enough.c revision 1.1.1.3
1/* enough.c -- determine the maximum size of inflate's Huffman code tables over
2 * all possible valid and complete prefix codes, subject to a length limit.
3 * Copyright (C) 2007, 2008, 2012, 2018 Mark Adler
4 * Version 1.5  5 August 2018  Mark Adler
5 */
6
7/* Version history:
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
20   1.5   5 Aug 2018  Clean up code style, formatting, and comments
21                     Show all the codes for the maximum, and only the maximum
22 */
23
24/*
25   Examine all possible prefix codes for a given number of symbols and a
26   maximum code length in bits to determine the maximum table size for zlib's
27   inflate. Only complete prefix codes are counted.
28
29   Two codes are considered distinct if the vectors of the number of codes per
30   length are not identical. So permutations of the symbol assignments result
31   in the same code for the counting, as do permutations of the assignments of
32   the bit values to the codes (i.e. only canonical codes are counted).
33
34   We build a code from shorter to longer lengths, determining how many symbols
35   are coded at each length. At each step, we have how many symbols remain to
36   be coded, what the last code length used was, and how many bit patterns of
37   that length remain unused. Then we add one to the code length and double the
38   number of unused patterns to graduate to the next code length. We then
39   assign all portions of the remaining symbols to that code length that
40   preserve the properties of a correct and eventually complete code. Those
41   properties are: we cannot use more bit patterns than are available; and when
42   all the symbols are used, there are exactly zero possible bit patterns left
43   unused.
44
45   The inflate Huffman decoding algorithm uses two-level lookup tables for
46   speed. There is a single first-level table to decode codes up to root bits
47   in length (root == 9 for literal/length codes and root == 6 for distance
48   codes, in the current inflate implementation). The base table has 1 << root
49   entries and is indexed by the next root bits of input. Codes shorter than
50   root bits have replicated table entries, so that the correct entry is
51   pointed to regardless of the bits that follow the short code. If the code is
52   longer than root bits, then the table entry points to a second-level table.
53   The size of that table is determined by the longest code with that root-bit
54   prefix. If that longest code has length len, then the table has size 1 <<
55   (len - root), to index the remaining bits in that set of codes. Each
56   subsequent root-bit prefix then has its own sub-table. The total number of
57   table entries required by the code is calculated incrementally as the number
58   of codes at each bit length is populated. When all of the codes are shorter
59   than root bits, then root is reduced to the longest code length, resulting
60   in a single, smaller, one-level table.
61
62   The inflate algorithm also provides for small values of root (relative to
63   the log2 of the number of symbols), where the shortest code has more bits
64   than root. In that case, root is increased to the length of the shortest
65   code. This program, by design, does not handle that case, so it is verified
66   that the number of symbols is less than 1 << (root + 1).
67
68   In order to speed up the examination (by about ten orders of magnitude for
69   the default arguments), the intermediate states in the build-up of a code
70   are remembered and previously visited branches are pruned. The memory
71   required for this will increase rapidly with the total number of symbols and
72   the maximum code length in bits. However this is a very small price to pay
73   for the vast speedup.
74
75   First, all of the possible prefix codes are counted, and reachable
76   intermediate states are noted by a non-zero count in a saved-results array.
77   Second, the intermediate states that lead to (root + 1) bit or longer codes
78   are used to look at all sub-codes from those junctures for their inflate
79   memory usage. (The amount of memory used is not affected by the number of
80   codes of root bits or less in length.)  Third, the visited states in the
81   construction of those sub-codes and the associated calculation of the table
82   size is recalled in order to avoid recalculating from the same juncture.
83   Beginning the code examination at (root + 1) bit codes, which is enabled by
84   identifying the reachable nodes, accounts for about six of the orders of
85   magnitude of improvement for the default arguments. About another four
86   orders of magnitude come from not revisiting previous states. Out of
87   approximately 2x10^16 possible prefix codes, only about 2x10^6 sub-codes
88   need to be examined to cover all of the possible table memory usage cases
89   for the default arguments of 286 symbols limited to 15-bit codes.
90
91   Note that the uintmax_t type is used for counting. It is quite easy to
92   exceed the capacity of an eight-byte integer with a large number of symbols
93   and a large maximum code length, so multiple-precision arithmetic would need
94   to replace the integer arithmetic in that case. This program will abort if
95   an overflow occurs. The big_t type identifies where the counting takes
96   place.
97
98   The uintmax_t type is also used for calculating the number of possible codes
99   remaining at the maximum length. This limits the maximum code length to the
100   number of bits in a long long minus the number of bits needed to represent
101   the symbols in a flat code. The code_t type identifies where the bit-pattern
102   counting takes place.
