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
22 * @brief Code instrumentation and execution count profiling.
23 * @author Adam M. Szalkowski
53 #include "besched_t.h"
63 #include "bechordal_t.h"
65 #include "beprofile.h"
67 /** An entry in the id-to-location map */
68 typedef struct loc_entry {
69 ir_entity *fname; /**< the entity holding the file name */
70 unsigned int lineno; /**< line number */
73 typedef struct _block_id_walker_data_t {
74 tarval **array; /**< the entity the holds the block counts */
75 unsigned int id; /**< current block id number */
76 ir_node *symconst; /**< the SymConst representing array */
77 pmap *fname_map; /**< set containing all found filenames */
78 loc_entry *locs; /**< locations */
79 ir_type *tp_char; /**< the character type */
80 unsigned flags; /**< profile flags */
81 } block_id_walker_data_t;
83 typedef struct _execcount_t {
89 cmp_execcount(const void * a, const void * b, size_t size)
91 return ((execcount_t*)a)->block != ((execcount_t*)b)->block;
95 block_counter(ir_node * bb, void * data)
97 unsigned int *count = data;
102 count_blocks(ir_graph * irg)
104 unsigned int count = 0;
106 irg_block_walk_graph(irg, block_counter, NULL, &count);
110 /* keep the execcounts here because they are only read once per compiler run */
111 static set * profile = NULL;
112 static hook_entry_t hook;
115 * Instrument a block with code needed for profiling
118 instrument_block(ir_node * bb, ir_node * address, unsigned int id)
120 ir_graph *irg = get_irn_irg(bb);
121 ir_node *start_block = get_irg_start_block(irg);
122 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
126 * We can't instrument the start and end block as there are no real
127 * instructions in these blocks
129 if(bb == start_block || bb == get_irg_end_block(irg))
132 unknown = new_r_Unknown(irg, mode_M);
133 cnst = new_r_Const_long(irg, start_block, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
134 offset = new_r_Add(irg, bb, address, cnst, mode_P);
135 load = new_r_Load(irg, bb, unknown, offset, mode_Iu);
136 projm = new_r_Proj(irg, bb, load, mode_M, pn_Load_M);
137 proji = new_r_Proj(irg, bb, load, mode_Iu, pn_Load_res);
138 cnst = new_r_Const_long(irg, start_block, mode_Iu, 1);
139 add = new_r_Add(irg, bb, proji, cnst, mode_Iu);
140 store = new_r_Store(irg, bb, projm, offset, add);
141 projm = new_r_Proj(irg, bb, store, mode_M, pn_Store_M);
142 set_irn_link(bb, projm);
143 set_irn_link(projm, load);
146 typedef struct fix_env {
147 ir_node *start_block;
152 * SSA Construction for instrumentation code memory
155 fix_ssa(ir_node * bb, void * data)
159 int arity = get_Block_n_cfgpreds(bb);
161 /* start and end block are not instrumented, skip! */
162 if (bb == env->start_block || bb == env->end_block)
166 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
170 ir_graph *irg = current_ir_graph;
172 NEW_ARR_A(ir_node*, ins, arity);
173 for (n = arity - 1; n >= 0; --n) {
174 ins[n] = get_irn_link(get_Block_cfgpred_block(bb, n));
176 mem = new_r_Phi(irg, bb, arity, ins, mode_M);
178 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
183 * Generates a new irg which calls the initializer
186 * void __firmprof_initializer(void) { __init_firmprof(ent_filename, bblock_id, bblock_counts, n_blocks); }
189 gen_initializer_irg(ir_entity * ent_filename, ir_entity * bblock_id, ir_entity * bblock_counts, int n_blocks)
191 ir_node *start_block;
194 ident *name = new_id_from_str("__firmprof_initializer");
195 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(name, 0, 0));
196 ir_node *ret, *call, *symconst;
199 ident *init_name = new_id_from_str("__init_firmprof");
200 ir_type *init_type = new_type_method(init_name, 4, 0);
201 ir_type *uint, *uintptr, *string;
205 ir_type *empty_frame_type;
207 set_entity_ld_ident(ent, name);
209 uint = new_type_primitive(new_id_from_str("__uint"), mode_Iu);
210 uintptr = new_type_pointer(new_id_from_str("__uintptr"), uint, mode_P);
211 string = new_type_pointer(new_id_from_str("__charptr"), new_type_primitive(new_id_from_str("__char"), mode_Bs), mode_P);
