1 /* enough.c -- determine the maximum size of inflate's Huffman code tables over
2 * all possible valid and complete Huffman codes, subject to a length limit.
3 * Copyright (C) 2007, 2008, 2012 Mark Adler
4 * Version 1.4 18 August 2012 Mark Adler
8 1.0 3 Jan 2007 First version (derived from codecount.c version 1.4)
9 1.1 4 Jan 2007 Use faster incremental table usage computation
10 Prune examine() search on previously visited states
11 1.2 5 Jan 2007 Comments clean up
12 As inflate does, decrease root for short codes
13 Refuse cases where inflate would increase root
14 1.3 17 Feb 2008 Add argument for initial root table size
15 Fix bug for initial root table size == max - 1
16 Use a macro to compute the history index
17 1.4 18 Aug 2012 Avoid shifts more than bits in type (caused endless loop!)
18 Clean up comparisons of different types
19 Clean up code indentation
23 Examine all possible Huffman codes for a given number of symbols and a
24 maximum code length in bits to determine the maximum table size for zilb's
25 inflate. Only complete Huffman codes are counted.
27 Two codes are considered distinct if the vectors of the number of codes per
28 length are not identical. So permutations of the symbol assignments result
29 in the same code for the counting, as do permutations of the assignments of
30 the bit values to the codes (i.e. only canonical codes are counted).
32 We build a code from shorter to longer lengths, determining how many symbols
33 are coded at each length. At each step, we have how many symbols remain to
34 be coded, what the last code length used was, and how many bit patterns of
35 that length remain unused. Then we add one to the code length and double the
36 number of unused patterns to graduate to the next code length. We then
37 assign all portions of the remaining symbols to that code length that
38 preserve the properties of a correct and eventually complete code. Those
39 properties are: we cannot use more bit patterns than are available; and when
40 all the symbols are used, there are exactly zero possible bit patterns
43 The inflate Huffman decoding algorithm uses two-level lookup tables for
44 speed. There is a single first-level table to decode codes up to root bits
45 in length (root == 9 in the current inflate implementation). The table
46 has 1 << root entries and is indexed by the next root bits of input. Codes
47 shorter than root bits have replicated table entries, so that the correct
48 entry is pointed to regardless of the bits that follow the short code. If
49 the code is longer than root bits, then the table entry points to a second-
50 level table. The size of that table is determined by the longest code with
51 that root-bit prefix. If that longest code has length len, then the table
52 has size 1 << (len - root), to index the remaining bits in that set of
53 codes. Each subsequent root-bit prefix then has its own sub-table. The
54 total number of table entries required by the code is calculated
55 incrementally as the number of codes at each bit length is populated. When
56 all of the codes are shorter than root bits, then root is reduced to the
57 longest code length, resulting in a single, smaller, one-level table.
59 The inflate algorithm also provides for small values of root (relative to
60 the log2 of the number of symbols), where the shortest code has more bits
61 than root. In that case, root is increased to the length of the shortest
62 code. This program, by design, does not handle that case, so it is verified
63 that the number of symbols is less than 2^(root + 1).
65 In order to speed up the examination (by about ten orders of magnitude for
66 the default arguments), the intermediate states in the build-up of a code
67 are remembered and previously visited branches are pruned. The memory
68 required for this will increase rapidly with the total number of symbols and
69 the maximum code length in bits. However this is a very small price to pay
72 First, all of the possible Huffman codes are counted, and reachable
73 intermediate states are noted by a non-zero count in a saved-results array.
74 Second, the intermediate states that lead to (root + 1) bit or longer codes
75 are used to look at all sub-codes from those junctures for their inflate
76 memory usage. (The amount of memory used is not affected by the number of
77 codes of root bits or less in length.) Third, the visited states in the
78 construction of those sub-codes and the associated calculation of the table
79 size is recalled in order to avoid recalculating from the same juncture.
