[PATCH] kallsyms data size reduction / lookup speedup

This patch is an improvement over my first kallsyms speedup patch posted about
2 weeks ago.

It changes scripts/kallsyms as to produce a different format for
kallsyms_names and extra data to speedup lookups.  The compression algorithm
is quite simple: it uses all the char codes not actually used in symbols to
build a lookup table that translates these codes into small strings.  For
instance, in my test runs the code 0xFE was being translated into "acpi_"
giving a 4 byte save on every translation.

The advantage of this algorithm is that to translate a symbol we only require
information that is stored on that symbol position, and never need to go back
on the compressed stream to get information from other symbols.

To give an idea about the benefits of this algorithm here are some benchmark
results on a P4 2.8GHz with a symbol table with 10000 entries:

kallsyms_lookup average time:
  vanilla           1346.0 us
  speedup             14.4 us
  with this patch      0.5 us

total data produced by scripts/kallsyms:
  uncompressed         169 Kb
  vanilla              134 Kb
  with this patch       91 Kb

(speedup was my latest patch, that only changed the way kallsyms_lookup worked
and not the data format)

I removed a cond_resched() from the proc/kallsyms handling code path, because
using stem compression, if the current position went backwards, the hole
stream would be uncompressed up to the current position.  It seemed that by
removing this loop it would be safe to remove the conditional reschedule
altogether.

There is just one catch with this patch: the time it takes to compile the
kernel goes up just a bit (about 0.8s on a P4 2.8GHz with defconfig).  If this
delay is not acceptable I can change the compression algorithm so that it can
use the previous table (calculating a new table is what consumes most of the
time, and not doing the actual compression) and check to see if it obtains a
similar compression ratio.  If it does, then this is a sign that the symbol
patterns haven't changed that much and this table is still good to use.  This
would not only cut the time down to half on any compilation (because of the 2
pass symbol build method), but in frequent cases where a developer is
compiling a single file and linking everything over and over again, the table
optimization process would never run.

I'm CC'ing Brent Casavant on this email, because last june he sent a patch
trying a different approach that used a 32 entry symbol cache, because there
was a problem with the time "top" took to read "proc/<pid>/wchan".  I was
hopping he would be willing to test this patch and comment on the results.
Signed-off-by: Paulo Marques <pmarques@grupopie.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

kernel/kallsyms.c

@@ -4,7 +4,12 @@
  * Rewritten and vastly simplified by Rusty Russell for in-kernel
  * module loader:
  *   Copyright 2002 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
- * Stem compression by Andi Kleen.
+ *
+ * ChangeLog:
+ *
+ * (25/Aug/2004) Paulo Marques <pmarques@grupopie.com>
+ *      Changed the compression method from stem compression to "table lookup"
+ *      compression (see scripts/kallsyms.c for a more complete description)
  */
 #include <linux/kallsyms.h>
 #include <linux/module.h>
@@ -17,7 +22,12 @@
 /* These will be re-linked against their real values during the second link stage */
 extern unsigned long kallsyms_addresses[] __attribute__((weak));
 extern unsigned long kallsyms_num_syms __attribute__((weak));
-extern char kallsyms_names[] __attribute__((weak));
+extern u8 kallsyms_names[] __attribute__((weak));
+
+extern u8 kallsyms_token_table[] __attribute__((weak));
+extern u16 kallsyms_token_index[] __attribute__((weak));
+
+extern unsigned long kallsyms_markers[] __attribute__((weak));
 
 /* Defined by the linker script. */
 extern char _stext[], _etext[], _sinittext[], _einittext[];
@@ -37,21 +47,88 @@ static inline int is_kernel_text(unsigned long addr)
 	return 0;
 }
 
