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
20 /** vim: set sw=4 ts=4:
23 * @author Adam M. Szalkowski
26 * Code instrumentation and execution count profiling
28 * Copyright (C) 2006 Universitaet Karlsruhe
29 * Released under the GPL
58 #include "besched_t.h"
68 #include "bechordal_t.h"
70 #include "beprofile.h"
72 /** An entry in the id-to-location map */
73 typedef struct loc_entry {
74 ir_entity *fname; /**< the entity holding the file name */
75 unsigned int lineno; /**< line number */
78 typedef struct _block_id_walker_data_t {
79 tarval **array; /**< the entity the holds the block counts */
80 unsigned int id; /**< current block id number */
81 ir_node *symconst; /**< the SymConst representing array */
82 pmap *fname_map; /**< set containing all found filenames */
83 loc_entry *locs; /**< locations */
84 ir_type *tp_char; /**< the character type */
85 unsigned flags; /**< profile flags */
86 } block_id_walker_data_t;
88 typedef struct _execcount_t {
94 cmp_execcount(const void * a, const void * b, size_t size)
96 return ((execcount_t*)a)->block != ((execcount_t*)b)->block;
100 block_counter(ir_node * bb, void * data)
102 unsigned int *count = data;
107 count_blocks(ir_graph * irg)
109 unsigned int count = 0;
111 irg_block_walk_graph(irg, block_counter, NULL, &count);
115 /* keep the execcounts here because they are only read once per compiler run */
116 static set * profile = NULL;
117 static hook_entry_t hook;
120 * Instrument a block with code needed for profiling
123 instrument_block(ir_node * bb, ir_node * address, unsigned int id)
125 ir_graph *irg = get_irn_irg(bb);
126 ir_node *start_block = get_irg_start_block(irg);
127 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
131 * We can't instrument the start and end block as there are no real
132 * instructions in these blocks
134 if(bb == start_block || bb == get_irg_end_block(irg))
137 unknown = new_r_Unknown(irg, mode_M);
138 cnst = new_r_Const_long(irg, start_block, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
139 offset = new_r_Add(irg, bb, address, cnst, mode_P);
140 load = new_r_Load(irg, bb, unknown, offset, mode_Iu);
141 projm = new_r_Proj(irg, bb, load, mode_M, pn_Load_M);
142 proji = new_r_Proj(irg, bb, load, mode_Iu, pn_Load_res);
143 cnst = new_r_Const_long(irg, start_block, mode_Iu, 1);
144 add = new_r_Add(irg, bb, proji, cnst, mode_Iu);
145 store = new_r_Store(irg, bb, projm, offset, add);
146 projm = new_r_Proj(irg, bb, store, mode_M, pn_Store_M);
147 set_irn_link(bb, projm);
148 set_irn_link(projm, load);
151 typedef struct fix_env {
152 ir_node *start_block;
157 * SSA Construction for instrumentation code memory
160 fix_ssa(ir_node * bb, void * data)
164 int arity = get_Block_n_cfgpreds(bb);
166 /* start and end block are not instrumented, skip! */
167 if (bb == env->start_block || bb == env->end_block)
171 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
175 ir_graph *irg = current_ir_graph;
177 NEW_ARR_A(ir_node*, ins, arity);
178 for (n = arity - 1; n >= 0; --n) {
179 ins[n] = get_irn_link(get_Block_cfgpred_block(bb, n));
181 mem = new_r_Phi(irg, bb, arity, ins, mode_M);
183 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
188 * Generates a new irg which calls the initializer
191 * void __firmprof_initializer(void) { __init_firmprof(ent_filename, bblock_id, bblock_counts, n_blocks); }
194 gen_initializer_irg(ir_entity * ent_filename, ir_entity * bblock_id, ir_entity * bblock_counts, int n_blocks)
196 ir_node *start_block;
199 ident *name = new_id_from_str("__firmprof_initializer");
200 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(name, 0, 0));
201 ir_node *ret, *call, *symconst;
204 ident *init_name = new_id_from_str("__init_firmprof");
205 ir_type *init_type = new_type_method(init_name, 4, 0);
206 ir_type *uint, *uintptr, *string;
210 ir_type *empty_frame_type;
212 set_entity_ld_ident(ent, name);
214 uint = new_type_primitive(new_id_from_str("__uint"), mode_Iu);
215 uintptr = new_type_pointer(new_id_from_str("__uintptr"), uint, mode_P);
