2 * Copyright (C) 1995-2011 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 #include "lc_printf.h"
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 size_t 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 = (const pattern_entry_t*)elt;
148 const pattern_entry_t *e2 = (const pattern_entry_t*)key;
150 if (e1->len == e2->len)
151 return memcmp(e1->buf, e2->buf, e1->len);
153 return e1->len < e2->len ? -1 : +1;
157 * Initialize a code buffer.
159 * @param buf the code buffer
160 * @param data a buffer address
161 * @param len the length of the data buffer
163 static void init_buf(CODE_BUFFER *buf, BYTE *data, size_t len)
167 buf->end = data + len;
168 buf->hash = 0x2BAD4; /* An arbitrary seed. */
173 * Put a byte into the buffer.
175 * @param buf the code buffer
176 * @param byte the byte to write
178 * The hash value for the buffer content is updated.
180 static inline void put_byte(CODE_BUFFER *buf, BYTE byte)
182 if (buf->next < buf->end) {
184 buf->hash = (buf->hash * 9) ^ byte;
191 * Returns the current length of a buffer.
193 * @param buf the code buffer
195 * @return the length of the buffer content
197 static size_t buf_lenght(const CODE_BUFFER *buf)
199 return buf->next - buf->start;
203 * Returns the current content of a buffer.
205 * @param buf the code buffer
207 * @return the start address of the buffer content
209 static const BYTE *buf_content(const CODE_BUFFER *buf)
215 * Returns the hash value of a buffer.
217 * @param buf the code buffer
219 * @return the hash value of the buffer content
221 static unsigned buf_hash(const CODE_BUFFER *buf)
227 * Returns non-zero if a buffer overrun has occurred.
229 * @param buf the code buffer
231 static unsigned buf_overrun(const CODE_BUFFER *buf)
237 * Returns the next byte from the buffer WITHOUT dropping.
239 * @param buf the code buffer
241 * @return the next byte from the code buffer
243 static inline BYTE look_byte(CODE_BUFFER *buf)
245 if (buf->next < buf->end)
251 * Returns the next byte from the buffer WITH dropping.
253 * @param buf the code buffer
255 * @return the next byte from the code buffer
257 static inline BYTE get_byte(CODE_BUFFER *buf)
259 if (buf->next < buf->end)
264 #define BITS(n) (1 << (n))
267 * Put a 32bit value into the buffer.
269 * @param buf the code buffer
270 * @param code the code to be written into the buffer
272 static void put_code(CODE_BUFFER *buf, unsigned code)
274 if (code < BITS(7)) {
275 put_byte(buf, VLC_7BIT | code);
276 } else if (code < BITS(6 + 8)) {
277 put_byte(buf, VLC_14BIT | (code >> 8));
279 } else if (code < BITS(5 + 8 + 8)) {
280 put_byte(buf, VLC_21BIT | (code >> 16));
281 put_byte(buf, code >> 8);
283 } else if (code < BITS(4 + 8 + 8 + 8)) {
284 put_byte(buf, VLC_28BIT | (code >> 24));
285 put_byte(buf, code >> 16);
286 put_byte(buf, code >> 8);
289 put_byte(buf, VLC_32BIT);
290 put_byte(buf, code >> 24);
291 put_byte(buf, code >> 16);
292 put_byte(buf, code >> 8);
297 #define BIT_MASK(n) ((1 << (n)) - 1)
300 * Get 32 bit from the buffer.
302 * @param buf the code buffer
304 * @return next 32bit value from the code buffer
306 static unsigned get_code(CODE_BUFFER *buf)
308 unsigned code = get_byte(buf);
310 if (code < VLC_14BIT)
312 if (code < VLC_21BIT)
313 return ((code & BIT_MASK(6)) << 8) | get_byte(buf);
314 if (code < VLC_28BIT) {
315 code = ((code & BIT_MASK(5)) << 16) | (get_byte(buf) << 8);
316 code |= get_byte(buf);
319 if (code < VLC_32BIT) {
320 code = ((code & BIT_MASK(4)) << 24) | (get_byte(buf) << 16);
321 code |= get_byte(buf) << 8;
322 code |= get_byte(buf);
325 if (code == VLC_32BIT) {
326 code = get_byte(buf) << 24;
327 code |= get_byte(buf) << 16;
328 code |= get_byte(buf) << 8;
329 code |= get_byte(buf);
332 /* should not happen */
333 panic("Wrong code in buffer");
337 * Put a tag into the buffer.