+/* expand a compressed symbol data into the resulting uncompressed string,
+   given the offset to where the symbol is in the compressed stream */
+static unsigned int kallsyms_expand_symbol(unsigned int off, char *result)
+{
+	int len, skipped_first = 0;
+	u8 *tptr, *data;
+
+	/* get the compressed symbol length from the first symbol byte */
+	data = &kallsyms_names[off];
+	len = *data;
+	data++;
+
+	/* update the offset to return the offset for the next symbol on
+	 * the compressed stream */
+	off += len + 1;
+
+	/* for every byte on the compressed symbol data, copy the table
+	   entry for that byte */
+	while(len) {
+		tptr = &kallsyms_token_table[ kallsyms_token_index[*data] ];
+		data++;
+		len--;
+
+		while (*tptr) {
+			if(skipped_first) {
+				*result = *tptr;
+				result++;
+			} else
+				skipped_first = 1;
+			tptr++;
+		}
+	}
+
+	*result = '\0';
+
+	/* return to offset to the next symbol */
+	return off;
+}
+
+/* get symbol type information. This is encoded as a single char at the
+ * begining of the symbol name */
+static char kallsyms_get_symbol_type(unsigned int off)
+{
+	/* get just the first code, look it up in the token table, and return the
+	 * first char from this token */
+	return kallsyms_token_table[ kallsyms_token_index[ kallsyms_names[off+1] ] ];
+}
+
+
+/* find the offset on the compressed stream given and index in the
+ * kallsyms array */
+static unsigned int get_symbol_offset(unsigned long pos)
+{
+	u8 *name;
+	int i;
+
+	/* use the closest marker we have. We have markers every 256 positions,
+	 * so that should be close enough */
+	name = &kallsyms_names[ kallsyms_markers[pos>>8] ];
+
+	/* sequentially scan all the symbols up to the point we're searching for.
+	 * Every symbol is stored in a [<len>][<len> bytes of data] format, so we
+	 * just need to add the len to the current pointer for every symbol we
+	 * wish to skip */
+	for(i = 0; i < (pos&0xFF); i++)
+		name = name + (*name) + 1;
+
+	return name - kallsyms_names;
+}
+
 /* Lookup the address for this symbol. Returns 0 if not found. */
 unsigned long kallsyms_lookup_name(const char *name)
 {
 	char namebuf[KSYM_NAME_LEN+1];
 	unsigned long i;
-	char *knames;
+	unsigned int off;
 
-	for (i = 0, knames = kallsyms_names; i < kallsyms_num_syms; i++) {
-		unsigned prefix = *knames++;
+	for (i = 0, off = 0; i < kallsyms_num_syms; i++) {
+		off = kallsyms_expand_symbol(off, namebuf);
 
-		strlcpy(namebuf + prefix, knames, KSYM_NAME_LEN - prefix);
 		if (strcmp(namebuf, name) == 0)
 			return kallsyms_addresses[i];
-
-		knames += strlen(knames) + 1;
 	}
 	return module_kallsyms_lookup_name(name);
 }
@@ -62,7 +139,7 @@ const char *kallsyms_lookup(unsigned long addr,
 			    unsigned long *offset,
 			    char **modname, char *namebuf)
 {
-	unsigned long i, best = 0;
+	unsigned long i, low, high, mid;
 
 	/* This kernel should never had been booted. */
 	BUG_ON(!kallsyms_addresses);
@@ -71,40 +148,45 @@ const char *kallsyms_lookup(unsigned long addr,
 	namebuf[0] = 0;
 
 	if (is_kernel_text(addr) || is_kernel_inittext(addr)) {
-		unsigned long symbol_end;
-		char *name = kallsyms_names;
-
-		/* They're sorted, we could be clever here, but who cares? */
-		for (i = 0; i < kallsyms_num_syms; i++) {
-			if (kallsyms_addresses[i] > kallsyms_addresses[best] &&
-			    kallsyms_addresses[i] <= addr)
-				best = i;
+		unsigned long symbol_end=0;
+
+		/* do a binary search on the sorted kallsyms_addresses array */
+		low = 0;
+		high = kallsyms_num_syms;
+
+		while (high-low > 1) {
+			mid = (low + high) / 2;
+			if (kallsyms_addresses[mid] <= addr) low = mid;
+			else high = mid;
 		}
 
+		/* search for the first aliased symbol. Aliased symbols are
+		   symbols with the same address */
+		while (low && kallsyms_addresses[low - 1] == kallsyms_addresses[low])
+			--low;
+
 		/* Grab name */
-		for (i = 0; i <= best; i++) { 
-			unsigned prefix = *name++;
-			strncpy(namebuf + prefix, name, KSYM_NAME_LEN - prefix);
-			name += strlen(name) + 1;
+		kallsyms_expand_symbol(get_symbol_offset(low), namebuf);
+
+		/* Search for next non-aliased symbol */
+		for (i = low + 1; i < kallsyms_num_syms; i++) {
+			if (kallsyms_addresses[i] > kallsyms_addresses[low]) {
+				symbol_end = kallsyms_addresses[i];
+				break;
+			}
 		}
 
