2 * Copyright (C) 1995-2011 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, Steven Schaefer
24 * @date 06.04.2006, 11.11.2010
46 #include "irprofile.h"
48 /* Instrument blocks walker. */
49 typedef struct block_id_walker_data_t {
50 unsigned int id; /**< current block id number */
51 ir_node *symconst; /**< the SymConst representing the counter array */
52 } block_id_walker_data_t;
54 /* Associate counters with blocks. */
55 typedef struct block_assoc_t {
56 unsigned int i; /**< current block id number */
57 unsigned int *counters; /**< block execution counts */
60 typedef struct intialize_execfreq_env_t {
62 ir_exec_freq *execfreqs;
64 } initialize_execfreq_env_t;
66 /* minimal execution frequency (an execfreq of 0 confuses algos) */
67 #define MIN_EXECFREQ 0.00001
69 /* keep the execcounts here because they are only read once per compiler run */
70 static set *profile = NULL;
72 /* Hook for vcg output. */
75 /* The debug module handle. */
76 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
78 /* Since the backend creates a new firm graph we cannot associate counts with
79 * blocks directly. Instead we associate them with the block ids, which are
82 typedef struct execcount_t {
83 unsigned long block; /**< block id */
84 unsigned int count; /**< execution count */
88 * Compare two execcount_t entries.
90 static int cmp_execcount(const void *a, const void *b, size_t size)
92 const execcount_t *ea = (const execcount_t*)a;
93 const execcount_t *eb = (const execcount_t*)b;
95 return ea->block != eb->block;
99 * Block walker, count number of blocks.
101 static void block_counter(ir_node *bb, void *data)
103 unsigned *count = (unsigned*) data;
109 * Returns the number of blocks the given graph.
111 static unsigned int get_irg_n_blocks(ir_graph *irg)
113 unsigned int count = 0;
114 irg_block_walk_graph(irg, block_counter, NULL, &count);
119 * Returns the number of basic blocks in the current ir program.
121 static unsigned int get_irp_n_blocks(void)
123 int i, n = get_irp_n_irgs();
124 unsigned int count = 0;
126 for (i = 0; i < n; i++) {
127 ir_graph *irg = get_irp_irg(i);
128 count += get_irg_n_blocks(irg);
135 static void dump_profile_node_info(void *ctx, FILE *f, const ir_node *irn)
139 unsigned int execcount = ir_profile_get_block_execcount(irn);
140 fprintf(f, "profiled execution count: %u\n", execcount);
145 * Add the given method entity as a constructor.
147 static void add_constructor(ir_entity *method)
149 ir_type *method_type = get_entity_type(method);
150 ir_type *ptr_type = new_type_pointer(method_type);
152 ir_type *constructors = get_segment_type(IR_SEGMENT_CONSTRUCTORS);
153 ident *ide = id_unique("constructor_ptr.%u");
154 ir_entity *ptr = new_entity(constructors, ide, ptr_type);
155 ir_graph *irg = get_const_code_irg();
156 ir_node *val = new_rd_SymConst_addr_ent(NULL, irg, mode_P_code, method);
158 set_entity_ld_ident(ptr, new_id_from_chars("", 0));
159 set_entity_compiler_generated(ptr, 1);
160 set_entity_linkage(ptr, IR_LINKAGE_CONSTANT | IR_LINKAGE_HIDDEN_USER);
161 set_entity_visibility(ptr, ir_visibility_private);
162 set_atomic_ent_value(ptr, val);
166 * Returns an entity representing the __init_firmprof function from libfirmprof
167 * This is the equivalent of:
168 * extern void __init_firmprof(char *filename, uint *counters, uint size)
170 static ir_entity *get_init_firmprof_ref(void)
172 ident *init_name = new_id_from_str("__init_firmprof");
173 ir_type *init_type = new_type_method(3, 0);
174 ir_type *uint = new_type_primitive(mode_Iu);
175 ir_type *uintptr = new_type_pointer(uint);
176 ir_type *string = new_type_pointer(new_type_primitive(mode_Bs));
179 set_method_param_type(init_type, 0, string);
180 set_method_param_type(init_type, 1, uintptr);
181 set_method_param_type(init_type, 2, uint);
183 result = new_entity(get_glob_type(), init_name, init_type);
184 set_entity_visibility(result, ir_visibility_external);