103 */
104
105#include <stdio.h>
106#include <stdlib.h>
107#include <string.h>
108#include <stdarg.h>
109#include <stdint.h>
110#include <assert.h>
111
112#define local static
113
114// Special data types.
115typedef uintmax_t big_t;    // type for code counting
116#define PRIbig "ju"         // printf format for big_t
117typedef uintmax_t code_t;   // type for bit pattern counting
118struct tab {                // type for been-here check
119    size_t len;             // allocated length of bit vector in octets
120    char *vec;              // allocated bit vector
121};
122
123/* The array for saving results, num[], is indexed with this triplet:
124
125      syms: number of symbols remaining to code
126      left: number of available bit patterns at length len
127      len: number of bits in the codes currently being assigned
128
129   Those indices are constrained thusly when saving results:
130
131      syms: 3..totsym (totsym == total symbols to code)
132      left: 2..syms - 1, but only the evens (so syms == 8 -> 2, 4, 6)
133      len: 1..max - 1 (max == maximum code length in bits)
134
135   syms == 2 is not saved since that immediately leads to a single code. left
136   must be even, since it represents the number of available bit patterns at
137   the current length, which is double the number at the previous length. left
138   ends at syms-1 since left == syms immediately results in a single code.
139   (left > sym is not allowed since that would result in an incomplete code.)
140   len is less than max, since the code completes immediately when len == max.
141
142   The offset into the array is calculated for the three indices with the first
143   one (syms) being outermost, and the last one (len) being innermost. We build
144   the array with length max-1 lists for the len index, with syms-3 of those
145   for each symbol. There are totsym-2 of those, with each one varying in
146   length as a function of sym. See the calculation of index in map() for the
147   index, and the calculation of size in main() for the size of the array.
148
149   For the deflate example of 286 symbols limited to 15-bit codes, the array
150   has 284,284 entries, taking up 2.17 MB for an 8-byte big_t. More than half
151   of the space allocated for saved results is actually used -- not all
152   possible triplets are reached in the generation of valid prefix codes.
153 */
154
155/* The array for tracking visited states, done[], is itself indexed identically
156   to the num[] array as described above for the (syms, left, len) triplet.
157   Each element in the array is further indexed by the (mem, rem) doublet,
158   where mem is the amount of inflate table space used so far, and rem is the
159   remaining unused entries in the current inflate sub-table. Each indexed
160   element is simply one bit indicating whether the state has been visited or
161   not. Since the ranges for mem and rem are not known a priori, each bit
162   vector is of a variable size, and grows as needed to accommodate the visited
163   states. mem and rem are used to calculate a single index in a triangular
164   array. Since the range of mem is expected in the default case to be about
165   ten times larger than the range of rem, the array is skewed to reduce the
166   memory usage, with eight times the range for mem than for rem. See the
167   calculations for offset and bit in been_here() for the details.
168
169   For the deflate example of 286 symbols limited to 15-bit codes, the bit
170   vectors grow to total 5.5 MB, in addition to the 4.3 MB done array itself.
171 */
172
173// Type for a variable-length, allocated string.
174typedef struct {
175    char *str;          // pointer to allocated string
176    size_t size;        // size of allocation
177    size_t len;         // length of string, not including terminating zero
178} string_t;
179
180// Clear a string_t.
181local void string_clear(string_t *s) {
182    s->str[0] = 0;
183    s->len = 0;
184}
185
186// Initialize a string_t.
187local void string_init(string_t *s) {
188    s->size = 16;
189    s->str = malloc(s->size);
190    assert(s->str != NULL && "out of memory");
191    string_clear(s);
192}
193
194// Release the allocation of a string_t.
195local void string_free(string_t *s) {
196    free(s->str);
197    s->str = NULL;
198    s->size = 0;
199    s->len = 0;
200}
201
202// Save the results of printf with fmt and the subsequent argument list to s.
203// Each call appends to s. The allocated space for s is increased as needed.
204local void string_printf(string_t *s, char *fmt, ...) {
205    va_list ap;
206    va_start(ap, fmt);
207    size_t len = s->len;
208    int ret = vsnprintf(s->str + len, s->size - len, fmt, ap);
209    assert(ret >= 0 && "out of memory");
210    s->len += ret;
211    if (s->size < s->len + 1) {
212        do {
213            s->size <<= 1;
214            assert(s->size != 0 && "overflow");
215        } while (s->size < s->len + 1);
216        s->str = realloc(s->str, s->size);
217        assert(s->str != NULL && "out of memory");
218        vsnprintf(s->str + len, s->size - len, fmt, ap);
219    }
220    va_end(ap);
221}
222
223// Globals to avoid propagating constants or constant pointers recursively.
