound, + struct arc **inarcsorig) { struct arc *a; s->tmp = lastfound; lastfound = s; - for (a = s->outs; a != NULL; a = a->outchain) { - if (a->type == EMPTY && a->to->tmp == NULL) { - lastfound = emptyreachable(a->to, lastfound); + for (a = inarcsorig[s->no]; a != NULL; a = a->inchain) { + if (a->type == EMPTY && a->from->tmp == NULL) { + lastfound = emptyreachable(nfa, a->from, lastfound, inarcsorig); } } return lastfound; } - + +/* + * isconstraintarc - detect whether an arc is of a constraint type + */ +static inline int +isconstraintarc(struct arc * a) +{ + switch (a->type) + { + case '^': + case '$': + case BEHIND: + case AHEAD: + case LACON: + return 1; + } + return 0; +} + /* - - replaceempty - replace an EMPTY arc chain with some non-empty arcs - * The EMPTY arc(s) should be deleted later, but we can't do it here because - * they may still be needed to identify other arc chains during fixempties(). - ^ static void replaceempty(struct nfa *, struct state *, struct state *); + * hasconstraintout - does state have a constraint out arc? + */ +static int +hasconstraintout(struct state * s) +{ + struct arc *a; + + for (a = s->outs; a != NULL; a = a->outchain) { + if (isconstraintarc(a)) { + return 1; + } + } + return 0; +} + +/* + * fixconstraintloops - get rid of loops containing only constraint arcs + * + * A loop of states that contains only constraint arcs is useless, since + * passing around the loop represents no forward progress. Moreover, it + * would cause infinite looping in pullback/pushfwd, so we need to get rid + * of such loops before doing that. */ static void -replaceempty( - struct nfa *nfa, - struct state *from, - struct state *to) +fixconstraintloops( + struct nfa * nfa, + FILE *f) /* for debug output; NULL none */ { - int fromouts; - int toins; + struct state *s; + struct state *nexts; + struct arc *a; + struct arc *nexta; + int hasconstraints; + + /* + * In the trivial case of a state that loops to itself, we can just drop + * the constraint arc altogether. This is worth special-casing because + * such loops are far more common than loops containing multiple states. + * While we're at it, note whether any constraint arcs survive. + */ + hasconstraints = 0; + for (s = nfa->states; s != NULL && !NISERR(); s = nexts) { + nexts = s->next; + /* while we're at it, ensure tmp fields are clear for next step */ + assert(s->tmp == NULL); + for (a = s->outs; a != NULL && !NISERR(); a = nexta) { + nexta = a->outchain; + if (isconstraintarc(a)) { + if (a->to == s) { + freearc(nfa, a); + } else { + hasconstraints = 1; + } + } + } + /* If we removed all the outarcs, the state is useless. */ + if (s->nouts == 0 && !s->flag) { + dropstate(nfa, s); + } + } + + /* Nothing to do if no remaining constraint arcs */ + if (NISERR() || !hasconstraints) { + return; + } + + /* + * Starting from each remaining NFA state, search outwards for a + * constraint loop. If we find a loop, break the loop, then start the + * search over. (We could possibly retain some state from the first scan, + * but it would complicate things greatly, and multi-state constraint + * loops are rare enough that it's not worth optimizing the case.) + */ + restart: + for (s = nfa->states; s != NULL && !NISERR(); s = s->next) { + if (findconstraintloop(nfa, s)) { + goto restart; + } + } - assert(from != to); + if (NISERR()) { + return; + } /* - * Create replacement arcs that bypass the need for the EMPTY chain. We - * can do this either by pushing arcs forward (linking directly from - * "from"'s predecessors to "to") or by pulling them back (linking - * directly from "from" to "to"'s successors). In general, we choose - * whichever way creates greater fan-out or fan-in, so as to improve the - * odds of reducing the other state to zero in-arcs or out-arcs and - * thereby being able to delete it. However, if "from" is doomed (has no - * non-EMPTY out-arcs), we must keep it so, so always push forward in that - * case. + * Now remove any states that have become useless. (This cleanup is not + * very thorough, and would be even less so if we tried to combine it with + * the previous step; but cleanup() will take care of anything we miss.) * - * The fan-out/fan-in comparison should count only non-EMPTY arcs. If - * "from" is doomed, we can skip counting "to"'s arcs, since we want to - * force taking the copynonemptyins path in that