2 * Copyright (C) 1995-2007 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
39 #include "pattern_dmp.h"
43 * just be make some things clear :-), the
46 #define HASH_MAP(type) pset_##type
48 typedef pset pset_pattern_entry_t;
50 typedef unsigned char BYTE;
52 /** Maximum size of the pattern store. */
53 #define PATTERN_STORE_SIZE 2048
59 typedef struct _code_buf_t {
60 BYTE *next; /**< Next byte address to be written. */
61 BYTE *end; /**< End address of the buffer. */
62 BYTE *start; /**< Start address of the buffer. */
63 unsigned hash; /**< The hash value for the buffer content. */
64 unsigned overrun; /**< flag set if the buffer was overrun */
71 VLC_7BIT = 0x00, /**< 8 bit code, carrying 7 bits payload */
72 VLC_14BIT = 0x80, /**< 16 bit code, carrying 14 bits payload */
73 VLC_21BIT = 0xC0, /**< 24 bit code, carrying 21 bits payload */
74 VLC_28BIT = 0xE0, /**< 32 bit code, carrying 28 bits payload */
75 VLC_32BIT = 0xF0, /**< 40 bit code, carrying 32 bits payload */
77 VLC_TAG_FIRST = 0xF1, /**< First possible tag value. */
78 VLC_TAG_ICONST = 0xFB, /**< Encodes an integer constant. */
79 VLC_TAG_EMPTY = 0xFC, /**< Encodes an empty entity. */
80 VLC_TAG_OPTION = 0xFD, /**< Options exists. */
81 VLC_TAG_REF = 0xFE, /**< Special tag, next code is an ID. */
82 VLC_TAG_END = 0xFF, /**< End tag. */
86 * An entry for holding one pattern.
88 typedef struct _pattern_entry_t {
89 counter_t count; /**< Amount of pattern occurance. */
90 unsigned len; /**< The length of the VLC encoded buffer. */
91 BYTE buf[1]; /**< The buffer containing the VLC encoded pattern. */
95 * Current options for the pattern matcher.
98 OPT_WITH_MODE = 0x00000001, /**< use modes */
99 OPT_ENC_DAG = 0x00000002, /**< encode DAGs, not terms */
100 OPT_WITH_ICONST = 0x00000004, /**< encode integer constants */
101 OPT_PERSIST_PATTERN = 0x00000008, /**< persistent pattern hash */
108 typedef struct _pattern_info_t {
109 int enable; /**< If non-zero, this module is enabled. */
110 struct obstack obst; /**< An obstack containing the counters. */
111 HASH_MAP(pattern_entry_t) *pattern_hash; /**< A hash map containing the pattern. */
112 unsigned bound; /**< Lowest value for pattern output. */
113 unsigned options; /**< Current option mask. */
114 unsigned min_depth; /**< Minimum pattern depth. */
115 unsigned max_depth; /**< Maximum pattern depth. */
121 static pattern_info_t _status, *status = &_status;
124 * Compare two pattern for its occurance counter.
126 static int pattern_count_cmp(const void *elt, const void *key) {
129 pattern_entry_t **e1 = (pattern_entry_t **)elt;
130 pattern_entry_t **e2 = (pattern_entry_t **)key;
132 /* we want it sorted in descending order */
133 cmp = cnt_cmp(&(*e2)->count, &(*e1)->count);
136 } /* pattern_count_cmp */
139 * Compare two pattern for its pattern hash.
141 static int pattern_cmp(const void *elt, const void *key) {
142 const pattern_entry_t *e1 = elt;
143 const pattern_entry_t *e2 = key;
144 int diff = e1->len - e2->len;
149 return memcmp(e1->buf, e2->buf, e1->len);
153 * Initialize a code buffer.
155 * @param buf the code buffer
156 * @param data a buffer address
157 * @param len the length of the data buffer
159 static void init_buf(CODE_BUFFER *buf, BYTE *data, unsigned len) {
162 buf->end = data + len;
163 buf->hash = 0x2BAD4; /* An arbitrary seed. */
168 * Put a byte into the buffer.
170 * @param buf the code buffer
171 * @param byte the byte to write
173 * The hash value for the buffer content is updated.