339 * @param buf the code buffer
340 * @param tag the tag to write to the code buffer
342 static void put_tag(CODE_BUFFER *buf, BYTE tag)
344 assert(tag >= VLC_TAG_FIRST && "invalid tag");
350 * Returns the next tag or zero if the next code isn't a tag.
352 * @param buf the code buffer
354 * @return the next tag in the code buffer
356 static BYTE next_tag(CODE_BUFFER *buf)
358 BYTE b = look_byte(buf);
360 if (b >= VLC_TAG_FIRST)
361 return get_byte(buf);
366 * An Environment for the pattern encoder.
368 typedef struct codec_enc_t {
369 CODE_BUFFER *buf; /**< The current code buffer. */
370 set *id_set; /**< A set containing all already seen Firm nodes. */
371 unsigned curr_id; /**< The current node id. */
372 unsigned options; /**< The encoding options. */
373 pattern_dumper_t *dmp; /**< The dumper for the decoder. */
379 typedef struct addr_entry_t {
380 void *addr; /**< the address */
381 unsigned id; /**< associated ID */
385 * Compare two addresses.
387 static int addr_cmp(const void *p1, const void *p2, size_t size)
389 const addr_entry_t *e1 = (const addr_entry_t*)p1;
390 const addr_entry_t *e2 = (const addr_entry_t*)p2;
393 return e1->addr != e2->addr;
397 * Return the index of a (existing) mode.
399 static size_t find_mode_index(const ir_mode *mode)
401 size_t i, n = get_irp_n_modes();
403 for (i = 0; i < n; ++i) {
404 if (get_irp_mode(i) == mode)
407 /* should really not happen */
408 assert(!"Cound not find index of mode in find_mode_index()");
413 * Encodes an IR-node, recursive worker.
415 * @return reached depth
417 static int _encode_node(ir_node *node, int max_depth, codec_env_t *env)
419 addr_entry_t entry, *r_entry;
424 unsigned code = get_irn_opcode(node);
426 /* insert the node into our ID map */
428 entry.id = env->curr_id;
430 s_entry = set_hinsert(env->id_set, &entry, sizeof(entry), HASH_PTR(node));
431 r_entry = (addr_entry_t *)s_entry->dptr;
433 if (r_entry->id != env->curr_id) {
434 /* already in the map, add an REF */
435 put_tag(env->buf, VLC_TAG_REF);
436 put_code(env->buf, r_entry->id);
440 /* a new entry, proceed */
444 put_code(env->buf, (unsigned)code);
446 /* do we need the mode ? */
447 if (env->options & OPT_WITH_MODE) {
448 ir_mode *mode = get_irn_mode(node);
451 put_code(env->buf, find_mode_index(mode));
453 put_tag(env->buf, VLC_TAG_EMPTY);
456 /* do we need integer constants */
457 if (env->options & OPT_WITH_ICONST) {
458 if (code == iro_Const) {
459 ir_tarval *tv = get_Const_tarval(node);
461 if (tarval_is_long(tv)) {
462 long v = get_tarval_long(tv);
464 put_tag(env->buf, VLC_TAG_ICONST);
465 put_code(env->buf, v);
472 if (max_depth <= 0) {
473 put_code(env->buf, 0);
477 preds = get_irn_arity(node);
478 put_code(env->buf, preds);
481 if (is_op_commutative(get_irn_op(node))) {
482 ir_node *l = get_binop_left(node);
483 ir_node *r = get_binop_right(node);
484 int opcode_diff = (int)get_irn_opcode(l) - (int)get_irn_opcode(r);
486 if (opcode_diff > 0) {
490 } else if (opcode_diff == 0 && l != r) {
491 /* Both nodes have the same opcode, but are different.
492 Need a better method here to decide which goes to the left side. */
495 /* special handling for commutative operators */
496 depth = _encode_node(l, max_depth, env);
499 depth = _encode_node(r, max_depth, env);
503 for (i = 0; i < preds; ++i) {
504 ir_node *n = get_irn_n(node, i);
506 depth = _encode_node(n, max_depth, env);
515 * Encode a DAG starting by the IR-node node.
517 * @param node The root node of the graph
518 * @param buf The code buffer to store the bitstring in
519 * @param max_depth The maximum depth for descending
521 * @return The depth of the encoded graph (without cycles)
523 static int encode_node(ir_node *node, CODE_BUFFER *buf, int max_depth)
528 /* initialize the encoder environment */
530 env.curr_id = 1; /* 0 is used for special purpose */
531 env.options = status->options;
534 if (env.options & OPT_ENC_DAG)
535 env.id_set = new_set(addr_cmp, 32);
539 /* encode options if any for the decoder */
541 put_tag(buf, VLC_TAG_OPTION);
542 put_code(buf, env.options);
545 res = _encode_node(node, max_depth, &env);
547 if (env.id_set != NULL)
550 return max_depth - res;
554 * Decode an IR-node, recursive walker.