-		/* At worst, symbol ends at end of section. */
+		/* if we found no next symbol, we use the end of the section */
+		if (!symbol_end) {
 			if (is_kernel_inittext(addr))
 				symbol_end = (unsigned long)_einittext;
 			else
 				symbol_end = (unsigned long)_etext;
-
-		/* Search for next non-aliased symbol */
-		for (i = best+1; i < kallsyms_num_syms; i++) {
-			if (kallsyms_addresses[i] > kallsyms_addresses[best]) {
-				symbol_end = kallsyms_addresses[i];
-				break;
-			}
 		}
 
-		*symbolsize = symbol_end - kallsyms_addresses[best];
+		*symbolsize = symbol_end - kallsyms_addresses[low];
 		*modname = NULL;
-		*offset = addr - kallsyms_addresses[best];
+		*offset = addr - kallsyms_addresses[low];
 		return namebuf;
 	}
 
@@ -135,7 +217,7 @@ void __print_symbol(const char *fmt, unsigned long address)
 	printk(fmt, buffer);
 }
 
-/* To avoid O(n^2) iteration, we carry prefix along. */
+/* To avoid using get_symbol_offset for every symbol, we carry prefix along. */
 struct kallsym_iter
 {
 	loff_t pos;
@@ -168,31 +250,23 @@ static int get_ksymbol_mod(struct kallsym_iter *iter)
 /* Returns space to next name. */
 static unsigned long get_ksymbol_core(struct kallsym_iter *iter)
 {
-	unsigned stemlen, off = iter->nameoff;
+	unsigned off = iter->nameoff;
 
-	/* First char of each symbol name indicates prefix length
-	   shared with previous name (stem compression). */
-	stemlen = kallsyms_names[off++];
-
-	strlcpy(iter->name+stemlen, kallsyms_names + off,
-		KSYM_NAME_LEN+1-stemlen);
-	off += strlen(kallsyms_names + off) + 1;
 	iter->owner = NULL;
 	iter->value = kallsyms_addresses[iter->pos];
-	if (is_kernel_text(iter->value) || is_kernel_inittext(iter->value))
-		iter->type = 't';
-	else
-		iter->type = 'd';
 
-	upcase_if_global(iter);
+	iter->type = kallsyms_get_symbol_type(off);
+
+	off = kallsyms_expand_symbol(off, iter->name);
+
 	return off - iter->nameoff;
 }
 
-static void reset_iter(struct kallsym_iter *iter)
+static void reset_iter(struct kallsym_iter *iter, loff_t new_pos)
 {
 	iter->name[0] = '\0';
-	iter->nameoff = 0;
-	iter->pos = 0;
+	iter->nameoff = get_symbol_offset(new_pos);
+	iter->pos = new_pos;
 }
 
 /* Returns false if pos at or past end of file. */
@@ -204,16 +278,13 @@ static int update_iter(struct kallsym_iter *iter, loff_t pos)
 		return get_ksymbol_mod(iter);
 	}
 	
-	/* If we're past the desired position, reset to start. */
-	if (pos < iter->pos)
-		reset_iter(iter);
+	/* If we're not on the desired position, reset to new position. */
+	if (pos != iter->pos)
+		reset_iter(iter, pos);
 
-	/* We need to iterate through the previous symbols: can be slow */
-	for (; iter->pos != pos; iter->pos++) {
 	iter->nameoff += get_ksymbol_core(iter);
-		cond_resched();
-	}
-	get_ksymbol_core(iter);
+	iter->pos++;
+
 	return 1;
 }
 
@@ -267,14 +338,15 @@ struct seq_operations kallsyms_op = {
 static int kallsyms_open(struct inode *inode, struct file *file)
 {
 	/* We keep iterator in m->private, since normal case is to
-	 * s_start from where we left off, so we avoid O(N^2). */
+	 * s_start from where we left off, so we avoid doing
+	 * using get_symbol_offset for every symbol */
 	struct kallsym_iter *iter;
 	int ret;
 