175 static INLINE void put_byte(CODE_BUFFER *buf, BYTE byte) {
176 if (buf->next < buf->end) {
178 buf->hash = (buf->hash * 9) ^ byte;
185 * Returns the current length of a buffer.
187 * @param buf the code buffer
189 * @return the length of the buffer content
191 static unsigned buf_lenght(const CODE_BUFFER *buf) {
192 return buf->next - buf->start;
196 * Returns the current content of a buffer.
198 * @param buf the code buffer
200 * @return the start address of the buffer content
202 static const BYTE *buf_content(const CODE_BUFFER *buf) {
207 * Returns the hash value of a buffer.
209 * @param buf the code buffer
211 * @return the hash value of the buffer content
213 static unsigned buf_hash(const CODE_BUFFER *buf) {
218 * Returns non-zero if a buffer overrun has occurred.
220 * @param buf the code buffer
222 static unsigned buf_overrun(const CODE_BUFFER *buf) {
227 * Returns the next byte from the buffer WITHOUT dropping.
229 * @param buf the code buffer
231 * @return the next byte from the code buffer
233 static INLINE BYTE look_byte(CODE_BUFFER *buf) {
234 if (buf->next < buf->end)
240 * Returns the next byte from the buffer WITH dropping.
242 * @param buf the code buffer
244 * @return the next byte from the code buffer
246 static INLINE BYTE get_byte(CODE_BUFFER *buf) {
247 if (buf->next < buf->end)
252 #define BITS(n) (1 << (n))
255 * Put a 32bit value into the buffer.
257 * @param buf the code buffer
258 * @param code the code to be written into the buffer
260 static void put_code(CODE_BUFFER *buf, unsigned code) {
261 if (code < BITS(7)) {
262 put_byte(buf, VLC_7BIT | code);
263 } else if (code < BITS(6 + 8)) {
264 put_byte(buf, VLC_14BIT | (code >> 8));
266 } else if (code < BITS(5 + 8 + 8)) {
267 put_byte(buf, VLC_21BIT | (code >> 16));
268 put_byte(buf, code >> 8);
270 } else if (code < BITS(4 + 8 + 8 + 8)) {
271 put_byte(buf, VLC_28BIT | (code >> 24));
272 put_byte(buf, code >> 16);
273 put_byte(buf, code >> 8);
276 put_byte(buf, VLC_32BIT);
277 put_byte(buf, code >> 24);
278 put_byte(buf, code >> 16);
279 put_byte(buf, code >> 8);
284 #define BIT_MASK(n) ((1 << (n)) - 1)
287 * Get 32 bit from the buffer.
289 * @param buf the code buffer
291 * @return next 32bit value from the code buffer
293 static unsigned get_code(CODE_BUFFER *buf) {
294 unsigned code = get_byte(buf);
296 if (code < VLC_14BIT)
298 if (code < VLC_21BIT)
299 return ((code & BIT_MASK(6)) << 8) | get_byte(buf);
300 if (code < VLC_28BIT) {
301 code = ((code & BIT_MASK(5)) << 16) | (get_byte(buf) << 8);
302 code |= get_byte(buf);
305 if (code < VLC_32BIT) {
306 code = ((code & BIT_MASK(4)) << 24) | (get_byte(buf) << 16);
307 code |= get_byte(buf) << 8;
308 code |= get_byte(buf);
311 if (code == VLC_32BIT) {
312 code = get_byte(buf) << 24;
313 code |= get_byte(buf) << 16;
314 code |= get_byte(buf) << 8;
315 code |= get_byte(buf);
318 /* should not happen */
319 assert(0 && "Wrong code in buffer");
325 * Put a tag into the buffer.
327 * @param buf the code buffer
328 * @param tag the tag to write to the code buffer
330 static void put_tag(CODE_BUFFER *buf, BYTE tag) {
331 assert(tag >= VLC_TAG_FIRST && "invalid tag");
337 * Returns the next tag or zero if the next code isn't a tag.
339 * @param buf the code buffer
341 * @return the next tag in the code buffer
343 static BYTE next_tag(CODE_BUFFER *buf) {
344 BYTE b = look_byte(buf);
346 if (b >= VLC_TAG_FIRST)
347 return get_byte(buf);
352 * An Environment for the pattern encoder.