213 set_method_param_type(init_type, 0, string);
214 set_method_param_type(init_type, 1, uintptr);
215 set_method_param_type(init_type, 2, uintptr);
216 set_method_param_type(init_type, 3, uint);
217 init_ent = new_entity(get_glob_type(), init_name, init_type);
218 set_entity_ld_ident(init_ent, init_name);
220 irg = new_ir_graph(ent, 0);
221 empty_frame_type = get_irg_frame_type(irg);
222 set_type_size_bytes(empty_frame_type, 0);
224 bb = get_cur_block();
226 start_block = get_irg_start_block(irg);
228 sym.entity_p = init_ent;
229 symconst = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
231 sym.entity_p = ent_filename;
232 ins[0] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
233 sym.entity_p = bblock_id;
234 ins[1] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
235 sym.entity_p = bblock_counts;
236 ins[2] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
237 ins[3] = new_r_Const_long(irg, start_block, mode_Iu, n_blocks);
239 call = new_r_Call(irg, bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
240 ret = new_r_Return(irg, bb, new_r_Proj(irg, bb, call, mode_M, pn_Call_M_regular), 0, NULL);
243 add_immBlock_pred(get_irg_end_block(irg), ret);
244 mature_immBlock(get_irg_end_block(irg));
246 irg_finalize_cons(irg);
252 * Create the location data for the given debug info.
254 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
257 const char *fname = be_retrieve_dbg_info(dbg, &lineno);
260 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
264 static unsigned nr = 0;
268 int i, len = strlen(fname) + 1;
269 tarval **tarval_string;
271 snprintf(buf, sizeof(buf), "firm_name_arr.%d", nr);
272 arr = new_type_array(new_id_from_str(buf), 1, wd->tp_char);
273 set_array_bounds_int(arr, 0, 0, len);
275 snprintf(buf, sizeof(buf), "__firm_name.%d", nr++);
276 id = new_id_from_str(buf);
277 ent = new_entity(get_glob_type(), id, arr);
278 set_entity_ld_ident(ent, id);
280 pmap_insert(wd->fname_map, (void *)fname, ent);
282 /* initialize file name string constant */
283 tarval_string = alloca(sizeof(*tarval_string) * (len));
284 for (i = 0; i < len; ++i) {
285 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
287 set_entity_variability(ent, variability_constant);
288 set_array_entity_values(ent, tarval_string, len);
292 wd->locs[wd->id].fname = ent;
293 wd->locs[wd->id].lineno = lineno;
295 wd->locs[wd->id].fname = NULL;
296 wd->locs[wd->id].lineno = 0;
301 * Walker: assigns an ID to every block.
302 * Builds the string table
305 block_id_walker(ir_node * bb, void * data)
307 block_id_walker_data_t *wd = data;
309 wd->array[wd->id] = new_tarval_from_long(get_irn_node_nr(bb), mode_Iu);
310 instrument_block(bb, wd->symconst, wd->id);
312 if (wd->flags & profile_with_locations) {
313 dbg_info *dbg = get_irn_dbg_info(bb);
314 create_location_data(dbg, wd);
319 #define IDENT(x) new_id_from_chars(x, sizeof(x) - 1)
322 be_profile_instrument(const char *filename, unsigned flags)
326 ir_entity *bblock_id;
327 ir_entity *bblock_counts;
328 ir_entity *ent_filename;
329 ir_entity *ent_locations = NULL;
330 ir_entity *loc_lineno = NULL;
331 ir_entity *loc_name = NULL;
335 ir_type *string_type;
336 ir_type *character_type;
337 ir_type *loc_type = NULL;
338 ir_type *charptr_type;
340 tarval **tarval_array;
341 tarval **tarval_string;
343 int filename_len = strlen(filename)+1;
345 int align_l, align_n, size;
347 block_id_walker_data_t wd;
350 /* count the number of block first */
351 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
352 ir_graph *irg = get_irp_irg(n);
354 n_blocks += count_blocks(irg);
357 /* create all the necessary types and entities. Note that the
358 types must have a fixed layout, because we already running in the
360 uint_type = new_type_primitive(IDENT("__uint"), mode_Iu);
361 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
362 array_type = new_type_array(IDENT("__block_info_array"), 1, uint_type);
363 set_array_bounds_int(array_type, 0, 0, n_blocks);