case. + * Because findconstraintloop intentionally doesn't reset all tmp fields, + * we have to clear them after it's done. This is a convenient place to + * do that, too. + */ + for (s = nfa->states; s != NULL; s = nexts) { + nexts = s->next; + s->tmp = NULL; + if ((s->nins == 0 || s->nouts == 0) && !s->flag) { + dropstate(nfa, s); + } + } + + if (f != NULL) { + dumpnfa(nfa, f); + } +} + +/* + * findconstraintloop - recursively find a loop of constraint arcs + * + * If we find a loop, break it by calling breakconstraintloop(), then + * return 1; otherwise return 0. + * + * State tmp fields are guaranteed all NULL on a success return, because + * breakconstraintloop does that. After a failure return, any state that + * is known not to be part of a loop is marked with s->tmp == s; this allows + * us not to have to re-prove that fact on later calls. (This convention is + * workable because we already eliminated single-state loops.) + * + * Note that the found loop doesn't necessarily include the first state we + * are called on. Any loop reachable from that state will do. + * + * The maximum recursion depth here is one more than the length of the longest + * loop-free chain of constraint arcs, which is surely no more than the size + * of the NFA, and in practice will be a lot less than that. + */ +static int +findconstraintloop(struct nfa * nfa, struct state * s) +{ + struct arc *a; + + /* Since this is recursive, it could be driven to stack overflow */ + if (STACK_TOO_DEEP(nfa->v->re)) { + NERR(REG_ETOOBIG); + return 1; /* to exit as quickly as possible */ + } + + if (s->tmp != NULL) { + /* Already proven uninteresting? */ + if (s->tmp == s) { + return 0; + } + /* Found a loop involving s */ + breakconstraintloop(nfa, s); + /* The tmp fields have been cleaned up by breakconstraintloop */ + return 1; + } + for (a = s->outs; a != NULL; a = a->outchain) { + if (isconstraintarc(a)) { + struct state *sto = a->to; + + assert(sto != s); + s->tmp = sto; + if (findconstraintloop(nfa, sto)) { + return 1; + } + } + } + + /* + * If we get here, no constraint loop exists leading out from s. Mark it + * with s->tmp == s so we need not rediscover that fact again later. + */ + s->tmp = s; + return 0; +} + +/* + * breakconstraintloop - break a loop of constraint arcs + * + * sinitial is any one member state of the loop. Each loop member's tmp + * field links to its successor within the loop. (Note that this function + * will reset all the tmp fields to NULL.) + * + * We can break the loop by, for any one state S1 in the loop, cloning its + * loop successor state S2 (and possibly following states), and then moving + * all S1->S2 constraint arcs to point to the cloned S2. The cloned S2 should + * copy any non-constraint outarcs of S2. Constraint outarcs should be + * dropped if they point back to S1, else they need to be copied as arcs to + * similarly cloned states S3, S4, etc. In general, each cloned state copies + * non-constraint outarcs, drops constraint outarcs that would lead to itself + * or any earlier cloned state, and sends other constraint outarcs to newly + * cloned states. No cloned state will have any inarcs that aren't constraint + * arcs or do not lead from S1 or earlier-cloned states. It's okay to drop + * constraint back-arcs since they would not take us to any state we've not + * already been in; therefore, no new constraint loop is created. In this way + * we generate a modified NFA that can still represent every useful state + * sequence, but not sequences that represent state loops with no consumption + * of input data. Note that the set of cloned states will certainly include + * all of the loop member states other than S1, and it may also include + * non-loop states that are reachable from S2 via constraint arcs. This is + * important because there is no guarantee that findconstraintloop found a + * maximal loop (and searching for one would be NP-hard, so don't try). + * Frequently the "non-loop states" are actually part of a larger loop that + * we didn't notice, and indeed there may be several overlapping loops. + * This technique ensures convergence in such cases, while considering only + * the originally-found loop does not. + * + * If there is only one S1->S2 constraint arc, then that constraint is + * certainly satisfied when we enter any of the clone states. This means that + * in the