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
51 #include "irprofile.h"
53 /** An entry in the id-to-location map */
54 typedef struct loc_entry {
55 ir_entity *fname; /**< the entity holding the file name */
56 unsigned int lineno; /**< line number */
59 typedef struct _block_id_walker_data_t {
60 tarval **array; /**< the entity the holds the block counts */
61 unsigned int id; /**< current block id number */
62 ir_node *symconst; /**< the SymConst representing array */
63 pmap *fname_map; /**< set containing all found filenames */
64 loc_entry *locs; /**< locations */
65 ir_type *tp_char; /**< the character type */
66 unsigned flags; /**< profile flags */
67 } block_id_walker_data_t;
69 typedef struct _execcount_t {
75 * Compare two execcount_t entries.
77 static int cmp_execcount(const void *a, const void *b, size_t size) {
78 const execcount_t *ea = a;
79 const execcount_t *eb = b;
81 return ea->block != eb->block;
85 * Block walker, count number of blocks.
87 static void block_counter(ir_node * bb, void * data) {
88 unsigned int *count = data;
94 * Return the number of blocks the given graph.
96 static unsigned int count_blocks(ir_graph *irg) {
97 unsigned int count = 0;
99 irg_block_walk_graph(irg, block_counter, NULL, &count);
103 /* keep the execcounts here because they are only read once per compiler run */
104 static set * profile = NULL;
105 static hook_entry_t hook;
108 * Instrument a block with code needed for profiling
111 instrument_block(ir_node *bb, ir_node *address, unsigned int id)
113 ir_graph *irg = get_irn_irg(bb);
114 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
118 * We can't instrument the end block as there are no real instructions there
120 if(bb == get_irg_end_block(irg))
123 unknown = new_r_Unknown(irg, mode_M);
124 cnst = new_r_Const_long(irg, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
125 offset = new_r_Add(bb, address, cnst, get_modeP_data());
126 load = new_r_Load(bb, unknown, offset, mode_Iu, 0);
127 projm = new_r_Proj(bb, load, mode_M, pn_Load_M);
128 proji = new_r_Proj(bb, load, mode_Iu, pn_Load_res);
129 cnst = new_r_Const_long(irg, mode_Iu, 1);
130 add = new_r_Add(bb, proji, cnst, mode_Iu);
131 store = new_r_Store(bb, projm, offset, add, 0);
132 projm = new_r_Proj(bb, store, mode_M, pn_Store_M);
133 set_irn_link(bb, projm);
134 set_irn_link(projm, load);
137 typedef struct fix_env {
142 * SSA Construction for instrumentation code memory
145 fix_ssa(ir_node * bb, void * data)
149 int arity = get_Block_n_cfgpreds(bb);
151 /* end block are not instrumented, skip! */
152 if (bb == env->end_block)
155 if (bb == get_irg_start_block(get_irn_irg(bb))) {
156 mem = get_irg_initial_mem(get_irn_irg(bb));
157 } else if (arity == 1) {
158 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
163 NEW_ARR_A(ir_node*, ins, arity);
164 for (n = arity - 1; n >= 0; --n) {
165 ins[n] = get_irn_link(get_Block_cfgpred_block(bb, n));
167 mem = new_r_Phi(bb, arity, ins, mode_M);
169 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
172 static void add_constructor(ir_entity *method)
174 ir_type *method_type = get_entity_type(method);
175 ident *id = id_unique("constructor_ptrt.%u");
176 ir_type *ptr_type = new_type_pointer(id, method_type, mode_P_code);
178 ir_type *constructors = get_segment_type(IR_SEGMENT_CONSTRUCTORS);
179 ident *ide = id_unique("constructor_ptr.%u");
180 ir_entity *ptr = new_entity(constructors, ide, ptr_type);
181 ir_graph *irg = get_const_code_irg();