365 character_type = new_type_primitive(IDENT("__char"), mode_Bs);
366 string_type = new_type_array(IDENT("__filename"), 1, character_type);
367 set_array_bounds_int(string_type, 0, 0, filename_len);
369 gtp = get_glob_type();
371 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
372 bblock_id = new_entity(gtp, cur_ident, array_type);
373 set_entity_ld_ident(bblock_id, cur_ident);
374 set_entity_variability(bblock_id, variability_initialized);
376 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
377 bblock_counts = new_entity(gtp, cur_ident, array_type);
378 set_entity_ld_ident(bblock_counts, cur_ident);
379 set_entity_variability(bblock_counts, variability_initialized);
381 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
382 ent_filename = new_entity(gtp, cur_ident, string_type);
383 set_entity_ld_ident(ent_filename, cur_ident);
385 if (flags & profile_with_locations) {
386 loc_type = new_type_struct(IDENT("__location"));
387 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
388 align_l = get_type_alignment_bytes(uint_type);
389 size = get_type_size_bytes(uint_type);
390 set_entity_offset(loc_lineno, 0);
392 charptr_type = new_type_pointer(IDENT("__charptr"), character_type, mode_P_data);
393 align_n = get_type_size_bytes(charptr_type);
394 set_type_alignment_bytes(charptr_type, align_n);
395 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
396 size = (size + align_n - 1) & -align_n;
397 set_entity_offset(loc_name, size);
400 if (align_n > align_l)
402 size = (size + align_l - 1) & -align_l;
403 set_type_size_bytes(loc_type, size);
404 set_type_state(loc_type, layout_fixed);
406 loc_type = new_type_array(IDENT("__locarray"), 1, loc_type);
407 set_array_bounds_int(string_type, 0, 0, n_blocks);
409 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
410 ent_locations = new_entity(gtp, cur_ident, loc_type);
411 set_entity_ld_ident(ent_locations, cur_ident);
414 /* initialize count array */
415 NEW_ARR_A(tarval *, tarval_array, n_blocks);
416 tv = get_tarval_null(mode_Iu);
417 for (i = 0; i < n_blocks; ++i) {
418 tarval_array[i] = tv;
420 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
422 /* initialize function name string constant */
423 tarval_string = alloca(sizeof(*tarval_string) * (filename_len));
424 for (i = 0; i < filename_len; ++i) {
425 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
427 set_entity_variability(ent_filename, variability_constant);
428 set_array_entity_values(ent_filename, tarval_string, filename_len);
430 /* initialize block id array and instrument blocks */
431 wd.array = tarval_array;
433 wd.tp_char = character_type;
435 if (flags & profile_with_locations) {
436 wd.fname_map = pmap_create();
437 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
440 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
441 ir_graph *irg = get_irp_irg(n);
443 ir_node *endbb = get_irg_end_block(irg);
446 set_current_ir_graph(irg);
448 /* generate a symbolic constant pointing to the count array */
449 sym.entity_p = bblock_counts;
450 wd.symconst = new_r_SymConst(irg, get_irg_start_block(irg), sym, symconst_addr_ent);
452 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
453 env.start_block = get_irg_start_block(irg);
454 env.end_block = get_irg_end_block(irg);
455 set_irn_link(env.start_block, get_irg_no_mem(irg));
456 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
457 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
458 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
459 ir_node *bb = get_Block_cfgpred_block(endbb, i);
463 switch (get_irn_opcode(node)) {
465 ins[0] = get_irn_link(bb);
466 ins[1] = get_Return_mem(node);
467 sync = new_r_Sync(irg, bb, 2, ins);
468 set_Return_mem(node, sync);
471 ins[0] = get_irn_link(bb);
472 ins[1] = get_Raise_mem(node);
473 sync = new_r_Sync(irg, bb, 2, ins);
474 set_Raise_mem(node, sync);
477 /* a fragile's op exception. There should be another path to End,
479 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