common case where many of the constraint arcs are identically + * labeled, we can merge together clone states linked by a similarly-labeled + * constraint: if we can get to the first one we can certainly get to the + * second, so there's no need to distinguish. This greatly reduces the number + * of new states needed, so we preferentially break the given loop at a state + * pair where this is true. + * + * Furthermore, it's fairly common to find that a cloned successor state has + * no outarcs, especially if we're a bit aggressive about removing unnecessary + * outarcs. If that happens, then there is simply not any interesting state + * that can be reached through the predecessor's loop arcs, which means we can + * break the loop just by removing those loop arcs, with no new states added. + */ +static void +breakconstraintloop(struct nfa * nfa, struct state * sinitial) +{ + struct state *s; + struct state *shead; + struct state *stail; + struct state *sclone; + struct state *nexts; + struct arc *refarc; + struct arc *a; + struct arc *nexta; + + /* + * Start by identifying which loop step we want to break at. + * Preferentially this is one with only one constraint arc. (XXX are + * there any other secondary heuristics we want to use here?) Set refarc + * to point to the selected lone constraint arc, if there is one. + */ + refarc = NULL; + s = sinitial; + do { + nexts = s->tmp; + assert(nexts != s); /* should not see any one-element loops */ + if (refarc == NULL) { + int narcs = 0; + + for (a = s->outs; a != NULL; a = a->outchain) { + if (a->to == nexts && isconstraintarc(a)) { + refarc = a; + narcs++; + } + } + assert(narcs > 0); + if (narcs > 1) { + refarc = NULL; /* multiple constraint arcs here, no good */ + } + } + s = nexts; + } while (s != sinitial); + + if (refarc) { + /* break at the refarc */ + shead = refarc->from; + stail = refarc->to; + assert(stail == shead->tmp); + } else { + /* for lack of a better idea, break after sinitial */ + shead = sinitial; + stail = sinitial->tmp; + } + + /* + * Reset the tmp fields so that we can use them for local storage in + * clonesuccessorstates. (findconstraintloop won't mind, since it's just + * going to abandon its search anyway.) */ - fromouts = nonemptyouts(from); - toins = (fromouts == 0) ? 1 : nonemptyins(to); + for (s = nfa->states; s != NULL; s = s->next) { + s->tmp = NULL; + } - if (fromouts > toins) { - copyouts(nfa, to, from, 0); + /* + * Recursively build clone state(s) as needed. + */ + sclone = newstate(nfa); + if (sclone == NULL) { + assert(NISERR()); return; } - if (fromouts < toins) { - copyins(nfa, from, to, 0); + + clonesuccessorstates(nfa, stail, sclone, shead, refarc, + NULL, NULL, nfa->nstates); + + if (NISERR()) { return; } /* - * fromouts == toins. Decide on secondary issue: copy fewest arcs. - * - * Doesn't seem to be worth the trouble to exclude empties from these - * comparisons; that takes extra time and doesn't seem to improve the - * resulting graph much. + * It's possible that sclone has no outarcs at all, in which case it's + * useless. (We don't try extremely hard to get rid of useless states + * here, but this is an easy and fairly common case.) */ - if (from->nins > to->nouts) { - copyouts(nfa, to, from, 0); + if (sclone->nouts == 0) { + freestate(nfa, sclone); + sclone = NULL; + } + + /* + * Move shead's constraint-loop arcs to point to sclone, or just drop them + * if we discovered we don't need sclone. + */ + for (a = shead->outs; a != NULL; a = nexta) { + nexta = a->outchain; + if (a->to == stail && isconstraintarc(a)) { + if (sclone) { + cparc(nfa, a, shead, sclone); + } + freearc(nfa, a); + if (NISERR()) { + break; + } + } + } +} + +/* + * clonesuccessorstates - create a tree of constraint-arc successor states + * + * ssource is the state to be cloned, and sclone is the state to copy its + * outarcs into. sclone's inarcs, if any, should already be set up. + * + * spredecessor is the original predecessor state that we are trying to build + * successors for (it may not be the immediate predecessor of ssource). + * refarc, if not NULL, is the original constraint arc that is known to have + * been traversed out of spredecessor to reach the successor(s). + * + * For each cloned successor state, we transiently create a "donemap" that is + * a boolean array showing which source states we've already visited for this + * clone state. This prevents infinite recursion as well as useless repeat + * visits to the same state subtree (which can add up fast, since typical NFAs + * have multiple redundant arc pathways). Each donemap is a char array + * indexed by state number. The donemaps are all of the same size "nstates", + * which is nfa->nstates as of the start of the recursion. This is enough to + * have entries for all pre-existing states, but *not* entries for clone + * states created during the recursion. That's okay since we have no need to + * mark those. + * + * curdonemap is NULL when recursing to a new sclone state, or sclone's + * donemap when we are recursing without having created a new state (which we + * do when we decide we can merge a successor state into the current clone + * state). outerdonemap is NULL at the top level and otherwise the parent + * clone state's donemap. + * + * The successor states we create and fill here form a strict tree structure, + * with each state having exactly one predecessor, except that the toplevel + * state has no inarcs as yet (breakconstraintloop will add its inarcs from + * spredecessor after we're done). Thus, we can examine sclone's inarcs back + * to the root, plus refarc if any, to identify the set of constraints already + * known valid at the current point. This allows us to avoid generating extra + * successor states. + */ +static void +clonesuccessorstates( + struct nfa * nfa, + struct state * ssource, + struct state * sclone, + struct state * spredecessor, + struct arc * refarc, + char *curdonemap, + char *outerdonemap, + int nstates) +{ + char *donemap; + struct arc *a; + + /* Since this is recursive, it could be driven to stack overflow */ + if (STACK_TOO_DEEP(nfa->v->re)) { + NERR(REG_ETOOBIG); return; } - copyins(nfa, from, to, 0); + /* If this state hasn't already got a donemap, create one */ + donemap = curdonemap; + if (donemap == NULL) { + donemap = (char *) MALLOC(nstates * sizeof(char)); + if (donemap == NULL) { + NERR(REG_ESPACE); + return; + } + + if (outerdonemap != NULL) { + /* + * Not at outermost recursion level, so copy the outer level's + * donemap; this ensures that we see states in process of being + * visited at outer levels, or already merged into predecessor + * states, as ones we shouldn't traverse back to. + */ + memcpy(donemap, outerdonemap, nstates * sizeof(char)); + } else { + /* At outermost level, only spredecessor is off-limits */ + memset(donemap, 0, nstates * sizeof(char)); + assert(spredecessor->no < nstates); + donemap[spredecessor->no] = 1; + } + } + + /* Mark ssource as visited in the donemap */ + assert(ssource->no < nstates); + assert(donemap[ssource->no] == 0); + donemap[ssource->no] = 1; + + /* + * We proceed by first cloning all of ssource's outarcs, creating new + * clone states as needed but not doing more with them than that. Then in + * a second pass, recurse to process the child clone states. This allows + * us to have only one child clone state per reachable source state, even + * when there are multiple outarcs leading to the same state. Also, when + * we do visit a child state, its set of inarcs is known exactly, which + * makes it safe to apply the constraint-is-already-checked optimization. + * Also, this ensures that we've merged all the states we can into the + * current clone before we recurse to any children, thus possibly saving + * them from making extra images of those states. + * + * While this function runs, child clone states of the current state are + * marked by setting their tmp fields to point to the original state they + * were cloned from. This makes it possible to detect multiple outarcs + * leading to the same state, and also makes it easy to distinguish clone + * states from original states (which will have tmp == NULL). + */ + for (a = ssource->outs; a != NULL && !NISERR(); a = a->outchain) { + struct state *sto = a->to; + + /* + * We do not consider cloning successor states that have no constraint + * outarcs; just link to them as-is. They cannot be part of a + * constraint loop so there is no need to make copies. In particular, + * this rule keeps us from trying to clone the post state, which would + * be a bad idea. + */ + if (isconstraintarc(a) && hasconstraintout(sto)) { + struct state *prevclone; + int canmerge; + struct arc *a2; + + /* + * Back-link constraint arcs must not