216 string = new_type_pointer(new_id_from_str("__charptr"), new_type_primitive(new_id_from_str("__char"), mode_Bs), mode_P);
218 set_method_param_type(init_type, 0, string);
219 set_method_param_type(init_type, 1, uintptr);
220 set_method_param_type(init_type, 2, uintptr);
221 set_method_param_type(init_type, 3, uint);
222 init_ent = new_entity(get_glob_type(), init_name, init_type);
223 set_entity_ld_ident(init_ent, init_name);
225 irg = new_ir_graph(ent, 0);
226 empty_frame_type = get_irg_frame_type(irg);
227 set_type_size_bytes(empty_frame_type, 0);
229 bb = get_cur_block();
231 start_block = get_irg_start_block(irg);
233 sym.entity_p = init_ent;
234 symconst = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
236 sym.entity_p = ent_filename;
237 ins[0] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
238 sym.entity_p = bblock_id;
239 ins[1] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
240 sym.entity_p = bblock_counts;
241 ins[2] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
242 ins[3] = new_r_Const_long(irg, start_block, mode_Iu, n_blocks);
244 call = new_r_Call(irg, bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
245 ret = new_r_Return(irg, bb, new_r_Proj(irg, bb, call, mode_M, pn_Call_M_regular), 0, NULL);
248 add_immBlock_pred(get_irg_end_block(irg), ret);
249 mature_immBlock(get_irg_end_block(irg));
251 irg_finalize_cons(irg);
257 * Create the location data for the given debug info.
259 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
262 const char *fname = be_retrieve_dbg_info(dbg, &lineno);
265 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
269 static unsigned nr = 0;
273 int i, len = strlen(fname) + 1;
274 tarval **tarval_string;
276 snprintf(buf, sizeof(buf), "firm_name_arr.%d", nr);
277 arr = new_type_array(new_id_from_str(buf), 1, wd->tp_char);
278 set_array_bounds_int(arr, 0, 0, len);
280 snprintf(buf, sizeof(buf), "__firm_name.%d", nr++);
281 id = new_id_from_str(buf);
282 ent = new_entity(get_glob_type(), id, arr);
283 set_entity_ld_ident(ent, id);
285 pmap_insert(wd->fname_map, (void *)fname, ent);
287 /* initialize file name string constant */
288 tarval_string = alloca(sizeof(*tarval_string) * (len));
289 for (i = 0; i < len; ++i) {
290 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
292 set_entity_variability(ent, variability_constant);
293 set_array_entity_values(ent, tarval_string, len);
297 wd->locs[wd->id].fname = ent;
298 wd->locs[wd->id].lineno = lineno;
300 wd->locs[wd->id].fname = NULL;
301 wd->locs[wd->id].lineno = 0;
306 * Walker: assigns an ID to every block.
307 * Builds the string table
310 block_id_walker(ir_node * bb, void * data)
312 block_id_walker_data_t *wd = data;
314 wd->array[wd->id] = new_tarval_from_long(get_irn_node_nr(bb), mode_Iu);
315 instrument_block(bb, wd->symconst, wd->id);
317 if (wd->flags & profile_with_locations) {
318 dbg_info *dbg = get_irn_dbg_info(bb);
319 create_location_data(dbg, wd);
324 #define IDENT(x) new_id_from_chars(x, sizeof(x) - 1)
327 be_profile_instrument(const char *filename, unsigned flags)
331 ir_entity *bblock_id;
332 ir_entity *bblock_counts;
333 ir_entity *ent_filename;
334 ir_entity *ent_locations = NULL;
335 ir_entity *loc_lineno = NULL;
336 ir_entity *loc_name = NULL;
340 ir_type *string_type;
341 ir_type *character_type;
342 ir_type *loc_type = NULL;
343 ir_type *charptr_type;
345 tarval **tarval_array;
346 tarval **tarval_string;
348 int filename_len = strlen(filename)+1;
350 int align_l, align_n, size;
352 block_id_walker_data_t wd;
355 /* count the number of block first */
356 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
357 ir_graph *irg = get_irp_irg(n);
359 n_blocks += count_blocks(irg);
362 /* create all the necessary types and entities. Note that the
363 types must have a fixed layout, because we already running in the
365 uint_type = new_type_primitive(IDENT("__uint"), mode_Iu);
366 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
367 array_type = new_type_array(IDENT("__block_info_array"), 1, uint_type);