80 Beginning the code examination at (root + 1) bit codes, which is enabled by
81 identifying the reachable nodes, accounts for about six of the orders of
82 magnitude of improvement for the default arguments. About another four
83 orders of magnitude come from not revisiting previous states. Out of
84 approximately 2x10^16 possible Huffman codes, only about 2x10^6 sub-codes
85 need to be examined to cover all of the possible table memory usage cases
86 for the default arguments of 286 symbols limited to 15-bit codes.
88 Note that an unsigned long long type is used for counting. It is quite easy
89 to exceed the capacity of an eight-byte integer with a large number of
90 symbols and a large maximum code length, so multiple-precision arithmetic
91 would need to replace the unsigned long long arithmetic in that case. This
92 program will abort if an overflow occurs. The big_t type identifies where
93 the counting takes place.
95 An unsigned long long type is also used for calculating the number of
96 possible codes remaining at the maximum length. This limits the maximum
97 code length to the number of bits in a long long minus the number of bits
98 needed to represent the symbols in a flat code. The code_t type identifies
99 where the bit pattern counting takes place.
109 /* special data types */
110 typedef unsigned long long big_t; /* type for code counting */
111 #define PRIbig "llu" /* printf format for big_t */
112 typedef unsigned long long code_t; /* type for bit pattern counting */
113 struct tab { /* type for been here check */
114 size_t len; /* length of bit vector in char's */
115 char *vec; /* allocated bit vector */
118 /* The array for saving results, num[], is indexed with this triplet:
120 syms: number of symbols remaining to code
121 left: number of available bit patterns at length len
122 len: number of bits in the codes currently being assigned
124 Those indices are constrained thusly when saving results:
126 syms: 3..totsym (totsym == total symbols to code)
127 left: 2..syms - 1, but only the evens (so syms == 8 -> 2, 4, 6)
128 len: 1..max - 1 (max == maximum code length in bits)
130 syms == 2 is not saved since that immediately leads to a single code. left
131 must be even, since it represents the number of available bit patterns at
132 the current length, which is double the number at the previous length.
133 left ends at syms-1 since left == syms immediately results in a single code.
134 (left > sym is not allowed since that would result in an incomplete code.)
135 len is less than max, since the code completes immediately when len == max.
137 The offset into the array is calculated for the three indices with the
138 first one (syms) being outermost, and the last one (len) being innermost.
139 We build the array with length max-1 lists for the len index, with syms-3
140 of those for each symbol. There are totsym-2 of those, with each one
141 varying in length as a function of sym. See the calculation of index in
142 count() for the index, and the calculation of size in main() for the size
145 For the deflate example of 286 symbols limited to 15-bit codes, the array
146 has 284,284 entries, taking up 2.17 MB for an 8-byte big_t. More than
147 half of the space allocated for saved results is actually used -- not all
148 possible triplets are reached in the generation of valid Huffman codes.
151 /* The array for tracking visited states, done[], is itself indexed identically
152 to the num[] array as described above for the (syms, left, len) triplet.
153 Each element in the array is further indexed by the (mem, rem) doublet,
154 where mem is the amount of inflate table space used so far, and rem is the
155 remaining unused entries in the current inflate sub-table. Each indexed
156 element is simply one bit indicating whether the state has been visited or
157 not. Since the ranges for mem and rem are not known a priori, each bit
158 vector is of a variable size, and grows as needed to accommodate the visited
159 states. mem and rem are used to calculate a single index in a triangular
160 array. Since the range of mem is expected in the default case to be about
161 ten times larger than the range of rem, the array is skewed to reduce the
162 memory usage, with eight times the range for mem than for rem. See the
163 calculations for offset and bit in beenhere() for the details.
