Avoid generating extra subre tree nodes for capturing parentheses.

tglsfdc · tglsfdc · commit ea1268f6301c · 2021-02-20T19:26:41.000-05:00
Previously, each pair of capturing parentheses gave rise to a separate subre tree node, whose only function was to identify that we ought to capture the match details for this particular sub-expression. In most cases we don't really need that, since we can perfectly well put a "capture this" annotation on the child node that does the real matching work. As with the two preceding commits, the main value of this is to avoid generating and optimizing an NFA for a tree node that's not really pulling its weight. The chosen data representation only allows one capture annotation per subre node. In the legal-per-spec, but seemingly not very useful, case where there are multiple capturing parens around the exact same bit of the regex (i.e. "((xyz))"), wrap the child node in N-1 capture nodes that act the same as before. We could work harder at that but I'll refrain, pending some evidence that such cases are worth troubling over. In passing, improve the comments in regex.h to say what all the different re_info bits mean. Some of them were pretty obvious but others not so much, so reverse-engineer some documentation. This is part of a patch series that in total reduces the regex engine's runtime by about a factor of four on a large corpus of real-world regexes. Patch by me, reviewed by Joel Jacobson Discussion: https://postgr.es/m/1340281.1613018383@sss.pgh.pa.us
diff --git a/src/backend/regex/README b/src/backend/regex/README
@@ -130,8 +130,8 @@ possibilities.
 
 As an example, consider the regex "(a[bc]+)\1".  The compiled
 representation will have a top-level concatenation subre node.  Its first
-child is a capture node, and the child of that is a plain DFA node for
-"a[bc]+".  The concatenation's second child is a backref node for \1.
+child is a plain DFA node for "a[bc]+" (which is marked as being a capture
+node).  The concatenation's second child is a backref node for \1.
 The DFA associated with the concatenation node will be "a[bc]+a[bc]+",
 where the backref has been replaced by a copy of the DFA for its referent
 expression.  When executed, the concatenation node will have to search for
@@ -147,6 +147,17 @@ run much faster than a pure NFA engine could do.  It is this behavior that
 justifies using the phrase "hybrid DFA/NFA engine" to describe Spencer's
 library.
 
+It's perhaps worth noting that separate capture subre nodes are a rarity:
+normally, we just mark a subre as capturing and that's it.  However, it's
+legal to write a regex like "((x))" in which the same substring has to be
+captured by multiple sets of parentheses.  Since a subre has room for only
+one "capno" field, a single subre can't handle that.  We handle such cases
+by wrapping the base subre (which captures the innermost parens) in a
+no-op capture node, or even more than one for "(((x)))" etc.  This is a
+little bit inefficient because we end up with multiple identical NFAs,
+but since the case is pointless and infrequent, it's not worth working
+harder.
+
 
 Colors and colormapping
 -----------------------
diff --git a/src/backend/regex/regcomp.c b/src/backend/regex/regcomp.c
@@ -452,7 +452,7 @@ pg_regcomp(regex_t *re,
 #endif
 
