unicode: introduce UTF-8 character database

The decomposition and casefolding of UTF-8 characters are described in a
prefix tree in utf8data.h, which is a generate from the Unicode
Character Database (UCD), published by the Unicode Consortium, and
should not be edited by hand.  The structures in utf8data.h are meant to
be used for lookup operations by the unicode subsystem, when decoding a
utf-8 string.

mkutf8data.c is the source for a program that generates utf8data.h. It
was written by Olaf Weber from SGI and originally proposed to be merged
into Linux in 2014.  The original proposal performed the compatibility
decomposition, NFKD, but the current version was modified by me to do
canonical decomposition, NFD, as suggested by the community.  The
changes from the original submission are:

  * Rebase to mainline.
  * Fix out-of-tree-build.
  * Update makefile to build 11.0.0 ucd files.
  * drop references to xfs.
  * Convert NFKD to NFD.
  * Merge back robustness fixes from original patch. Requested by
    Dave Chinner.

The original submission is archived at:

<https://linux-xfs.oss.sgi.narkive.com/Xx10wjVY/rfc-unicode-utf-8-support-for-xfs>

The utf8data.h file can be regenerated using the instructions in
fs/unicode/README.utf8data.

- Notes on the update from 8.0.0 to 11.0:

The structure of the ucd files and special cases have not experienced
any changes between versions 8.0.0 and 11.0.0.  8.0.0 saw the addition
of Cherokee LC characters, which is an interesting case for
case-folding.  The update is accompanied by new tests on the test_ucd
module to catch specific cases.  No changes to mkutf8data script were
required for the updates.

Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.co.uk>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

2 files changed, 3191 insertions, 0 deletions

diff --git a/scripts/Makefile b/scripts/Makefile
index 9d442ee050bd..b87e3e0ade4d 100644
--- a/scripts/Makefile
+++ b/scripts/Makefile
@@ -20,6 +20,7 @@ hostprogs-$(CONFIG_ASN1)	 += asn1_compiler
 hostprogs-$(CONFIG_MODULE_SIG)	 += sign-file
 hostprogs-$(CONFIG_SYSTEM_TRUSTED_KEYRING) += extract-cert
 hostprogs-$(CONFIG_SYSTEM_EXTRA_CERTIFICATE) += insert-sys-cert
+hostprogs-$(CONFIG_UNICODE) += mkutf8data
 
 HOSTCFLAGS_sortextable.o = -I$(srctree)/tools/include
 HOSTCFLAGS_asn1_compiler.o = -I$(srctree)/include
diff --git a/scripts/mkutf8data.c b/scripts/mkutf8data.c
new file mode 100644
index 000000000000..bf593695350e
--- /dev/null
+++ b/scripts/mkutf8data.c
@@ -0,0 +1,3190 @@
+/*
+ * Copyright (c) 2014 SGI.
+ * All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it would be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write the Free Software Foundation,
+ * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
+ */
+
+/* Generator for a compact trie for unicode normalization */
+
+#include <sys/types.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <assert.h>
+#include <string.h>
+#include <unistd.h>
+#include <errno.h>
+
+/* Default names of the in- and output files. */
+
+#define AGE_NAME	"DerivedAge.txt"
+#define CCC_NAME	"DerivedCombiningClass.txt"
+#define PROP_NAME	"DerivedCoreProperties.txt"
+#define DATA_NAME	"UnicodeData.txt"
+#define FOLD_NAME	"CaseFolding.txt"
+#define NORM_NAME	"NormalizationCorrections.txt"
+#define TEST_NAME	"NormalizationTest.txt"
+#define UTF8_NAME	"utf8data.h"
+
+const char	*age_name  = AGE_NAME;
+const char	*ccc_name  = CCC_NAME;
+const char	*prop_name = PROP_NAME;
+const char	*data_name = DATA_NAME;
+const char	*fold_name = FOLD_NAME;
+const char	*norm_name = NORM_NAME;
+const char	*test_name = TEST_NAME;
+const char	*utf8_name = UTF8_NAME;
+
+int verbose = 0;
+
+/* An arbitrary line size limit on input lines. */
+
+#define LINESIZE	1024
+char line[LINESIZE];
+char buf0[LINESIZE];
+char buf1[LINESIZE];
+char buf2[LINESIZE];
+char buf3[LINESIZE];
+
+const char *argv0;
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+
+/* ------------------------------------------------------------------ */
+
+/*
+ * Unicode version numbers consist of three parts: major, minor, and a
+ * revision.  These numbers are packed into an unsigned int to obtain
+ * a single version number.