182 ir_node *val = new_rd_SymConst_addr_ent(NULL, irg, mode_P_code,
185 set_entity_compiler_generated(ptr, 1);
186 set_entity_variability(ptr, variability_constant);
187 set_atomic_ent_value(ptr, val);
191 * Generates a new irg which calls the initializer
194 * void __firmprof_initializer(void) { __init_firmprof(ent_filename, bblock_id, bblock_counts, n_blocks); }
197 gen_initializer_irg(ir_entity *ent_filename, ir_entity *bblock_id, ir_entity *bblock_counts, int n_blocks)
200 ident *name = new_id_from_str("__firmprof_initializer");
201 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(name, 0, 0));
202 ir_node *ret, *call, *symconst;
205 ident *init_name = new_id_from_str("__init_firmprof");
206 ir_type *init_type = new_type_method(init_name, 4, 0);
207 ir_type *uint, *uintptr, *string;
211 ir_type *empty_frame_type;
213 set_entity_ld_ident(ent, name);
215 uint = new_type_primitive(new_id_from_str("__uint"), mode_Iu);
216 uintptr = new_type_pointer(new_id_from_str("__uintptr"), uint, get_modeP_data());
217 string = new_type_pointer(new_id_from_str("__charptr"), new_type_primitive(new_id_from_str("__char"), mode_Bs), get_modeP_data());
219 set_method_param_type(init_type, 0, string);
220 set_method_param_type(init_type, 1, uintptr);
221 set_method_param_type(init_type, 2, uintptr);
222 set_method_param_type(init_type, 3, uint);
223 init_ent = new_entity(get_glob_type(), init_name, init_type);
224 set_entity_ld_ident(init_ent, init_name);
226 irg = new_ir_graph(ent, 0);
227 empty_frame_type = get_irg_frame_type(irg);
228 set_type_size_bytes(empty_frame_type, 0);
229 set_type_state(empty_frame_type, layout_fixed);
231 bb = get_cur_block();
233 sym.entity_p = init_ent;
234 symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
236 sym.entity_p = ent_filename;
237 ins[0] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
238 sym.entity_p = bblock_id;
239 ins[1] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
240 sym.entity_p = bblock_counts;
241 ins[2] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
242 ins[3] = new_r_Const_long(irg, mode_Iu, n_blocks);
244 call = new_r_Call(bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
245 ret = new_r_Return(bb, new_r_Proj(bb, call, mode_M, pn_Call_M), 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);
253 add_constructor(ent);
259 * Create the location data for the given debug info.
261 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
264 const char *fname = ir_retrieve_dbg_info(dbg, &lineno);
267 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
271 static unsigned nr = 0;
275 int i, len = strlen(fname) + 1;
276 tarval **tarval_string;
278 snprintf(buf, sizeof(buf), "firm_name_arr.%d", nr);
279 arr = new_type_array(new_id_from_str(buf), 1, wd->tp_char);
280 set_array_bounds_int(arr, 0, 0, len);
282 snprintf(buf, sizeof(buf), "__firm_name.%d", nr++);
283 id = new_id_from_str(buf);
284 ent = new_entity(get_glob_type(), id, arr);
285 set_entity_ld_ident(ent, id);
287 pmap_insert(wd->fname_map, (void *)fname, ent);
289 /* initialize file name string constant */
290 tarval_string = ALLOCAN(tarval*, len);
291 for (i = 0; i < len; ++i) {
292 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
294 set_entity_variability(ent, variability_constant);
295 set_array_entity_values(ent, tarval_string, len);
299 wd->locs[wd->id].fname = ent;
300 wd->locs[wd->id].lineno = lineno;
302 wd->locs[wd->id].fname = NULL;
303 wd->locs[wd->id].lineno = 0;
308 * Walker: assigns an ID to every block.
