1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski
7 * Code instrumentation and execution count profiling
9 * Copyright (C) 2006 Universitaet Karlsruhe
10 * Released under the GPL
39 #include "besched_t.h"
49 #include "bechordal_t.h"
52 #include <libcore/lc_opts.h>
53 #include <libcore/lc_opts_enum.h>
54 #endif /* WITH_LIBCORE */
56 #include "beprofile.h"
58 /** An entry in the id-to-location map */
59 typedef struct loc_entry {
60 ir_entity *fname; /**< the entity holding the file name */
61 unsigned int lineno; /**< line number */
64 typedef struct _block_id_walker_data_t {
65 tarval **array; /**< the entity the holds the block counts */
66 unsigned int id; /**< current block id number */
67 ir_node *symconst; /**< the SymConst representing array */
68 pmap *fname_map; /**< set containing all found filenames */
69 loc_entry *locs; /**< locations */
70 ir_type *tp_char; /**< the character type */
71 unsigned flags; /**< profile flags */
72 } block_id_walker_data_t;
74 typedef struct _execcount_t {
80 cmp_execcount(const void * a, const void * b, size_t size)
82 return ((execcount_t*)a)->block != ((execcount_t*)b)->block;
86 block_counter(ir_node * bb, void * data)
88 unsigned int *count = data;
93 count_blocks(ir_graph * irg)
95 unsigned int count = 0;
97 irg_block_walk_graph(irg, block_counter, NULL, &count);
101 /* keep the execcounts here because they are only read once per compiler run */
102 static set * profile = NULL;
103 static hook_entry_t hook;
106 * Instrument a block with code needed for profiling
109 instrument_block(ir_node * bb, ir_node * address, unsigned int id)
111 ir_graph *irg = get_irn_irg(bb);
112 ir_node *start_block = get_irg_start_block(irg);
113 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown;
117 * We can't instrument the start and end block as there are no real
118 * instructions in these blocks
120 if(bb == start_block || bb == get_irg_end_block(irg))
123 unknown = new_r_Unknown(irg, mode_M);
124 cnst = new_r_Const_long(irg, start_block, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
125 offset = new_r_Add(irg, bb, address, cnst, mode_P);
126 load = new_r_Load(irg, bb, unknown, offset, mode_Iu);
127 projm = new_r_Proj(irg, bb, load, mode_M, pn_Load_M);
128 proji = new_r_Proj(irg, bb, load, mode_Iu, pn_Load_res);
129 cnst = new_r_Const_long(irg, start_block, mode_Iu, 1);
130 add = new_r_Add(irg, bb, proji, cnst, mode_Iu);
131 store = new_r_Store(irg, bb, projm, offset, add);
132 projm = new_r_Proj(irg, bb, store, mode_M, pn_Store_M);
133 set_irn_link(bb, projm);
134 set_irn_link(projm, load);
137 typedef struct fix_env {
138 ir_node *start_block;
143 * SSA Construction for instrumentation code memory
146 fix_ssa(ir_node * bb, void * data)
150 int arity = get_Block_n_cfgpreds(bb);
152 /* start and end block are not instrumented, skip! */
153 if (bb == env->start_block || bb == env->end_block)
157 mem = get_irn_link(get_Block_cfgpred_block(bb, 0));
161 ir_graph *irg = current_ir_graph;
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(irg, bb, arity, ins, mode_M);
169 set_Load_mem(get_irn_link(get_irn_link(bb)), mem);
174 * Generates a new irg which calls the initializer
177 * void __firmprof_initializer(void) { __init_firmprof(ent_filename, bblock_id, bblock_counts, n_blocks); }
180 gen_initializer_irg(ir_entity * ent_filename, ir_entity * bblock_id, ir_entity * bblock_counts, int n_blocks)
182 ir_node *start_block;
185 ident *name = new_id_from_str("__firmprof_initializer");
186 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(name, 0, 0));
187 ir_node *ret, *call, *symconst;
190 ident *init_name = new_id_from_str("__init_firmprof");
191 ir_type *init_type = new_type_method(init_name, 4, 0);
192 ir_type *uint, *uintptr, *string;
196 ir_type *empty_frame_type;
198 set_entity_ld_ident(ent, name);
200 uint = new_type_primitive(new_id_from_str("__uint"), mode_Iu);
201 uintptr = new_type_pointer(new_id_from_str("__uintptr"), uint, mode_P);
202 string = new_type_pointer(new_id_from_str("__charptr"), new_type_primitive(new_id_from_str("__char"), mode_Bs), mode_P);
204 set_method_param_type(init_type, 0, string);
205 set_method_param_type(init_type, 1, uintptr);
206 set_method_param_type(init_type, 2, uintptr);
207 set_method_param_type(init_type, 3, uint);
208 init_ent = new_entity(get_glob_type(), init_name, init_type);
209 set_entity_ld_ident(init_ent, init_name);
211 irg = new_ir_graph(ent, 0);
212 empty_frame_type = get_irg_frame_type(irg);
213 set_type_size_bytes(empty_frame_type, 0);
215 bb = get_cur_block();
217 start_block = get_irg_start_block(irg);
219 sym.entity_p = init_ent;
220 symconst = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
222 sym.entity_p = ent_filename;
223 ins[0] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
224 sym.entity_p = bblock_id;
225 ins[1] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
226 sym.entity_p = bblock_counts;
227 ins[2] = new_r_SymConst(irg, start_block, sym, symconst_addr_ent);
228 ins[3] = new_r_Const_long(irg, start_block, mode_Iu, n_blocks);
230 call = new_r_Call(irg, bb, get_irg_initial_mem(irg), symconst, 4, ins, init_type);
231 ret = new_r_Return(irg, bb, new_r_Proj(irg, bb, call, mode_M, pn_Call_M_regular), 0, NULL);
234 add_immBlock_pred(get_irg_end_block(irg), ret);
235 mature_immBlock(get_irg_end_block(irg));
237 irg_finalize_cons(irg);
243 * Create the location data for the given debug info.