368 set_array_bounds_int(array_type, 0, 0, n_blocks);
370 character_type = new_type_primitive(IDENT("__char"), mode_Bs);
371 string_type = new_type_array(IDENT("__filename"), 1, character_type);
372 set_array_bounds_int(string_type, 0, 0, filename_len);
374 gtp = get_glob_type();
376 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
377 bblock_id = new_entity(gtp, cur_ident, array_type);
378 set_entity_ld_ident(bblock_id, cur_ident);
379 set_entity_variability(bblock_id, variability_initialized);
381 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
382 bblock_counts = new_entity(gtp, cur_ident, array_type);
383 set_entity_ld_ident(bblock_counts, cur_ident);
384 set_entity_variability(bblock_counts, variability_initialized);
386 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
387 ent_filename = new_entity(gtp, cur_ident, string_type);
388 set_entity_ld_ident(ent_filename, cur_ident);
390 if (flags & profile_with_locations) {
391 loc_type = new_type_struct(IDENT("__location"));
392 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
393 align_l = get_type_alignment_bytes(uint_type);
394 size = get_type_size_bytes(uint_type);
395 set_entity_offset(loc_lineno, 0);
397 charptr_type = new_type_pointer(IDENT("__charptr"), character_type, mode_P_data);
398 align_n = get_type_size_bytes(charptr_type);
399 set_type_alignment_bytes(charptr_type, align_n);
400 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
401 size = (size + align_n - 1) & -align_n;
402 set_entity_offset(loc_name, size);
405 if (align_n > align_l)
407 size = (size + align_l - 1) & -align_l;
408 set_type_size_bytes(loc_type, size);
409 set_type_state(loc_type, layout_fixed);
411 loc_type = new_type_array(IDENT("__locarray"), 1, loc_type);
412 set_array_bounds_int(string_type, 0, 0, n_blocks);
414 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
415 ent_locations = new_entity(gtp, cur_ident, loc_type);
416 set_entity_ld_ident(ent_locations, cur_ident);
419 /* initialize count array */
420 NEW_ARR_A(tarval *, tarval_array, n_blocks);
421 tv = get_tarval_null(mode_Iu);
422 for (i = 0; i < n_blocks; ++i) {
423 tarval_array[i] = tv;
425 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
427 /* initialize function name string constant */
428 tarval_string = alloca(sizeof(*tarval_string) * (filename_len));
429 for (i = 0; i < filename_len; ++i) {
430 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
432 set_entity_variability(ent_filename, variability_constant);
433 set_array_entity_values(ent_filename, tarval_string, filename_len);
435 /* initialize block id array and instrument blocks */
436 wd.array = tarval_array;
438 wd.tp_char = character_type;
440 if (flags & profile_with_locations) {
441 wd.fname_map = pmap_create();
442 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
445 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
446 ir_graph *irg = get_irp_irg(n);
448 ir_node *endbb = get_irg_end_block(irg);
451 set_current_ir_graph(irg);
453 /* generate a symbolic constant pointing to the count array */
454 sym.entity_p = bblock_counts;
455 wd.symconst = new_r_SymConst(irg, get_irg_start_block(irg), sym, symconst_addr_ent);
457 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
458 env.start_block = get_irg_start_block(irg);
459 env.end_block = get_irg_end_block(irg);
460 set_irn_link(env.start_block, get_irg_no_mem(irg));
461 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
462 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
463 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
464 ir_node *bb = get_Block_cfgpred_block(endbb, i);
468 switch (get_irn_opcode(node)) {
470 ins[0] = get_irn_link(bb);
471 ins[1] = get_Return_mem(node);
472 sync = new_r_Sync(irg, bb, 2, ins);
473 set_Return_mem(node, sync);
476 ins[0] = get_irn_link(bb);
477 ins[1] = get_Raise_mem(node);
478 sync = new_r_Sync(irg, bb, 2, ins);
479 set_Raise_mem(node, sync);
482 /* a fragile's op exception. There should be another path to End,
484 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