 		/* Prepend .* to pattern if it's a lookbehind LACON */
-		nfanode(v, lasub, !LATYPE_IS_AHEAD(lasub->subno), debug);
+		nfanode(v, lasub, !LATYPE_IS_AHEAD(lasub->latype), debug);
 	}
 	CNOERR();
 	if (v->tree->flags & SHORTER)
@@ -944,7 +944,13 @@ parseqatom(struct vars *v,
 			else
 				atomtype = PLAIN;	/* something that's not '(' */
 			NEXT();
-			/* need new endpoints because tree will contain pointers */
+
+			/*
+			 * Make separate endpoints to ensure we keep this sub-NFA cleanly
+			 * separate from what surrounds it.  We need to be sure that when
+			 * we duplicate the sub-NFA for a backref, we get the right states
+			 * and no others.
+			 */
 			s = newstate(v->nfa);
 			s2 = newstate(v->nfa);
 			NOERR();
@@ -959,11 +965,21 @@ parseqatom(struct vars *v,
 			{
 				assert(v->subs[subno] == NULL);
 				v->subs[subno] = atom;
-				t = subre(v, '(', atom->flags | CAP, lp, rp);
-				NOERR();
-				t->subno = subno;
-				t->child = atom;
-				atom = t;
+				if (atom->capno == 0)
+				{
+					/* normal case: just mark the atom as capturing */
+					atom->flags |= CAP;
+					atom->capno = subno;
+				}
+				else
+				{
+					/* generate no-op wrapper node to handle "((x))" */
+					t = subre(v, '(', atom->flags | CAP, lp, rp);
+					NOERR();
+					t->capno = subno;
+					t->child = atom;
+					atom = t;
+				}
 			}
 			/* postpone everything else pending possible {0} */
 			break;
@@ -976,7 +992,7 @@ parseqatom(struct vars *v,
 			atom = subre(v, 'b', BACKR, lp, rp);
 			NOERR();
 			subno = v->nextvalue;
-			atom->subno = subno;
+			atom->backno = subno;
 			EMPTYARC(lp, rp);	/* temporarily, so there's something */
 			NEXT();
 			break;
@@ -1276,8 +1292,10 @@ parseqatom(struct vars *v,
 			freesubre(v, top->child);
 			top->op = t->op;
 			top->flags = t->flags;
+			top->latype = t->latype;
 			top->id = t->id;
-			top->subno = t->subno;
+			top->capno = t->capno;
+			top->backno = t->backno;
 			top->min = t->min;
 			top->max = t->max;
 			top->child = t->child;
@@ -1790,8 +1808,10 @@ subre(struct vars *v,
 
 	ret->op = op;
 	ret->flags = flags;
+	ret->latype = (char) -1;
 	ret->id = 0;				/* will be assigned later */
-	ret->subno = 0;
+	ret->capno = 0;
+	ret->backno = 0;
 	ret->min = ret->max = 1;
 	ret->child = NULL;
 	ret->sibling = NULL;
@@ -1893,7 +1913,7 @@ numst(struct subre *t,
 	assert(t != NULL);
 
 	i = start;
-	t->id = (short) i++;
+	t->id = i++;
 	for (t2 = t->child; t2 != NULL; t2 = t2->sibling)
 		i = numst(t2, i);
 	return i;
@@ -2040,7 +2060,7 @@ newlacon(struct vars *v,
 	sub = &v->lacons[n];
 	sub->begin = begin;
 	sub->end = end;
-	sub->subno = latype;
+	sub->latype = latype;
 	ZAPCNFA(sub->cnfa);
 	return n;
 }
@@ -2163,7 +2183,7 @@ dump(regex_t *re,
 		struct subre *lasub = &g->lacons[i];
 		const char *latype;
 
-		switch (lasub->subno)
+		switch (lasub->latype)
 		{
 			case LATYPE_AHEAD_POS:
 				latype = "positive lookahead";
@@ -2227,8 +2247,12 @@ stdump(struct subre *t,
 		fprintf(f, " hasbackref");
 	if (!(t->flags & INUSE))
 		fprintf(f, " UNUSED");
-	if (t->subno != 0)
-		fprintf(f, " (#%d)", t->subno);
+	if (t->latype != (char) -1)
+		fprintf(f, " latype(%d)", t->latype);
+	if (t->capno != 0)
+		fprintf(f, " capture(%d)", t->capno);
+	if (t->backno != 0)
+		fprintf(f, " backref(%d)", t->backno);
 	if (t->min != 1 || t->max != 1)
 	{
 		fprintf(f, " {%d,", t->min);
diff --git a/src/backend/regex/rege_dfa.c b/src/backend/regex/rege_dfa.c
@@ -825,12 +825,12 @@ lacon(struct vars *v,
 	d = getladfa(v, n);
 	if (d == NULL)
 		return 0;
-	if (LATYPE_IS_AHEAD(sub->subno))
+	if (LATYPE_IS_AHEAD(sub->latype))
 	{
 		/* used to use longest() here, but shortest() could be much cheaper */
 		end = shortest(v, d, cp, cp, v->stop,
 					   (chr **) NULL, (int *) NULL);
-		satisfied = LATYPE_IS_POS(sub->subno) ? (end != NULL) : (end == NULL);
+		satisfied = LATYPE_IS_POS(sub->latype) ? (end != NULL) : (end == NULL);
 	}
 	else
 	{
@@ -843,7 +843,7 @@ lacon(struct vars *v,
 		 * nominal match.
 		 */
 		satisfied = matchuntil(v, d, cp, &v->lblastcss[n], &v->lblastcp[n]);
-		if (!LATYPE_IS_POS(sub->subno))
+		if (!LATYPE_IS_POS(sub->latype))
 			satisfied = !satisfied;
 	}
 	FDEBUG(("=== lacon %d satisfied %d\n", n, satisfied));
diff --git a/src/backend/regex/regexec.c b/src/backend/regex/regexec.c
@@ -640,13 +640,11 @@ static void
 zaptreesubs(struct vars *v,
 			struct subre *t)
 {
+	int			n = t->capno;
 	struct subre *t2;
 