354 typedef struct _codec_enc_t {
355 CODE_BUFFER *buf; /**< The current code buffer. */
356 set *id_set; /**< A set containing all already seen Firm nodes. */
357 unsigned curr_id; /**< The current node id. */
358 unsigned options; /**< The encoding options. */
359 pattern_dumper_t *dmp; /**< The dumper for the decoder. */
365 typedef struct _addr_entry_t {
366 void *addr; /**< the address */
367 unsigned id; /**< associated ID */
371 * Compare two addresses.
373 static int addr_cmp(const void *p1, const void *p2, size_t size) {
374 const addr_entry_t *e1 = p1;
375 const addr_entry_t *e2 = p2;
377 return e1->addr != e2->addr;
381 * Encodes an IR-node, recursive worker.
383 * @return reached depth
385 static int _encode_node(ir_node *node, int max_depth, codec_env_t *env) {
386 addr_entry_t entry, *r_entry;
391 ir_opcode code = get_irn_opcode(node);
393 /* insert the node into our ID map */
395 entry.id = env->curr_id;
397 s_entry = set_hinsert(env->id_set, &entry, sizeof(entry), HASH_PTR(node));
398 r_entry = (addr_entry_t *)s_entry->dptr;
400 if (r_entry->id != env->curr_id) {
401 /* already in the map, add an REF */
402 put_tag(env->buf, VLC_TAG_REF);
403 put_code(env->buf, r_entry->id);
407 /* a new entry, proceed */
411 put_code(env->buf, (unsigned)code);
413 /* do we need the mode ? */
414 if (env->options & OPT_WITH_MODE) {
415 ir_mode *mode = get_irn_mode(node);
418 /* FIXME: not 64bit save */
419 put_code(env->buf, (unsigned)mode);
421 put_tag(env->buf, VLC_TAG_EMPTY);
424 /* do we need integer constants */
425 if (env->options & OPT_WITH_ICONST) {
426 if (code == iro_Const) {
427 tarval *tv = get_Const_tarval(node);
429 if (tarval_is_long(tv)) {
430 long v = get_tarval_long(tv);
432 put_tag(env->buf, VLC_TAG_ICONST);
433 put_code(env->buf, v);
440 if (max_depth <= 0) {
441 put_code(env->buf, 0);
445 preds = get_irn_arity(node);
446 put_code(env->buf, preds);
449 if (is_op_commutative(get_irn_op(node))) {
450 ir_node *l = get_binop_left(node);
451 ir_node *r = get_binop_right(node);
452 int opcode_diff = (int)get_irn_opcode(l) - (int)get_irn_opcode(r);
454 if (opcode_diff > 0) {
458 } else if (opcode_diff == 0 && l != r) {
459 /* Both nodes have the same opcode, but are different.
460 Need a better method here to decide which goes to the left side. */
463 /* special handling for commutative operators */
464 depth = _encode_node(l, max_depth, env);
467 depth = _encode_node(r, max_depth, env);
471 for (i = 0; i < preds; ++i) {
472 ir_node *n = get_irn_n(node, i);
474 depth = _encode_node(n, max_depth, env);
483 * Encode a DAG starting by the IR-node node.
485 * @param node The root node of the graph
486 * @param buf The code buffer to store the bitstring in
487 * @param max_depth The maximum depth for descending
489 * @return The depth of the encoded graph (without cycles)
491 static int encode_node(ir_node *node, CODE_BUFFER *buf, int max_depth) {
495 /* initialize the encoder environment */
497 env.curr_id = 1; /* 0 is used for special purpose */
498 env.options = status->options;
501 if (env.options & OPT_ENC_DAG)
502 env.id_set = new_set(addr_cmp, 32);
506 /* encode options if any for the decoder */
508 put_tag(buf, VLC_TAG_OPTION);
509 put_code(buf, env.options);
512 res = _encode_node(node, max_depth, &env);
514 if (env.id_set != NULL)
517 return max_depth - res;
521 * Decode an IR-node, recursive walker.
523 static void _decode_node(unsigned parent, int position, codec_env_t *env) {
526 unsigned mode_code = 0;
530 code = next_tag(env->buf);
531 if (code == VLC_TAG_REF) { /* it's a REF */
532 code = get_code(env->buf);
538 * the mode of a Firm edge can be either computed from its target or
539 * from its source and position. We must take the second approach because
540 * we don't know the target here, it's a ref.