224struct {
225    int max;            // maximum allowed bit length for the codes
226    int root;           // size of base code table in bits
227    int large;          // largest code table so far
228    size_t size;        // number of elements in num and done
229    big_t tot;          // total number of codes with maximum tables size
230    string_t out;       // display of subcodes for maximum tables size
231    int *code;          // number of symbols assigned to each bit length
232    big_t *num;         // saved results array for code counting
233    struct tab *done;   // states already evaluated array
234} g;
235
236// Index function for num[] and done[].
237local inline size_t map(int syms, int left, int len) {
238    return ((size_t)((syms - 1) >> 1) * ((syms - 2) >> 1) +
239            (left >> 1) - 1) * (g.max - 1) +
240           len - 1;
241}
242
243// Free allocated space in globals.
244local void cleanup(void) {
245    if (g.done != NULL) {
246        for (size_t n = 0; n < g.size; n++)
247            if (g.done[n].len)
248                free(g.done[n].vec);
249        g.size = 0;
250        free(g.done);   g.done = NULL;
251    }
252    free(g.num);    g.num = NULL;
253    free(g.code);   g.code = NULL;
254    string_free(&g.out);
255}
256
257// Return the number of possible prefix codes using bit patterns of lengths len
258// through max inclusive, coding syms symbols, with left bit patterns of length
259// len unused -- return -1 if there is an overflow in the counting. Keep a
260// record of previous results in num to prevent repeating the same calculation.
261local big_t count(int syms, int left, int len) {
262    // see if only one possible code
263    if (syms == left)
264        return 1;
265
266    // note and verify the expected state
267    assert(syms > left && left > 0 && len < g.max);
268
269    // see if we've done this one already
270    size_t index = map(syms, left, len);
271    big_t got = g.num[index];
272    if (got)
273        return got;         // we have -- return the saved result
274
275    // we need to use at least this many bit patterns so that the code won't be
276    // incomplete at the next length (more bit patterns than symbols)
277    int least = (left << 1) - syms;
278    if (least < 0)
279        least = 0;
280
281    // we can use at most this many bit patterns, lest there not be enough
282    // available for the remaining symbols at the maximum length (if there were
283    // no limit to the code length, this would become: most = left - 1)
284    int most = (((code_t)left << (g.max - len)) - syms) /
285               (((code_t)1 << (g.max - len)) - 1);
286
287    // count all possible codes from this juncture and add them up
288    big_t sum = 0;
289    for (int use = least; use <= most; use++) {
290        got = count(syms - use, (left - use) << 1, len + 1);
291        sum += got;
292        if (got == (big_t)-1 || sum < got)      // overflow
293            return (big_t)-1;
294    }
295
296    // verify that all recursive calls are productive
297    assert(sum != 0);
298
299    // save the result and return it
300    g.num[index] = sum;
301    return sum;
302}
303
304// Return true if we've been here before, set to true if not. Set a bit in a
305// bit vector to indicate visiting this state. Each (syms,len,left) state has a
306// variable size bit vector indexed by (mem,rem). The bit vector is lengthened
307// as needed to allow setting the (mem,rem) bit.
308local int been_here(int syms, int left, int len, int mem, int rem) {
309    // point to vector for (syms,left,len), bit in vector for (mem,rem)
310    size_t index = map(syms, left, len);
311    mem -= 1 << g.root;             // mem always includes the root table
312    mem >>= 1;                      // mem and rem are always even
313    rem >>= 1;
314    size_t offset = (mem >> 3) + rem;
315    offset = ((offset * (offset + 1)) >> 1) + rem;
316    int bit = 1 << (mem & 7);
317
318    // see if we've been here
319    size_t length = g.done[index].len;
320    if (offset < length && (g.done[index].vec[offset] & bit) != 0)
321        return 1;       // done this!
322
323    // we haven't been here before -- set the bit to show we have now
324
325    // see if we need to lengthen the vector in order to set the bit
326    if (length <= offset) {
327        // if we have one already, enlarge it, zero out the appended space
328        char *vector;
329        if (length) {
330            do {
331                length <<= 1;
332            } while (length <= offset);
333            vector = realloc(g.done[index].vec, length);
334            assert(vector != NULL && "out of memory");
335            memset(vector + g.done[index].len, 0, length - g.done[index].len);
336        }
337
338        // otherwise we need to make a new vector and zero it out
339        else {
340            length = 16;
341            while (length <= offset)
342                length <<= 1;
343            vector = calloc(length, 1);
344            assert(vector != NULL && "out of memory");
345        }
346
347        // install the new vector
348        g.done[index].len = length;
349        g.done[index].vec = vector;
350    }
351
352    // set the bit
353    g.done[index].vec[offset] |= bit;
354    return 0;
355}
356
357// Examine all possible codes from the given node (syms, len, left). Compute
358// the amount of memory required to build inflate's decoding tables, where the
359// number of code structures used so far is mem, and the number remaining in
360// the current sub-table is rem.