+ *
+ * To save space in the generated trie, the unicode version is not
+ * stored directly, instead we calculate a generation number from the
+ * unicode versions seen in the DerivedAge file, and use that as an
+ * index into a table of unicode versions.
+ */
+#define UNICODE_MAJ_SHIFT		(16)
+#define UNICODE_MIN_SHIFT		(8)
+
+#define UNICODE_MAJ_MAX			((unsigned short)-1)
+#define UNICODE_MIN_MAX			((unsigned char)-1)
+#define UNICODE_REV_MAX			((unsigned char)-1)
+
+#define UNICODE_AGE(MAJ,MIN,REV)			\
+	(((unsigned int)(MAJ) << UNICODE_MAJ_SHIFT) |	\
+	 ((unsigned int)(MIN) << UNICODE_MIN_SHIFT) |	\
+	 ((unsigned int)(REV)))
+
+unsigned int *ages;
+int ages_count;
+
+unsigned int unicode_maxage;
+
+static int age_valid(unsigned int major, unsigned int minor,
+		     unsigned int revision)
+{
+	if (major > UNICODE_MAJ_MAX)
+		return 0;
+	if (minor > UNICODE_MIN_MAX)
+		return 0;
+	if (revision > UNICODE_REV_MAX)
+		return 0;
+	return 1;
+}
+
+/* ------------------------------------------------------------------ */
+
+/*
+ * utf8trie_t
+ *
+ * A compact binary tree, used to decode UTF-8 characters.
+ *
+ * Internal nodes are one byte for the node itself, and up to three
+ * bytes for an offset into the tree.  The first byte contains the
+ * following information:
+ *  NEXTBYTE  - flag        - advance to next byte if set
+ *  BITNUM    - 3 bit field - the bit number to tested
+ *  OFFLEN    - 2 bit field - number of bytes in the offset
+ * if offlen == 0 (non-branching node)
+ *  RIGHTPATH - 1 bit field - set if the following node is for the
+ *                            right-hand path (tested bit is set)
+ *  TRIENODE  - 1 bit field - set if the following node is an internal
+ *                            node, otherwise it is a leaf node
+ * if offlen != 0 (branching node)
+ *  LEFTNODE  - 1 bit field - set if the left-hand node is internal
+ *  RIGHTNODE - 1 bit field - set if the right-hand node is internal
+ *
+ * Due to the way utf8 works, there cannot be branching nodes with
+ * NEXTBYTE set, and moreover those nodes always have a righthand
+ * descendant.
+ */
+typedef unsigned char utf8trie_t;
+#define BITNUM		0x07
+#define NEXTBYTE	0x08
+#define OFFLEN		0x30
+#define OFFLEN_SHIFT	4
+#define RIGHTPATH	0x40
+#define TRIENODE	0x80
+#define RIGHTNODE	0x40
+#define LEFTNODE	0x80
+
+/*
+ * utf8leaf_t
+ *
+ * The leaves of the trie are embedded in the trie, and so the same
+ * underlying datatype, unsigned char.
+ *
+ * leaf[0]: The unicode version, stored as a generation number that is
+ *          an index into utf8agetab[].  With this we can filter code
+ *          points based on the unicode version in which they were
+ *          defined.  The CCC of a non-defined code point is 0.
+ * leaf[1]: Canonical Combining Class. During normalization, we need
+ *          to do a stable sort into ascending order of all characters
+ *          with a non-zero CCC that occur between two characters with
+ *          a CCC of 0, or at the begin or end of a string.
+ *          The unicode standard guarantees that all CCC values are
+ *          between 0 and 254 inclusive, which leaves 255 available as
+ *          a special value.
+ *          Code points with CCC 0 are known as stoppers.
+ * leaf[2]: Decomposition. If leaf[1] == 255, then leaf[2] is the
+ *          start of a NUL-terminated string that is the decomposition
+ *          of the character.
+ *          The CCC of a decomposable character is the same as the CCC
+ *          of the first character of its decomposition.
+ *          Some characters decompose as the empty string: these are
+ *          characters with the Default_Ignorable_Code_Point property.
+ *          These do affect normalization, as they all have CCC 0.
+ *
+ * The decompositions in the trie have been fully expanded.
+ *
+ * Casefolding, if applicable, is also done using decompositions.
