2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Statistics for Firm. Pattern history.
23 * @author Michael Beck
40 #include "pattern_dmp.h"
43 #ifdef FIRM_STATISTICS
46 * just be make some things clear :-), the
49 #define HASH_MAP(type) pset_##type
51 typedef pset pset_pattern_entry_t;
53 typedef unsigned char BYTE;
55 /** Maximum size of the pattern store. */
56 #define PATTERN_STORE_SIZE 2048
62 typedef struct _code_buf_t {
63 BYTE *next; /**< Next byte address to be written. */
64 BYTE *end; /**< End address of the buffer. */
65 BYTE *start; /**< Start address of the buffer. */
66 unsigned hash; /**< The hash value for the buffer content. */
67 unsigned overrun; /**< flag set if the buffer was overrun */
74 VLC_7BIT = 0x00, /**< 8 bit code, carrying 7 bits payload */
75 VLC_14BIT = 0x80, /**< 16 bit code, carrying 14 bits payload */
76 VLC_21BIT = 0xC0, /**< 24 bit code, carrying 21 bits payload */
77 VLC_28BIT = 0xE0, /**< 32 bit code, carrying 28 bits payload */
78 VLC_32BIT = 0xF0, /**< 40 bit code, carrying 32 bits payload */
80 VLC_TAG_FIRST = 0xF1, /**< First possible tag value. */
81 VLC_TAG_ICONST = 0xFB, /**< Encodes an integer constant. */
82 VLC_TAG_EMPTY = 0xFC, /**< Encodes an empty entity. */
83 VLC_TAG_OPTION = 0xFD, /**< Options exists. */
84 VLC_TAG_REF = 0xFE, /**< Special tag, next code is an ID. */
85 VLC_TAG_END = 0xFF, /**< End tag. */
89 * An entry for holding one pattern.
91 typedef struct _pattern_entry_t {
92 counter_t count; /**< Amount of pattern occurance. */
93 unsigned len; /**< The length of the VLC encoded buffer. */
94 BYTE buf[1]; /**< The buffer containing the VLC encoded pattern. */
98 * Current options for the pattern matcher.
101 OPT_WITH_MODE = 0x00000001, /**< use modes */
102 OPT_ENC_DAG = 0x00000002, /**< encode DAGs, not terms */
103 OPT_WITH_ICONST = 0x00000004, /**< encode integer constants */
104 OPT_PERSIST_PATTERN = 0x00000008, /**< persistent pattern hash */
111 typedef struct _pattern_info_t {
112 int enable; /**< If non-zero, this module is enabled. */
113 struct obstack obst; /**< An obstack containing the counters. */
114 HASH_MAP(pattern_entry_t) *pattern_hash; /**< A hash map containing the pattern. */
115 unsigned bound; /**< Lowest value for pattern output. */
116 unsigned options; /**< Current option mask. */
117 unsigned min_depth; /**< Minimum pattern depth. */
118 unsigned max_depth; /**< Maximum pattern depth. */
124 static pattern_info_t _status, *status = &_status;
127 * Compare two pattern for its occurance counter.
129 static int pattern_count_cmp(const void *elt, const void *key)
133 pattern_entry_t **e1 = (pattern_entry_t **)elt;
134 pattern_entry_t **e2 = (pattern_entry_t **)key;
136 /* we want it sorted in descending order */
137 cmp = cnt_cmp(&(*e2)->count, &(*e1)->count);
140 } /* pattern_count_cmp */
143 * Compare two pattern for its pattern hash.
145 static int pattern_cmp(const void *elt, const void *key)
147 const pattern_entry_t *e1 = elt;
148 const pattern_entry_t *e2 = key;
149 int diff = e1->len - e2->len;
154 return memcmp(e1->buf, e2->buf, e1->len);
158 * Initialize a code buffer.
160 * @param buf the code buffer
161 * @param data a buffer address
162 * @param len the length of the data buffer
164 static void init_buf(CODE_BUFFER *buf, BYTE *data, unsigned len)
168 buf->end = data + len;
169 buf->hash = 0x2BAD4; /* An arbitrary seed. */
174 * Put a byte into the buffer.
176 * @param buf the code buffer
177 * @param byte the byte to write
179 * The hash value for the buffer content is updated.