309 * Builds the string table
312 block_id_walker(ir_node * bb, void * data)
314 block_id_walker_data_t *wd = data;
316 wd->array[wd->id] = new_tarval_from_long(get_irn_node_nr(bb), mode_Iu);
317 instrument_block(bb, wd->symconst, wd->id);
319 if (wd->flags & profile_with_locations) {
320 dbg_info *dbg = get_irn_dbg_info(bb);
321 create_location_data(dbg, wd);
326 #define IDENT(x) new_id_from_chars(x, sizeof(x) - 1)
329 ir_profile_instrument(const char *filename, unsigned flags)
333 ir_entity *bblock_id;
334 ir_entity *bblock_counts;
335 ir_entity *ent_filename;
336 ir_entity *ent_locations = NULL;
337 ir_entity *loc_lineno = NULL;
338 ir_entity *loc_name = NULL;
342 ir_type *string_type;
343 ir_type *character_type;
344 ir_type *loc_type = NULL;
345 ir_type *charptr_type;
347 ir_node *start_block;
348 tarval **tarval_array;
349 tarval **tarval_string;
351 int filename_len = strlen(filename)+1;
353 unsigned align_l, align_n, size;
355 block_id_walker_data_t wd;
358 /* count the number of block first */
359 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
360 ir_graph *irg = get_irp_irg(n);
362 n_blocks += count_blocks(irg);
365 /* create all the necessary types and entities. Note that the
366 types must have a fixed layout, because we already running in the
368 uint_type = new_type_primitive(IDENT("__uint"), mode_Iu);
369 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
371 array_type = new_type_array(IDENT("__block_info_array"), 1, uint_type);
372 set_array_bounds_int(array_type, 0, 0, n_blocks);
373 set_type_size_bytes(array_type, n_blocks * get_mode_size_bytes(mode_Iu));
374 set_type_alignment_bytes(array_type, get_mode_size_bytes(mode_Iu));
375 set_type_state(array_type, layout_fixed);
377 character_type = new_type_primitive(IDENT("__char"), mode_Bs);
378 string_type = new_type_array(IDENT("__filename"), 1, character_type);
379 set_array_bounds_int(string_type, 0, 0, filename_len);
380 set_type_size_bytes(string_type, filename_len);
381 set_type_alignment_bytes(string_type, 1);
382 set_type_state(string_type, layout_fixed);
384 gtp = get_glob_type();
386 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
387 bblock_id = new_entity(gtp, cur_ident, array_type);
388 set_entity_ld_ident(bblock_id, cur_ident);
389 set_entity_variability(bblock_id, variability_initialized);
391 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
392 bblock_counts = new_entity(gtp, cur_ident, array_type);
393 set_entity_ld_ident(bblock_counts, cur_ident);
394 set_entity_variability(bblock_counts, variability_initialized);
396 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
397 ent_filename = new_entity(gtp, cur_ident, string_type);
398 set_entity_ld_ident(ent_filename, cur_ident);
400 if (flags & profile_with_locations) {
401 loc_type = new_type_struct(IDENT("__location"));
402 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
403 align_l = get_type_alignment_bytes(uint_type);
404 size = get_type_size_bytes(uint_type);
405 set_entity_offset(loc_lineno, 0);
407 charptr_type = new_type_pointer(IDENT("__charptr"), character_type, mode_P_data);
408 align_n = get_type_size_bytes(charptr_type);
409 set_type_alignment_bytes(charptr_type, align_n);
410 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
411 size = (size + align_n - 1) & ~(align_n - 1);
412 set_entity_offset(loc_name, size);
415 if (align_n > align_l)
417 size = (size + align_l - 1) & ~(align_l - 1);
418 set_type_size_bytes(loc_type, size);
419 set_type_state(loc_type, layout_fixed);
421 loc_type = new_type_array(IDENT("__locarray"), 1, loc_type);
422 set_array_bounds_int(string_type, 0, 0, n_blocks);
424 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
425 ent_locations = new_entity(gtp, cur_ident, loc_type);
426 set_entity_ld_ident(ent_locations, cur_ident);
429 /* initialize count array */
430 NEW_ARR_A(tarval *, tarval_array, n_blocks);
431 tv = get_tarval_null(mode_Iu);
432 for (i = 0; i < n_blocks; ++i) {
433 tarval_array[i] = tv;
435 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
437 /* initialize function name string constant */
438 tarval_string = ALLOCAN(tarval*, filename_len);
439 for (i = 0; i < filename_len; ++i) {
440 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
442 set_entity_variability(ent_filename, variability_constant);
443 set_array_entity_values(ent_filename, tarval_string, filename_len);
445 /* initialize block id array and instrument blocks */
446 wd.array = tarval_array;
448 wd.tp_char = character_type;
450 if (flags & profile_with_locations) {
451 wd.fname_map = pmap_create();
452 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
455 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
456 ir_graph *irg = get_irp_irg(n);
458 ir_node *endbb = get_irg_end_block(irg);