483 set_array_entity_values(bblock_id, tarval_array, n_blocks);
485 if (flags & profile_with_locations) {
486 /* build the initializer for the locations */
487 rem = current_ir_graph;
488 current_ir_graph = get_const_code_irg();
489 ent = get_array_element_entity(loc_type);
490 set_entity_variability(ent_locations, variability_constant);
491 for (i = 0; i < n_blocks; ++i) {
492 compound_graph_path *path;
497 path = new_compound_graph_path(loc_type, 2);
498 set_compound_graph_path_array_index(path, 0, i);
499 set_compound_graph_path_node(path, 0, ent);
500 set_compound_graph_path_node(path, 1, loc_lineno);
501 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
502 add_compound_ent_value_w_path(ent_locations, new_Const(mode_Iu, tv), path);
505 path = new_compound_graph_path(loc_type, 2);
506 set_compound_graph_path_array_index(path, 0, i);
507 set_compound_graph_path_node(path, 0, ent);
508 set_compound_graph_path_node(path, 1, loc_name);
509 if (wd.locs[i].fname) {
510 sym.entity_p = wd.locs[i].fname;
511 n = new_SymConst(sym, symconst_addr_ent);
513 n = new_Const(mode_P_data, get_mode_null(mode_P_data));
515 add_compound_ent_value_w_path(ent_locations, n, path);
517 pmap_destroy(wd.fname_map);
519 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
523 profile_node_info(void *ctx, FILE *f, const ir_node *irn)
526 fprintf(f, "profiled execution count: %u\n", be_profile_get_block_execcount(irn));
531 register_vcg_hook(void)
533 memset(&hook, 0, sizeof(hook));
534 hook.hook._hook_node_info = profile_node_info;
535 register_hook(hook_node_info, &hook);
539 unregister_vcg_hook(void)
541 unregister_hook(hook_node_info, &hook);
545 * Reads the corresponding profile info file if it exists and returns a
546 * profile info struct
549 be_profile_read(const char *filename)
555 f = fopen(filename, "r");
559 printf("found profile data '%s'.\n", filename);
562 ret = fread(buf, 8, 1, f);
563 if(ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
567 if(profile) be_profile_free();
568 profile = new_set(cmp_execcount, 16);
572 ret = fread(&query, sizeof(unsigned int), 2, f);
576 set_insert(profile, &query, sizeof(query), query.block);
584 * Frees the profile info
587 be_profile_free(void)
590 unregister_vcg_hook();
596 * Tells whether profile module has acquired data
599 be_profile_has_data(void)
601 return (profile != NULL);
605 * Get block execution count as determined be profiling
608 be_profile_get_block_execcount(const ir_node *block)
610 execcount_t *ec, query;
615 query.block = get_irn_node_nr(block);
616 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
621 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
627 typedef struct _intialize_execfreq_env_t {
629 ir_exec_freq *execfreqs;
631 } initialize_execfreq_env_t;
633 // minimal execution frequency (an execfreq of 0 confuses algos)
634 static const double MIN_EXECFREQ = 0.00001;
636 static void initialize_execfreq(ir_node *block, void *data) {
637 initialize_execfreq_env_t *env = data;
640 if(block == get_irg_start_block(env->irg)
641 || block == get_irg_end_block(env->irg)) {
644 freq = be_profile_get_block_execcount(block);
645 freq *= env->freq_factor;
646 if(freq < MIN_EXECFREQ)
650 set_execfreq(env->execfreqs, block, freq);
653 ir_exec_freq *be_create_execfreqs_from_profile(ir_graph *irg)
655 ir_node *block2 = NULL;
656 ir_node *start_block;
657 const ir_edge_t *edge;
658 initialize_execfreq_env_t env;
662 env.execfreqs = create_execfreq(irg);
664 // find the successor to the start block
665 start_block = get_irg_start_block(irg);
666 foreach_block_succ(start_block, edge) {
667 ir_node *succ = get_edge_src_irn(edge);
668 if(succ != start_block) {
673 assert(block2 != NULL);
675 count = be_profile_get_block_execcount(block2);
677 // the function was never executed, so fallback to estimated freqs
678 free_execfreq(env.execfreqs);
680 return compute_execfreq(irg, 10);
683 env.freq_factor = 1.0 / count;
684 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
686 return env.execfreqs;