542 pattern_dump_edge(env->dmp, code, parent, position, edge_mode);
545 /* dump the node ref */
546 pattern_dump_ref(env->dmp, code);
552 op_code = get_code(env->buf);
554 /* get the mode if encoded */
555 if (env->options & OPT_WITH_MODE) {
556 if (next_tag(env->buf) != VLC_TAG_EMPTY) {
557 mode_code = get_code(env->buf);
561 /* check, if a ICONST attribute is given */
562 if (next_tag(env->buf) == VLC_TAG_ICONST) {
563 iconst = get_code(env->buf);
572 * the mode of a Firm edge can be either computed from its target or
573 * from its source and position. We take the second approach because
574 * we need it anyway for ref's.
576 pattern_dump_edge(env->dmp, env->curr_id, parent, position, edge_mode);
580 parent = env->curr_id;
581 pattern_dump_node(env->dmp, parent, op_code, mode_code, attr);
583 /* ok, we have a new ID */
586 code = next_tag(env->buf);
587 if (code != VLC_TAG_END) {
588 /* more info, do recursion */
591 preds = get_code(env->buf);
593 pattern_start_children(env->dmp, parent);
594 for (i = 0; i < preds; ++i) {
595 _decode_node(parent, i, env);
597 pattern_finish_children(env->dmp, parent);
605 static void decode_node(BYTE *b, unsigned len, pattern_dumper_t *dump) {
608 unsigned code, options = 0;
610 init_buf(&buf, b, len);
613 env.curr_id = 1; /* 0 is used for special purpose */
617 code = next_tag(&buf);
618 if (code == VLC_TAG_OPTION) {
619 options = get_code(&buf);
621 env.options = options;
623 _decode_node(0, 0, &env);
627 * The environment for the pattern calculation.
629 typedef struct _pattern_env {
630 int max_depth; /**< maximum depth for pattern generation. */
634 * Returns the associates pattern_entry_t for a CODE_BUF.
636 * @param buf the code buffer
637 * @param set the hash table containing all pattern entries
639 * @return the associated pattern_entry_t for the given code buffer
641 * If the code content was never seen before, a new pattern_entry is created
644 static pattern_entry_t *pattern_get_entry(CODE_BUFFER *buf, pset *set) {
645 pattern_entry_t *key, *elem;
646 unsigned len = buf_lenght(buf);
649 key = obstack_alloc(&status->obst, offsetof(pattern_entry_t, buf) + len);
653 memcpy(key->buf, buf_content(buf), len);
655 hash = buf_hash(buf);
657 elem = pset_find(set, key, hash);
659 obstack_free(&status->obst, key);
663 cnt_clr(&key->count);
664 return pset_insert(set, key, hash);
665 } /* pattern_get_entry */
668 * Increase the count for a pattern.
670 * @param buf the code buffer containing the pattern
671 * @param depth the pattern depth
673 * @note Single node patterns are ignored
675 static void count_pattern(CODE_BUFFER *buf, int depth) {
676 pattern_entry_t *entry;
678 /* ignore single node pattern (i.e. constants) */
680 entry = pattern_get_entry(buf, status->pattern_hash);
683 cnt_inc(&entry->count);
685 } /* count_pattern */
688 * Pre-walker for nodes pattern calculation.
690 static void calc_nodes_pattern(ir_node *node, void *ctx) {
691 pattern_env_t *env = ctx;
692 BYTE buffer[PATTERN_STORE_SIZE];
696 init_buf(&buf, buffer, sizeof(buffer));
697 depth = encode_node(node, &buf, env->max_depth);
699 if (buf_overrun(&buf)) {
700 fprintf(stderr, "Pattern store: buffer overrun at size %d. Pattern ignored.\n", sizeof(buffer));
702 count_pattern(&buf, depth);
703 } /* calc_nodes_pattern */
706 * Store all collected patterns.
708 * @param fname filename for storage
710 static void store_pattern(const char *fname) {
712 pattern_entry_t *entry;
713 int i, count = pset_count(status->pattern_hash);
718 f = fopen(fname, "wb");
724 fwrite("FPS1", 4, 1, f);
725 fwrite(&count, sizeof(count), 1, f);
727 for (i = 0, entry = pset_first(status->pattern_hash);
729 entry = pset_next(status->pattern_hash), ++i) {
730 fwrite(entry, offsetof(pattern_entry_t, buf) + entry->len, 1, f);
733 } /* store_pattern */
736 * Read collected patterns from a file.