556 static void _decode_node(unsigned parent, int position, codec_env_t *env)
560 unsigned mode_code = 0;
564 code = next_tag(env->buf);
565 if (code == VLC_TAG_REF) { /* it's a REF */
566 code = 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 must take the second approach because
574 * we don't know the target here, it's a ref.
576 pattern_dump_edge(env->dmp, code, parent, position, edge_mode);
579 /* dump the node ref */
580 pattern_dump_ref(env->dmp, code);
586 op_code = get_code(env->buf);
588 /* get the mode if encoded */
589 if (env->options & OPT_WITH_MODE) {
590 if (next_tag(env->buf) != VLC_TAG_EMPTY) {
591 mode_code = get_code(env->buf);
595 /* check, if a ICONST attribute is given */
596 if (next_tag(env->buf) == VLC_TAG_ICONST) {
597 iconst = get_code(env->buf);
606 * the mode of a Firm edge can be either computed from its target or
607 * from its source and position. We take the second approach because
608 * we need it anyway for ref's.
610 pattern_dump_edge(env->dmp, env->curr_id, parent, position, edge_mode);
614 parent = env->curr_id;
615 pattern_dump_node(env->dmp, parent, op_code, mode_code, attr);
617 /* ok, we have a new ID */
620 code = next_tag(env->buf);
621 if (code != VLC_TAG_END) {
622 /* more info, do recursion */
625 preds = get_code(env->buf);
627 pattern_start_children(env->dmp, parent);
628 for (i = 0; i < preds; ++i) {
629 _decode_node(parent, i, env);
631 pattern_finish_children(env->dmp, parent);
639 static void decode_node(BYTE *b, size_t len, pattern_dumper_t *dump)
643 unsigned code, options = 0;
645 init_buf(&buf, b, len);
648 env.curr_id = 1; /* 0 is used for special purpose */
652 code = next_tag(&buf);
653 if (code == VLC_TAG_OPTION) {
654 options = get_code(&buf);
656 env.options = options;
658 _decode_node(0, 0, &env);
662 * The environment for the pattern calculation.
664 typedef struct pattern_env {
665 int max_depth; /**< maximum depth for pattern generation. */
669 * Returns the associates pattern_entry_t for a CODE_BUF.
671 * @param buf the code buffer
672 * @param set the hash table containing all pattern entries
674 * @return the associated pattern_entry_t for the given code buffer
676 * If the code content was never seen before, a new pattern_entry is created
679 static pattern_entry_t *pattern_get_entry(CODE_BUFFER *buf, pset *set)
681 pattern_entry_t *key, *elem;
682 size_t len = buf_lenght(buf);
685 key = OALLOCF(&status->obst, pattern_entry_t, buf, len);
687 memcpy(key->buf, buf_content(buf), len);
689 hash = buf_hash(buf);
691 elem = (pattern_entry_t*)pset_find(set, key, hash);
693 obstack_free(&status->obst, key);
697 cnt_clr(&key->count);
698 return (pattern_entry_t*)pset_insert(set, key, hash);
699 } /* pattern_get_entry */
702 * Increase the count for a pattern.
704 * @param buf the code buffer containing the pattern
705 * @param depth the pattern depth
707 * @note Single node patterns are ignored
709 static void count_pattern(CODE_BUFFER *buf, int depth)
711 pattern_entry_t *entry;
713 /* ignore single node pattern (i.e. constants) */
715 entry = pattern_get_entry(buf, status->pattern_hash);
718 cnt_inc(&entry->count);
720 } /* count_pattern */
723 * Pre-walker for nodes pattern calculation.
725 static void calc_nodes_pattern(ir_node *node, void *ctx)
727 pattern_env_t *env = (pattern_env_t*)ctx;
728 BYTE buffer[PATTERN_STORE_SIZE];
732 init_buf(&buf, buffer, sizeof(buffer));
733 depth = encode_node(node, &buf, env->max_depth);
735 if (buf_overrun(&buf)) {
736 lc_fprintf(stderr, "Pattern store: buffer overrun at size %zu. Pattern ignored.\n", sizeof(buffer));
738 count_pattern(&buf, depth);
739 } /* calc_nodes_pattern */
742 * Store all collected patterns.
