2 * Copyright (C) 1995-2008 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 "irprofile.h"
55 /** An entry in the id-to-location map */
56 typedef struct loc_entry {
57 ir_entity *fname; /**< the entity holding the file name */
58 unsigned int lineno; /**< line number */
61 typedef struct _block_id_walker_data_t {
62 tarval **array; /**< the entity the holds the block counts */
63 unsigned int id; /**< current block id number */
64 ir_node *symconst; /**< the SymConst representing array */
65 pmap *fname_map; /**< set containing all found filenames */
66 loc_entry *locs; /**< locations */
67 ir_type *tp_char; /**< the character type */
68 unsigned flags; /**< profile flags */
69 } block_id_walker_data_t;
71 typedef struct _execcount_t {
77 * Compare two execcount_t entries.
79 static int cmp_execcount(const void *a, const void *b, size_t size) {
80 const execcount_t *ea = a;
81 const execcount_t *eb = b;
83 return ea->block != eb->block;
87 * Block walker, count number of blocks.
89 static void block_counter(ir_node * bb, void * data) {
90 unsigned int *count = data;
96 * Return the number of blocks the given graph.
98 static unsigned int count_blocks(ir_graph *irg) {
99 unsigned int count = 0;
101 irg_block_walk_graph(irg, block_counter, NULL, &count);
105 /* keep the execcounts here because they are only read once per compiler run */
106 static set * profile = NULL;
107 static hook_entry_t hook;
110 * Instrument a block with code needed for profiling
113 instrument_block(ir_node *bb, ir_node *address, unsigned int id)
115 ir_graph *irg = get_irn_irg(bb);
116 ir_node *start_block = get_irg_start_block(irg);
117 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
121 * We can't instrument the end block as there are no real instructions there
123 if(bb == get_irg_end_block(irg))
126 unknown = new_r_Unknown(irg, mode_M);
127 cnst = new_r_Const_long(irg, start_block, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
128 offset = new_r_Add(irg, bb, address, cnst, mode_P);
129 load = new_r_Load(irg, bb, unknown, offset, mode_Iu);
130 projm = new_r_Proj(irg, bb, load, mode_M, pn_Load_M);
131 proji = new_r_Proj(irg, bb, load, mode_Iu, pn_Load_res);
132 cnst = new_r_Const_long(irg, start_block, mode_Iu, 1);
133 add = new_r_Add(irg, bb, proji, cnst, mode_Iu);
134 store = new_r_Store(irg, bb, projm, offset, add);
135 projm = new_r_Proj(irg, bb, store, mode_M, pn_Store_M);
136 set_irn_link(bb, projm);
137 set_irn_link(projm, load);
140 typedef struct fix_env {
145 * SSA Construction for instrumentation code memory
148 fix_ssa(ir_node * bb, void * data)
152 int arity = get_Block_n_cfgpreds(bb);
154 /* start and end block are not instrumented, skip! */
155 if (bb == env->end_block)
158 if (bb == get_irg_start_block(get_irn_irg(bb))) {
160 } else if (arity == 1) {
161 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
165 ir_graph *irg = current_ir_graph;
167 NEW_ARR_A(ir_node*, ins, arity);
168 for (n = arity - 1; n >= 0; --n) {
169 ins[n] = get_irn_link(get_Block_cfgpred_block(bb, n));
171 mem = new_r_Phi(irg, bb, arity, ins, mode_M);
173 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
176 static void add_constructor(ir_entity *method)
178 ir_type *method_type = get_entity_type(method);
179 ident *id = id_unique("constructor_ptrt.%u");
180 ir_type *ptr_type = new_type_pointer(id, method_type, mode_P_code);
182 ir_type *constructors = get_segment_type(IR_SEGMENT_CONSTRUCTORS);
183 ident *ide = id_unique("constructor_ptr.%u");
184 ir_entity *ptr = new_entity(constructors, ide, ptr_type);
185 ir_graph *irg = get_const_code_irg();
186 ir_node *val = new_rd_SymConst_addr_ent(NULL, irg, mode_P_code,