+ */
+typedef unsigned char utf8leaf_t;
+
+#define LEAF_GEN(LEAF)	((LEAF)[0])
+#define LEAF_CCC(LEAF)	((LEAF)[1])
+#define LEAF_STR(LEAF)	((const char*)((LEAF) + 2))
+
+#define MAXGEN		(255)
+
+#define MINCCC		(0)
+#define MAXCCC		(254)
+#define STOPPER		(0)
+#define DECOMPOSE	(255)
+
+struct tree;
+static utf8leaf_t *utf8nlookup(struct tree *, const char *, size_t);
+static utf8leaf_t *utf8lookup(struct tree *, const char *);
+
+unsigned char *utf8data;
+size_t utf8data_size;
+
+utf8trie_t *nfdi;
+utf8trie_t *nfdicf;
+
+/* ------------------------------------------------------------------ */
+
+/*
+ * UTF8 valid ranges.
+ *
+ * The UTF-8 encoding spreads the bits of a 32bit word over several
+ * bytes. This table gives the ranges that can be held and how they'd
+ * be represented.
+ *
+ * 0x00000000 0x0000007F: 0xxxxxxx
+ * 0x00000000 0x000007FF: 110xxxxx 10xxxxxx
+ * 0x00000000 0x0000FFFF: 1110xxxx 10xxxxxx 10xxxxxx
+ * 0x00000000 0x001FFFFF: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+ * 0x00000000 0x03FFFFFF: 111110xx 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx
+ * 0x00000000 0x7FFFFFFF: 1111110x 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx
+ *
+ * There is an additional requirement on UTF-8, in that only the
+ * shortest representation of a 32bit value is to be used.  A decoder
+ * must not decode sequences that do not satisfy this requirement.
+ * Thus the allowed ranges have a lower bound.
+ *
+ * 0x00000000 0x0000007F: 0xxxxxxx
+ * 0x00000080 0x000007FF: 110xxxxx 10xxxxxx
+ * 0x00000800 0x0000FFFF: 1110xxxx 10xxxxxx 10xxxxxx
+ * 0x00010000 0x001FFFFF: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+ * 0x00200000 0x03FFFFFF: 111110xx 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx
+ * 0x04000000 0x7FFFFFFF: 1111110x 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx 10xxxxxx
+ *
+ * Actual unicode characters are limited to the range 0x0 - 0x10FFFF,
+ * 17 planes of 65536 values.  This limits the sequences actually seen
+ * even more, to just the following.
+ *
+ *          0 -     0x7f: 0                     0x7f
+ *       0x80 -    0x7ff: 0xc2 0x80             0xdf 0xbf
+ *      0x800 -   0xffff: 0xe0 0xa0 0x80        0xef 0xbf 0xbf
+ *    0x10000 - 0x10ffff: 0xf0 0x90 0x80 0x80   0xf4 0x8f 0xbf 0xbf
+ *
+ * Even within those ranges not all values are allowed: the surrogates
+ * 0xd800 - 0xdfff should never be seen.
+ *
+ * Note that the longest sequence seen with valid usage is 4 bytes,
+ * the same a single UTF-32 character.  This makes the UTF-8
+ * representation of Unicode strictly smaller than UTF-32.
+ *
+ * The shortest sequence requirement was introduced by:
+ *    Corrigendum #1: UTF-8 Shortest Form
+ * It can be found here:
+ *    http://www.unicode.org/versions/corrigendum1.html
+ *
+ */
+
+#define UTF8_2_BITS     0xC0
+#define UTF8_3_BITS     0xE0
+#define UTF8_4_BITS     0xF0
+#define UTF8_N_BITS     0x80
+#define UTF8_2_MASK     0xE0
+#define UTF8_3_MASK     0xF0
+#define UTF8_4_MASK     0xF8
+#define UTF8_N_MASK     0xC0
+#define UTF8_V_MASK     0x3F
+#define UTF8_V_SHIFT    6
+
+static int utf8encode(char *str, unsigned int val)
+{
