1 /* Generate assembler source containing symbol information
3 * Copyright 2002 by Kai Germaschewski
5 * This software may be used and distributed according to the terms
6 * of the GNU General Public License, incorporated herein by reference.
8 * Usage: nm -n vmlinux | scripts/kallsyms [--all-symbols] > symbols.S
12 * (25/Aug/2004) Paulo Marques <pmarques@grupopie.com>
13 * Changed the compression method from stem compression to "table lookup"
16 * Table compression uses all the unused char codes on the symbols and
17 * maps these to the most used substrings (tokens). For instance, it might
18 * map char code 0xF7 to represent "write_" and then in every symbol where
19 * "write_" appears it can be replaced by 0xF7, saving 5 bytes.
20 * The used codes themselves are also placed in the table so that the
21 * decompresion can work without "special cases".
22 * Applied to kernel symbols, this usually produces a compression ratio
32 /* maximum token length used. It doesn't pay to increase it a lot, because
33 * very long substrings probably don't repeat themselves too often. */
34 #define MAX_TOK_SIZE 11
35 #define KSYM_NAME_LEN 127
37 /* we use only a subset of the complete symbol table to gather the token count,
38 * to speed up compression, at the expense of a little compression ratio */
39 #define WORKING_SET 1024
41 /* first find the best token only on the list of tokens that would profit more
42 * than GOOD_BAD_THRESHOLD. Only if this list is empty go to the "bad" list.
43 * Increasing this value will put less tokens on the "good" list, so the search
44 * is faster. However, if the good list runs out of tokens, we must painfully
45 * search the bad list. */
46 #define GOOD_BAD_THRESHOLD 10
48 /* token hash parameters */
50 #define HASH_TABLE_SIZE (1 << HASH_BITS)
51 #define HASH_MASK (HASH_TABLE_SIZE - 1)
52 #define HASH_BASE_OFFSET 2166136261U
53 #define HASH_FOLD(a) ((a)&(HASH_MASK))
55 /* flags to mark symbols */
56 #define SYM_FLAG_VALID 1
57 #define SYM_FLAG_SAMPLED 2
60 unsigned long long addr
;
68 static struct sym_entry
*table
;
70 static unsigned long long _stext
, _etext
, _sinittext
, _einittext
;
71 static int all_symbols
= 0;
74 unsigned char data
[MAX_TOK_SIZE
];
76 /* profit: the number of bytes that could be saved by inserting this
77 * token into the table */
79 struct token
*next
; /* next token on the hash list */
80 struct token
*right
; /* next token on the good/bad list */
81 struct token
*left
; /* previous token on the good/bad list */
82 struct token
*smaller
; /* token that is less one letter than this one */
85 struct token bad_head
, good_head
;
86 struct token
*hash_table
[HASH_TABLE_SIZE
];
88 /* the table that holds the result of the compression */
89 unsigned char best_table
[256][MAX_TOK_SIZE
+1];
90 unsigned char best_table_len
[256];
96 fprintf(stderr
, "Usage: kallsyms [--all-symbols] < in.map > out.S\n");
101 read_symbol(FILE *in
, struct sym_entry
*s
)
106 rc
= fscanf(in
, "%llx %c %499s\n", &s
->addr
, &s
->type
, str
);
115 /* Ignore most absolute/undefined (?) symbols. */
116 if (strcmp(str
, "_stext") == 0)
118 else if (strcmp(str
, "_etext") == 0)
120 else if (strcmp(str
, "_sinittext") == 0)
121 _sinittext
= s
->addr
;
122 else if (strcmp(str
, "_einittext") == 0)
123 _einittext
= s
->addr
;
124 else if (toupper(s
->type
) == 'A' || toupper(s
->type
) == 'U')
127 /* include the type field in the symbol name, so that it gets
128 * compressed together */
129 s
->len
= strlen(str
) + 1;
130 s
->sym
= (char *) malloc(s
->len
+ 1);
131 strcpy(s
->sym
+ 1, str
);
138 symbol_valid(struct sym_entry
*s
)
140 /* Symbols which vary between passes. Passes 1 and 2 must have
141 * identical symbol lists. The kallsyms_* symbols below are only added
142 * after pass 1, they would be included in pass 2 when --all-symbols is
143 * specified so exclude them to get a stable symbol list.