 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
 	if (!iter)
 		return -ENOMEM;
-	reset_iter(iter);
+	reset_iter(iter, 0);
 
 	ret = seq_open(file, &kallsyms_op);
 	if (ret == 0)

scripts/kallsyms.c

@@ -6,6 +6,22 @@
  * of the GNU General Public License, incorporated herein by reference.
  *
  * Usage: nm -n vmlinux | scripts/kallsyms [--all-symbols] > symbols.S
+ *
+ * ChangeLog:
+ *
+ * (25/Aug/2004) Paulo Marques <pmarques@grupopie.com>
+ *      Changed the compression method from stem compression to "table lookup"
+ *      compression
+ *
+ *      Table compression uses all the unused char codes on the symbols and
+ *  maps these to the most used substrings (tokens). For instance, it might
+ *  map char code 0xF7 to represent "write_" and then in every symbol where
+ *  "write_" appears it can be replaced by 0xF7, saving 5 bytes.
+ *      The used codes themselves are also placed in the table so that the
+ *  decompresion can work without "special cases".
+ *      Applied to kernel symbols, this usually produces a compression ratio
+ *  of about 50%.
+ *
  */
 
 #include <stdio.h>
@@ -13,10 +29,39 @@
 #include <string.h>
 #include <ctype.h>
 
+/* maximum token length used. It doesn't pay to increase it a lot, because
+ * very long substrings probably don't repeat themselves too often. */
+#define MAX_TOK_SIZE		11
+#define KSYM_NAME_LEN		127
+
+/* we use only a subset of the complete symbol table to gather the token count,
+ * to speed up compression, at the expense of a little compression ratio */
+#define WORKING_SET		1024
+
+/* first find the best token only on the list of tokens that would profit more
+ * than GOOD_BAD_THRESHOLD. Only if this list is empty go to the "bad" list.
+ * Increasing this value will put less tokens on the "good" list, so the search
+ * is faster. However, if the good list runs out of tokens, we must painfully
+ * search the bad list. */
+#define GOOD_BAD_THRESHOLD	10
+
+/* token hash parameters */
+#define HASH_BITS		18
+#define HASH_TABLE_SIZE		(1 << HASH_BITS)
+#define HASH_MASK		(HASH_TABLE_SIZE - 1)
+#define HASH_BASE_OFFSET	2166136261U
+#define HASH_FOLD(a)		((a)&(HASH_MASK))
+
+/* flags to mark symbols */
+#define SYM_FLAG_VALID		1
+#define SYM_FLAG_SAMPLED	2
+
 struct sym_entry {
 	unsigned long long addr;
 	char type;
-	char *sym;
+	unsigned char flags;
+	unsigned char len;
+	unsigned char *sym;
 };
 
 
@@ -25,6 +70,26 @@ static int size, cnt;
 static unsigned long long _stext, _etext, _sinittext, _einittext;
 static int all_symbols = 0;
 
+struct token {
+	unsigned char data[MAX_TOK_SIZE];
+	unsigned char len;
+	/* profit: the number of bytes that could be saved by inserting this
+	 * token into the table */
+	int profit;
+	struct token *next;	/* next token on the hash list */
+	struct token *right;	/* next token on the good/bad list */
+	struct token *left;    /* previous token on the good/bad list */
+	struct token *smaller; /* token that is less one letter than this one */
+	};
+
+struct token bad_head, good_head;
+struct token *hash_table[HASH_TABLE_SIZE];
+
+/* the table that holds the result of the compression */
+unsigned char best_table[256][MAX_TOK_SIZE+1];
+unsigned char best_table_len[256];
+
+
 static void
 usage(void)
 {
@@ -59,33 +124,52 @@ read_symbol(FILE *in, struct sym_entry *s)
 	else if (toupper(s->type) == 'A' || toupper(s->type) == 'U')
 		return -1;
 