+	int len;
+
+	if (val < 0x80) {
+		str[0] = val;
+		len = 1;
+	} else if (val < 0x800) {
+		str[1] = val & UTF8_V_MASK;
+		str[1] |= UTF8_N_BITS;
+		val >>= UTF8_V_SHIFT;
+		str[0] = val;
+		str[0] |= UTF8_2_BITS;
+		len = 2;
+	} else if (val < 0x10000) {
+		str[2] = val & UTF8_V_MASK;
+		str[2] |= UTF8_N_BITS;
+		val >>= UTF8_V_SHIFT;
+		str[1] = val & UTF8_V_MASK;
+		str[1] |= UTF8_N_BITS;
+		val >>= UTF8_V_SHIFT;
+		str[0] = val;
+		str[0] |= UTF8_3_BITS;
+		len = 3;
+	} else if (val < 0x110000) {
+		str[3] = val & UTF8_V_MASK;
+		str[3] |= UTF8_N_BITS;
+		val >>= UTF8_V_SHIFT;
+		str[2] = val & UTF8_V_MASK;
+		str[2] |= UTF8_N_BITS;
+		val >>= UTF8_V_SHIFT;
+		str[1] = val & UTF8_V_MASK;
+		str[1] |= UTF8_N_BITS;
+		val >>= UTF8_V_SHIFT;
+		str[0] = val;
+		str[0] |= UTF8_4_BITS;
+		len = 4;
+	} else {
+		printf("%#x: illegal val\n", val);
+		len = 0;
+	}
+	return len;
+}
+
+static unsigned int utf8decode(const char *str)
+{
+	const unsigned char *s = (const unsigned char*)str;
+	unsigned int unichar = 0;
+
+	if (*s < 0x80) {
+		unichar = *s;
+	} else if (*s < UTF8_3_BITS) {
+		unichar = *s++ & 0x1F;
+		unichar <<= UTF8_V_SHIFT;
+		unichar |= *s & 0x3F;
+	} else if (*s < UTF8_4_BITS) {
+		unichar = *s++ & 0x0F;
+		unichar <<= UTF8_V_SHIFT;
+		unichar |= *s++ & 0x3F;
+		unichar <<= UTF8_V_SHIFT;
+		unichar |= *s & 0x3F;
+	} else {
+		unichar = *s++ & 0x0F;
+		unichar <<= UTF8_V_SHIFT;
+		unichar |= *s++ & 0x3F;
+		unichar <<= UTF8_V_SHIFT;
+		unichar |= *s++ & 0x3F;
+		unichar <<= UTF8_V_SHIFT;
+		unichar |= *s & 0x3F;
+	}
+	return unichar;
+}
+
+static int utf32valid(unsigned int unichar)
+{
+	return unichar < 0x110000;
+}
+
+#define NODE 1
+#define LEAF 0
+
+struct tree {
+	void *root;
+	int childnode;
+	const char *type;
+	unsigned int maxage;
+	struct tree *next;
+	int (*leaf_equal)(void *, void *);
+	void (*leaf_print)(void *, int);
+	int (*leaf_mark)(void *);
+	int (*leaf_size)(void *);
+	int *(*leaf_index)(struct tree *, void *);
+	unsigned char *(*leaf_emit)(void *, unsigned char *);
+	int leafindex[0x110000];
+	int index;
+};
+
+struct node {
+	int index;
+	int offset;
+	int mark;
+	int size;
+	struct node *parent;
+	void *left;
+	void *right;
+	unsigned char bitnum;
+	unsigned char nextbyte;
+	unsigned char leftnode;
+	unsigned char rightnode;
+	unsigned int keybits;
+	unsigned int keymask;
+};
+
+/*
+ * Example lookup function for a tree.
+ */
+static void *lookup(struct tree *tree, const char *key)
+{
+	struct node *node;
+	void *leaf = NULL;
+
+	node = tree->root;
+	while (!leaf && node) {
+		if (node->nextbyte)
+			key++;
+		if (*key & (1 << (node->bitnum & 7))) {
+			/* Right leg */
+			if (node->rightnode == NODE) {
+				node = node->right;
+			} else if (node->rightnode == LEAF) {
+				leaf = node->right;
+			} else {
+				node = NULL;
+			}
+		} else {
+			/* Left leg */
+			if (node->leftnode == NODE) {
+				node = node->left;
+			} else if (node->leftnode == LEAF) {
+				leaf = node->left;
+			} else {
+				node = NULL;
+			}
+		}
+	}
+
+	return leaf;
+}
+
+/*
+ * A simple non-recursive tree walker: keep track of visits to the
+ * left and right branches in the leftmask and rightmask.
