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 ir_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)
79 const execcount_t *ea = a;
80 const execcount_t *eb = b;
82 return ea->block != eb->block;
86 * Block walker, count number of blocks.
88 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)
100 unsigned int count = 0;
102 irg_block_walk_graph(irg, block_counter, NULL, &count);
106 /* keep the execcounts here because they are only read once per compiler run */
107 static set * profile = NULL;
108 static hook_entry_t hook;
111 * Instrument a block with code needed for profiling
113 static void instrument_block(ir_node *bb, ir_node *address, unsigned int id)
115 ir_graph *irg = get_irn_irg(bb);
116 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
120 * We can't instrument the end block as there are no real instructions there
122 if (bb == get_irg_end_block(irg))
125 unknown = new_r_Unknown(irg, mode_M);
126 cnst = new_r_Const_long(irg, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
127 offset = new_r_Add(bb, address, cnst, get_modeP_data());
128 load = new_r_Load(bb, unknown, offset, mode_Iu, 0);
129 projm = new_r_Proj(load, mode_M, pn_Load_M);
130 proji = new_r_Proj(load, mode_Iu, pn_Load_res);
131 cnst = new_r_Const_long(irg, mode_Iu, 1);
132 add = new_r_Add(bb, proji, cnst, mode_Iu);
133 store = new_r_Store(bb, projm, offset, add, 0);
134 projm = new_r_Proj(store, mode_M, pn_Store_M);
135 set_irn_link(bb, projm);
136 set_irn_link(projm, load);
139 typedef struct fix_env {
144 * SSA Construction for instrumentation code memory
146 static void fix_ssa(ir_node *bb, void *data)
150 int arity = get_Block_n_cfgpreds(bb);
152 /* end block are not instrumented, skip! */
153 if (bb == env->end_block)
156 if (bb == get_irg_start_block(get_irn_irg(bb))) {
157 mem = get_irg_initial_mem(get_irn_irg(bb));
158 } else if (arity == 1) {
159 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
164 NEW_ARR_A(ir_node*, ins, arity);
165 for (n = arity - 1; n >= 0; --n) {
166 ins[n] = get_irn_link(get_Block_cfgpred_block(bb, n));
168 mem = new_r_Phi(bb, arity, ins, mode_M);
170 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
173 static void add_constructor(ir_entity *method)
175 ir_type *method_type = get_entity_type(method);
176 ir_type *ptr_type = new_type_pointer(method_type);
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_linkage(ptr, IR_LINKAGE_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); }
196 static ir_graph *gen_initializer_irg(ir_entity *ent_filename, ir_entity *bblock_id, ir_entity *bblock_counts, int n_blocks)
199 ident *name = new_id_from_str("__firmprof_initializer");
200 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(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(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(mode_Iu);
215 uintptr = new_type_pointer(uint);
216 string = new_type_pointer(new_type_primitive(mode_Bs));
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);
228 set_type_state(empty_frame_type, layout_fixed);
230 bb = get_cur_block();
232 sym.entity_p = init_ent;
233 symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
235 sym.entity_p = ent_filename;
236 ins[0] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
237 sym.entity_p = bblock_id;
238 ins[1] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
239 sym.entity_p = bblock_counts;
240 ins[2] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
241 ins[3] = new_r_Const_long(irg, mode_Iu, n_blocks);
243 call = new_r_Call(bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
244 ret = new_r_Return(bb, new_r_Proj(call, mode_M, pn_Call_M), 0, NULL);
247 add_immBlock_pred(get_irg_end_block(irg), ret);
248 mature_immBlock(get_irg_end_block(irg));
250 irg_finalize_cons(irg);
252 add_constructor(ent);
258 * Create the location data for the given debug info.
260 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
263 const char *fname = ir_retrieve_dbg_info(dbg, &lineno);
266 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
270 static unsigned nr = 0;
274 int i, len = strlen(fname) + 1;
275 ir_tarval **tarval_string;
277 snprintf(buf, sizeof(buf), "firm_name_arr.%u", nr);
278 arr = new_type_array(1, wd->tp_char);
279 set_array_bounds_int(arr, 0, 0, len);
281 snprintf(buf, sizeof(buf), "__firm_name.%u", nr++);
282 id = new_id_from_str(buf);
283 ent = new_entity(get_glob_type(), id, arr);
284 set_entity_ld_ident(ent, id);
286 pmap_insert(wd->fname_map, (void *)fname, ent);
288 /* initialize file name string constant */
289 tarval_string = ALLOCAN(ir_tarval*, len);
290 for (i = 0; i < len; ++i) {
291 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
293 set_entity_linkage(ent, IR_LINKAGE_CONSTANT);
294 set_array_entity_values(ent, tarval_string, len);
298 wd->locs[wd->id].fname = ent;
299 wd->locs[wd->id].lineno = lineno;
301 wd->locs[wd->id].fname = NULL;
302 wd->locs[wd->id].lineno = 0;
307 * Walker: assigns an ID to every block.