245 static void create_location_data(dbg_info *dbg, block_id_walker_data_t *wd)
248 const char *fname = be_retrieve_dbg_info(dbg, &lineno);
251 pmap_entry *entry = pmap_find(wd->fname_map, (void *)fname);
255 static unsigned nr = 0;
259 int i, len = strlen(fname) + 1;
260 tarval **tarval_string;
262 snprintf(buf, sizeof(buf), "firm_name_arr.%d", nr);
263 arr = new_type_array(new_id_from_str(buf), 1, wd->tp_char);
264 set_array_bounds_int(arr, 0, 0, len);
266 snprintf(buf, sizeof(buf), "__firm_name.%d", nr++);
267 id = new_id_from_str(buf);
268 ent = new_entity(get_glob_type(), id, arr);
269 set_entity_ld_ident(ent, id);
271 pmap_insert(wd->fname_map, (void *)fname, ent);
273 /* initialize file name string constant */
274 tarval_string = alloca(sizeof(*tarval_string) * (len));
275 for (i = 0; i < len; ++i) {
276 tarval_string[i] = new_tarval_from_long(fname[i], mode_Bs);
278 set_entity_variability(ent, variability_constant);
279 set_array_entity_values(ent, tarval_string, len);
283 wd->locs[wd->id].fname = ent;
284 wd->locs[wd->id].lineno = lineno;
286 wd->locs[wd->id].fname = NULL;
287 wd->locs[wd->id].lineno = 0;
292 * Walker: assigns an ID to every block.
293 * Builds the string table
296 block_id_walker(ir_node * bb, void * data)
298 block_id_walker_data_t *wd = data;
300 wd->array[wd->id] = new_tarval_from_long(get_irn_node_nr(bb), mode_Iu);
301 instrument_block(bb, wd->symconst, wd->id);
303 if (wd->flags & profile_with_locations) {
304 dbg_info *dbg = get_irn_dbg_info(bb);
305 create_location_data(dbg, wd);
310 #define IDENT(x) new_id_from_chars(x, sizeof(x) - 1)
313 be_profile_instrument(const char *filename, unsigned flags)
316 unsigned int n_blocks = 0;
317 ir_entity *bblock_id;
318 ir_entity *bblock_counts;
319 ir_entity *ent_filename;
320 ir_entity *ent_locations = NULL;
321 ir_entity *loc_lineno = NULL;
322 ir_entity *loc_name = NULL;
326 ir_type *string_type;
327 ir_type *character_type;
328 ir_type *loc_type = NULL;
329 ir_type *charptr_type;
331 tarval **tarval_array;
332 tarval **tarval_string;
334 int filename_len = strlen(filename)+1;
336 int align_l, align_n, size;
338 block_id_walker_data_t wd;
341 /* count the number of block first */
342 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
343 ir_graph *irg = get_irp_irg(n);
345 n_blocks += count_blocks(irg);
348 /* create all the necessary types and entities. Note that the
349 types must have a fixed layout, because we already running in the
351 uint_type = new_type_primitive(IDENT("__uint"), mode_Iu);
352 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
353 array_type = new_type_array(IDENT("__block_info_array"), 1, uint_type);
354 set_array_bounds_int(array_type, 0, 0, n_blocks);
356 character_type = new_type_primitive(IDENT("__char"), mode_Bs);
357 string_type = new_type_array(IDENT("__filename"), 1, character_type);
358 set_array_bounds_int(string_type, 0, 0, filename_len);
360 gtp = get_glob_type();
362 cur_ident = IDENT("__FIRMPROF__BLOCK_IDS");
363 bblock_id = new_entity(gtp, cur_ident, array_type);
364 set_entity_ld_ident(bblock_id, cur_ident);
365 set_entity_variability(bblock_id, variability_initialized);
367 cur_ident = IDENT("__FIRMPROF__BLOCK_COUNTS");
368 bblock_counts = new_entity(gtp, cur_ident, array_type);
369 set_entity_ld_ident(bblock_counts, cur_ident);
370 set_entity_variability(bblock_counts, variability_initialized);
372 cur_ident = IDENT("__FIRMPROF__FILE_NAME");
373 ent_filename = new_entity(gtp, cur_ident, string_type);
374 set_entity_ld_ident(ent_filename, cur_ident);
376 if (flags & profile_with_locations) {