145 static char *special_symbols
[] = {
146 "kallsyms_addresses",
150 "kallsyms_token_table",
151 "kallsyms_token_index",
153 /* Exclude linker generated symbols which vary between passes */
154 "_SDA_BASE_", /* ppc */
155 "_SDA2_BASE_", /* ppc */
159 /* if --all-symbols is not specified, then symbols outside the text
160 * and inittext sections are discarded */
162 if ((s
->addr
< _stext
|| s
->addr
> _etext
)
163 && (s
->addr
< _sinittext
|| s
->addr
> _einittext
))
167 /* Exclude symbols which vary between passes. */
168 if (strstr(s
->sym
+ 1, "_compiled."))
171 for (i
= 0; special_symbols
[i
]; i
++)
172 if( strcmp(s
->sym
+ 1, special_symbols
[i
]) == 0 )
184 table
= realloc(table
, sizeof(*table
) * size
);
186 fprintf(stderr
, "out of memory\n");
190 if (read_symbol(in
, &table
[cnt
]) == 0)
195 static void output_label(char *label
)
197 printf(".globl %s\n",label
);
199 printf("%s:\n",label
);
202 /* uncompress a compressed symbol. When this function is called, the best table
203 * might still be compressed itself, so the function needs to be recursive */
204 static int expand_symbol(unsigned char *data
, int len
, char *result
)
206 int c
, rlen
, total
=0;
210 /* if the table holds a single char that is the same as the one
211 * we are looking for, then end the search */
212 if (best_table
[c
][0]==c
&& best_table_len
[c
]==1) {
216 /* if not, recurse and expand */
217 rlen
= expand_symbol(best_table
[c
], best_table_len
[c
], result
);
232 int i
, k
, off
, valid
;
233 unsigned int best_idx
[256];
234 unsigned int *markers
;
235 char buf
[KSYM_NAME_LEN
+1];
237 printf("#include <asm/types.h>\n");
238 printf("#if BITS_PER_LONG == 64\n");
239 printf("#define PTR .quad\n");
240 printf("#define ALGN .align 8\n");
242 printf("#define PTR .long\n");
243 printf("#define ALGN .align 4\n");
248 output_label("kallsyms_addresses");
250 for (i
= 0; i
< cnt
; i
++) {
251 if (table
[i
].flags
& SYM_FLAG_VALID
) {
252 printf("\tPTR\t%#llx\n", table
[i
].addr
);
258 output_label("kallsyms_num_syms");
259 printf("\tPTR\t%d\n", valid
);
262 /* table of offset markers, that give the offset in the compressed stream
263 * every 256 symbols */
264 markers
= (unsigned int *) malloc(sizeof(unsigned int)*((valid
+ 255) / 256));
266 output_label("kallsyms_names");
269 for (i
= 0; i
< cnt
; i
++) {
271 if (!table
[i
].flags
& SYM_FLAG_VALID
)
274 if ((valid
& 0xFF) == 0)
275 markers
[valid
>> 8] = off
;
277 printf("\t.byte 0x%02x", table
[i
].len
);
278 for (k
= 0; k
< table
[i
].len
; k
++)
279 printf(", 0x%02x", table
[i
].sym
[k
]);
282 off
+= table
[i
].len
+ 1;
287 output_label("kallsyms_markers");
288 for (i
= 0; i
< ((valid
+ 255) >> 8); i
++)
289 printf("\tPTR\t%d\n", markers
[i
]);
294 output_label("kallsyms_token_table");
296 for (i
= 0; i
< 256; i
++) {
298 expand_symbol(best_table
[i
],best_table_len
[i