308 * Builds the string table
310 static void 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)
326 ir_graph *ir_profile_instrument(const char *filename, unsigned flags)
330 ir_entity *bblock_id;
331 ir_entity *bblock_counts;
332 ir_entity *ent_filename;
333 ir_entity *ent_locations = NULL;
334 ir_entity *loc_lineno = NULL;
335 ir_entity *loc_name = NULL;
339 ir_type *string_type;
340 ir_type *character_type;
341 ir_type *loc_type = NULL;
342 ir_type *charptr_type;
344 ir_tarval **tarval_array;
345 ir_tarval **tarval_string;
347 int filename_len = strlen(filename)+1;
349 unsigned align_l, align_n, size;
350 block_id_walker_data_t wd;
353 /* count the number of block first */
354 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
355 ir_graph *irg = get_irp_irg(n);
357 n_blocks += count_blocks(irg);
360 /* create all the necessary types and entities. Note that the
361 types must have a fixed layout, because we already running in the
363 uint_type = new_type_primitive(mode_Iu);
364 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
366 array_type = new_type_array(1, uint_type);
367 set_array_bounds_int(array_type, 0, 0, n_blocks);
368 set_type_size_bytes(array_type, n_blocks * get_mode_size_bytes(mode_Iu));
369 set_type_alignment_bytes(array_type, get_mode_size_bytes(mode_Iu));
370 set_type_state(array_type, layout_fixed);
372 character_type = new_type_primitive(mode_Bs);
373 string_type = new_type_array(1, character_type);
374 set_array_bounds_int(string_type, 0, 0, filename_len);
375 set_type_size_bytes(string_type, filename_len);
376 set_type_alignment_bytes(string_type, 1);
377 set_type_state(string_type, layout_fixed);
379 gtp = get_glob_type();
381 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
382 bblock_id = new_entity(gtp, cur_ident, array_type);
383 set_entity_ld_ident(bblock_id, cur_ident);
385 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
386 bblock_counts = new_entity(gtp, cur_ident, array_type);
387 set_entity_ld_ident(bblock_counts, cur_ident);
389 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
390 ent_filename = new_entity(gtp, cur_ident, string_type);
391 set_entity_ld_ident(ent_filename, cur_ident);
393 if (flags & profile_with_locations) {
394 loc_type = new_type_struct(IDENT("__location"));
395 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
396 align_l = get_type_alignment_bytes(uint_type);
397 size = get_type_size_bytes(uint_type);
398 set_entity_offset(loc_lineno, 0);
400 charptr_type = new_type_pointer(character_type);
401 align_n = get_type_size_bytes(charptr_type);
402 set_type_alignment_bytes(charptr_type, align_n);
403 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
404 size = (size + align_n - 1) & ~(align_n - 1);
405 set_entity_offset(loc_name, size);
408 if (align_n > align_l)
410 size = (size + align_l - 1) & ~(align_l - 1);
411 set_type_size_bytes(loc_type, size);
412 set_type_state(loc_type, layout_fixed);
414 loc_type = new_type_array(1, loc_type);
415 set_array_bounds_int(string_type, 0, 0, n_blocks);
417 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
418 ent_locations = new_entity(gtp, cur_ident, loc_type);
419 set_entity_ld_ident(ent_locations, cur_ident);
422 /* initialize count array */
423 NEW_ARR_A(ir_tarval *, tarval_array, n_blocks);
424 tv = get_mode_null(mode_Iu);
425 for (i = 0; i < n_blocks; ++i) {
426 tarval_array[i] = tv;
428 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
430 /* initialize function name string constant */
431 tarval_string = ALLOCAN(ir_tarval*, filename_len);
432 for (i = 0; i < filename_len; ++i) {
433 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
435 set_entity_linkage(ent_filename, IR_LINKAGE_CONSTANT);
436 set_array_entity_values(ent_filename, tarval_string, filename_len);
438 /* initialize block id array and instrument blocks */
439 wd.array = tarval_array;
441 wd.tp_char = character_type;
443 if (flags & profile_with_locations) {
444 wd.fname_map = pmap_create();
445 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
448 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
449 ir_graph *irg = get_irp_irg(i);
450 ir_node *endbb = get_irg_end_block(irg);
454 /* generate a symbolic constant pointing to the count array */