461 set_current_ir_graph(irg);
463 /* generate a symbolic constant pointing to the count array */
464 sym.entity_p = bblock_counts;
465 wd.symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
467 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
468 start_block = get_irg_start_block(irg);
469 env.end_block = get_irg_end_block(irg);
470 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
471 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
472 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
473 ir_node *bb = get_Block_cfgpred_block(endbb, i);
477 switch (get_irn_opcode(node)) {
479 ins[0] = get_irn_link(bb);
480 ins[1] = get_Return_mem(node);
481 sync = new_r_Sync(bb, 2, ins);
482 set_Return_mem(node, sync);
485 ins[0] = get_irn_link(bb);
486 ins[1] = get_Raise_mem(node);
487 sync = new_r_Sync(bb, 2, ins);
488 set_Raise_mem(node, sync);
491 /* a fragile's op exception. There should be another path to End,
493 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
497 set_array_entity_values(bblock_id, tarval_array, n_blocks);
499 if (flags & profile_with_locations) {
500 /* build the initializer for the locations */
501 rem = current_ir_graph;
502 current_ir_graph = get_const_code_irg();
503 ent = get_array_element_entity(loc_type);
504 set_entity_variability(ent_locations, variability_constant);
505 for (i = 0; i < n_blocks; ++i) {
506 compound_graph_path *path;
511 path = new_compound_graph_path(loc_type, 2);
512 set_compound_graph_path_array_index(path, 0, i);
513 set_compound_graph_path_node(path, 0, ent);
514 set_compound_graph_path_node(path, 1, loc_lineno);
515 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
516 add_compound_ent_value_w_path(ent_locations, new_Const(tv), path);
519 path = new_compound_graph_path(loc_type, 2);
520 set_compound_graph_path_array_index(path, 0, i);
521 set_compound_graph_path_node(path, 0, ent);
522 set_compound_graph_path_node(path, 1, loc_name);
523 if (wd.locs[i].fname) {
524 sym.entity_p = wd.locs[i].fname;
525 n = new_SymConst(mode_P_data, sym, symconst_addr_ent);
527 n = new_Const(get_mode_null(mode_P_data));
529 add_compound_ent_value_w_path(ent_locations, n, path);
531 pmap_destroy(wd.fname_map);
533 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
537 profile_node_info(void *ctx, FILE *f, const ir_node *irn)
541 fprintf(f, "profiled execution count: %u\n", ir_profile_get_block_execcount(irn));
546 register_vcg_hook(void)
548 memset(&hook, 0, sizeof(hook));
549 hook.hook._hook_node_info = profile_node_info;
550 register_hook(hook_node_info, &hook);
554 unregister_vcg_hook(void)
556 unregister_hook(hook_node_info, &hook);
560 * Reads the corresponding profile info file if it exists and returns a
561 * profile info struct
564 ir_profile_read(const char *filename)
570 f = fopen(filename, "r");
574 printf("found profile data '%s'.\n", filename);
577 ret = fread(buf, 8, 1, f);
578 if(ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
582 if(profile) ir_profile_free();
583 profile = new_set(cmp_execcount, 16);
587 ret = fread(&query, sizeof(unsigned int), 2, f);
591 set_insert(profile, &query, sizeof(query), query.block);
599 * Frees the profile info
602 ir_profile_free(void)
605 unregister_vcg_hook();
611 * Tells whether profile module has acquired data
614 ir_profile_has_data(void)
616 return (profile != NULL);
620 * Get block execution count as determined be profiling
623 ir_profile_get_block_execcount(const ir_node *block)
625 execcount_t *ec, query;
630 query.block = get_irn_node_nr(block);
631 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
636 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
642 typedef struct _intialize_execfreq_env_t {
644 ir_exec_freq *execfreqs;
646 } initialize_execfreq_env_t;
648 // minimal execution frequency (an execfreq of 0 confuses algos)
649 static const double MIN_EXECFREQ = 0.00001;
651 static void initialize_execfreq(ir_node *block, void *data) {
652 initialize_execfreq_env_t *env = data;
655 if(block == get_irg_start_block(env->irg)
656 || block == get_irg_end_block(env->irg)) {
659 freq = ir_profile_get_block_execcount(block);
660 freq *= env->freq_factor;
661 if(freq < MIN_EXECFREQ)
665 set_execfreq(env->execfreqs, block, freq);
668 ir_exec_freq *ir_create_execfreqs_from_profile(ir_graph *irg)
670 ir_node *start_block;
671 initialize_execfreq_env_t env;
675 env.execfreqs = create_execfreq(irg);
676 start_block = get_irg_start_block(irg);
678 count = ir_profile_get_block_execcount(start_block);
680 // the function was never executed, so fallback to estimated freqs
681 free_execfreq(env.execfreqs);
683 return compute_execfreq(irg, 10);
686 env.freq_factor = 1.0 / count;
687 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
689 return env.execfreqs;