181 static inline void put_byte(CODE_BUFFER *buf, BYTE byte)
183 if (buf->next < buf->end) {
185 buf->hash = (buf->hash * 9) ^ byte;
192 * Returns the current length of a buffer.
194 * @param buf the code buffer
196 * @return the length of the buffer content
198 static unsigned buf_lenght(const CODE_BUFFER *buf)
200 return buf->next - buf->start;
204 * Returns the current content of a buffer.
206 * @param buf the code buffer
208 * @return the start address of the buffer content
210 static const BYTE *buf_content(const CODE_BUFFER *buf)
216 * Returns the hash value of a buffer.
218 * @param buf the code buffer
220 * @return the hash value of the buffer content
222 static unsigned buf_hash(const CODE_BUFFER *buf)
228 * Returns non-zero if a buffer overrun has occurred.
230 * @param buf the code buffer
232 static unsigned buf_overrun(const CODE_BUFFER *buf)
238 * Returns the next byte from the buffer WITHOUT dropping.
240 * @param buf the code buffer
242 * @return the next byte from the code buffer
244 static inline BYTE look_byte(CODE_BUFFER *buf)
246 if (buf->next < buf->end)
252 * Returns the next byte from the buffer WITH dropping.
254 * @param buf the code buffer
256 * @return the next byte from the code buffer
258 static inline BYTE get_byte(CODE_BUFFER *buf)
260 if (buf->next < buf->end)
265 #define BITS(n) (1 << (n))
268 * Put a 32bit value into the buffer.
270 * @param buf the code buffer
271 * @param code the code to be written into the buffer
273 static void put_code(CODE_BUFFER *buf, unsigned code)
275 if (code < BITS(7)) {
276 put_byte(buf, VLC_7BIT | code);
277 } else if (code < BITS(6 + 8)) {
278 put_byte(buf, VLC_14BIT | (code >> 8));
280 } else if (code < BITS(5 + 8 + 8)) {
281 put_byte(buf, VLC_21BIT | (code >> 16));
282 put_byte(buf, code >> 8);
284 } else if (code < BITS(4 + 8 + 8 + 8)) {
285 put_byte(buf, VLC_28BIT | (code >> 24));
286 put_byte(buf, code >> 16);
287 put_byte(buf, code >> 8);
290 put_byte(buf, VLC_32BIT);
291 put_byte(buf, code >> 24);
292 put_byte(buf, code >> 16);
293 put_byte(buf, code >> 8);
298 #define BIT_MASK(n) ((1 << (n)) - 1)
301 * Get 32 bit from the buffer.
303 * @param buf the code buffer
305 * @return next 32bit value from the code buffer
307 static unsigned get_code(CODE_BUFFER *buf)
309 unsigned code = get_byte(buf);
311 if (code < VLC_14BIT)
313 if (code < VLC_21BIT)
314 return ((code & BIT_MASK(6)) << 8) | get_byte(buf);
315 if (code < VLC_28BIT) {
316 code = ((code & BIT_MASK(5)) << 16) | (get_byte(buf) << 8);
317 code |= get_byte(buf);
320 if (code < VLC_32BIT) {
321 code = ((code & BIT_MASK(4)) << 24) | (get_byte(buf) << 16);
322 code |= get_byte(buf) << 8;
323 code |= get_byte(buf);
326 if (code == VLC_32BIT) {
327 code = get_byte(buf) << 24;
328 code |= get_byte(buf) << 16;
329 code |= get_byte(buf) << 8;
330 code |= get_byte(buf);
333 /* should not happen */
334 assert(0 && "Wrong code in buffer");
340 * Put a tag into the buffer.
342 * @param buf the code buffer
343 * @param tag the tag to write to the code buffer
345 static void put_tag(CODE_BUFFER *buf, BYTE tag)
347 assert(tag >= VLC_TAG_FIRST && "invalid tag");
353 * Returns the next tag or zero if the next code isn't a tag.
355 * @param buf the code buffer
357 * @return the next tag in the code buffer
359 static BYTE next_tag(CODE_BUFFER *buf)
361 BYTE b = look_byte(buf);
363 if (b >= VLC_TAG_FIRST)
364 return get_byte(buf);
369 * An Environment for the pattern encoder.
