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
54 #include "besched_t.h"
64 #include "bechordal_t.h"
66 #include "beprofile.h"
68 /** An entry in the id-to-location map */
69 typedef struct loc_entry {
70 ir_entity *fname; /**< the entity holding the file name */
71 unsigned int lineno; /**< line number */
74 typedef struct _block_id_walker_data_t {
75 tarval **array; /**< the entity the holds the block counts */
76 unsigned int id; /**< current block id number */
77 ir_node *symconst; /**< the SymConst representing array */
78 pmap *fname_map; /**< set containing all found filenames */
79 loc_entry *locs; /**< locations */
80 ir_type *tp_char; /**< the character type */
81 unsigned flags; /**< profile flags */
82 } block_id_walker_data_t;
84 typedef struct _execcount_t {
90 cmp_execcount(const void * a, const void * b, size_t size)
92 return ((execcount_t*)a)->block != ((execcount_t*)b)->block;
96 block_counter(ir_node * bb, void * data)
98 unsigned int *count = data;
103 count_blocks(ir_graph * irg)
105 unsigned int count = 0;
107 irg_block_walk_graph(irg, block_counter, NULL, &count);
111 /* keep the execcounts here because they are only read once per compiler run */
112 static set * profile = NULL;
113 static hook_entry_t hook;
116 * Instrument a block with code needed for profiling
119 instrument_block(ir_node * bb, ir_node * address, unsigned int id)
121 ir_graph *irg = get_irn_irg(bb);
122 ir_node *start_block = get_irg_start_block(irg);
123 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
127 * We can't instrument the start and end block as there are no real
128 * instructions in these blocks
130 if(bb == start_block || bb == get_irg_end_block(irg))
133 unknown = new_r_Unknown(irg, mode_M);
134 cnst = new_r_Const_long(irg, start_block, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
135 offset = new_r_Add(irg, bb, address, cnst, mode_P);
136 load = new_r_Load(irg, bb, unknown, offset, mode_Iu);
137 projm = new_r_Proj(irg, bb, load, mode_M, pn_Load_M);
138 proji = new_r_Proj(irg, bb, load, mode_Iu, pn_Load_res);
139 cnst = new_r_Const_long(irg, start_block, mode_Iu, 1);
140 add = new_r_Add(irg, bb, proji, cnst, mode_Iu);
141 store = new_r_Store(irg, bb, projm, offset, add);
142 projm = new_r_Proj(irg, bb, store, mode_M, pn_Store_M);
143 set_irn_link(bb, projm);
144 set_irn_link(projm, load);
147 typedef struct fix_env {
148 ir_node *start_block;
153 * SSA Construction for instrumentation code memory
156 fix_ssa(ir_node * bb, void * data)
160 int arity = get_Block_n_cfgpreds(bb);
162 /* start and end block are not instrumented, skip! */
163 if (bb == env->start_block || bb == env->end_block)
167 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
171 ir_graph *irg = current_ir_graph;
173 NEW_ARR_A(ir_node*, ins, arity);
174 for (n = arity - 1; n >= 0; --n) {
175 ins[n] = get_irn_link(get_Block_cfgpred_block(bb, n));
177 mem = new_r_Phi(irg, bb, arity, ins, mode_M);
179 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
184 * Generates a new irg which calls the initializer
187 * void __firmprof_initializer(void) { __init_firmprof(ent_filename, bblock_id, bblock_counts, n_blocks); }
190 gen_initializer_irg(ir_entity * ent_filename, ir_entity * bblock_id, ir_entity * bblock_counts, int n_blocks)
192 ir_node *start_block;
195 ident *name = new_id_from_str("__firmprof_initializer");
196 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(name, 0, 0));
197 ir_node *ret, *call, *symconst;
200 ident *init_name = new_id_from_str("__init_firmprof");
201 ir_type *init_type = new_type_method(init_name, 4, 0);
202 ir_type *uint, *uintptr, *string;
206 ir_type *empty_frame_type;
208 set_entity_ld_ident(ent, name);
210 uint = new_type_primitive(new_id_from_str("__uint"), mode_Iu);
211 uintptr = new_type_pointer(new_id_from_str("__uintptr"), uint, mode_P);
212 string = new_type_pointer(new_id_from_str("__charptr"), new_type_primitive(new_id_from_str("__char"), mode_Bs), mode_P);