377 loc_type = new_type_struct(IDENT("__location"));
378 loc_lineno = new_entity(loc_type, IDENT("lineno"), uint_type);
379 align_l = get_type_alignment_bytes(uint_type);
380 size = get_type_size_bytes(uint_type);
381 set_entity_offset(loc_lineno, 0);
383 charptr_type = new_type_pointer(IDENT("__charptr"), character_type, mode_P_data);
384 align_n = get_type_size_bytes(charptr_type);
385 set_type_alignment_bytes(charptr_type, align_n);
386 loc_name = new_entity(loc_type, IDENT("name"), charptr_type);
387 size = (size + align_n - 1) & -align_n;
388 set_entity_offset(loc_name, size);
391 if (align_n > align_l)
393 size = (size + align_l - 1) & -align_l;
394 set_type_size_bytes(loc_type, size);
395 set_type_state(loc_type, layout_fixed);
397 loc_type = new_type_array(IDENT("__locarray"), 1, loc_type);
398 set_array_bounds_int(string_type, 0, 0, n_blocks);
400 cur_ident = IDENT("__FIRMPROF__LOCATIONS");
401 ent_locations = new_entity(gtp, cur_ident, loc_type);
402 set_entity_ld_ident(ent_locations, cur_ident);
405 /* initialize count array */
406 NEW_ARR_A(tarval *, tarval_array, n_blocks);
407 tv = get_tarval_null(mode_Iu);
408 for (i = 0; i < n_blocks; ++i) {
409 tarval_array[i] = tv;
411 set_array_entity_values(bblock_counts, tarval_array, n_blocks);
413 /* initialize function name string constant */
414 tarval_string = alloca(sizeof(*tarval_string) * (filename_len));
415 for (i = 0; i < filename_len; ++i) {
416 tarval_string[i] = new_tarval_from_long(filename[i], mode_Bs);
418 set_entity_variability(ent_filename, variability_constant);
419 set_array_entity_values(ent_filename, tarval_string, filename_len);
421 /* initialize block id array and instrument blocks */
422 wd.array = tarval_array;
424 wd.tp_char = character_type;
426 if (flags & profile_with_locations) {
427 wd.fname_map = pmap_create();
428 NEW_ARR_A(loc_entry, wd.locs, n_blocks);
431 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
432 ir_graph *irg = get_irp_irg(n);
434 ir_node *endbb = get_irg_end_block(irg);
437 set_current_ir_graph(irg);
439 /* generate a symbolic constant pointing to the count array */
440 sym.entity_p = bblock_counts;
441 wd.symconst = new_r_SymConst(irg, get_irg_start_block(irg), sym, symconst_addr_ent);
443 irg_block_walk_graph(irg, block_id_walker, NULL, &wd);
444 env.start_block = get_irg_start_block(irg);
445 env.end_block = get_irg_end_block(irg);
446 set_irn_link(env.start_block, get_irg_no_mem(irg));
447 irg_block_walk_graph(irg, fix_ssa, NULL, &env);
448 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
449 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
450 ir_node *bb = get_Block_cfgpred_block(endbb, i);
454 switch (get_irn_opcode(node)) {
456 ins[0] = get_irn_link(bb);
457 ins[1] = get_Return_mem(node);
458 sync = new_r_Sync(irg, bb, 2, ins);
459 set_Return_mem(node, sync);
462 ins[0] = get_irn_link(bb);
463 ins[1] = get_Raise_mem(node);
464 sync = new_r_Sync(irg, bb, 2, ins);
465 set_Raise_mem(node, sync);
468 /* a fragile's op exception. There should be another path to End,
470 assert(is_fragile_op(node) && "unexpected End control flow predecessor");
474 set_array_entity_values(bblock_id, tarval_array, n_blocks);
476 if (flags & profile_with_locations) {
477 /* build the initializer for the locations */
478 rem = current_ir_graph;
479 current_ir_graph = get_const_code_irg();
480 ent = get_array_element_entity(loc_type);
481 set_entity_variability(ent_locations, variability_constant);
482 for (i = 0; i < n_blocks; ++i) {
483 compound_graph_path *path;
488 path = new_compound_graph_path(loc_type, 2);