371 typedef struct _codec_enc_t {
372 CODE_BUFFER *buf; /**< The current code buffer. */
373 set *id_set; /**< A set containing all already seen Firm nodes. */
374 unsigned curr_id; /**< The current node id. */
375 unsigned options; /**< The encoding options. */
376 pattern_dumper_t *dmp; /**< The dumper for the decoder. */
382 typedef struct _addr_entry_t {
383 void *addr; /**< the address */
384 unsigned id; /**< associated ID */
388 * Compare two addresses.
390 static int addr_cmp(const void *p1, const void *p2, size_t size)
392 const addr_entry_t *e1 = p1;
393 const addr_entry_t *e2 = p2;
396 return e1->addr != e2->addr;
400 * Encodes an IR-node, recursive worker.
402 * @return reached depth
404 static int _encode_node(ir_node *node, int max_depth, codec_env_t *env)
406 addr_entry_t entry, *r_entry;
411 ir_opcode code = get_irn_opcode(node);
413 /* insert the node into our ID map */
415 entry.id = env->curr_id;
417 s_entry = set_hinsert(env->id_set, &entry, sizeof(entry), HASH_PTR(node));
418 r_entry = (addr_entry_t *)s_entry->dptr;
420 if (r_entry->id != env->curr_id) {
421 /* already in the map, add an REF */
422 put_tag(env->buf, VLC_TAG_REF);
423 put_code(env->buf, r_entry->id);
427 /* a new entry, proceed */
431 put_code(env->buf, (unsigned)code);
433 /* do we need the mode ? */
434 if (env->options & OPT_WITH_MODE) {
435 ir_mode *mode = get_irn_mode(node);
438 put_code(env->buf, stat_find_mode_index(mode));
440 put_tag(env->buf, VLC_TAG_EMPTY);
443 /* do we need integer constants */
444 if (env->options & OPT_WITH_ICONST) {
445 if (code == iro_Const) {
446 tarval *tv = get_Const_tarval(node);
448 if (tarval_is_long(tv)) {
449 long v = get_tarval_long(tv);
451 put_tag(env->buf, VLC_TAG_ICONST);
452 put_code(env->buf, v);
459 if (max_depth <= 0) {
460 put_code(env->buf, 0);
464 preds = get_irn_arity(node);
465 put_code(env->buf, preds);
468 if (is_op_commutative(get_irn_op(node))) {
469 ir_node *l = get_binop_left(node);
470 ir_node *r = get_binop_right(node);
471 int opcode_diff = (int)get_irn_opcode(l) - (int)get_irn_opcode(r);
473 if (opcode_diff > 0) {
477 } else if (opcode_diff == 0 && l != r) {
478 /* Both nodes have the same opcode, but are different.
479 Need a better method here to decide which goes to the left side. */
482 /* special handling for commutative operators */
483 depth = _encode_node(l, max_depth, env);
486 depth = _encode_node(r, max_depth, env);
490 for (i = 0; i < preds; ++i) {
491 ir_node *n = get_irn_n(node, i);
493 depth = _encode_node(n, max_depth, env);
502 * Encode a DAG starting by the IR-node node.
504 * @param node The root node of the graph
505 * @param buf The code buffer to store the bitstring in
506 * @param max_depth The maximum depth for descending
508 * @return The depth of the encoded graph (without cycles)
510 static int encode_node(ir_node *node, CODE_BUFFER *buf, int max_depth)
515 /* initialize the encoder environment */
517 env.curr_id = 1; /* 0 is used for special purpose */
518 env.options = status->options;
521 if (env.options & OPT_ENC_DAG)
522 env.id_set = new_set(addr_cmp, 32);
526 /* encode options if any for the decoder */
528 put_tag(buf, VLC_TAG_OPTION);
529 put_code(buf, env.options);
532 res = _encode_node(node, max_depth, &env);
534 if (env.id_set != NULL)
537 return max_depth - res;
541 * Decode an IR-node, recursive walker.
543 static void _decode_node(unsigned parent, int position, codec_env_t *env)
547 unsigned mode_code = 0;
551 code = next_tag(env->buf);
552 if (code == VLC_TAG_REF) { /* it's a REF */
553 code = get_code(env->buf);
559 * the mode of a Firm edge can be either computed from its target or
560 * from its source and position. We must take the second approach because
561 * we don't know the target here, it's a ref.