-	s->sym = strdup(str);
+	/* include the type field in the symbol name, so that it gets
+	 * compressed together */
+	s->len = strlen(str) + 1;
+	s->sym = (char *) malloc(s->len + 1);
+	strcpy(s->sym + 1, str);
+	s->sym[0] = s->type;
+
 	return 0;
 }
 
 static int
 symbol_valid(struct sym_entry *s)
 {
+	/* Symbols which vary between passes.  Passes 1 and 2 must have
+	 * identical symbol lists.  The kallsyms_* symbols below are only added
+	 * after pass 1, they would be included in pass 2 when --all-symbols is
+	 * specified so exclude them to get a stable symbol list.
+	 */
+	static char *special_symbols[] = {
+		"kallsyms_addresses",
+		"kallsyms_num_syms",
+		"kallsyms_names",
+		"kallsyms_markers",
+		"kallsyms_token_table",
+		"kallsyms_token_index",
+
+	/* Exclude linker generated symbols which vary between passes */
+		"_SDA_BASE_",		/* ppc */
+		"_SDA2_BASE_",		/* ppc */
+		NULL };
+	int i;
+
+	/* if --all-symbols is not specified, then symbols outside the text
+	 * and inittext sections are discarded */
 	if (!all_symbols) {
 		if ((s->addr < _stext || s->addr > _etext)
 		    && (s->addr < _sinittext || s->addr > _einittext))
 			return 0;
 	}
 
-	/* Exclude symbols which vary between passes.  Passes 1 and 2 must have
-	 * identical symbol lists.  The kallsyms_* symbols below are only added
-	 * after pass 1, they would be included in pass 2 when --all-symbols is
-	 * specified so exclude them to get a stable symbol list.
-	 */
-	if (strstr(s->sym, "_compiled.") ||
-	    strcmp(s->sym, "kallsyms_addresses") == 0 ||
-	    strcmp(s->sym, "kallsyms_num_syms") == 0 ||
-	    strcmp(s->sym, "kallsyms_names") == 0)
+	/* Exclude symbols which vary between passes. */
+	if (strstr(s->sym + 1, "_compiled."))
 		return 0;
 
-	/* Exclude linker generated symbols which vary between passes */
-	if (strcmp(s->sym, "_SDA_BASE_") == 0 ||	/* ppc */
-	    strcmp(s->sym, "_SDA2_BASE_") == 0)		/* ppc */
+	for (i = 0; special_symbols[i]; i++)
+		if( strcmp(s->sym + 1, special_symbols[i]) == 0 )
 			return 0;
 
 	return 1;
@@ -108,11 +192,47 @@ read_map(FILE *in)
 	}
 }
 
+static void output_label(char *label)
+{
+	printf(".globl %s\n",label);
+	printf("\tALGN\n");
+	printf("%s:\n",label);
+}
+
+/* uncompress a compressed symbol. When this function is called, the best table
+ * might still be compressed itself, so the function needs to be recursive */
+static int expand_symbol(unsigned char *data, int len, char *result)
+{
+	int c, rlen, total=0;
+
+	while (len) {
+		c = *data;
+		/* if the table holds a single char that is the same as the one
+		 * we are looking for, then end the search */
+		if (best_table[c][0]==c && best_table_len[c]==1) {
+			*result++ = c;
+			total++;
+		} else {
+			/* if not, recurse and expand */
+			rlen = expand_symbol(best_table[c], best_table_len[c], result);
+			total += rlen;
+			result += rlen;
+		}
+		data++;
+		len--;
+	}
+	*result=0;
+
+	return total;
+}
+
 static void
 write_src(void)
 {
-	int i, valid = 0;
-	char *prev;
+	int i, k, off, valid;
+	unsigned int best_idx[256];
+	unsigned int *markers;
+	char buf[KSYM_NAME_LEN+1];
 
 	printf("#include <asm/types.h>\n");
 	printf("#if BITS_PER_LONG == 64\n");
@@ -125,43 +245,399 @@ write_src(void)
 
 	printf(".data\n");
 
-	printf(".globl kallsyms_addresses\n");
-	printf("\tALGN\n");
-	printf("kallsyms_addresses:\n");
+	output_label("kallsyms_addresses");
+	valid = 0;
 	for (i = 0; i < cnt; i++) {
-		if (!symbol_valid(&table[i]))
-			continue;
-
+		if (table[i].flags & SYM_FLAG_VALID) {
 			printf("\tPTR\t%#llx\n", table[i].addr);
 			valid++;
 		}
+	}
 	printf("\n");
 