361local void examine(int syms, int left, int len, int mem, int rem) {
362    // see if we have a complete code
363    if (syms == left) {
364        // set the last code entry
365        g.code[len] = left;
366
367        // complete computation of memory used by this code
368        while (rem < left) {
369            left -= rem;
370            rem = 1 << (len - g.root);
371            mem += rem;
372        }
373        assert(rem == left);
374
375        // if this is at the maximum, show the sub-code
376        if (mem >= g.large) {
377            // if this is a new maximum, update the maximum and clear out the
378            // printed sub-codes from the previous maximum
379            if (mem > g.large) {
380                g.large = mem;
381                string_clear(&g.out);
382            }
383
384            // compute the starting state for this sub-code
385            syms = 0;
386            left = 1 << g.max;
387            for (int bits = g.max; bits > g.root; bits--) {
388                syms += g.code[bits];
389                left -= g.code[bits];
390                assert((left & 1) == 0);
391                left >>= 1;
392            }
393
394            // print the starting state and the resulting sub-code to g.out
395            string_printf(&g.out, "<%u, %u, %u>:",
396                          syms, g.root + 1, ((1 << g.root) - left) << 1);
397            for (int bits = g.root + 1; bits <= g.max; bits++)
398                if (g.code[bits])
399                    string_printf(&g.out, " %d[%d]", g.code[bits], bits);
400            string_printf(&g.out, "\n");
401        }
402
403        // remove entries as we drop back down in the recursion
404        g.code[len] = 0;
405        return;
406    }
407
408    // prune the tree if we can
409    if (been_here(syms, left, len, mem, rem))
410        return;
411
412    // we need to use at least this many bit patterns so that the code won't be
413    // incomplete at the next length (more bit patterns than symbols)
414    int least = (left << 1) - syms;
415    if (least < 0)
416        least = 0;
417
418    // we can use at most this many bit patterns, lest there not be enough
419    // available for the remaining symbols at the maximum length (if there were
420    // no limit to the code length, this would become: most = left - 1)
421    int most = (((code_t)left << (g.max - len)) - syms) /
422               (((code_t)1 << (g.max - len)) - 1);
423
424    // occupy least table spaces, creating new sub-tables as needed
425    int use = least;
426    while (rem < use) {
427        use -= rem;
428        rem = 1 << (len - g.root);
429        mem += rem;
430    }
431    rem -= use;
432
433    // examine codes from here, updating table space as we go
434    for (use = least; use <= most; use++) {
435        g.code[len] = use;
436        examine(syms - use, (left - use) << 1, len + 1,
437                mem + (rem ? 1 << (len - g.root) : 0), rem << 1);
438        if (rem == 0) {
439            rem = 1 << (len - g.root);
440            mem += rem;
441        }
442        rem--;
443    }
444
445    // remove entries as we drop back down in the recursion
446    g.code[len] = 0;
447}
448
449// Look at all sub-codes starting with root + 1 bits. Look at only the valid
450// intermediate code states (syms, left, len). For each completed code,
451// calculate the amount of memory required by inflate to build the decoding
452// tables. Find the maximum amount of memory required and show the codes that
453// require that maximum.
454local void enough(int syms) {
455    // clear code
456    for (int n = 0; n <= g.max; n++)
457        g.code[n] = 0;
458
459    // look at all (root + 1) bit and longer codes
460    string_clear(&g.out);           // empty saved results
461    g.large = 1 << g.root;          // base table
462    if (g.root < g.max)             // otherwise, there's only a base table
463        for (int n = 3; n <= syms; n++)
464            for (int left = 2; left < n; left += 2) {
465                // look at all reachable (root + 1) bit nodes, and the
466                // resulting codes (complete at root + 2 or more)
467                size_t index = map(n, left, g.root + 1);
468                if (g.root + 1 < g.max && g.num[index]) // reachable node
469                    examine(n, left, g.root + 1, 1 << g.root, 0);
470
471                // also look at root bit codes with completions at root + 1
472                // bits (not saved in num, since complete), just in case
473                if (g.num[index - 1] && n <= left << 1)
474                    examine((n - left) << 1, (n - left) << 1, g.root + 1,
475                            1 << g.root, 0);
476            }
477
478    // done
479    printf("maximum of %d table entries for root = %d\n", g.large, g.root);
480    fputs(g.out.str, stdout);
481}
482
483// Examine and show the total number of possible prefix codes for a given
484// maximum number of symbols, initial root table size, and maximum code length
485// in bits -- those are the command arguments in that order. The default values
486// are 286, 9, and 15 respectively, for the deflate literal/length code. The
487// possible codes are counted for each number of coded symbols from two to the
488// maximum. The counts for each of those and the total number of codes are
489// shown. The maximum number of inflate table entires is then calculated across
490// all possible codes. Each new maximum number of table entries and the
491// associated sub-code (starting at root + 1 == 10 bits) is shown.