+ */
+static void tree_walk(struct tree *tree)
+{
+	struct node *node;
+	unsigned int leftmask;
+	unsigned int rightmask;
+	unsigned int bitmask;
+	int indent = 1;
+	int nodes, singletons, leaves;
+
+	nodes = singletons = leaves = 0;
+
+	printf("%s_%x root %p\n", tree->type, tree->maxage, tree->root);
+	if (tree->childnode == LEAF) {
+		assert(tree->root);
+		tree->leaf_print(tree->root, indent);
+		leaves = 1;
+	} else {
+		assert(tree->childnode == NODE);
+		node = tree->root;
+		leftmask = rightmask = 0;
+		while (node) {
+			printf("%*snode @ %p bitnum %d nextbyte %d"
+			       " left %p right %p mask %x bits %x\n",
+				indent, "", node,
+				node->bitnum, node->nextbyte,
+				node->left, node->right,
+				node->keymask, node->keybits);
+			nodes += 1;
+			if (!(node->left && node->right))
+				singletons += 1;
+
+			while (node) {
+				bitmask = 1 << node->bitnum;
+				if ((leftmask & bitmask) == 0) {
+					leftmask |= bitmask;
+					if (node->leftnode == LEAF) {
+						assert(node->left);
+						tree->leaf_print(node->left,
+								 indent+1);
+						leaves += 1;
+					} else if (node->left) {
+						assert(node->leftnode == NODE);
+						indent += 1;
+						node = node->left;
+						break;
+					}
+				}
+				if ((rightmask & bitmask) == 0) {
+					rightmask |= bitmask;
+					if (node->rightnode == LEAF) {
+						assert(node->right);
+						tree->leaf_print(node->right,
+								 indent+1);
+						leaves += 1;
+					} else if (node->right) {
+						assert(node->rightnode==NODE);
+						indent += 1;
+						node = node->right;
+						break;
+					}
+				}
+				leftmask &= ~bitmask;
+				rightmask &= ~bitmask;
+				node = node->parent;
+				indent -= 1;
+			}
+		}
+	}
+	printf("nodes %d leaves %d singletons %d\n",
+	       nodes, leaves, singletons);
+}
+
+/*
+ * Allocate an initialize a new internal node.
+ */
+static struct node *alloc_node(struct node *parent)
+{
+	struct node *node;
+	int bitnum;
+
+	node = malloc(sizeof(*node));
+	node->left = node->right = NULL;
+	node->parent = parent;
+	node->leftnode = NODE;
+	node->rightnode = NODE;
+	node->keybits = 0;
+	node->keymask = 0;
+	node->mark = 0;
+	node->index = 0;
+	node->offset = -1;
+	node->size = 4;
+
+	if (node->parent) {
+		bitnum = parent->bitnum;
+		if ((bitnum & 7) == 0) {
+			node->bitnum = bitnum + 7 + 8;
+			node->nextbyte = 1;
+		} else {
+			node->bitnum = bitnum - 1;
+			node->nextbyte = 0;
+		}
+	} else {
+		node->bitnum = 7;
+		node->nextbyte = 0;
+	}
+
+	return node;
+}
+
+/*
+ * Insert a new leaf into the tree, and collapse any subtrees that are
+ * fully populated and end in identical leaves. A nextbyte tagged
+ * internal node will not be removed to preserve the tree's integrity.
+ * Note that due to the structure of utf8, no nextbyte tagged node
+ * will be a candidate for removal.
+ */
+static int insert(struct tree *tree, char *key, int keylen, void *leaf)
+{
+	struct node *node;
+	struct node *parent;
+	void **cursor;
+	int keybits;
+
+	assert(keylen >= 1 && keylen <= 4);
+
+	node = NULL;
+	cursor = &tree->root;
+	keybits = 8 * keylen;
+
+	/* Insert, creating path along the way. */
+	while (keybits) {
+		if (!*cursor)
+			*cursor = alloc_node(node);
+		node = *cursor;
+		if (node->nextbyte)
+			key++;
+		if (*key & (1 << (node->bitnum & 7)))
+			cursor = &node->right;
+		else
+			cursor = &node->left;
+		keybits--;
+	}
+	*cursor = leaf;
+
+	/* Merge subtrees if possible. */
+	while (node) {
+		if (*key & (1 << (node->bitnum & 7)))
+			node->rightnode = LEAF;
+		else
+			node->leftnode = LEAF;
+		if (node->nextbyte)
+			break;
+		if (node->leftnode == NODE || node->rightnode == NODE)
+			break;
+		assert(node->left);
+		assert(node->right);
+		/* Compare */
+		if (! tree->leaf_equal(node->left, node->right))
+			break;
+		/* Keep left, drop right leaf. */
+		leaf = node->left;
+		/* Check in parent */
+		parent = node->parent;
+		if (!parent) {
+			/* root of tree! */
+			tree->root = leaf;
+			tree->childnode = LEAF;
+		} else if (parent->left == node) {
+			parent->left = leaf;
+			parent->leftnode = LEAF;
+			if (parent->right) {
+				parent->keymask = 0;
+				parent->keybits = 0;
+			} else {
+				parent->keymask |= (1 << node->bitnum);
+			}
+		} else if (parent->right == node) {
+			parent->right = leaf;
+			parent->rightnode = LEAF;
+			if (parent->left) {
+				parent->keymask = 0;
+				parent->keybits = 0;
+			} else {
+				parent->keymask |= (1 << node->bitnum);
+				parent->keybits |= (1 << node->bitnum);
+			}
+		} else {
+			/* internal tree error */
+			assert(0);
+		}
+		free(node);
+		node = parent;
+	}
+
+	/* Propagate keymasks up along singleton chains. */
+	while (node) {
+		parent = node->parent;
+		if (!parent)
+			break;
+		/* Nix the mask for parents with two children. */
+		if (node->keymask == 0) {
+			parent->keymask = 0;
+			parent->keybits = 0;
+		} else if (parent->left && parent->right) {
+			parent->keymask = 0;
+			parent->keybits = 0;
+		} else {
+			assert((parent->keymask & node->keymask) == 0);
+			parent->keymask |= node->keymask;
+			parent->keymask |= (1 << parent->bitnum);
+			parent->keybits |= node->keybits;
+			if (parent->right)
+				parent->keybits |= (1 << parent->bitnum);
+		}
+		node = parent;
+	}
+
+	return 0;
+}
+
+/*
+ * Prune internal nodes.