488 set_array_entity_values(bblock_id, tarval_array, n_blocks);
490 if (flags & profile_with_locations) {
491 /* build the initializer for the locations */
492 rem = current_ir_graph;
493 current_ir_graph = get_const_code_irg();
494 ent = get_array_element_entity(loc_type);
495 set_entity_variability(ent_locations, variability_constant);
496 for (i = 0; i < n_blocks; ++i) {
497 compound_graph_path *path;
502 path = new_compound_graph_path(loc_type, 2);
503 set_compound_graph_path_array_index(path, 0, i);
504 set_compound_graph_path_node(path, 0, ent);
505 set_compound_graph_path_node(path, 1, loc_lineno);
506 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
507 add_compound_ent_value_w_path(ent_locations, new_Const(mode_Iu, tv), path);
510 path = new_compound_graph_path(loc_type, 2);
511 set_compound_graph_path_array_index(path, 0, i);
512 set_compound_graph_path_node(path, 0, ent);
513 set_compound_graph_path_node(path, 1, loc_name);
514 if (wd.locs[i].fname) {
515 sym.entity_p = wd.locs[i].fname;
516 n = new_SymConst(sym, symconst_addr_ent);
518 n = new_Const(mode_P_data, get_mode_null(mode_P_data));
520 add_compound_ent_value_w_path(ent_locations, n, path);
522 pmap_destroy(wd.fname_map);
524 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
528 profile_node_info(void *ctx, FILE *f, const ir_node *irn)
531 fprintf(f, "profiled execution count: %u\n", be_profile_get_block_execcount(irn));
536 register_vcg_hook(void)
538 memset(&hook, 0, sizeof(hook));
539 hook.hook._hook_node_info = profile_node_info;
540 register_hook(hook_node_info, &hook);
544 unregister_vcg_hook(void)
546 unregister_hook(hook_node_info, &hook);
550 * Reads the corresponding profile info file if it exists and returns a
551 * profile info struct
554 be_profile_read(const char *filename)
560 f = fopen(filename, "r");
564 printf("found profile data '%s'.\n", filename);
567 ret = fread(buf, 8, 1, f);
568 if(ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
572 if(profile) be_profile_free();
573 profile = new_set(cmp_execcount, 16);
577 ret = fread(&query, sizeof(unsigned int), 2, f);
581 set_insert(profile, &query, sizeof(query), query.block);
589 * Frees the profile info
592 be_profile_free(void)
595 unregister_vcg_hook();
601 * Tells whether profile module has acquired data
604 be_profile_has_data(void)
606 return (profile != NULL);
610 * Get block execution count as determined be profiling
613 be_profile_get_block_execcount(const ir_node *block)
615 execcount_t *ec, query;
620 query.block = get_irn_node_nr(block);
621 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
626 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
632 typedef struct _intialize_execfreq_env_t {
634 ir_exec_freq *execfreqs;
636 } initialize_execfreq_env_t;
638 // minimal execution frequency (an execfreq of 0 confuses algos)
639 static const double MIN_EXECFREQ = 0.00001;
641 static void initialize_execfreq(ir_node *block, void *data) {
642 initialize_execfreq_env_t *env = data;
645 if(block == get_irg_start_block(env->irg)
646 || block == get_irg_end_block(env->irg)) {
649 freq = be_profile_get_block_execcount(block);
650 freq *= env->freq_factor;
651 if(freq < MIN_EXECFREQ)
655 set_execfreq(env->execfreqs, block, freq);
658 ir_exec_freq *be_create_execfreqs_from_profile(ir_graph *irg)
660 ir_node *block2 = NULL;
661 ir_node *start_block;
662 const ir_edge_t *edge;
663 initialize_execfreq_env_t env;
667 env.execfreqs = create_execfreq(irg);
669 // find the successor to the start block
670 start_block = get_irg_start_block(irg);
671 foreach_block_succ(start_block, edge) {
672 ir_node *succ = get_edge_src_irn(edge);
673 if(succ != start_block) {
678 assert(block2 != NULL);
680 count = be_profile_get_block_execcount(block2);
682 // the function was never executed, so fallback to estimated freqs
683 free_execfreq(env.execfreqs);
685 return compute_execfreq(irg, 10);
688 env.freq_factor = 1.0 / count;
689 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
691 return env.execfreqs;