738 * @param fname filename
740 static HASH_MAP(pattern_entry_t) *read_pattern(const char *fname) {
742 pattern_entry_t *entry, tmp;
746 HASH_MAP(pattern_entry_t) *pattern_hash = new_pset(pattern_cmp, 8);
747 BYTE buffer[PATTERN_STORE_SIZE];
750 f = fopen(fname, "rb");
756 fread(magic, 4, 1, f);
758 fread(&count, sizeof(count), 1, f);
759 if (memcmp(magic, "FPS1", 4) != 0 || count <= 0) {
760 fprintf(stderr, "Error: %s is not a Firm pattern store. Ignored.\n", fname);
765 /* read all pattern entries and put them into the hash table. */
766 for (i = 0; i < count; ++i) {
767 init_buf(&buf, buffer, sizeof(buffer));
768 fread(&tmp, offsetof(pattern_entry_t, buf), 1, f);
769 for (j = 0; j < tmp.len; ++j)
770 put_byte(&buf, fgetc(f));
771 entry = pattern_get_entry(&buf, pattern_hash);
772 memcpy(&entry->count, &tmp.count, sizeof(entry->count));
776 printf("Read %d pattern from %s\n", count, fname);
777 assert(pset_count(pattern_hash) == count);
783 * Write the collected patterns to a VCG file for inspection.
785 * @param fname name of the VCG file to create
787 static void pattern_output(const char *fname) {
788 pattern_entry_t *entry;
789 pattern_entry_t **pattern_arr;
790 pattern_dumper_t *dump;
791 int i, count = pset_count(status->pattern_hash);
793 printf("\n%d pattern detected\n", count);
798 /* creates a dumper */
799 dump = new_vcg_dumper(fname, 100);
801 pattern_arr = xmalloc(sizeof(*pattern_arr) * count);
802 for (i = 0, entry = pset_first(status->pattern_hash);
804 entry = pset_next(status->pattern_hash), ++i) {
805 pattern_arr[i] = entry;
811 qsort(pattern_arr, count, sizeof(*pattern_arr), pattern_count_cmp);
813 for (i = 0; i < count; ++i) {
814 entry = pattern_arr[i];
815 if (cnt_to_uint(&entry->count) < status->bound)
819 pattern_dump_new_pattern(dump, &entry->count);
820 decode_node(entry->buf, entry->len, dump);
821 pattern_dump_finish_pattern(dump);
826 } /* pattern_output */
829 * Calculates the pattern history.
831 void stat_calc_pattern_history(ir_graph *irg) {
835 if (! status->enable)
838 /* do NOT count the const code IRG */
839 if (irg == get_const_code_irg())
842 for (i = status->min_depth; i <= status->max_depth; ++i) {
844 irg_walk_graph(irg, calc_nodes_pattern, NULL, &env);
846 } /* stat_calc_pattern_history */
849 * Initializes the pattern history.
851 void stat_init_pattern_history(int enable) {
852 HASH_MAP(pattern_entry_t) *pattern_hash = NULL;
854 status->enable = enable;
859 status->options = /* OPT_WITH_MODE | */ OPT_ENC_DAG | OPT_WITH_ICONST | OPT_PERSIST_PATTERN;
860 status->min_depth = 3;
861 status->max_depth = 5;
863 obstack_init(&status->obst);
865 /* create the hash-table */
866 if (status->options & OPT_PERSIST_PATTERN)
867 pattern_hash = read_pattern("pattern.fps");
868 if (pattern_hash == NULL)
869 pattern_hash = new_pset(pattern_cmp, 8);
870 status->pattern_hash = pattern_hash;
871 } /* stat_init_pattern_history */
874 * Finish the pattern history.
876 void stat_finish_pattern_history(const char *fname) {
877 if (! status->enable)
880 store_pattern("pattern.fps");
881 pattern_output("pattern.vcg");
883 del_pset(status->pattern_hash);
884 obstack_free(&status->obst, NULL);
887 } /* stat_finish_pattern_history */