214 set_method_param_type(init_type, 0, string);
215 set_method_param_type(init_type, 1, uintptr);
216 set_method_param_type(init_type, 2, uintptr);
217 set_method_param_type(init_type, 3, uint);
218 init_ent = new_entity(get_glob_type(), init_name, init_type);
219 set_entity_ld_ident(init_ent, init_name);
221 irg = new_ir_graph(ent, 0);
222 empty_frame_type = get_irg_frame_type(irg);
223 set_type_size_bytes(empty_frame_type, 0);
225 bb = get_cur_block();
227 start_block = get_irg_start_block(irg);
229 sym.entity_p = init_ent;
230 symconst = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
232 sym.entity_p = ent_filename;
233 ins[0] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
234 sym.entity_p = bblock_id;
235 ins[1] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
236 sym.entity_p = bblock_counts;
237 ins[2] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
238 ins[3] = new_r_Const_long(irg, start_block, mode_Iu, n_blocks);
240 call = new_r_Call(irg, bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
241 ret = new_r_Return(irg, bb, new_r_Proj(irg, bb, call, mode_M, pn_Call_M_regular), 0, NULL);
244 add_immBlock_pred(get_irg_end_block(irg), ret);
245 mature_immBlock(get_irg_end_block(irg));
247 irg_finalize_cons(irg);
253 * Create the location data for the given debug info.
255 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
258 const char *fname = be_retrieve_dbg_info(dbg, &lineno);
261 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
265 static unsigned nr = 0;
269 int i, len = strlen(fname) + 1;
270 tarval **tarval_string;
272 snprintf(buf, sizeof(buf), "firm_name_arr.%d", nr);
273 arr = new_type_array(new_id_from_str(buf), 1, wd->tp_char);
274 set_array_bounds_int(arr, 0, 0, len);
276 snprintf(buf, sizeof(buf), "__firm_name.%d", nr++);
277 id = new_id_from_str(buf);
278 ent = new_entity(get_glob_type(), id, arr);
279 set_entity_ld_ident(ent, id);
281 pmap_insert(wd->fname_map, (void *)fname, ent);
283 /* initialize file name string constant */
284 tarval_string = alloca(sizeof(*tarval_string) * (len));
285 for (i = 0; i < len; ++i) {
286 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
288 set_entity_variability(ent, variability_constant);
289 set_array_entity_values(ent, tarval_string, len);
293 wd->locs[wd->id].fname = ent;
294 wd->locs[wd->id].lineno = lineno;
296 wd->locs[wd->id].fname = NULL;
297 wd->locs[wd->id].lineno = 0;
302 * Walker: assigns an ID to every block.
303 * Builds the string table
306 block_id_walker(ir_node * bb, void * data)
308 block_id_walker_data_t *wd = data;
310 wd->array[wd->id] = new_tarval_from_long(get_irn_node_nr(bb), mode_Iu);
311 instrument_block(bb, wd->symconst, wd->id);
313 if (wd->flags & profile_with_locations) {
314 dbg_info *dbg = get_irn_dbg_info(bb);
315 create_location_data(dbg, wd);
320 #define IDENT(x) new_id_from_chars(x, sizeof(x) - 1)
323 be_profile_instrument(const char *filename, unsigned flags)
327 ir_entity *bblock_id;
328 ir_entity *bblock_counts;
329 ir_entity *ent_filename;
330 ir_entity *ent_locations = NULL;
331 ir_entity *loc_lineno = NULL;
332 ir_entity *loc_name = NULL;
336 ir_type *string_type;
337 ir_type *character_type;
338 ir_type *loc_type = NULL;
339 ir_type *charptr_type;
341 tarval **tarval_array;
342 tarval **tarval_string;
344 int filename_len = strlen(filename)+1;
346 int align_l, align_n, size;
348 block_id_walker_data_t wd;
351 /* count the number of block first */
352 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
353 ir_graph *irg = get_irp_irg(n);
355 n_blocks += count_blocks(irg);
358 /* create all the necessary types and entities. Note that the
359 types must have a fixed layout, because we already running in the
361 uint_type = new_type_primitive(IDENT("__uint"), mode_Iu);
362 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
363 array_type = new_type_array(IDENT("__block_info_array"), 1, uint_type);
364 set_array_bounds_int(array_type, 0, 0, n_blocks);
366 character_type = new_type_primitive(IDENT("__char"), mode_Bs);