744 * @param fname filename for storage
746 static void store_pattern(const char *fname)
749 pattern_entry_t *entry;
750 size_t i, count = pset_count(status->pattern_hash);
755 f = fopen(fname, "wb");
761 fwrite("FPS1", 4, 1, f);
762 fwrite(&count, sizeof(count), 1, f);
764 for (i = 0, entry = (pattern_entry_t*)pset_first(status->pattern_hash);
766 entry = (pattern_entry_t*)pset_next(status->pattern_hash), ++i) {
767 fwrite(entry, offsetof(pattern_entry_t, buf) + entry->len, 1, f);
770 } /* store_pattern */
773 * Read collected patterns from a file.
775 * @param fname filename
777 static HASH_MAP(pattern_entry_t) *read_pattern(const char *fname)
780 pattern_entry_t *entry, tmp;
784 HASH_MAP(pattern_entry_t) *pattern_hash = new_pset(pattern_cmp, 8);
785 BYTE buffer[PATTERN_STORE_SIZE];
788 f = fopen(fname, "rb");
794 fread(magic, 4, 1, f);
796 fread(&count, sizeof(count), 1, f);
797 if (memcmp(magic, "FPS1", 4) != 0 || count <= 0) {
798 fprintf(stderr, "Error: %s is not a Firm pattern store. Ignored.\n", fname);
803 /* read all pattern entries and put them into the hash table. */
804 for (i = 0; i < count; ++i) {
805 init_buf(&buf, buffer, sizeof(buffer));
806 fread(&tmp, offsetof(pattern_entry_t, buf), 1, f);
807 for (j = 0; j < tmp.len; ++j)
808 put_byte(&buf, fgetc(f));
809 entry = pattern_get_entry(&buf, pattern_hash);
810 entry->count = tmp.count;
814 lc_printf("Read %zu pattern from %s\n", count, fname);
815 assert(pset_count(pattern_hash) == count);
821 * Write the collected patterns to a VCG file for inspection.
823 * @param fname name of the VCG file to create
825 static void pattern_output(const char *fname)
827 pattern_entry_t *entry;
828 pattern_entry_t **pattern_arr;
829 pattern_dumper_t *dump;
830 size_t i, count = pset_count(status->pattern_hash);
832 lc_printf("\n%zu pattern detected\n", count);
837 /* creates a dumper */
838 dump = new_vcg_dumper(fname, 100);
840 pattern_arr = XMALLOCN(pattern_entry_t*, count);
841 for (i = 0, entry = (pattern_entry_t*)pset_first(status->pattern_hash);
843 entry = (pattern_entry_t*)pset_next(status->pattern_hash), ++i) {
844 pattern_arr[i] = entry;
850 qsort(pattern_arr, count, sizeof(*pattern_arr), pattern_count_cmp);
852 for (i = 0; i < count; ++i) {
853 entry = pattern_arr[i];
854 if (cnt_to_uint(&entry->count) < status->bound)
858 pattern_dump_new_pattern(dump, &entry->count);
859 decode_node(entry->buf, entry->len, dump);
860 pattern_dump_finish_pattern(dump);
865 } /* pattern_output */
868 * Calculates the pattern history.
870 void stat_calc_pattern_history(ir_graph *irg)
875 if (! status->enable)
878 /* do NOT count the const code IRG */
879 if (irg == get_const_code_irg())
882 for (i = status->min_depth; i <= status->max_depth; ++i) {
884 irg_walk_graph(irg, calc_nodes_pattern, NULL, &env);
886 } /* stat_calc_pattern_history */
889 * Initializes the pattern history.
891 void stat_init_pattern_history(int enable)
893 HASH_MAP(pattern_entry_t) *pattern_hash = NULL;
895 status->enable = enable;
900 status->options = /* OPT_WITH_MODE | */ OPT_ENC_DAG | OPT_WITH_ICONST | OPT_PERSIST_PATTERN;
901 status->min_depth = 3;
902 status->max_depth = 5;
904 obstack_init(&status->obst);
906 /* create the hash-table */
907 if (status->options & OPT_PERSIST_PATTERN)
908 pattern_hash = read_pattern("pattern.fps");
909 if (pattern_hash == NULL)
910 pattern_hash = new_pset(pattern_cmp, 8);
911 status->pattern_hash = pattern_hash;
912 } /* stat_init_pattern_history */
915 * Finish the pattern history.
917 void stat_finish_pattern_history(const char *fname)
920 if (! status->enable)
923 store_pattern("pattern.fps");
924 pattern_output("pattern.vcg");
926 del_pset(status->pattern_hash);
927 obstack_free(&status->obst, NULL);
930 } /* stat_finish_pattern_history */