+ *
+ * Fully populated subtrees that end at the same leaf have already
+ * been collapsed.  There are still internal nodes that have for both
+ * their left and right branches a sequence of singletons that make
+ * identical choices and end in identical leaves.  The keymask and
+ * keybits collected in the nodes describe the choices made in these
+ * singleton chains.  When they are identical for the left and right
+ * branch of a node, and the two leaves comare identical, the node in
+ * question can be removed.
+ *
+ * Note that nodes with the nextbyte tag set will not be removed by
+ * this to ensure tree integrity.  Note as well that the structure of
+ * utf8 ensures that these nodes would not have been candidates for
+ * removal in any case.
+ */
+static void prune(struct tree *tree)
+{
+	struct node *node;
+	struct node *left;
+	struct node *right;
+	struct node *parent;
+	void *leftleaf;
+	void *rightleaf;
+	unsigned int leftmask;
+	unsigned int rightmask;
+	unsigned int bitmask;
+	int count;
+
+	if (verbose > 0)
+		printf("Pruning %s_%x\n", tree->type, tree->maxage);
+
+	count = 0;
+	if (tree->childnode == LEAF)
+		return;
+	if (!tree->root)
+		return;
+
+	leftmask = rightmask = 0;
+	node = tree->root;
+	while (node) {
+		if (node->nextbyte)
+			goto advance;
+		if (node->leftnode == LEAF)
+			goto advance;
+		if (node->rightnode == LEAF)
+			goto advance;
+		if (!node->left)
+			goto advance;
+		if (!node->right)
+			goto advance;
+		left = node->left;
+		right = node->right;
+		if (left->keymask == 0)
+			goto advance;
+		if (right->keymask == 0)
+			goto advance;
+		if (left->keymask != right->keymask)
+			goto advance;
+		if (left->keybits != right->keybits)
+			goto advance;
+		leftleaf = NULL;
+		while (!leftleaf) {
+			assert(left->left || left->right);
+			if (left->leftnode == LEAF)
+				leftleaf = left->left;
+			else if (left->rightnode == LEAF)
+				leftleaf = left->right;
+			else if (left->left)
+				left = left->left;
+			else if (left->right)
+				left = left->right;
+			else
+				assert(0);
+		}
+		rightleaf = NULL;
+		while (!rightleaf) {
+			assert(right->left || right->right);
+			if (right->leftnode == LEAF)
+				rightleaf = right->left;
+			else if (right->rightnode == LEAF)
+				rightleaf = right->right;
+			else if (right->left)
+				right = right->left;
+			else if (right->right)
+				right = right->right;
+			else
+				assert(0);
+		}
+		if (! tree->leaf_equal(leftleaf, rightleaf))
+			goto advance;
+		/*
+		 * This node has identical singleton-only subtrees.
+		 * Remove it.