165 For the deflate example of 286 symbols limited to 15-bit codes, the bit
166 vectors grow to total approximately 21 MB, in addition to the 4.3 MB done[]
170 /* Globals to avoid propagating constants or constant pointers recursively */
172 int max; /* maximum allowed bit length for the codes */
173 int root; /* size of base code table in bits */
174 int large; /* largest code table so far */
175 size_t size; /* number of elements in num and done */
176 int *code; /* number of symbols assigned to each bit length */
177 big_t *num; /* saved results array for code counting */
178 struct tab *done; /* states already evaluated array */
181 /* Index function for num[] and done[] */
182 #define INDEX(i,j,k) (((size_t)((i-1)>>1)*((i-2)>>1)+(j>>1)-1)*(g.max-1)+k-1)
184 /* Free allocated space. Uses globals code, num, and done. */
185 local void cleanup(void)
189 if (g.done != NULL) {
190 for (n = 0; n < g.size; n++)
201 /* Return the number of possible Huffman codes using bit patterns of lengths
202 len through max inclusive, coding syms symbols, with left bit patterns of
203 length len unused -- return -1 if there is an overflow in the counting.
204 Keep a record of previous results in num to prevent repeating the same
205 calculation. Uses the globals max and num. */
206 local big_t count(int syms, int len, int left)
208 big_t sum; /* number of possible codes from this juncture */
209 big_t got; /* value returned from count() */
210 int least; /* least number of syms to use at this juncture */
211 int most; /* most number of syms to use at this juncture */
212 int use; /* number of bit patterns to use in next call */
213 size_t index; /* index of this case in *num */
215 /* see if only one possible code */
219 /* note and verify the expected state */
220 assert(syms > left && left > 0 && len < g.max);
222 /* see if we've done this one already */
223 index = INDEX(syms, left, len);
226 return got; /* we have -- return the saved result */
228 /* we need to use at least this many bit patterns so that the code won't be
229 incomplete at the next length (more bit patterns than symbols) */
230 least = (left << 1) - syms;
234 /* we can use at most this many bit patterns, lest there not be enough
235 available for the remaining symbols at the maximum length (if there were
236 no limit to the code length, this would become: most = left - 1) */
237 most = (((code_t)left << (g.max - len)) - syms) /
238 (((code_t)1 << (g.max - len)) - 1);
240 /* count all possible codes from this juncture and add them up */
242 for (use = least; use <= most; use++) {
243 got = count(syms - use, len + 1, (left - use) << 1);
245 if (got == (big_t)0 - 1 || sum < got) /* overflow */
249 /* verify that all recursive calls are productive */
252 /* save the result and return it */
257 /* Return true if we've been here before, set to true if not. Set a bit in a
258 bit vector to indicate visiting this state. Each (syms,len,left) state
259 has a variable size bit vector indexed by (mem,rem). The bit vector is
260 lengthened if needed to allow setting the (mem,rem) bit. */
261 local int beenhere(int syms, int len, int left, int mem, int rem)
263 size_t index; /* index for this state's bit vector */
264 size_t offset; /* offset in this state's bit vector */
265 int bit; /* mask for this state's bit */
266 size_t length; /* length of the bit vector in bytes */
267 char *vector; /* new or enlarged bit vector */
269 /* point to vector for (syms,left,len), bit in vector for (mem,rem) */
270 index = INDEX(syms, left, len);
272 offset = (mem >> 3) + rem;
273 offset = ((offset * (offset + 1)) >> 1) + rem;
274 bit = 1 << (mem & 7);
276 /* see if we've been here */
277 length = g.done[index].len;
278 if (offset < length && (g.done[index].vec[offset] & bit) != 0)
279 return 1; /* done this! */
281 /* we haven't been here before -- set the bit to show we have now */
283 /* see if we need to lengthen the vector in order to set the bit */
284 if (length <= offset) {
285 /* if we have one already, enlarge it, zero out the appended space */
289 } while (length <= offset);
290 vector = realloc(g.done[index].vec, length);
292 memset(vector + g.done[index].len, 0,