563 pattern_dump_edge(env->dmp, code, parent, position, edge_mode);
566 /* dump the node ref */
567 pattern_dump_ref(env->dmp, code);
573 op_code = get_code(env->buf);
575 /* get the mode if encoded */
576 if (env->options & OPT_WITH_MODE) {
577 if (next_tag(env->buf) != VLC_TAG_EMPTY) {
578 mode_code = get_code(env->buf);
582 /* check, if a ICONST attribute is given */
583 if (next_tag(env->buf) == VLC_TAG_ICONST) {
584 iconst = get_code(env->buf);
593 * the mode of a Firm edge can be either computed from its target or
594 * from its source and position. We take the second approach because
595 * we need it anyway for ref's.
597 pattern_dump_edge(env->dmp, env->curr_id, parent, position, edge_mode);
601 parent = env->curr_id;
602 pattern_dump_node(env->dmp, parent, op_code, mode_code, attr);
604 /* ok, we have a new ID */
607 code = next_tag(env->buf);
608 if (code != VLC_TAG_END) {
609 /* more info, do recursion */
612 preds = get_code(env->buf);
614 pattern_start_children(env->dmp, parent);
615 for (i = 0; i < preds; ++i) {
616 _decode_node(parent, i, env);
618 pattern_finish_children(env->dmp, parent);
626 static void decode_node(BYTE *b, unsigned len, pattern_dumper_t *dump)
630 unsigned code, options = 0;
632 init_buf(&buf, b, len);
635 env.curr_id = 1; /* 0 is used for special purpose */
639 code = next_tag(&buf);
640 if (code == VLC_TAG_OPTION) {
641 options = get_code(&buf);
643 env.options = options;
645 _decode_node(0, 0, &env);
649 * The environment for the pattern calculation.
651 typedef struct _pattern_env {
652 int max_depth; /**< maximum depth for pattern generation. */
656 * Returns the associates pattern_entry_t for a CODE_BUF.
658 * @param buf the code buffer
659 * @param set the hash table containing all pattern entries
661 * @return the associated pattern_entry_t for the given code buffer
663 * If the code content was never seen before, a new pattern_entry is created
666 static pattern_entry_t *pattern_get_entry(CODE_BUFFER *buf, pset *set)
668 pattern_entry_t *key, *elem;
669 unsigned len = buf_lenght(buf);
672 key = OALLOCF(&status->obst, pattern_entry_t, buf, len);
674 memcpy(key->buf, buf_content(buf), len);
676 hash = buf_hash(buf);
678 elem = pset_find(set, key, hash);
680 obstack_free(&status->obst, key);
684 cnt_clr(&key->count);
685 return pset_insert(set, key, hash);
686 } /* pattern_get_entry */
689 * Increase the count for a pattern.
691 * @param buf the code buffer containing the pattern
692 * @param depth the pattern depth
694 * @note Single node patterns are ignored
696 static void count_pattern(CODE_BUFFER *buf, int depth)
698 pattern_entry_t *entry;
700 /* ignore single node pattern (i.e. constants) */
702 entry = pattern_get_entry(buf, status->pattern_hash);
705 cnt_inc(&entry->count);
707 } /* count_pattern */
710 * Pre-walker for nodes pattern calculation.
712 static void calc_nodes_pattern(ir_node *node, void *ctx)
714 pattern_env_t *env = ctx;
715 BYTE buffer[PATTERN_STORE_SIZE];
719 init_buf(&buf, buffer, sizeof(buffer));
720 depth = encode_node(node, &buf, env->max_depth);
722 if (buf_overrun(&buf)) {
723 fprintf(stderr, "Pattern store: buffer overrun at size %u. Pattern ignored.\n", (unsigned) sizeof(buffer));
725 count_pattern(&buf, depth);
726 } /* calc_nodes_pattern */
729 * Store all collected patterns.
731 * @param fname filename for storage
733 static void store_pattern(const char *fname)
736 pattern_entry_t *entry;
737 int i, count = pset_count(status->pattern_hash);
742 f = fopen(fname, "wb");
748 fwrite("FPS1", 4, 1, f);
749 fwrite(&count, sizeof(count), 1, f);
751 for (i = 0, entry = pset_first(status->pattern_hash);
753 entry = pset_next(status->pattern_hash), ++i) {
754 fwrite(entry, offsetof(pattern_entry_t, buf) + entry->len, 1, f);
757 } /* store_pattern */
760 * Read collected patterns from a file.