455 sym.entity_p = bblock_counts;
456 wd.symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
458 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
459 env.end_block = get_irg_end_block(irg);
460 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
461 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
462 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
463 ir_node *bb = get_Block_cfgpred_block(endbb, i);
467 switch (get_irn_opcode(node)) {
469 ins[0] = get_irn_link(bb);
470 ins[1] = get_Return_mem(node);
471 sync = new_r_Sync(bb, 2, ins);
472 set_Return_mem(node, sync);
475 ins[0] = get_irn_link(bb);
476 ins[1] = get_Raise_mem(node);
477 sync = new_r_Sync(bb, 2, ins);
478 set_Raise_mem(node, sync);
481 /* a fragile's op exception. There should be another path to End,
483 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
487 set_array_entity_values(bblock_id, tarval_array, n_blocks);
489 if (flags & profile_with_locations) {
490 ir_graph *irg = get_const_code_irg();
491 /* build the initializer for the locations */
492 ent = get_array_element_entity(loc_type);
493 set_entity_linkage(ent_locations, IR_LINKAGE_CONSTANT);
494 for (i = 0; i < n_blocks; ++i) {
495 compound_graph_path *path;
500 path = new_compound_graph_path(loc_type, 2);
501 set_compound_graph_path_array_index(path, 0, i);
502 set_compound_graph_path_node(path, 0, ent);
503 set_compound_graph_path_node(path, 1, loc_lineno);
504 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
505 add_compound_ent_value_w_path(ent_locations, new_r_Const(irg, tv), path);
508 path = new_compound_graph_path(loc_type, 2);
509 set_compound_graph_path_array_index(path, 0, i);
510 set_compound_graph_path_node(path, 0, ent);
511 set_compound_graph_path_node(path, 1, loc_name);
512 if (wd.locs[i].fname) {
513 sym.entity_p = wd.locs[i].fname;
514 n = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
516 n = new_r_Const(irg, get_mode_null(mode_P_data));
518 add_compound_ent_value_w_path(ent_locations, n, path);
520 pmap_destroy(wd.fname_map);
522 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
525 static void profile_node_info(void *ctx, FILE *f, const ir_node *irn)
529 fprintf(f, "profiled execution count: %u\n", ir_profile_get_block_execcount(irn));
533 static void register_vcg_hook(void)
535 memset(&hook, 0, sizeof(hook));
536 hook.hook._hook_node_info = profile_node_info;
537 register_hook(hook_node_info, &hook);
540 static void 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
549 void ir_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) ir_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
586 void ir_profile_free(void)
589 unregister_vcg_hook();
595 * Tells whether profile module has acquired data
597 int ir_profile_has_data(void)
599 return (profile != NULL);
603 * Get block execution count as determined be profiling
606 ir_profile_get_block_execcount(const ir_node *block)
608 execcount_t *ec, query;
613 query.block = get_irn_node_nr(block);
614 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
619 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
625 typedef struct intialize_execfreq_env_t {
627 ir_exec_freq *execfreqs;
629 } initialize_execfreq_env_t;
631 // minimal execution frequency (an execfreq of 0 confuses algos)
632 static const double MIN_EXECFREQ = 0.00001;
634 static void initialize_execfreq(ir_node *block, void *data)
636 initialize_execfreq_env_t *env = data;
639 if (block == get_irg_start_block(env->irg)
640 || block == get_irg_end_block(env->irg)) {
643 freq = ir_profile_get_block_execcount(block);
644 freq *= env->freq_factor;
645 if (freq < MIN_EXECFREQ)
649 set_execfreq(env->execfreqs, block, freq);
652 ir_exec_freq *ir_create_execfreqs_from_profile(ir_graph *irg)
654 ir_node *start_block;
655 initialize_execfreq_env_t env;
659 env.execfreqs = create_execfreq(irg);
660 start_block = get_irg_start_block(irg);
662 count = ir_profile_get_block_execcount(start_block);
664 // the function was never executed, so fallback to estimated freqs
665 free_execfreq(env.execfreqs);
667 return compute_execfreq(irg, 10);
670 env.freq_factor = 1.0 / count;
671 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
673 return env.execfreqs;