be followed. Nor is there a + * need to revisit states previously merged into this clone. + */ + assert(sto->no < nstates); + if (donemap[sto->no] != 0) { + continue; + } + + /* + * Check whether we already have a child clone state for this + * source state. + */ + prevclone = NULL; + for (a2 = sclone->outs; a2 != NULL; a2 = a2->outchain) { + if (a2->to->tmp == sto) { + prevclone = a2->to; + break; + } + } + + /* + * If this arc is labeled the same as refarc, or the same as any + * arc we must have traversed to get to sclone, then no additional + * constraints need to be met to get to sto, so we should just + * merge its outarcs into sclone. + */ + if (refarc && a->type == refarc->type && a->co == refarc->co) { + canmerge = 1; + } else { + struct state *s; + + canmerge = 0; + for (s = sclone; s->ins; s = s->ins->from) { + if (s->nins == 1 && + a->type == s->ins->type && a->co == s->ins->co) { + canmerge = 1; + break; + } + } + } + + if (canmerge) { + /* + * We can merge into sclone. If we previously made a child + * clone state, drop it; there's no need to visit it. (This + * can happen if ssource has multiple pathways to sto, and we + * only just now found one that is provably a no-op.) + */ + if (prevclone) { + dropstate(nfa, prevclone); /* kills our outarc, too */ + } + + /* Recurse to merge sto's outarcs into sclone */ + clonesuccessorstates(nfa, sto, sclone, spredecessor, refarc, + donemap, outerdonemap, nstates); + /* sto should now be marked as previously visited */ + assert(NISERR() || donemap[sto->no] == 1); + } else if (prevclone) { + /* + * We already have a clone state for this successor, so just + * make another arc to it. + */ + cparc(nfa, a, sclone, prevclone); + } else { + /* + * We need to create a new successor clone state. + */ + struct state *stoclone; + + stoclone = newstate(nfa); + if (stoclone == NULL) { + assert(NISERR()); + break; + } + /* Mark it as to what it's a clone of */ + stoclone->tmp = sto; + /* ... and add the outarc leading to it */ + cparc(nfa, a, sclone, stoclone); + } + } else { + /* + * Non-constraint outarcs just get copied to sclone, as do outarcs + * leading to states with no constraint outarc. + */ + cparc(nfa, a, sclone, sto); + } + } + + /* + * If we are at outer level for this clone state, recurse to all its child + * clone states, clearing their tmp fields as we go. (If we're not + * outermost for sclone, leave this to be done by the outer call level.) + * Note that if we have multiple outarcs leading to the same clone state, + * it will only be recursed-to once. + */ + if (curdonemap == NULL) { + for (a = sclone->outs; a != NULL && !NISERR(); a = a->outchain) { + struct state *stoclone = a->to; + struct state *sto = stoclone->tmp; + + if (sto != NULL) { + stoclone->tmp = NULL; + clonesuccessorstates(nfa, sto, stoclone, spredecessor, refarc, + NULL, donemap, nstates); + } + } + + /* Don't forget to free sclone's donemap when done with it */ + FREE(donemap); + } } /* @@ -1631,7 +2795,7 @@ analyze( } /* - - compact - compact an NFA + - compact - construct the compact representation of an NFA ^ static void compact(struct nfa *, struct cnfa *); */ static void @@ -1702,11 +2866,11 @@ compact( cnfa->flags |= HASLACONS; break; default: - assert(NOTREACHED); + NERR(REG_ASSERT); break; } } - carcsort(first, ca-1); + carcsort(first, ca - first); ca->co = COLORLESS; ca->to = 0; ca++; @@ -1726,33 +2890,39 @@ compact( /* - carcsort - sort compacted-NFA arcs by color - * Really dumb algorithm, but if the list is long enough for that to matter, - * you're in real trouble anyway. ^ static void carcsort(struct carc *, struct carc *); */ static void carcsort( struct carc *first, - struct carc *last) + size_t n) { - struct carc *p; - struct carc *q; - struct carc tmp; - - if (last - first <= 1) { - return; + if (n > 1) { + qsort(first, n, sizeof(struct carc), carc_cmp); } +} - for (p = first; p <= last; p++) { - for (q = p; q <= last; q++) { - if (p->co > q->co || (p->co == q->co && p->to > q->to)) { - assert(p != q); - tmp = *p; - *p = *q; - *q = tmp; - } - } +static int +carc_cmp( + const void *a, + const void *b) +{ + const struct carc *aa = (const struct carc *) a; + const struct carc *bb = (const struct carc *) b; + + if (aa->co < bb->co) { + return -1; + } + if (aa->co > bb->co) { + return +1; + } + if (aa->to < bb->to) { + return -1; + } + if (aa->to > bb->to) { + return +1; } + return 0; } /* @@ -1781,6 +2951,8 @@ dumpnfa( { #ifdef REG_DEBUG struct state *s; + int nstates = 0; + int narcs = 0; fprintf(f, "pre %d, post %d", nfa->pre->no, nfa->post->no); if (nfa->bos[0] != COLORLESS) { @@ -1798,7 +2970,10 @@ dumpnfa( fprintf(f, "\n"); for (s = nfa->states; s != NULL; s = s->next) { dumpstate(s, f); + nstates++; + narcs += s->nouts; } + fprintf(f, "total of %d states, %d arcs\n", nstates, narcs); if (nfa->parent == NULL) { dumpcolors(nfa->cm, f); } @@ -1851,37 +3026,28 @@ dumparcs( FILE *f) { int pos; + struct arc *a; - assert(s->nouts > 0); - /* printing arcs in reverse order is usually clearer */ - pos = dumprarcs(s->outs, s, f, 1); - if (pos != 1) { - fprintf(f, "\n"); - } -} - -/* - - dumprarcs - dump remaining outarcs, recursively, in reverse order - ^ static int dumprarcs(struct arc *, struct state *, FILE *, int); - */ -static int /* resulting print position */ -dumprarcs( - struct arc *a, - struct state *s, - FILE *f, - int pos) /* initial print position */ -{ - if (a->outchain != NULL) { - pos = dumprarcs(a->outchain, s, f, pos); + /* printing oldest arcs first is usually clearer */ + a = s->outs; + assert(a != NULL); + while (a->outchain != NULL) { + a = a->outchain; } - dumparc(a, s, f); - if (pos == 5) { + pos = 1; + do { + dumparc(a, s, f); + if (pos == 5) { + fprintf(f, "\n"); + pos = 1; + } else { + pos++; + } + a = a->outchainRev; + } while (a != NULL); + if (pos != 1) { fprintf(f, "\n"); - pos = 1; - } else { - pos++; } - return pos; } /* diff --git a/generic/regcomp.c b/generic/regcomp.c index 32d4212..211cd70 100644 --- a/generic/regcomp.c +++ b/generic/regcomp.c @@ -116,17 +116,22 @@ static void dropstate(struct nfa *, struct state *); static void freestate(struct nfa *, struct state *); static void destroystate(struct nfa *, struct state *); static void newarc(struct nfa *, int, pcolor, struct state *, struct state *); +static void createarc(struct nfa *, int, pcolor, struct state *, struct state *); static struct arc *allocarc(struct nfa *, struct state *); static void freearc(struct nfa *, struct arc *); +static void changearctarget(struct arc *, struct state *); static int hasnonemptyout(struct state *); -static int nonemptyouts(struct state *); -static int nonemptyins(struct state *); static struct arc *findarc(struct state *, int, pcolor); static void cparc(struct nfa *, struct arc *, struct state *, struct state *); +static void sortins(struct nfa *, struct state *); +static int sortins_cmp(const void *, const void *); +static void sortouts(struct nfa *, struct state *); +static int sortouts_cmp(const void *, const void *); static void moveins(struct nfa *, struct state *, struct state *); -static void copyins(struct nfa *, struct state *, struct state *, int); +static void copyins(struct nfa *, struct state *, struct state *); +static void mergeins(struct nfa *, struct state *, struct arc **, int); static void moveouts(struct nfa *, struct state *, struct state *); -static void copyouts(struct nfa *, struct state *, struct state *, int); +static void copyouts(struct nfa *, struct state *, struct state *); static void cloneouts(struct nfa *, struct state *, struct state *, struct state *, int); static void delsub(struct nfa *, struct state *, struct state *); static void deltraverse(struct nfa *, struct state *, struct state *); @@ -136,28 +141,35 @@ static void cleartraverse(struct nfa *, struct state *); static void specialcolors(struct nfa *); static long optimize(struct nfa *, FILE *); static void pullback(struct nfa *, FILE *); -static int pull(struct nfa *, struct arc *); +static int pull(struct nfa *, struct arc *, struct state **); static void pushfwd(struct nfa *, FILE *); -static int push(struct nfa *, struct arc *); +static int push(struct nfa *, struct arc *, struct state **); #define INCOMPATIBLE 1 /* destroys arc */ #define SATISFIED 2 /* constraint satisfied */ #define COMPATIBLE 3 /* compatible but not satisfied yet */ static int combine(struct arc *, struct arc *); static void fixempties(struct nfa *, FILE *); -static struct state *emptyreachable(struct state *, struct state *); -static void replaceempty(struct nfa *, struct state *, struct state *); +static struct state *emptyreachable(struct nfa *, struct state *, + struct state *, struct arc **); +static int isconstraintarc(struct arc *); +static int