-	printf(".globl kallsyms_num_syms\n");
-	printf("\tALGN\n");
-	printf("kallsyms_num_syms:\n");
+	output_label("kallsyms_num_syms");
 	printf("\tPTR\t%d\n", valid);
 	printf("\n");
 
-	printf(".globl kallsyms_names\n");
-	printf("\tALGN\n");
-	printf("kallsyms_names:\n");
-	prev = ""; 
+	/* table of offset markers, that give the offset in the compressed stream
+	 * every 256 symbols */
+	markers = (unsigned int *) malloc(sizeof(unsigned int)*((valid + 255) / 256));
+
+	output_label("kallsyms_names");
+	valid = 0;
+	off = 0;
 	for (i = 0; i < cnt; i++) {
-		int k;
 
-		if (!symbol_valid(&table[i]))
+		if (!table[i].flags & SYM_FLAG_VALID)
 			continue;
 
-		for (k = 0; table[i].sym[k] && table[i].sym[k] == prev[k]; ++k)
-			; 
+		if ((valid & 0xFF) == 0)
+			markers[valid >> 8] = off;
 
-		printf("\t.byte 0x%02x\n\t.asciz\t\"%s\"\n", k, table[i].sym + k);
-		prev = table[i].sym;
+		printf("\t.byte 0x%02x", table[i].len);
+		for (k = 0; k < table[i].len; k++)
+			printf(", 0x%02x", table[i].sym[k]);
+		printf("\n");
+
+		off += table[i].len + 1;
+		valid++;
 	}
 	printf("\n");
+
+	output_label("kallsyms_markers");
+	for (i = 0; i < ((valid + 255) >> 8); i++)
+		printf("\tPTR\t%d\n", markers[i]);
+	printf("\n");
+
+	free(markers);
+
+	output_label("kallsyms_token_table");
+	off = 0;
+	for (i = 0; i < 256; i++) {
+		best_idx[i] = off;
+		expand_symbol(best_table[i],best_table_len[i],buf);
+		printf("\t.asciz\t\"%s\"\n", buf);
+		off += strlen(buf) + 1;
+	}
+	printf("\n");
+
+	output_label("kallsyms_token_index");
+	for (i = 0; i < 256; i++)
+		printf("\t.short\t%d\n", best_idx[i]);
+	printf("\n");
+}
+
+
+/* table lookup compression functions */
+
+static inline unsigned int rehash_token(unsigned int hash, unsigned char data)
+{
+	return ((hash * 16777619) ^ data);
+}
+
+static unsigned int hash_token(unsigned char *data, int len)
+{
+	unsigned int hash=HASH_BASE_OFFSET;
+	int i;
+
+	for (i = 0; i < len; i++)
+		hash = rehash_token(hash, data[i]);
+
+	return HASH_FOLD(hash);
+}
+
+/* find a token given its data and hash value */
+static struct token *find_token_hash(unsigned char *data, int len, unsigned int hash)
+{
+	struct token *ptr;
+
+	ptr = hash_table[hash];
+
+	while (ptr) {
+		if ((ptr->len == len) && (memcmp(ptr->data, data, len) == 0))
+			return ptr;
+		ptr=ptr->next;
+	}
+
+	return NULL;
+}
+
+static inline void insert_token_in_group(struct token *head, struct token *ptr)
+{
+	ptr->right = head->right;
+	ptr->right->left = ptr;
+	head->right = ptr;
+	ptr->left = head;
+}
+
+static inline void remove_token_from_group(struct token *ptr)
+{
+	ptr->left->right = ptr->right;
+	ptr->right->left = ptr->left;
 }
 