],buf
);
299 printf("\t.asciz\t\"%s\"\n", buf
);
300 off
+= strlen(buf
) + 1;
304 output_label("kallsyms_token_index");
305 for (i
= 0; i
< 256; i
++)
306 printf("\t.short\t%d\n", best_idx
[i
]);
311 /* table lookup compression functions */
313 static inline unsigned int rehash_token(unsigned int hash
, unsigned char data
)
315 return ((hash
* 16777619) ^ data
);
318 static unsigned int hash_token(unsigned char *data
, int len
)
320 unsigned int hash
=HASH_BASE_OFFSET
;
323 for (i
= 0; i
< len
; i
++)
324 hash
= rehash_token(hash
, data
[i
]);
326 return HASH_FOLD(hash
);
329 /* find a token given its data and hash value */
330 static struct token
*find_token_hash(unsigned char *data
, int len
, unsigned int hash
)
334 ptr
= hash_table
[hash
];
337 if ((ptr
->len
== len
) && (memcmp(ptr
->data
, data
, len
) == 0))
345 static inline void insert_token_in_group(struct token
*head
, struct token
*ptr
)
347 ptr
->right
= head
->right
;
348 ptr
->right
->left
= ptr
;
353 static inline void remove_token_from_group(struct token
*ptr
)
355 ptr
->left
->right
= ptr
->right
;
356 ptr
->right
->left
= ptr
->left
;
360 /* build the counts for all the tokens that start with "data", and have lenghts
362 static void learn_token(unsigned char *data
, int len
)
364 struct token
*ptr
,*last_ptr
;
366 unsigned int hash
= HASH_BASE_OFFSET
;
367 unsigned int hashes
[MAX_TOK_SIZE
+ 1];
369 if (len
> MAX_TOK_SIZE
)
372 /* calculate and store the hash values for all the sub-tokens */
373 hash
= rehash_token(hash
, data
[0]);
374 for (i
= 2; i
<= len
; i
++) {
375 hash
= rehash_token(hash
, data
[i
-1]);
376 hashes
[i
] = HASH_FOLD(hash
);
382 for (i
= len
; i
>= 2; i
--) {
385 if (!ptr
) ptr
= find_token_hash(data
, i
, hash
);
388 /* create a new token entry */
389 ptr
= (struct token
*) malloc(sizeof(*ptr
));
391 memcpy(ptr
->data
, data
, i
);
394 /* when we create an entry, it's profit is 0 because
395 * we also take into account the size of the token on
396 * the compressed table. We then subtract GOOD_BAD_THRESHOLD
397 * so that the test to see if this token belongs to
398 * the good or bad list, is a comparison to zero */
399 ptr
->profit
= -GOOD_BAD_THRESHOLD
;
401 ptr
->next
= hash_table
[hash
];
402 hash_table
[hash
] = ptr
;
404 insert_token_in_group(&bad_head
, ptr
);
408 newprofit
= ptr
->profit
+ (ptr
->len
- 1);
409 /* check to see if this token needs to be moved to a
411 if((ptr
->profit
< 0) && (newprofit
>= 0)) {
412 remove_token_from_group(ptr
);
413 insert_token_in_group(&good_head
,ptr
);
415 ptr
->profit
= newprofit
;
418 if (last_ptr
) last_ptr
->smaller
= ptr
;
425 /* decrease the counts for all the tokens that start with "data", and have lenghts
426 * from 2 to "len". This function is much simpler than learn_token because we have
427 * more guarantees (tho tokens exist, the ->smaller pointer is set, etc.)