-	if (t->op == '(')
+	if (n > 0)
 	{
-		int			n = t->subno;
-
-		assert(n > 0);
 		if ((size_t) n < v->nmatch)
 		{
 			v->pmatch[n].rm_so = -1;
@@ -667,7 +665,7 @@ subset(struct vars *v,
 	   chr *begin,
 	   chr *end)
 {
-	int			n = sub->subno;
+	int			n = sub->capno;
 
 	assert(n > 0);
 	if ((size_t) n >= v->nmatch)
@@ -739,12 +737,10 @@ cdissect(struct vars *v,
 			else
 				er = citerdissect(v, t, begin, end);
 			break;
-		case '(':				/* capturing */
+		case '(':				/* no-op capture node */
 			assert(t->child != NULL);
-			assert(t->subno > 0);
+			assert(t->capno > 0);
 			er = cdissect(v, t->child, begin, end);
-			if (er == REG_OKAY)
-				subset(v, t, begin, end);
 			break;
 		default:
 			er = REG_ASSERT;
@@ -758,6 +754,12 @@ cdissect(struct vars *v,
 	 */
 	assert(er != REG_NOMATCH || (t->flags & BACKR));
 
+	/*
+	 * If this node is marked as capturing, save successful match's location.
+	 */
+	if (t->capno > 0 && er == REG_OKAY)
+		subset(v, t, begin, end);
+
 	return er;
 }
 