492//
493// To count and examine prefix codes that are not length-limited, provide a
494// maximum length equal to the number of symbols minus one.
495//
496// For the deflate literal/length code, use "enough". For the deflate distance
497// code, use "enough 30 6".
498int main(int argc, char **argv) {
499    // set up globals for cleanup()
500    g.code = NULL;
501    g.num = NULL;
502    g.done = NULL;
503    string_init(&g.out);
504
505    // get arguments -- default to the deflate literal/length code
506    int syms = 286;
507    g.root = 9;
508    g.max = 15;
509    if (argc > 1) {
510        syms = atoi(argv[1]);
511        if (argc > 2) {
512            g.root = atoi(argv[2]);
513            if (argc > 3)
514                g.max = atoi(argv[3]);
515        }
516    }
517    if (argc > 4 || syms < 2 || g.root < 1 || g.max < 1) {
518        fputs("invalid arguments, need: [sym >= 2 [root >= 1 [max >= 1]]]\n",
519              stderr);
520        return 1;
521    }
522
523    // if not restricting the code length, the longest is syms - 1
524    if (g.max > syms - 1)
525        g.max = syms - 1;
526
527    // determine the number of bits in a code_t
528    int bits = 0;
529    for (code_t word = 1; word; word <<= 1)
530        bits++;
531
532    // make sure that the calculation of most will not overflow
533    if (g.max > bits || (code_t)(syms - 2) >= ((code_t)-1 >> (g.max - 1))) {
534        fputs("abort: code length too long for internal types\n", stderr);
535        return 1;
536    }
537
538    // reject impossible code requests
539    if ((code_t)(syms - 1) > ((code_t)1 << g.max) - 1) {
540        fprintf(stderr, "%d symbols cannot be coded in %d bits\n",
541                syms, g.max);
542        return 1;
543    }
544
545    // allocate code vector
546    g.code = calloc(g.max + 1, sizeof(int));
547    assert(g.code != NULL && "out of memory");
548
549    // determine size of saved results array, checking for overflows,
550    // allocate and clear the array (set all to zero with calloc())
551    if (syms == 2)              // iff max == 1
552        g.num = NULL;           // won't be saving any results
553    else {
554        g.size = syms >> 1;
555        int n = (syms - 1) >> 1;
556        assert(g.size <= (size_t)-1 / n && "overflow");
557        g.size *= n;
558        n = g.max - 1;
559        assert(g.size <= (size_t)-1 / n && "overflow");
560        g.size *= n;
561        g.num = calloc(g.size, sizeof(big_t));
562        assert(g.num != NULL && "out of memory");
563    }
564
565    // count possible codes for all numbers of symbols, add up counts
566    big_t sum = 0;
567    for (int n = 2; n <= syms; n++) {
568        big_t got = count(n, 2, 1);
569        sum += got;
570        assert(got != (big_t)-1 && sum >= got && "overflow");
571    }
572    printf("%"PRIbig" total codes for 2 to %d symbols", sum, syms);
573    if (g.max < syms - 1)
574        printf(" (%d-bit length limit)\n", g.max);
575    else
576        puts(" (no length limit)");
577
578    // allocate and clear done array for been_here()
579    if (syms == 2)
580        g.done = NULL;
581    else {
582        g.done = calloc(g.size, sizeof(struct tab));
583        assert(g.done != NULL && "out of memory");
584    }
585
586    // find and show maximum inflate table usage
587    if (g.root > g.max)             // reduce root to max length
588        g.root = g.max;
589    if ((code_t)syms < ((code_t)1 << (g.root + 1)))
590        enough(syms);
591    else
592        fputs("cannot handle minimum code lengths > root", stderr);
593
594    // done
595    cleanup();
596    return 0;
597}
598