428 * The two separate functions exist only because of compression performance */
429 static void forget_token(unsigned char *data
, int len
)
435 if (len
> MAX_TOK_SIZE
) len
= MAX_TOK_SIZE
;
437 hash
= hash_token(data
, len
);
438 ptr
= find_token_hash(data
, len
, hash
);
440 for (i
= len
; i
>= 2; i
--) {
442 newprofit
= ptr
->profit
- (ptr
->len
- 1);
443 if ((ptr
->profit
>= 0) && (newprofit
< 0)) {
444 remove_token_from_group(ptr
);
445 insert_token_in_group(&bad_head
, ptr
);
447 ptr
->profit
=newprofit
;
453 /* count all the possible tokens in a symbol */
454 static void learn_symbol(unsigned char *symbol
, int len
)
458 for (i
= 0; i
< len
- 1; i
++)
459 learn_token(symbol
+ i
, len
- i
);
462 /* decrease the count for all the possible tokens in a symbol */
463 static void forget_symbol(unsigned char *symbol
, int len
)
467 for (i
= 0; i
< len
- 1; i
++)
468 forget_token(symbol
+ i
, len
- i
);
471 /* set all the symbol flags and do the initial token count */
472 static void build_initial_tok_table(void)
474 int i
, use_it
, valid
;
477 for (i
= 0; i
< cnt
; i
++) {
479 if ( symbol_valid(&table
[i
]) ) {
480 table
[i
].flags
|= SYM_FLAG_VALID
;
486 for (i
= 0; i
< cnt
; i
++) {
488 /* subsample the available symbols. This method is almost like
489 * a Bresenham's algorithm to get uniformly distributed samples
490 * across the symbol table */
491 if (table
[i
].flags
& SYM_FLAG_VALID
) {
493 use_it
+= WORKING_SET
;
495 if (use_it
>= valid
) {
496 table
[i
].flags
|= SYM_FLAG_SAMPLED
;
500 if (table
[i
].flags
& SYM_FLAG_SAMPLED
)
501 learn_symbol(table
[i
].sym
, table
[i
].len
);
505 /* replace a given token in all the valid symbols. Use the sampled symbols
506 * to update the counts */
507 static void compress_symbols(unsigned char *str
, int tlen
, int idx
)
509 int i
, len
, learn
, size
;
512 for (i
= 0; i
< cnt
; i
++) {
514 if (!(table
[i
].flags
& SYM_FLAG_VALID
)) continue;
521 /* find the token on the symbol */
522 p
= (unsigned char *) strstr((char *) p
, (char *) str
);
526 /* if this symbol was used to count, decrease it */
527 if (table
[i
].flags
& SYM_FLAG_SAMPLED
)
528 forget_symbol(table
[i
].sym
, len
);
533 size
= (len
- (p
- table
[i
].sym
)) - tlen
+ 1;
534 memmove(p
+ 1, p
+ tlen
, size
);
538 } while (size
>= tlen
);
542 /* if this symbol was used to count, learn it again */
543 if(table
[i
].flags
& SYM_FLAG_SAMPLED
)
544 learn_symbol(table
[i
].sym
, len
);
549 /* search the token with the maximum profit */
550 static struct token
*find_best_token(void)
552 struct token
*ptr
,*best
,*head
;
557 /* failsafe: if the "good" list is empty search from the "bad" list */
558 if(good_head
.right
== &good_head
) head
= &bad_head
;
559 else head
= &good_head
;
563 while (ptr
!= head
) {
564 if (ptr
->profit
> bestprofit
) {
565 bestprofit
= ptr
->profit
;
574 /* this is the core of the algorithm: calculate the "best" table */
575 static void optimize_result(void)
580 /* using the '\0' symbol last allows compress_symbols to use standard
581 * fast string functions */
582 for (i
= 255; i
>= 0; i
--) {
584 /* if this table slot is empty (it is not used by an actual
585 * original char code */
586 if (!best_table_len
[i
]) {
588 /* find the token with the breates profit value */
589 best
= find_best_token();
591 /* place it in the "best" table */
592 best_table_len
[i
] = best
->len
;
593 memcpy(best_table
[i
], best
->data
, best_table_len
[i
]);
594 /* zero terminate the token so that we can use strstr
595 in compress_symbols */
596 best_table
[i
][best_table_len
[i
]]='\0';
598 /* replace this token in all the valid symbols */
599 compress_symbols(best_table
[i
], best_table_len
[i
], i
);
604 /* start by placing the symbols that are actually used on the table */
605 static void insert_real_symbols_in_table(void)
609 memset(best_table
, 0, sizeof(best_table
));
610 memset(best_table_len
, 0, sizeof(best_table_len
));
612 for (i
= 0; i
< cnt
; i
++) {
613 if (table
[i
].flags
& SYM_FLAG_VALID
) {
614 for (j
= 0; j
< table
[i
].len
; j
++) {
623 static void optimize_token_table(void)
625 memset(hash_table
, 0, sizeof(hash_table
));
627 good_head
.left
= &good_head
;
628 good_head
.right
= &good_head
;
630 bad_head
.left
= &bad_head
;
631 bad_head
.right
= &bad_head
;
633 build_initial_tok_table();
635 insert_real_symbols_in_table();
642 main(int argc
, char **argv
)
644 if (argc
== 2 && strcmp(argv
[1], "--all-symbols") == 0)
650 optimize_token_table();