+
+/* build the counts for all the tokens that start with "data", and have lenghts
+ * from 2 to "len" */
+static void learn_token(unsigned char *data, int len)
+{
+	struct token *ptr,*last_ptr;
+	int i, newprofit;
+	unsigned int hash = HASH_BASE_OFFSET;
+	unsigned int hashes[MAX_TOK_SIZE + 1];
+
+	if (len > MAX_TOK_SIZE)
+		len = MAX_TOK_SIZE;
+
+	/* calculate and store the hash values for all the sub-tokens */
+	hash = rehash_token(hash, data[0]);
+	for (i = 2; i <= len; i++) {
+		hash = rehash_token(hash, data[i-1]);
+		hashes[i] = HASH_FOLD(hash);
+	}
+
+	last_ptr = NULL;
+	ptr = NULL;
+
+	for (i = len; i >= 2; i--) {
+		hash = hashes[i];
+
+		if (!ptr) ptr = find_token_hash(data, i, hash);
+
+		if (!ptr) {
+			/* create a new token entry */
+			ptr = (struct token *) malloc(sizeof(*ptr));
+
+			memcpy(ptr->data, data, i);
+			ptr->len = i;
+
+			/* when we create an entry, it's profit is 0 because
+			 * we also take into account the size of the token on
+			 * the compressed table. We then subtract GOOD_BAD_THRESHOLD
+			 * so that the test to see if this token belongs to
+			 * the good or bad list, is a comparison to zero */
+			ptr->profit = -GOOD_BAD_THRESHOLD;
+
+			ptr->next = hash_table[hash];
+			hash_table[hash] = ptr;
+
+			insert_token_in_group(&bad_head, ptr);
+
+			ptr->smaller = NULL;
+		} else {
+			newprofit = ptr->profit + (ptr->len - 1);
+			/* check to see if this token needs to be moved to a
+			 * different list */
+			if((ptr->profit < 0) && (newprofit >= 0)) {
+				remove_token_from_group(ptr);
+				insert_token_in_group(&good_head,ptr);
+			}
+			ptr->profit = newprofit;
+		}
+
+		if (last_ptr) last_ptr->smaller = ptr;
+		last_ptr = ptr;
+
+		ptr = ptr->smaller;
+	}
+}
+
+/* decrease the counts for all the tokens that start with "data", and have lenghts
+ * from 2 to "len". This function is much simpler than learn_token because we have
+ * more guarantees (tho tokens exist, the ->smaller pointer is set, etc.)
+ * The two separate functions exist only because of compression performance */
+static void forget_token(unsigned char *data, int len)
+{
+	struct token *ptr;
+	int i, newprofit;
+	unsigned int hash=0;
+
+	if (len > MAX_TOK_SIZE) len = MAX_TOK_SIZE;
+
+	hash = hash_token(data, len);
+	ptr = find_token_hash(data, len, hash);
+
+	for (i = len; i >= 2; i--) {
+
+		newprofit = ptr->profit - (ptr->len - 1);
+		if ((ptr->profit >= 0) && (newprofit < 0)) {
+			remove_token_from_group(ptr);
+			insert_token_in_group(&bad_head, ptr);
+		}
+		ptr->profit=newprofit;
+
+		ptr=ptr->smaller;
+	}
+}
+
+/* count all the possible tokens in a symbol */
+static void learn_symbol(unsigned char *symbol, int len)
+{
+	int i;
+
+	for (i = 0; i < len - 1; i++)
+		learn_token(symbol + i, len - i);
+}
+
+/* decrease the count for all the possible tokens in a symbol */
+static void forget_symbol(unsigned char *symbol, int len)
+{
+	int i;
+
+	for (i = 0; i < len - 1; i++)
+		forget_token(symbol + i, len - i);
+}
+
+/* set all the symbol flags and do the initial token count */
+static void build_initial_tok_table(void)
+{
+	int i, use_it, valid;
+
+	valid = 0;
+	for (i = 0; i < cnt; i++) {
+		table[i].flags = 0;
+		if ( symbol_valid(&table[i]) ) {
+			table[i].flags |= SYM_FLAG_VALID;
+			valid++;
+		}
+	}
+
+	use_it = 0;
+	for (i = 0; i < cnt; i++) {
+
+		/* subsample the available symbols. This method is almost like
+		 * a Bresenham's algorithm to get uniformly distributed samples
+		 * across the symbol table */
+		if (table[i].flags & SYM_FLAG_VALID) {
+
+			use_it += WORKING_SET;
+