489 set_compound_graph_path_array_index(path, 0, i);
490 set_compound_graph_path_node(path, 0, ent);
491 set_compound_graph_path_node(path, 1, loc_lineno);
492 tv = new_tarval_from_long(wd.locs[i].lineno, mode_Iu);
493 add_compound_ent_value_w_path(ent_locations, new_Const(mode_Iu, tv), path);
496 path = new_compound_graph_path(loc_type, 2);
497 set_compound_graph_path_array_index(path, 0, i);
498 set_compound_graph_path_node(path, 0, ent);
499 set_compound_graph_path_node(path, 1, loc_name);
500 if (wd.locs[i].fname) {
501 sym.entity_p = wd.locs[i].fname;
502 n = new_SymConst(sym, symconst_addr_ent);
504 n = new_Const(mode_P_data, get_mode_null(mode_P_data));
506 add_compound_ent_value_w_path(ent_locations, n, path);
508 pmap_destroy(wd.fname_map);
510 return gen_initializer_irg(ent_filename, bblock_id, bblock_counts, n_blocks);
514 profile_node_info(void *ctx, FILE *f, const ir_node *irn)
517 fprintf(f, "profiled execution count: %u\n", be_profile_get_block_execcount(irn));
522 register_vcg_hook(void)
524 memset(&hook, 0, sizeof(hook));
525 hook.hook._hook_node_info = profile_node_info;
526 register_hook(hook_node_info, &hook);
530 unregister_vcg_hook(void)
532 unregister_hook(hook_node_info, &hook);
536 * Reads the corresponding profile info file if it exists and returns a
537 * profile info struct
540 be_profile_read(const char *filename)
546 f = fopen(filename, "r");
550 printf("found profile data '%s'.\n", filename);
553 ret = fread(buf, 8, 1, f);
554 if(ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
558 if(profile) be_profile_free();
559 profile = new_set(cmp_execcount, 16);
563 ret = fread(&query, sizeof(unsigned int), 2, f);
567 set_insert(profile, &query, sizeof(query), query.block);
575 * Frees the profile info
578 be_profile_free(void)
581 unregister_vcg_hook();
587 * Tells whether profile module has acquired data
590 be_profile_has_data(void)
592 return (profile != NULL);
596 * Get block execution count as determined be profiling
599 be_profile_get_block_execcount(const ir_node *block)
601 execcount_t *ec, query;
606 query.block = get_irn_node_nr(block);
607 ec = set_find(profile, &query, sizeof(query), get_irn_node_nr(block));
612 ir_fprintf(stderr, "Warning: Profile contains no data for %+F\n",
618 typedef struct _intialize_execfreq_env_t {
620 ir_exec_freq *execfreqs;
622 } initialize_execfreq_env_t;
624 // minimal execution frequency (an execfreq of 0 confuses algos)
625 static const double MIN_EXECFREQ = 0.00001;
627 static void initialize_execfreq(ir_node *block, void *data) {
628 initialize_execfreq_env_t *env = data;
631 if(block == get_irg_start_block(env->irg)
632 || block == get_irg_end_block(env->irg)) {
635 freq = be_profile_get_block_execcount(block);
636 freq *= env->freq_factor;
637 if(freq < MIN_EXECFREQ)
641 set_execfreq(env->execfreqs, block, freq);
644 ir_exec_freq *be_create_execfreqs_from_profile(ir_graph *irg)
646 ir_node *block2 = NULL;
647 ir_node *start_block;
648 const ir_edge_t *edge;
649 initialize_execfreq_env_t env;
653 env.execfreqs = create_execfreq(irg);
655 // find the successor to the start block
656 start_block = get_irg_start_block(irg);
657 foreach_block_succ(start_block, edge) {
658 ir_node *succ = get_edge_src_irn(edge);
659 if(succ != start_block) {
664 assert(block2 != NULL);
666 count = be_profile_get_block_execcount(block2);
668 // the function was never executed, so fallback to estimated freqs
669 free_execfreq(env.execfreqs);
671 return compute_execfreq(irg, 10);
674 env.freq_factor = 1.0 / count;
675 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
677 return env.execfreqs;