189 set_entity_compiler_generated(ptr, 1);
190 set_entity_variability(ptr, variability_constant);
191 set_atomic_ent_value(ptr, val);
195 * Generates a new irg which calls the initializer
198 * void __firmprof_initializer(void) { __init_firmprof(ent_filename, bblock_id, bblock_counts, n_blocks); }
201 gen_initializer_irg(ir_entity * ent_filename, ir_entity * bblock_id, ir_entity * bblock_counts, int n_blocks)
203 ir_node *start_block;
206 ident *name = new_id_from_str("__firmprof_initializer");
207 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(name, 0, 0));
208 ir_node *ret, *call, *symconst;
211 ident *init_name = new_id_from_str("__init_firmprof");
212 ir_type *init_type = new_type_method(init_name, 4, 0);
213 ir_type *uint, *uintptr, *string;
217 ir_type *empty_frame_type;
219 set_entity_ld_ident(ent, name);
221 uint = new_type_primitive(new_id_from_str("__uint"), mode_Iu);
222 uintptr = new_type_pointer(new_id_from_str("__uintptr"), uint, mode_P);
223 string = new_type_pointer(new_id_from_str("__charptr"), new_type_primitive(new_id_from_str("__char"), mode_Bs), mode_P);
225 set_method_param_type(init_type, 0, string);
226 set_method_param_type(init_type, 1, uintptr);
227 set_method_param_type(init_type, 2, uintptr);
228 set_method_param_type(init_type, 3, uint);
229 init_ent = new_entity(get_glob_type(), init_name, init_type);
230 set_entity_ld_ident(init_ent, init_name);
232 irg = new_ir_graph(ent, 0);
233 empty_frame_type = get_irg_frame_type(irg);
234 set_type_size_bytes(empty_frame_type, 0);
236 bb = get_cur_block();
238 start_block = get_irg_start_block(irg);
240 sym.entity_p = init_ent;
241 symconst = new_r_SymConst(irg, start_block, mode_P_data, sym, symconst_addr_ent);
243 sym.entity_p = ent_filename;
244 ins[0] = new_r_SymConst(irg, start_block, mode_P_data, sym, symconst_addr_ent);
245 sym.entity_p = bblock_id;
246 ins[1] = new_r_SymConst(irg, start_block, mode_P_data, sym, symconst_addr_ent);
247 sym.entity_p = bblock_counts;
248 ins[2] = new_r_SymConst(irg, start_block, mode_P_data, sym, symconst_addr_ent);
249 ins[3] = new_r_Const_long(irg, start_block, mode_Iu, n_blocks);
251 call = new_r_Call(irg, bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
252 ret = new_r_Return(irg, bb, new_r_Proj(irg, bb, call, mode_M, pn_Call_M_regular), 0, NULL);
255 add_immBlock_pred(get_irg_end_block(irg), ret);
256 mature_immBlock(get_irg_end_block(irg));
258 irg_finalize_cons(irg);
260 add_constructor(ent);
266 * Create the location data for the given debug info.
268 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
271 const char *fname = ir_retrieve_dbg_info(dbg, &lineno);
274 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
278 static unsigned nr = 0;
282 int i, len = strlen(fname) + 1;
283 tarval **tarval_string;
285 snprintf(buf, sizeof(buf), "firm_name_arr.%d", nr);
286 arr = new_type_array(new_id_from_str(buf), 1, wd->tp_char);
287 set_array_bounds_int(arr, 0, 0, len);
289 snprintf(buf, sizeof(buf), "__firm_name.%d", nr++);
290 id = new_id_from_str(buf);
291 ent = new_entity(get_glob_type(), id, arr);
292 set_entity_ld_ident(ent, id);
294 pmap_insert(wd->fname_map, (void *)fname, ent);
296 /* initialize file name string constant */
297 tarval_string = alloca(sizeof(*tarval_string) * (len));
298 for (i = 0; i < len; ++i) {
299 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
301 set_entity_variability(ent, variability_constant);
302 set_array_entity_values(ent, tarval_string, len);
306 wd->locs[wd->id].fname = ent;
307 wd->locs[wd->id].lineno = lineno;
309 wd->locs[wd->id].fname = NULL;
310 wd->locs[wd->id].lineno = 0;
315 * Walker: assigns an ID to every block.