367 string_type = new_type_array(IDENT("__filename"), 1, character_type);
368 set_array_bounds_int(string_type, 0, 0, filename_len);
370 gtp = get_glob_type();
372 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
373 bblock_id = new_entity(gtp, cur_ident, array_type);
374 set_entity_ld_ident(bblock_id, cur_ident);
375 set_entity_variability(bblock_id, variability_initialized);
377 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
378 bblock_counts = new_entity(gtp, cur_ident, array_type);
379 set_entity_ld_ident(bblock_counts, cur_ident);
380 set_entity_variability(bblock_counts, variability_initialized);
382 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
383 ent_filename = new_entity(gtp, cur_ident, string_type);
384 set_entity_ld_ident(ent_filename, cur_ident);
386 if (flags & profile_with_locations) {
387 loc_type = new_type_struct(IDENT("__location"));
388 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
389 align_l = get_type_alignment_bytes(uint_type);
390 size = get_type_size_bytes(uint_type);
391 set_entity_offset(loc_lineno, 0);
393 charptr_type = new_type_pointer(IDENT("__charptr"), character_type, mode_P_data);
394 align_n = get_type_size_bytes(charptr_type);
395 set_type_alignment_bytes(charptr_type, align_n);
396 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
397 size = (size + align_n - 1) & -align_n;
398 set_entity_offset(loc_name, size);
401 if (align_n > align_l)
403 size = (size + align_l - 1) & -align_l;
404 set_type_size_bytes(loc_type, size);
405 set_type_state(loc_type, layout_fixed);
407 loc_type = new_type_array(IDENT("__locarray"), 1, loc_type);
408 set_array_bounds_int(string_type, 0, 0, n_blocks);
410 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
411 ent_locations = new_entity(gtp, cur_ident, loc_type);
412 set_entity_ld_ident(ent_locations, cur_ident);
415 /* initialize count array */
416 NEW_ARR_A(tarval *, tarval_array, n_blocks);
417 tv = get_tarval_null(mode_Iu);
418 for (i = 0; i < n_blocks; ++i) {
419 tarval_array[i] = tv;
421 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
423 /* initialize function name string constant */
424 tarval_string = alloca(sizeof(*tarval_string) * (filename_len));
425 for (i = 0; i < filename_len; ++i) {
426 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
428 set_entity_variability(ent_filename, variability_constant);
429 set_array_entity_values(ent_filename, tarval_string, filename_len);
431 /* initialize block id array and instrument blocks */
432 wd.array = tarval_array;
434 wd.tp_char = character_type;
436 if (flags & profile_with_locations) {
437 wd.fname_map = pmap_create();
438 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
441 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
442 ir_graph *irg = get_irp_irg(n);
444 ir_node *endbb = get_irg_end_block(irg);
447 set_current_ir_graph(irg);
449 /* generate a symbolic constant pointing to the count array */
450 sym.entity_p = bblock_counts;
451 wd.symconst = new_r_SymConst(irg, get_irg_start_block(irg), sym, symconst_addr_ent);
453 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
454 env.start_block = get_irg_start_block(irg);
455 env.end_block = get_irg_end_block(irg);
456 set_irn_link(env.start_block, get_irg_no_mem(irg));
457 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
458 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
459 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
460 ir_node *bb = get_Block_cfgpred_block(endbb, i);
464 switch (get_irn_opcode(node)) {
466 ins[0] = get_irn_link(bb);
467 ins[1] = get_Return_mem(node);
468 sync = new_r_Sync(irg, bb, 2, ins);
469 set_Return_mem(node, sync);
472 ins[0] = get_irn_link(bb);
473 ins[1] = get_Raise_mem(node);
474 sync = new_r_Sync(irg, bb, 2, ins);
475 set_Raise_mem(node, sync);
478 /* a fragile's op exception. There should be another path to End,
480 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