+		 */
+		parent = node->parent;
+		left = node->left;
+		right = node->right;
+		if (parent->left == node)
+			parent->left = left;
+		else if (parent->right == node)
+			parent->right = left;
+		else
+			assert(0);
+		left->parent = parent;
+		left->keymask |= (1 << node->bitnum);
+		node->left = NULL;
+		while (node) {
+			bitmask = 1 << node->bitnum;
+			leftmask &= ~bitmask;
+			rightmask &= ~bitmask;
+			if (node->leftnode == NODE && node->left) {
+				left = node->left;
+				free(node);
+				count++;
+				node = left;
+			} else if (node->rightnode == NODE && node->right) {
+				right = node->right;
+				free(node);
+				count++;
+				node = right;
+			} else {
+				node = NULL;
+			}
+		}
+		/* Propagate keymasks up along singleton chains. */
+		node = parent;
+		/* Force re-check */
+		bitmask = 1 << node->bitnum;
+		leftmask &= ~bitmask;
+		rightmask &= ~bitmask;
+		for (;;) {
+			if (node->left && node->right)
+				break;
+			if (node->left) {
+				left = node->left;
+				node->keymask |= left->keymask;
+				node->keybits |= left->keybits;
+			}
+			if (node->right) {
+				right = node->right;
+				node->keymask |= right->keymask;
+				node->keybits |= right->keybits;
+			}
+			node->keymask |= (1 << node->bitnum);
+			node = node->parent;
+			/* Force re-check */
+			bitmask = 1 << node->bitnum;
+			leftmask &= ~bitmask;
+			rightmask &= ~bitmask;
+		}
+	advance:
+		bitmask = 1 << node->bitnum;
+		if ((leftmask & bitmask) == 0 &&
+		    node->leftnode == NODE &&
+		    node->left) {
+			leftmask |= bitmask;
+			node = node->left;
+		} else if ((rightmask & bitmask) == 0 &&
+			   node->rightnode == NODE &&
+			   node->right) {
+			rightmask |= bitmask;
+			node = node->right;
+		} else {
+			leftmask &= ~bitmask;
+			rightmask &= ~bitmask;
+			node = node->parent;
+		}
+	}
+	if (verbose > 0)
+		printf("Pruned %d nodes\n", count);
+}
+
+/*
+ * Mark the nodes in the tree that lead to leaves that must be
+ * emitted.
+ */
+static void mark_nodes(struct tree *tree)
+{
+	struct node *node;
+	struct node *n;
+	unsigned int leftmask;
+	unsigned int rightmask;
+	unsigned int bitmask;
+	int marked;
+
+	marked = 0;
+	if (verbose > 0)
+		printf("Marking %s_%x\n", tree->type, tree->maxage);
+	if (tree->childnode == LEAF)
+		goto done;
+
+	assert(tree->childnode == NODE);
+	node = tree->root;
+	leftmask = rightmask = 0;
+	while (node) {
+		bitmask = 1 << node->bitnum;
+		if ((leftmask & bitmask) == 0) {
+			leftmask |= bitmask;
+			if (node->leftnode == LEAF) {
+				assert(node->left);
+				if (tree->leaf_mark(node->left)) {
+					n = node;
+					while (n && !n->mark) {
+						marked++;
+						n->mark = 1;
+						n = n->parent;
+					}
+				}
+			} else if (node->left) {
+				assert(node->leftnode == NODE);
+				node = node->left;
+				continue;
+			}
+		}
+		if ((rightmask & bitmask) == 0) {
+			rightmask |= bitmask;
+			if (node->rightnode == LEAF) {
+				assert(node->right);
+				if (tree->leaf_mark(node->right)) {
+					n = node;
+					while (n && !n->mark) {
+						marked++;
+						n->mark = 1;
+						n = n->parent;
+					}
+				}
+			} else if (node->right) {
+				assert(node->rightnode==NODE);
+				node = node->right;
+				continue;
+			}
+		}
+		leftmask &= ~bitmask;
+		rightmask &= ~bitmask;
+		node = node->parent;
+	}
+
+	/* second pass: left siblings and singletons */
+
+	assert(tree->childnode == NODE);
+	node = tree->root;
+	leftmask = rightmask = 0;
+	while (node) {
+		bitmask = 1 << node->bitnum;
+		if ((leftmask & bitmask) == 0) {
+			leftmask |= bitmask;
+			if (node->leftnode == LEAF) {
+				assert(node->left);
+				if (tree->leaf_mark(node->left)) {
+					n = node;
+					while (n && !n->mark) {
+						marked++;
+						n->mark = 1;
+						n = n->parent;
+					}
+				}
+			} else if (node->left) {
+				assert(node->leftnode == NODE);
+				node = node->left;
+				if (!node->mark && node->parent->mark) {
+					marked++;
+					node->mark = 1;
+				}
+				continue;
+			}
+		}
+		if ((rightmask & bitmask) == 0) {
+			rightmask |= bitmask;
+			if (node->rightnode == LEAF) {
+				assert(node->right);
+				if (tree->leaf_mark(node->right)) {
+					n = node;
+					while (n && !n->mark) {
+						marked++;
+						n->mark = 1;
+						n = n->parent;
+					}
+				}
+			} else if (node->right) {
+				assert(node->rightnode==NODE);
+				node = node->right;
+				if (!node->mark && node->parent->mark &&
+				    !node->parent->left) {
+					marked++;
+					node->mark = 1;
+				}
+				continue;
+			}
+		}
+		leftmask &= ~bitmask;
+		rightmask &= ~bitmask;
+		node = node->parent;
+	}
+done:
+	if (verbose > 0)
+		printf("Marked %d nodes\n", marked);
+}
+
+/*
+ * Compute the index of each node and leaf, which is the offset in the
+ * emitted trie.  These values must be pre-computed because relative
+ * offsets between nodes are used to navigate the tree.