293 length - g.done[index].len);
296 /* otherwise we need to make a new vector and zero it out */
298 length = 1 << (len - g.root);
299 while (length <= offset)
301 vector = calloc(length, sizeof(char));
304 /* in either case, bail if we can't get the memory */
305 if (vector == NULL) {
306 fputs("abort: unable to allocate enough memory\n", stderr);
311 /* install the new vector */
312 g.done[index].len = length;
313 g.done[index].vec = vector;
317 g.done[index].vec[offset] |= bit;
321 /* Examine all possible codes from the given node (syms, len, left). Compute
322 the amount of memory required to build inflate's decoding tables, where the
323 number of code structures used so far is mem, and the number remaining in
324 the current sub-table is rem. Uses the globals max, code, root, large, and
326 local void examine(int syms, int len, int left, int mem, int rem)
328 int least; /* least number of syms to use at this juncture */
329 int most; /* most number of syms to use at this juncture */
330 int use; /* number of bit patterns to use in next call */
332 /* see if we have a complete code */
334 /* set the last code entry */
337 /* complete computation of memory used by this code */
340 rem = 1 << (len - g.root);
345 /* if this is a new maximum, show the entries used and the sub-code */
348 printf("max %d: ", mem);
349 for (use = g.root + 1; use <= g.max; use++)
351 printf("%d[%d] ", g.code[use], use);
356 /* remove entries as we drop back down in the recursion */
361 /* prune the tree if we can */
362 if (beenhere(syms, len, left, mem, rem))
365 /* we need to use at least this many bit patterns so that the code won't be
366 incomplete at the next length (more bit patterns than symbols) */
367 least = (left << 1) - syms;
371 /* we can use at most this many bit patterns, lest there not be enough
372 available for the remaining symbols at the maximum length (if there were
373 no limit to the code length, this would become: most = left - 1) */
374 most = (((code_t)left << (g.max - len)) - syms) /
375 (((code_t)1 << (g.max - len)) - 1);
377 /* occupy least table spaces, creating new sub-tables as needed */
381 rem = 1 << (len - g.root);
386 /* examine codes from here, updating table space as we go */
387 for (use = least; use <= most; use++) {
389 examine(syms - use, len + 1, (left - use) << 1,
390 mem + (rem ? 1 << (len - g.root) : 0), rem << 1);
392 rem = 1 << (len - g.root);
398 /* remove entries as we drop back down in the recursion */
402 /* Look at all sub-codes starting with root + 1 bits. Look at only the valid
403 intermediate code states (syms, left, len). For each completed code,
404 calculate the amount of memory required by inflate to build the decoding
405 tables. Find the maximum amount of memory required and show the code that
406 requires that maximum. Uses the globals max, root, and num. */
407 local void enough(int syms)
409 int n; /* number of remaing symbols for this node */
410 int left; /* number of unused bit patterns at this length */
411 size_t index; /* index of this case in *num */
414 for (n = 0; n <= g.max; n++)
417 /* look at all (root + 1) bit and longer codes */
418 g.large = 1 << g.root; /* base table */
419 if (g.root < g.max) /* otherwise, there's only a base table */
420 for (n = 3; n <= syms; n++)
421 for (left = 2; left < n; left += 2)
423 /* look at all reachable (root + 1) bit nodes, and the
424 resulting codes (complete at root + 2 or more) */
425 index = INDEX(n, left, g.root + 1);
426 if (g.root + 1 < g.max && g.num[index]) /* reachable node */
427 examine(n, g.root + 1, left, 1 << g.root, 0);
429 /* also look at root bit codes with completions at root + 1
430 bits (not saved in num, since complete), just in case */
431 if (g.num[index - 1] && n <= left << 1)
432 examine((n - left) << 1, g.root + 1, (n - left) << 1,
437 printf("done: maximum of %d table entries\n", g.large);
441 Examine and show the total number of possible Huffman codes for a given
442 maximum number of symbols, initial root table size, and maximum code length
443 in bits -- those are the command arguments in that order. The default
444 values are 286, 9, and 15 respectively, for the deflate literal/length code.
