cpython.git - https://github.com/python/cpython.git

	Commit message (Expand)	Author	Age	Files	Lines
*	excise the remains of STOP_CODE, which hasn't done anything useful for years	Benjamin Peterson	2011-07-18	1	-3/+0

*	Recorded merge of revisions 81029 via svnmerge from	Antoine Pitrou	2010-05-09	1	-1/+1

*	fix bootstrapping on machines with only 2.x installed	Benjamin Peterson	2010-03-11	1	-1/+1

*	convert shebang lines: python -> python3	Benjamin Peterson	2010-03-11	1	-1/+1

*	Make Python/makeopcodetargets runnable with Python 2.3. With any luck, this	Mark Dickinson	2009-01-31	1	-1/+4

*	Add executable property to Python/makeopcodetargets.py	Antoine Pitrou	2009-01-27	1	-0/+0

*	Issue #4753: By enabling a configure option named '--with-computed-gotos'	Antoine Pitrou	2009-01-25	1	-0/+43

ore 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; + } + } + + if (NISERR()) { + return; + } + + /* + * 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.) + * + * 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.) + */ + for (s = nfa->states; s != NULL; s = s->next) { + s->tmp = NULL; + } + + /* + * Recursively build clone state(s) as needed. + */ + sclone = newstate(nfa); + if (sclone == NULL) { + assert(NISERR()); + return; + } + + clonesuccessorstates(nfa, stail, sclone, shead, refarc, + NULL, NULL, nfa->nstates); + + if (NISERR()) { + return; + } + + /* + * 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 (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; + } + + /* 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); + } +} /* - cleanup - clean up NFA after optimizations @@ -1618,7 +2165,7 @@ analyze( } /* - - compact - compact an NFA + - compact - construct the compact representation of an NFA ^ static VOID compact(struct nfa *, struct cnfa *); */ static void @@ -1687,7 +2234,7 @@ compact( cnfa->flags |= HASLACONS; break; default: - assert(NOTREACHED); + NERR(REG_ASSERT); break; } } diff --git a/generic/regcomp.c b/generic/regcomp.c index b1c02d8..27bb736 100644 --- a/generic/regcomp.c +++ b/generic/regcomp.c @@ -149,6 +149,13 @@ 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 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 *); diff --git a/tests/reg.test b/tests/reg.test index 0ebfa11..b259ce5 100644 --- a/tests/reg.test +++ b/tests/reg.test @@ -1068,6 +1068,31 @@ test reg-33.13 {Bug 1810264 - infinite loop} { test reg-33.14 {Bug 1810264 - super-expensive expression} nonPortable { regexp {(x{200}){200}$y} {x} } 0 + +test reg-33.15 {constraint fixes} { + regexp {(^)+^} x +} 1 +test reg-33.16 {constraint fixes} { + regexp {($^)+} 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 # cleanup ::tcltest::cleanupTests -- cgit v0.12 From 8ed7672ab71b54afa94e164e73fdc274b0b39771 Mon Sep 17 00:00:00 2001 From: dgp Date: Mon, 19 Oct 2015 19:32:04 +0000 Subject: Adaptation of re-oNsquared.patch from Tom Lane @ postgres. --- generic/regc_nfa.c | 805 ++++++++++++++++++++++++++++++++++++++++++++++------- generic/regcomp.c | 10 +- generic/regguts.h | 4 +- 3 files changed, 720 insertions(+), 99 deletions(-) diff --git a/generic/regc_nfa.c b/generic/regc_nfa.c index 20eb3ba..0f572b8 100644 --- a/generic/regc_nfa.c +++ b/generic/regc_nfa.c @@ -35,6 +35,8 @@ #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 @@ -322,6 +324,10 @@ 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, @@ -334,16 +340,42 @@ 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; @@ -356,15 +388,21 @@ 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++; @@ -437,7 +475,7 @@ freearc( { struct state *from = victim->from; struct state *to = victim->to; - struct arc *a; + struct arc *predecessor; assert(victim->type != 0); @@ -454,16 +492,17 @@ 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--; @@ -472,31 +511,78 @@ 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; } + +/* + * 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 void +changearctarget(struct arc * a, struct state * newto) +{ + struct state *oldto = a->to; + struct arc *predecessor; + + 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; + } + if (a->inchain != NULL) { + assert(a->inchain->inchainRev == a); + a->inchain->inchainRev = predecessor; + } + oldto->nins--; + + 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; + } + newto->ins = a; + newto->nins++; +} /* - hasnonemptyout - Does state have a non-EMPTY out arc? @@ -589,13 +675,181 @@ 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 @@ -604,14 +858,79 @@ 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); } @@ -628,14 +947,178 @@ copyins( struct state *newState, int all) { - struct arc *a; - assert(oldState != newState); - for (a=oldState->ins ; a!=NULL ; a=a->inchain) { - if (all || a->type != EMPTY) { - cparc(nfa, a, a->from, newState); + 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) { + if (all || a->type != EMPTY) { + 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; + + if (!all && a->type == EMPTY) { + oa = oa->inchain; + continue; + } + + 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; + + if (!all && a->type == EMPTY) { + oa = oa->inchain; + continue; + } + + 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++; } } @@ -649,14 +1132,78 @@ 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); } /* @@ -671,13 +1218,80 @@ copyouts( struct state *newState, int all) { - struct arc *a; - assert(oldState != newState); - for (a=oldState->outs ; a!=NULL ; a=a->outchain) { - if (all || a->type != EMPTY) { - cparc(nfa, a, newState, a->to); + 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) { + if (all || a->type != EMPTY) { + 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; + + if (!all && a->type == EMPTY) { + oa = oa->outchain; + continue; + } + + 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; + + if (!all && a->type == EMPTY){ + oa = oa->outchain; + continue; + } + + oa = oa->outchain; + createarc(nfa, a->type, a->co, newState, a->to); } } } @@ -2238,7 +2852,7 @@ compact( break; } } - carcsort(first, ca-1); + carcsort(first, ca - first); ca->co = COLORLESS; ca->to = 0; ca++; @@ -2258,33 +2872,39 @@ 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; } /* @@ -2382,37 +3002,28 @@ 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 27bb736..b01311b 100644 --- a/generic/regcomp.c +++ b/generic/regcomp.c @@ -119,15 +119,22 @@ 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 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 cloneouts(struct nfa *, struct state *, struct state *, struct state *, int); @@ -161,7 +168,8 @@ static void markreachable(struct nfa *