762 * @param fname filename
764 static HASH_MAP(pattern_entry_t) *read_pattern(const char *fname)
767 pattern_entry_t *entry, tmp;
771 HASH_MAP(pattern_entry_t) *pattern_hash = new_pset(pattern_cmp, 8);
772 BYTE buffer[PATTERN_STORE_SIZE];
775 f = fopen(fname, "rb");
781 fread(magic, 4, 1, f);
783 fread(&count, sizeof(count), 1, f);
784 if (memcmp(magic, "FPS1", 4) != 0 || count <= 0) {
785 fprintf(stderr, "Error: %s is not a Firm pattern store. Ignored.\n", fname);
790 /* read all pattern entries and put them into the hash table. */
791 for (i = 0; i < count; ++i) {
792 init_buf(&buf, buffer, sizeof(buffer));
793 fread(&tmp, offsetof(pattern_entry_t, buf), 1, f);
794 for (j = 0; j < tmp.len; ++j)
795 put_byte(&buf, fgetc(f));
796 entry = pattern_get_entry(&buf, pattern_hash);
797 memcpy(&entry->count, &tmp.count, sizeof(entry->count));
801 printf("Read %d pattern from %s\n", count, fname);
802 assert(pset_count(pattern_hash) == count);
808 * Write the collected patterns to a VCG file for inspection.
810 * @param fname name of the VCG file to create
812 static void pattern_output(const char *fname)
814 pattern_entry_t *entry;
815 pattern_entry_t **pattern_arr;
816 pattern_dumper_t *dump;
817 int i, count = pset_count(status->pattern_hash);
819 printf("\n%d pattern detected\n", count);
824 /* creates a dumper */
825 dump = new_vcg_dumper(fname, 100);
827 pattern_arr = XMALLOCN(pattern_entry_t*, count);
828 for (i = 0, entry = pset_first(status->pattern_hash);
830 entry = pset_next(status->pattern_hash), ++i) {
831 pattern_arr[i] = entry;
837 qsort(pattern_arr, count, sizeof(*pattern_arr), pattern_count_cmp);
839 for (i = 0; i < count; ++i) {
840 entry = pattern_arr[i];
841 if (cnt_to_uint(&entry->count) < status->bound)
845 pattern_dump_new_pattern(dump, &entry->count);
846 decode_node(entry->buf, entry->len, dump);
847 pattern_dump_finish_pattern(dump);
852 } /* pattern_output */
855 * Calculates the pattern history.
857 void stat_calc_pattern_history(ir_graph *irg)
862 if (! status->enable)
865 /* do NOT count the const code IRG */
866 if (irg == get_const_code_irg())
869 for (i = status->min_depth; i <= status->max_depth; ++i) {
871 irg_walk_graph(irg, calc_nodes_pattern, NULL, &env);
873 } /* stat_calc_pattern_history */
876 * Initializes the pattern history.
878 void stat_init_pattern_history(int enable)
880 HASH_MAP(pattern_entry_t) *pattern_hash = NULL;
882 status->enable = enable;
887 status->options = /* OPT_WITH_MODE | */ OPT_ENC_DAG | OPT_WITH_ICONST | OPT_PERSIST_PATTERN;
888 status->min_depth = 3;
889 status->max_depth = 5;
891 obstack_init(&status->obst);
893 /* create the hash-table */
894 if (status->options & OPT_PERSIST_PATTERN)
895 pattern_hash = read_pattern("pattern.fps");
896 if (pattern_hash == NULL)
897 pattern_hash = new_pset(pattern_cmp, 8);
898 status->pattern_hash = pattern_hash;
899 } /* stat_init_pattern_history */
902 * Finish the pattern history.
904 void stat_finish_pattern_history(const char *fname)
907 if (! status->enable)
910 store_pattern("pattern.fps");
911 pattern_output("pattern.vcg");
913 del_pset(status->pattern_hash);
914 obstack_free(&status->obst, NULL);
917 } /* stat_finish_pattern_history */
919 #endif /* FIRM_STATISTICS */