445 The possible codes are counted for each number of coded symbols from two to
446 the maximum. The counts for each of those and the total number of codes are
447 shown. The maximum number of inflate table entires is then calculated
448 across all possible codes. Each new maximum number of table entries and the
449 associated sub-code (starting at root + 1 == 10 bits) is shown.
451 To count and examine Huffman codes that are not length-limited, provide a
452 maximum length equal to the number of symbols minus one.
454 For the deflate literal/length code, use "enough". For the deflate distance
455 code, use "enough 30 6".
457 int main(int argc, char **argv)
459 int syms; /* total number of symbols to code */
460 int n; /* number of symbols to code for this run */
461 big_t got; /* return value of count() */
462 big_t sum; /* accumulated number of codes over n */
463 code_t word; /* for counting bits in code_t */
465 /* set up globals for cleanup() */
470 /* get arguments -- default to the deflate literal/length code */
475 syms = atoi(argv[1]);
477 g.root = atoi(argv[2]);
479 g.max = atoi(argv[3]);
482 if (argc > 4 || syms < 2 || g.root < 1 || g.max < 1) {
483 fputs("invalid arguments, need: [sym >= 2 [root >= 1 [max >= 1]]]\n",
488 /* if not restricting the code length, the longest is syms - 1 */
489 if (g.max > syms - 1)
492 /* determine the number of bits in a code_t */
493 for (n = 0, word = 1; word; n++, word <<= 1)
496 /* make sure that the calculation of most will not overflow */
497 if (g.max > n || (code_t)(syms - 2) >= (((code_t)0 - 1) >> (g.max - 1))) {
498 fputs("abort: code length too long for internal types\n", stderr);
502 /* reject impossible code requests */
503 if ((code_t)(syms - 1) > ((code_t)1 << g.max) - 1) {
504 fprintf(stderr, "%d symbols cannot be coded in %d bits\n",
509 /* allocate code vector */
510 g.code = calloc(g.max + 1, sizeof(int));
511 if (g.code == NULL) {
512 fputs("abort: unable to allocate enough memory\n", stderr);
516 /* determine size of saved results array, checking for overflows,
517 allocate and clear the array (set all to zero with calloc()) */
518 if (syms == 2) /* iff max == 1 */
519 g.num = NULL; /* won't be saving any results */
522 if (g.size > ((size_t)0 - 1) / (n = (syms - 1) >> 1) ||
523 (g.size *= n, g.size > ((size_t)0 - 1) / (n = g.max - 1)) ||
524 (g.size *= n, g.size > ((size_t)0 - 1) / sizeof(big_t)) ||
525 (g.num = calloc(g.size, sizeof(big_t))) == NULL) {
526 fputs("abort: unable to allocate enough memory\n", stderr);
532 /* count possible codes for all numbers of symbols, add up counts */
534 for (n = 2; n <= syms; n++) {
535 got = count(n, 1, 2);
537 if (got == (big_t)0 - 1 || sum < got) { /* overflow */
538 fputs("abort: can't count that high!\n", stderr);
542 printf("%"PRIbig" %d-codes\n", got, n);
544 printf("%"PRIbig" total codes for 2 to %d symbols", sum, syms);
545 if (g.max < syms - 1)
546 printf(" (%d-bit length limit)\n", g.max);
548 puts(" (no length limit)");
550 /* allocate and clear done array for beenhere() */
553 else if (g.size > ((size_t)0 - 1) / sizeof(struct tab) ||
554 (g.done = calloc(g.size, sizeof(struct tab))) == NULL) {
555 fputs("abort: unable to allocate enough memory\n", stderr);
560 /* find and show maximum inflate table usage */
561 if (g.root > g.max) /* reduce root to max length */
563 if ((code_t)syms < ((code_t)1 << (g.root + 1)))
566 puts("cannot handle minimum code lengths > root");