+			if (use_it >= valid) {
+				table[i].flags |= SYM_FLAG_SAMPLED;
+				use_it -= valid;
+			}
+		}
+		if (table[i].flags & SYM_FLAG_SAMPLED)
+			learn_symbol(table[i].sym, table[i].len);
+	}
+}
+
+/* replace a given token in all the valid symbols. Use the sampled symbols
+ * to update the counts */
+static void compress_symbols(unsigned char *str, int tlen, int idx)
+{
+	int i, len, learn, size;
+	unsigned char *p;
+
+	for (i = 0; i < cnt; i++) {
+
+		if (!(table[i].flags & SYM_FLAG_VALID)) continue;
+
+		len = table[i].len;
+		learn = 0;
+		p = table[i].sym;
+
+		do {
+			/* find the token on the symbol */
+			p = (unsigned char *) strstr((char *) p, (char *) str);
+			if (!p) break;
+
+			if (!learn) {
+				/* if this symbol was used to count, decrease it */
+				if (table[i].flags & SYM_FLAG_SAMPLED)
+					forget_symbol(table[i].sym, len);
+				learn = 1;
+			}
+
+			*p = idx;
+			size = (len - (p - table[i].sym)) - tlen + 1;
+			memmove(p + 1, p + tlen, size);
+			p++;
+			len -= tlen - 1;
+
+		} while (size >= tlen);
+
+		if(learn) {
+			table[i].len = len;
+			/* if this symbol was used to count, learn it again */
+			if(table[i].flags & SYM_FLAG_SAMPLED)
+				learn_symbol(table[i].sym, len);
+		}
+	}
+}
+
+/* search the token with the maximum profit */
+static struct token *find_best_token(void)
+{
+	struct token *ptr,*best,*head;
+	int bestprofit;
+
+	bestprofit=-10000;
+
+	/* failsafe: if the "good" list is empty search from the "bad" list */
+	if(good_head.right == &good_head) head = &bad_head;
+	else head = &good_head;
+
+	ptr = head->right;
+	best = NULL;
+	while (ptr != head) {
+		if (ptr->profit > bestprofit) {
+			bestprofit = ptr->profit;
+			best = ptr;
+		}
+		ptr = ptr->right;
+	}
+
+	return best;
+}
+
+/* this is the core of the algorithm: calculate the "best" table */
+static void optimize_result(void)
+{
+	struct token *best;
+	int i;
+
+	/* using the '\0' symbol last allows compress_symbols to use standard
+	 * fast string functions */
+	for (i = 255; i >= 0; i--) {
+
+		/* if this table slot is empty (it is not used by an actual
+		 * original char code */
+		if (!best_table_len[i]) {
+
+			/* find the token with the breates profit value */
+			best = find_best_token();
+
+			/* place it in the "best" table */
+			best_table_len[i] = best->len;
+			memcpy(best_table[i], best->data, best_table_len[i]);
+			/* zero terminate the token so that we can use strstr
+			   in compress_symbols */
+			best_table[i][best_table_len[i]]='\0';
+
+			/* replace this token in all the valid symbols */
+			compress_symbols(best_table[i], best_table_len[i], i);
+		}
+	}
+}
+
+/* start by placing the symbols that are actually used on the table */
+static void insert_real_symbols_in_table(void)
+{
+	int i, j, c;
+
+	memset(best_table, 0, sizeof(best_table));
+	memset(best_table_len, 0, sizeof(best_table_len));
+
+	for (i = 0; i < cnt; i++) {
+		if (table[i].flags & SYM_FLAG_VALID) {
+			for (j = 0; j < table[i].len; j++) {
+				c = table[i].sym[j];
+				best_table[c][0]=c;
+				best_table_len[c]=1;
+			}
+		}
+	}
+}
+
+static void optimize_token_table(void)
+{
+	memset(hash_table, 0, sizeof(hash_table));
+
+	good_head.left = &good_head;
+	good_head.right = &good_head;
+
+	bad_head.left = &bad_head;
+	bad_head.right = &bad_head;
+
+	build_initial_tok_table();
+
+	insert_real_symbols_in_table();
+
+	optimize_result();
+}
+
+
 int
 main(int argc, char **argv)
 {
@@ -171,6 +647,7 @@ main(int argc, char **argv)
 		usage();
 
 	read_map(stdin);
+	optimize_token_table();
 	write_src();
 
 	return 0;