+ */
+static int index_nodes(struct tree *tree, int index)
+{
+	struct node *node;
+	unsigned int leftmask;
+	unsigned int rightmask;
+	unsigned int bitmask;
+	int count;
+	int indent;
+
+	/* Align to a cache line (or half a cache line?). */
+	while (index % 64)
+		index++;
+	tree->index = index;
+	indent = 1;
+	count = 0;
+
+	if (verbose > 0)
+		printf("Indexing %s_%x: %d\n", tree->type, tree->maxage, index);
+	if (tree->childnode == LEAF) {
+		index += tree->leaf_size(tree->root);
+		goto done;
+	}
+
+	assert(tree->childnode == NODE);
+	node = tree->root;
+	leftmask = rightmask = 0;
+	while (node) {
+		if (!node->mark)
+			goto skip;
+		count++;
+		if (node->index != index)
+			node->index = index;
+		index += node->size;
+skip:
+		while (node) {
+			bitmask = 1 << node->bitnum;
+			if (node->mark && (leftmask & bitmask) == 0) {
+				leftmask |= bitmask;
+				if (node->leftnode == LEAF) {
+					assert(node->left);
+					*tree->leaf_index(tree, node->left) =
+									index;
+					index += tree->leaf_size(node->left);
+					count++;
+				} else if (node->left) {
+					assert(node->leftnode == NODE);
+					indent += 1;
+					node = node->left;
+					break;
+				}
+			}
+			if (node->mark && (rightmask & bitmask) == 0) {
+				rightmask |= bitmask;
+				if (node->rightnode == LEAF) {
+					assert(node->right);
+					*tree->leaf_index(tree, node->right) = index;
+					index += tree->leaf_size(node->right);
+					count++;
+				} else if (node->right) {
+					assert(node->rightnode==NODE);
+					indent += 1;
+					node = node->right;
+					break;
+				}
+			}
+			leftmask &= ~bitmask;
+			rightmask &= ~bitmask;
+			node = node->parent;
+			indent -= 1;
+		}
+	}
+done:
+	/* Round up to a multiple of 16 */
+	while (index % 16)
+		index++;
+	if (verbose > 0)
+		printf("Final index %d\n", index);
+	return index;
+}
+
+/*
+ * Compute the size of nodes and leaves. We start by assuming that
+ * each node needs to store a three-byte offset. The indexes of the
+ * nodes are calculated based on that, and then this function is
+ * called to see if the sizes of some nodes can be reduced.  This is
+ * repeated until no more changes are seen.
+ */
+static int size_nodes(struct tree *tree)
+{
+	struct tree *next;
+	struct node *node;
+	struct node *right;
+	struct node *n;
+	unsigned int leftmask;
+	unsigned int rightmask;
+	unsigned int bitmask;
+	unsigned int pathbits;
+	unsigned int pathmask;
+	int changed;
+	int offset;
+	int size;
+	int indent;
+
+	indent = 1;
+	changed = 0;
+	size = 0;
+
+	if (verbose > 0)
+		printf("Sizing %s_%x\n", tree->type, tree->maxage);
+	if (tree->childnode == LEAF)
+		goto done;
+
+	assert(tree->childnode == NODE);
+	pathbits = 0;
+	pathmask = 0;
+	node = tree->root;
+	leftmask = rightmask = 0;
+	while (node) {
+		if (!node->mark)
+			goto skip;
+		offset = 0;
+		if (!node->left || !node->right) {
+			size = 1;
+		} else {
+			if (node->rightnode == NODE) {
+				right = node->right;
+				next = tree->next;
+				while (!right->mark) {
+					assert(next);
+					n = next->root;
+					while (n->bitnum != node->bitnum) {
+						if (pathbits & (1<<n->bitnum))
+							n = n->right;
+						else
+							n = n->left;
+					}
+					n = n->right;
+					assert(right->bitnum == n->bitnum);
+					right = n;
+					next = next->next;
+				}
+				offset = right->index - node->index;
+			} else {
+				offset = *tree->leaf_index(tree, node->right);
+				offset -= node->index;
+			}
+			assert(offset >= 0);
+			assert(offset <= 0xffffff);
+			if (offset <= 0xff) {
+				size = 2;
+			} else if (offset <= 0xffff) {
+				s