316 * Builds the string table
319 block_id_walker(ir_node * bb, void * data)
321 block_id_walker_data_t *wd = data;
323 wd->array[wd->id] = new_tarval_from_long(get_irn_node_nr(bb), mode_Iu);
324 instrument_block(bb, wd->symconst, wd->id);
326 if (wd->flags & profile_with_locations) {
327 dbg_info *dbg = get_irn_dbg_info(bb);
328 create_location_data(dbg, wd);
333 #define IDENT(x) new_id_from_chars(x, sizeof(x) - 1)
336 ir_profile_instrument(const char *filename, unsigned flags)
340 ir_entity *bblock_id;
341 ir_entity *bblock_counts;
342 ir_entity *ent_filename;
343 ir_entity *ent_locations = NULL;
344 ir_entity *loc_lineno = NULL;
345 ir_entity *loc_name = NULL;
349 ir_type *string_type;
350 ir_type *character_type;
351 ir_type *loc_type = NULL;
352 ir_type *charptr_type;
354 ir_node *start_block;
355 tarval **tarval_array;
356 tarval **tarval_string;
358 int filename_len = strlen(filename)+1;
360 unsigned align_l, align_n, size;
362 block_id_walker_data_t wd;
365 /* count the number of block first */
366 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
367 ir_graph *irg = get_irp_irg(n);
369 n_blocks += count_blocks(irg);
372 /* create all the necessary types and entities. Note that the
373 types must have a fixed layout, because we already running in the
375 uint_type = new_type_primitive(IDENT("__uint"), mode_Iu);
376 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
378 array_type = new_type_array(IDENT("__block_info_array"), 1, uint_type);
379 set_array_bounds_int(array_type, 0, 0, n_blocks);
380 set_type_size_bytes(array_type, n_blocks * get_mode_size_bytes(mode_Iu));
381 set_type_alignment_bytes(array_type, get_mode_size_bytes(mode_Iu));
382 set_type_state(array_type, layout_fixed);
384 character_type = new_type_primitive(IDENT("__char"), mode_Bs);
385 string_type = new_type_array(IDENT("__filename"), 1, character_type);
386 set_array_bounds_int(string_type, 0, 0, filename_len);
387 set_type_size_bytes(string_type, filename_len);
388 set_type_alignment_bytes(string_type, 1);
389 set_type_state(string_type, layout_fixed);
391 gtp = get_glob_type();
393 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
394 bblock_id = new_entity(gtp, cur_ident, array_type);
395 set_entity_ld_ident(bblock_id, cur_ident);
396 set_entity_variability(bblock_id, variability_initialized);
398 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
399 bblock_counts = new_entity(gtp, cur_ident, array_type);
400 set_entity_ld_ident(bblock_counts, cur_ident);
401 set_entity_variability(bblock_counts, variability_initialized);
403 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
404 ent_filename = new_entity(gtp, cur_ident, string_type);
405 set_entity_ld_ident(ent_filename, cur_ident);
407 if (flags & profile_with_locations) {
408 loc_type = new_type_struct(IDENT("__location"));
409 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
410 align_l = get_type_alignment_bytes(uint_type);
411 size = get_type_size_bytes(uint_type);
412 set_entity_offset(loc_lineno, 0);
414 charptr_type = new_type_pointer(IDENT("__charptr"), character_type, mode_P_data);
415 align_n = get_type_size_bytes(charptr_type);
416 set_type_alignment_bytes(charptr_type, align_n);
417 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
418 size = (size + align_n - 1) & ~(align_n - 1);
419 set_entity_offset(loc_name, size);
422 if (align_n > align_l)
424 size = (size + align_l - 1) & ~(align_l - 1);
425 set_type_size_bytes(loc_type, size);
426 set_type_state(loc_type, layout_fixed);
428 loc_type = new_type_array(IDENT("__locarray"), 1, loc_type);
429 set_array_bounds_int(string_type, 0, 0, n_blocks);
431 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
432 ent_locations = new_entity(gtp, cur_ident, loc_type);
433 set_entity_ld_ident(ent_locations, cur_ident);
436 /* initialize count array */
437 NEW_ARR_A(tarval *, tarval_array, n_blocks);
438 tv = get_tarval_null(mode_Iu);
439 for (i = 0; i < n_blocks; ++i) {
440 tarval_array[i] = tv;
442 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
444 /* initialize function name string constant */
445 tarval_string = alloca(sizeof(*tarval_string) * (filename_len));
446 for (i = 0; i < filename_len; ++i) {
447 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
449 set_entity_variability(ent_filename, variability_constant);
450 set_array_entity_values(ent_filename, tarval_string, filename_len);
452 /* initialize block id array and instrument blocks */
453 wd.array = tarval_array;
455 wd.tp_char = character_type;
457 if (flags & profile_with_locations) {
458 wd.fname_map = pmap_create();
459 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
462 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
463 ir_graph *irg = get_irp_irg(n);
465 ir_node *endbb = get_irg_end_block(irg);
468 set_current_ir_graph(irg);