484 set_array_entity_values(bblock_id, tarval_array, n_blocks);
486 if (flags & profile_with_locations) {
487 /* build the initializer for the locations */
488 rem = current_ir_graph;
489 current_ir_graph = get_const_code_irg();
490 ent = get_array_element_entity(loc_type);
491 set_entity_variability(ent_locations, variability_constant);
492 for (i = 0; i < n_blocks; ++i) {
493 compound_graph_path *path;
498 path = new_compound_graph_path(loc_type, 2);
499 set_compound_graph_path_array_index(path, 0, i);
500 set_compound_graph_path_node(path, 0, ent);
501 set_compound_graph_path_node(path, 1, loc_lineno);
502 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
503 add_compound_ent_value_w_path(ent_locations, new_Const(mode_Iu, tv), path);
506 path = new_compound_graph_path(loc_type, 2);
507 set_compound_graph_path_array_index(path, 0, i);
508 set_compound_graph_path_node(path, 0, ent);
509 set_compound_graph_path_node(path, 1, loc_name);
510 if (wd.locs[i].fname) {
511 sym.entity_p = wd.locs[i].fname;
512 n = new_SymConst(sym, symconst_addr_ent);
514 n = new_Const(mode_P_data, get_mode_null(mode_P_data));
516 add_compound_ent_value_w_path(ent_locations, n, path);
518 pmap_destroy(wd.fname_map);
520 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
524 profile_node_info(void *ctx, FILE *f, const ir_node *irn)
527 fprintf(f, "profiled execution count: %u\n", be_profile_get_block_execcount(irn));
532 register_vcg_hook(void)
534 memset(&hook, 0, sizeof(hook));
535 hook.hook._hook_node_info = profile_node_info;
536 register_hook(hook_node_info, &hook);
540 unregister_vcg_hook(void)
542 unregister_hook(hook_node_info, &hook);
546 * Reads the corresponding profile info file if it exists and returns a
547 * profile info struct
550 be_profile_read(const char *filename)
556 f = fopen(filename, "r");
560 printf("found profile data '%s'.\n", filename);
563 ret = fread(buf, 8, 1, f);
564 if(ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
568 if(profile) be_profile_free();
569 profile = new_set(cmp_execcount, 16);
573 ret = fread(&query, sizeof(unsigned int), 2, f);
577 set_insert(profile, &query, sizeof(query), query.block);
585 * Frees the profile info
588 be_profile_free(void)
591 unregister_vcg_hook();
597 * Tells whether profile module has acquired data
600 be_profile_has_data(void)
602 return (profile != NULL);
606 * Get block execution count as determined be profiling
609 be_profile_get_block_execcount(const ir_node *block)
611 execcount_t *ec, query;
616 query.block = get_irn_node_nr(block);
617 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
622 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
628 typedef struct _intialize_execfreq_env_t {
630 ir_exec_freq *execfreqs;
632 } initialize_execfreq_env_t;
634 // minimal execution frequency (an execfreq of 0 confuses algos)
635 static const double MIN_EXECFREQ = 0.00001;
637 static void initialize_execfreq(ir_node *block, void *data) {
638 initialize_execfreq_env_t *env = data;
641 if(block == get_irg_start_block(env->irg)
642 || block == get_irg_end_block(env->irg)) {
645 freq = be_profile_get_block_execcount(block);
646 freq *= env->freq_factor;
647 if(freq < MIN_EXECFREQ)
651 set_execfreq(env->execfreqs, block, freq);
654 ir_exec_freq *be_create_execfreqs_from_profile(ir_graph *irg)
656 ir_node *block2 = NULL;
657 ir_node *start_block;
658 const ir_edge_t *edge;
659 initialize_execfreq_env_t env;
663 env.execfreqs = create_execfreq(irg);
665 // find the successor to the start block
666 start_block = get_irg_start_block(irg);
667 foreach_block_succ(start_block, edge) {
668 ir_node *succ = get_edge_src_irn(edge);
669 if(succ != start_block) {
674 assert(block2 != NULL);
676 count = be_profile_get_block_execcount(block2);
678 // the function was never executed, so fallback to estimated freqs
679 free_execfreq(env.execfreqs);
681 return compute_execfreq(irg, 10);
684 env.freq_factor = 1.0 / count;
685 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
687 return env.execfreqs;