@@ -932,7 +934,7 @@ cbrdissect(struct vars *v,
 		   chr *begin,			/* beginning of relevant substring */
 		   chr *end)			/* end of same */
 {
-	int			n = t->subno;
+	int			n = t->backno;
 	size_t		numreps;
 	size_t		tlen;
 	size_t		brlen;
diff --git a/src/include/regex/regex.h b/src/include/regex/regex.h
@@ -56,21 +56,23 @@ typedef struct
 {
 	int			re_magic;		/* magic number */
 	size_t		re_nsub;		/* number of subexpressions */
-	long		re_info;		/* information about RE */
-#define  REG_UBACKREF		 000001
-#define  REG_ULOOKAROUND	 000002
-#define  REG_UBOUNDS	 000004
-#define  REG_UBRACES	 000010
-#define  REG_UBSALNUM		 000020
-#define  REG_UPBOTCH	 000040
-#define  REG_UBBS		 000100
-#define  REG_UNONPOSIX		 000200
-#define  REG_UUNSPEC	 000400
-#define  REG_UUNPORT	 001000
-#define  REG_ULOCALE	 002000
-#define  REG_UEMPTYMATCH	 004000
-#define  REG_UIMPOSSIBLE	 010000
-#define  REG_USHORTEST		 020000
+	long		re_info;		/* bitmask of the following flags: */
+#define  REG_UBACKREF		000001	/* has back-reference (\n) */
+#define  REG_ULOOKAROUND	000002	/* has lookahead/lookbehind constraint */
+#define  REG_UBOUNDS		000004	/* has bounded quantifier ({m,n}) */
+#define  REG_UBRACES		000010	/* has { that doesn't begin a quantifier */
+#define  REG_UBSALNUM		000020	/* has backslash-alphanumeric in non-ARE */
+#define  REG_UPBOTCH		000040	/* has unmatched right paren in ERE (legal
+									 * per spec, but that was a mistake) */
+#define  REG_UBBS			000100	/* has backslash within bracket expr */
+#define  REG_UNONPOSIX		000200	/* has any construct that extends POSIX */
+#define  REG_UUNSPEC		000400	/* has any case disallowed by POSIX, e.g.
+									 * an empty branch */
+#define  REG_UUNPORT		001000	/* has numeric character code dependency */
+#define  REG_ULOCALE		002000	/* has locale dependency */
+#define  REG_UEMPTYMATCH	004000	/* can match a zero-length string */
+#define  REG_UIMPOSSIBLE	010000	/* provably cannot match anything */
+#define  REG_USHORTEST		020000	/* has non-greedy quantifier */
 	int			re_csize;		/* sizeof(character) */
 	char	   *re_endp;		/* backward compatibility kludge */
 	Oid			re_collation;	/* Collation that defines LC_CTYPE behavior */
diff --git a/src/include/regex/regguts.h b/src/include/regex/regguts.h
@@ -422,7 +422,7 @@ struct cnfa
  * "op" is one of:
  *		'='  plain regex without interesting substructure (implemented as DFA)
  *		'b'  back-reference (has no substructure either)
- *		'('  capture node: captures the match of its single child
+ *		'('  no-op capture node: captures the match of its single child
  *		'.'  concatenation: matches a match for first child, then second child
  *		'|'  alternation: matches a match for any of its children
  *		'*'  iteration: matches some number of matches of its single child
@@ -446,8 +446,8 @@ struct subre
 #define  LONGER  01				/* prefers longer match */
 #define  SHORTER 02				/* prefers shorter match */
 #define  MIXED	 04				/* mixed preference below */
-#define  CAP	 010			/* capturing parens below */
-#define  BACKR	 020			/* back reference below */
+#define  CAP	 010			/* capturing parens here or below */
+#define  BACKR	 020			/* back reference here or below */
 #define  INUSE	 0100			/* in use in final tree */
 #define  NOPROP  03				/* bits which may not propagate up */
 #define  LMIX(f) ((f)<<2)		/* LONGER -> MIXED */
@@ -457,9 +457,10 @@ struct subre
 #define  PREF(f) ((f)&NOPROP)
 #define  PREF2(f1, f2)	 ((PREF(f1) != 0) ? PREF(f1) : PREF(f2))
 #define  COMBINE(f1, f2) (UP((f1)|(f2)) | PREF2(f1, f2))
-	short		id;				/* ID of subre (1..ntree-1) */
-	int			subno;			/* subexpression number for 'b' and '(', or
-								 * LATYPE code for lookaround constraint */
+	char		latype;			/* LATYPE code, if lookaround constraint */
+	int			id;				/* ID of subre (1..ntree-1) */
+	int			capno;			/* if capture node, subno to capture into */
+	int			backno;			/* if backref node, subno it refers to */
 	short		min;			/* min repetitions for iteration or backref */
 	short		max;			/* max repetitions for iteration or backref */
 	struct subre *child;		/* first child, if any (also freelist chain) */

Original file line number	Diff line number	Diff line change
`@@ -825,12 +825,12 @@ lacon(struct vars *v,`
`825`	`825`	`d = getladfa(v, n);`
`826`	`826`	`if (d == NULL)`
`827`	`827`	`return 0;`
`828`		`- if (LATYPE_IS_AHEAD(sub->subno))`
	`828`	`+ if (LATYPE_IS_AHEAD(sub->latype))`
`829`	`829`	`{`
`830`	`830`	`/* used to use longest() here, but shortest() could be much cheaper */`
`831`	`831`	`end = shortest(v, d, cp, cp, v->stop,`
`832`	`832`	`(chr *) NULL, (int ) NULL);`
`833`		`- satisfied = LATYPE_IS_POS(sub->subno) ? (end != NULL) : (end == NULL);`
	`833`	`+ satisfied = LATYPE_IS_POS(sub->latype) ? (end != NULL) : (end == NULL);`
`834`	`834`	`}`
`835`	`835`	`else`
`836`	`836`	`{`
`@@ -843,7 +843,7 @@ lacon(struct vars *v,`
`843`	`843`	`* nominal match.`
`844`	`844`	`*/`
`845`	`845`	`satisfied = matchuntil(v, d, cp, &v->lblastcss[n], &v->lblastcp[n]);`
`846`		`- if (!LATYPE_IS_POS(sub->subno))`
	`846`	`+ if (!LATYPE_IS_POS(sub->latype))`
`847`	`847`	`satisfied = !satisfied;`
`848`	`848`	`}`
`849`	`849`	`FDEBUG(("=== lacon %d satisfied %d\n", n, satisfied));`