470 /* generate a symbolic constant pointing to the count array */
471 sym.entity_p = bblock_counts;
472 wd.symconst = new_r_SymConst(irg, get_irg_start_block(irg), mode_P_data, sym, symconst_addr_ent);
474 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
475 start_block = get_irg_start_block(irg);
476 env.end_block = get_irg_end_block(irg);
477 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
478 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
479 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
480 ir_node *bb = get_Block_cfgpred_block(endbb, i);
484 switch (get_irn_opcode(node)) {
486 ins[0] = get_irn_link(bb);
487 ins[1] = get_Return_mem(node);
488 sync = new_r_Sync(irg, bb, 2, ins);
489 set_Return_mem(node, sync);
492 ins[0] = get_irn_link(bb);
493 ins[1] = get_Raise_mem(node);
494 sync = new_r_Sync(irg, bb, 2, ins);
495 set_Raise_mem(node, sync);
498 /* a fragile's op exception. There should be another path to End,
500 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
504 set_array_entity_values(bblock_id, tarval_array, n_blocks);
506 if (flags & profile_with_locations) {
507 /* build the initializer for the locations */
508 rem = current_ir_graph;
509 current_ir_graph = get_const_code_irg();
510 ent = get_array_element_entity(loc_type);
511 set_entity_variability(ent_locations, variability_constant);
512 for (i = 0; i < n_blocks; ++i) {
513 compound_graph_path *path;
518 path = new_compound_graph_path(loc_type, 2);
519 set_compound_graph_path_array_index(path, 0, i);
520 set_compound_graph_path_node(path, 0, ent);
521 set_compound_graph_path_node(path, 1, loc_lineno);
522 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
523 add_compound_ent_value_w_path(ent_locations, new_Const(mode_Iu, tv), path);
526 path = new_compound_graph_path(loc_type, 2);
527 set_compound_graph_path_array_index(path, 0, i);
528 set_compound_graph_path_node(path, 0, ent);
529 set_compound_graph_path_node(path, 1, loc_name);
530 if (wd.locs[i].fname) {
531 sym.entity_p = wd.locs[i].fname;
532 n = new_SymConst(mode_P_data, sym, symconst_addr_ent);
534 n = new_Const(mode_P_data, get_mode_null(mode_P_data));
536 add_compound_ent_value_w_path(ent_locations, n, path);
538 pmap_destroy(wd.fname_map);
540 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
544 profile_node_info(void *ctx, FILE *f, const ir_node *irn)
548 fprintf(f, "profiled execution count: %u\n", ir_profile_get_block_execcount(irn));
553 register_vcg_hook(void)
555 memset(&hook, 0, sizeof(hook));
556 hook.hook._hook_node_info = profile_node_info;
557 register_hook(hook_node_info, &hook);
561 unregister_vcg_hook(void)
563 unregister_hook(hook_node_info, &hook);
567 * Reads the corresponding profile info file if it exists and returns a
568 * profile info struct
571 ir_profile_read(const char *filename)
577 f = fopen(filename, "r");
581 printf("found profile data '%s'.\n", filename);
584 ret = fread(buf, 8, 1, f);
585 if(ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
589 if(profile) ir_profile_free();
590 profile = new_set(cmp_execcount, 16);
594 ret = fread(&query, sizeof(unsigned int), 2, f);
598 set_insert(profile, &query, sizeof(query), query.block);
606 * Frees the profile info
609 ir_profile_free(void)
612 unregister_vcg_hook();
618 * Tells whether profile module has acquired data
621 ir_profile_has_data(void)
623 return (profile != NULL);
627 * Get block execution count as determined be profiling
630 ir_profile_get_block_execcount(const ir_node *block)
632 execcount_t *ec, query;
637 query.block = get_irn_node_nr(block);
638 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
643 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
649 typedef struct _intialize_execfreq_env_t {
651 ir_exec_freq *execfreqs;
653 } initialize_execfreq_env_t;
655 // minimal execution frequency (an execfreq of 0 confuses algos)
656 static const double MIN_EXECFREQ = 0.00001;
658 static void initialize_execfreq(ir_node *block, void *data) {
659 initialize_execfreq_env_t *env = data;
662 if(block == get_irg_start_block(env->irg)
663 || block == get_irg_end_block(env->irg)) {
666 freq = ir_profile_get_block_execcount(block);
667 freq *= env->freq_factor;
668 if(freq < MIN_EXECFREQ)
672 set_execfreq(env->execfreqs, block, freq);
675 ir_exec_freq *ir_create_execfreqs_from_profile(ir_graph *irg)
677 ir_node *start_block;
678 initialize_execfreq_env_t env;
682 env.execfreqs = create_execfreq(irg);
683 start_block = get_irg_start_block(irg);
685 count = ir_profile_get_block_execcount(start_block);
687 // the function was never executed, so fallback to estimated freqs
688 free_execfreq(env.execfreqs);
690 return compute_execfreq(irg, 10);
693 env.freq_factor = 1.0 / count;
694 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
696 return env.execfreqs;