hasconstraintout(struct state *); +static void fixconstraintloops(struct nfa *, FILE *); +static int findconstraintloop(struct nfa *, struct state *); +static void breakconstraintloop(struct nfa *, struct state *); +static void clonesuccessorstates(struct nfa *, struct state *, struct state *, + struct state *, struct arc *, char *, char *, int); static void cleanup(struct nfa *); static void markreachable(struct nfa *, struct state *, struct state *, struct state *); static void markcanreach(struct nfa *, struct state *, struct state *, struct state *); static long analyze(struct nfa *); static void compact(struct nfa *, struct cnfa *); -static void carcsort(struct carc *, struct carc *); +static void carcsort(struct carc *, size_t); +static int carc_cmp(const void *, const void *); static void freecnfa(struct cnfa *); static void dumpnfa(struct nfa *, FILE *); #ifdef REG_DEBUG static void dumpstate(struct state *, FILE *); static void dumparcs(struct state *, FILE *); -static int dumprarcs(struct arc *, struct state *, FILE *, int); static void dumparc(struct arc *, struct state *, FILE *); #endif static void dumpcnfa(struct cnfa *, FILE *); @@ -212,6 +224,7 @@ struct vars { struct cvec *cv2; /* utility cvec */ struct subre *lacons; /* lookahead-constraint vector */ int nlacons; /* size of lacons */ + size_t spaceused; /* approx. space used for compilation */ }; /* parsing macros; most know that `v' is the struct vars pointer */ @@ -323,6 +336,7 @@ compile( v->cv2 = NULL; v->lacons = NULL; v->nlacons = 0; + v->spaceused = 0; re->re_magic = REMAGIC; re->re_info = 0; /* bits get set during parse */ re->re_csize = sizeof(chr); @@ -611,8 +625,9 @@ makesearch( for (s=slist ; s!=NULL ; s=s2) { s2 = newstate(nfa); - - copyouts(nfa, s, s2, 1); + NOERR(); + copyouts(nfa, s, s2); + NOERR(); for (a=s->ins ; a!=NULL ; a=b) { b = a->inchain; @@ -2076,7 +2091,7 @@ dump( dumpcolors(&g->cmap, f); if (!NULLCNFA(g->search)) { - printf("\nsearch:\n"); + fprintf(f, "\nsearch:\n"); dumpcnfa(&g->search, f); } for (i = 1; i < g->nlacons; i++) { diff --git a/generic/regerrs.h b/generic/regerrs.h index 72548ff..ee203d5 100644 --- a/generic/regerrs.h +++ b/generic/regerrs.h @@ -16,5 +16,5 @@ { REG_INVARG, "REG_INVARG", "invalid argument to regex function" }, { REG_MIXED, "REG_MIXED", "character widths of regex and string differ" }, { REG_BADOPT, "REG_BADOPT", "invalid embedded option" }, -{ REG_ETOOBIG, "REG_ETOOBIG", "nfa has too many states" }, +{ REG_ETOOBIG, "REG_ETOOBIG", "regular expression is too complex" }, { REG_ECOLORS, "REG_ECOLORS", "too many colors" }, diff --git a/generic/regex.h b/generic/regex.h index 9466fbb..53450e5 100644 --- a/generic/regex.h +++ b/generic/regex.h @@ -280,7 +280,7 @@ typedef struct { #define REG_INVARG 16 /* invalid argument to regex function */ #define REG_MIXED 17 /* character widths of regex and string differ */ #define REG_BADOPT 18 /* invalid embedded option */ -#define REG_ETOOBIG 19 /* nfa has too many states */ +#define REG_ETOOBIG 19 /* regular expression is too complex */ #define REG_ECOLORS 20 /* too many colors */ /* two specials for debugging and testing */ #define REG_ATOI 101 /* convert error-code name to number */ diff --git a/generic/regguts.h b/generic/regguts.h index de760b0..1ac2465 100644 --- a/generic/regguts.h +++ b/generic/regguts.h @@ -254,9 +254,11 @@ struct arc { struct state *from; /* where it's from (and contained within) */ struct state *to; /* where it's to */ struct arc *outchain; /* link in *from's outs chain or free chain */ -#define freechain outchain - struct arc *inchain; /* link in *to's ins chain */ - struct arc *colorchain; /* link in color's arc chain */ + struct arc *outchainRev; /* back-link in *from's outs chain */ +#define freechain outchain /* we do not maintain "freechainRev" */ + struct arc *inchain; /* *to's ins chain */ + struct arc *inchainRev; /* back-link in *to's ins chain */ + struct arc *colorchain; /* color's arc chain */ struct arc *colorchainRev; /* back-link in color's arc chain */ }; @@ -294,9 +296,6 @@ struct nfa { struct colormap *cm; /* the color map */ color bos[2]; /* colors, if any, assigned to BOS and BOL */ color eos[2]; /* colors, if any, assigned to EOS and EOL */ - size_t size; /* Current NFA size; differs from nstates as - * it also counts the number of states created - * by children of this state. */ struct vars *v; /* simplifies compile error reporting */ struct nfa *parent; /* parent NFA, if any */ }; @@ -340,11 +339,16 @@ struct cnfa { #define NULLCNFA(cnfa) ((cnfa).nstates == 0) /* - * Used to limit the maximum NFA size to something sane. [Bug 1810264] + * This symbol limits the transient heap space used by the regex compiler, + * and thereby also the maximum complexity of NFAs that we'll deal with. + * Currently we only count NFA states and arcs against this; the other + * transient data is generally not large enough to notice compared to those. + * Note that we do not charge anything for the final output data structures + * (the compacted NFA and the colormap). */ - -#ifndef REG_MAX_STATES -# define REG_MAX_STATES 100000 +#ifndef REG_MAX_COMPILE_SPACE +#define REG_MAX_COMPILE_SPACE \ + (100000 * sizeof(struct state) + 100000 * sizeof(struct arcbatch)) #endif /* diff --git a/tests/reg.test b/tests/reg.test index 647bba8..d040632 100644 --- a/tests/reg.test +++ b/tests/reg.test @@ -1088,7 +1088,8 @@ test reg-33.13 {Bug 1810264 - infinite loop} { test reg-33.14 {Bug 1810264 - super-expensive expression} nonPortable { regexp {(x{200}){200}$y} {x} } 0 -test reg-33.15 {Bug 3603557 - an "in the wild" RE} { + +test reg-33.15.1 {Bug 3603557 - an "in the wild" RE} { lindex [regexp -expanded -about { ^TETRA_MODE_CMD # Message Type ([[:blank:]]+) # Pad @@ -1163,10 +1164,62 @@ test reg-33.15 {Bug 3603557 - an "in the wild" RE} { (.*) # ConditionalFields }] 0 } 68 -test reg-33.16 {Bug [8d2c0da36d]- another "in the wild" RE} { +test reg-33.16.1 {Bug [8d2c0da36d]- another "in the wild" RE} { lindex [regexp -about "^MRK:client1: =1339 14HKelly Talisman 10011000 (\[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]*) \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 8 0 8 0 0 0 77 77 1 1 2 0 11 { 1 3 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 13HC6 My Creator 2 3 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 31HC7 Slightly offensive name, huh 3 8 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 23HE-mail:kelly@hotbox.com 4 9 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 17Hcompface must die 5 10 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 0 3HAir 6 12 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 14HPGP public key 7 13 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 16Hkelly@hotbox.com 8 30 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 0 12H2 text/plain 9 30 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 0 13H2 x-kom/basic 10 33 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 1H0 11 14 8 \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* \[0-9\]* 00000000 1 1H3 }\r?"] 0 } 1 - + +test reg-33.15 {constraint fixes} { + regexp {(^)+^} x +} 1 +test reg-33.16 {constraint fixes} { + regexp {($^)+} x +} 0 +test reg-33.17 {constraint fixes} { + regexp {(^$)*} x +} 1 +test reg-33.18 {constraint fixes} { + regexp {(^(?!aa))+} {aa bb cc} +} 0 +test reg-33.19 {constraint fixes} { + regexp {(^(?!aa)(?!bb)(?!cc))+} {aa x} +} 0 +test reg-33.20 {constraint fixes} { + regexp {(^(?!aa)(?!bb)(?!cc))+} {bb x} +} 0 +test reg-33.21 {constraint fixes} { + regexp {(^(?!aa)(?!bb)(?!cc))+} {cc x} +} 0 +test reg-33.22 {constraint fixes} { + regexp {(^(?!aa)(?!bb)(?!cc))+} {dd x} +} 1 + +test reg-33.23 {} { + regexp {abcd(\m)+xyz} x +} 0 +test reg-33.24 {} { + regexp {abcd(\m)+xyz} a +} 0 +test reg-33.25 {} { + regexp {^abcd*(((((^(a c(e?d)a+|)+|)+|)+|)+|a)+|)} x +} 0 +test reg-33.26 {} { + regexp {a^(^)bcd*xy(((((($a+|)+|)+|)+$|)+|)+|)^$} x +} 0 +test reg-33.27 {} { + regexp {xyz(\Y\Y)+} x +} 0 +test reg-33.28 {} { + regexp {x|(?:\M)+} x +} 1 +test reg-33.29 {} { + # This is near the limits of the RE engine + regexp [string repeat x*y*z* 480] x +} 1 + +test reg-33.30 {Bug 1080042} { + regexp {(\Y)+} foo +} 1 + # cleanup ::tcltest::cleanupTests return diff --git a/tests/regexp.test b/tests/regexp.test index a83c99b..9fff262 100644 --- a/tests/regexp.test +++ b/tests/regexp.test @@ -864,7 +864,7 @@ test regexp-22.4 {Bug 3606139} -setup { [a 668]([a 55])[a 668]([a 55])[a 668]([a 55])[a 511]] {}] a } -cleanup { rename a {} -} -returnCodes 1 -result {couldn't compile regular expression pattern: nfa has too many states} +} -returnCodes 1 -match glob -result {couldn't compile regular expression pattern: *} test regexp-22.5 {Bug 3610026} -setup { set e {} set cp 99 |