190 * Generates a new irg which calls the initializer
193 * static void __firmprof_initializer(void) __attribute__ ((constructor))
195 * __init_firmprof(ent_filename, bblock_counts, n_blocks);
198 static ir_graph *gen_initializer_irg(ir_entity *ent_filename,
199 ir_entity *bblock_counts, int n_blocks)
203 ir_node *bb, *ret, *call, *symconst;
204 ir_type *empty_frame_type;
207 ir_entity *init_ent = get_init_firmprof_ref();
209 ident *name = new_id_from_str("__firmprof_initializer");
210 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(0, 0));
211 set_entity_visibility(ent, ir_visibility_local);
212 set_entity_ld_ident(ent, name);
214 /* create the new ir_graph */
215 irg = new_ir_graph(ent, 0);
216 set_current_ir_graph(irg);
217 empty_frame_type = get_irg_frame_type(irg);
218 set_type_size_bytes(empty_frame_type, 0);
219 set_type_state(empty_frame_type, layout_fixed);
221 bb = get_r_cur_block(irg);
223 sym.entity_p = init_ent;
224 symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
226 sym.entity_p = ent_filename;
227 ins[0] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
228 sym.entity_p = bblock_counts;
229 ins[1] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
230 ins[2] = new_r_Const_long(irg, mode_Iu, n_blocks);
232 call = new_r_Call(bb, get_irg_initial_mem(irg), symconst, 3, ins,
233 get_entity_type(init_ent));
234 ret = new_r_Return(bb, new_r_Proj(call, mode_M, pn_Call_M), 0, NULL);
237 add_immBlock_pred(get_irg_end_block(irg), ret);
238 mature_immBlock(get_irg_end_block(irg));
240 irg_finalize_cons(irg);
242 /* add a pointer to the new function in the constructor section */
243 add_constructor(ent);
249 * Instrument a block with code needed for profiling.
250 * This just inserts the instruction nodes, it doesn't connect the memory
251 * nodes in a meaningful way.
253 static void instrument_block(ir_node *bb, ir_node *address, unsigned int id)
255 ir_graph *irg = get_irn_irg(bb);
256 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown, *cnst;
258 /* We can't instrument the end block */
259 if (bb == get_irg_end_block(irg))
262 unknown = new_r_Unknown(irg, mode_M);
263 cnst = new_r_Const_long(irg, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
264 offset = new_r_Add(bb, address, cnst, get_modeP_data());
265 load = new_r_Load(bb, unknown, offset, mode_Iu, cons_none);
266 projm = new_r_Proj(load, mode_M, pn_Load_M);
267 proji = new_r_Proj(load, mode_Iu, pn_Load_res);
268 cnst = new_r_Const(irg, get_mode_one(mode_Iu));
269 add = new_r_Add(bb, proji, cnst, mode_Iu);
270 store = new_r_Store(bb, projm, offset, add, cons_none);
271 projm = new_r_Proj(store, mode_M, pn_Store_M);
273 set_irn_link(bb, projm);
274 set_irn_link(projm, load);
278 * SSA Construction for instrumentation code memory.
280 * This introduces a new memory node and connects it to the instrumentation
281 * codes, inserting phiM nodes as necessary. Note that afterwards, the new
282 * memory is not connected to any return nodes and thus still dead.
284 static void fix_ssa(ir_node *bb, void *data)
286 ir_graph *irg = get_irn_irg(bb);
287 ir_node *mem, *proj, *load;
288 int n, arity = get_Block_n_cfgpreds(bb);
292 /* end blocks are not instrumented, skip! */
293 if (bb == get_irg_end_block(irg))
296 if (bb == get_irg_start_block(irg)) {
297 mem = get_irg_initial_mem(irg);
298 } else if (arity == 1) {
299 ir_node *pred = get_Block_cfgpred_block(bb, 0);
301 mem = (ir_node*) get_irn_link(pred);
303 mem = new_r_NoMem(irg);
305 ir_node **ins = ALLOCAN(ir_node*, arity);
306 for (n = arity - 1; n >= 0; --n) {
307 ir_node *pred = get_Block_cfgpred_block(bb, n);
309 ins[n] = (ir_node*) get_irn_link(pred);
311 ins[n] = new_r_NoMem(irg);
313 mem = new_r_Phi(bb, arity, ins, mode_M);
316 /* The block link fields point to the projm from the instrumentation code,
317 * the projm in turn links to the initial load which lacks a memory
318 * argument at this point. */
319 proj = (ir_node*) get_irn_link(bb);
320 load = (ir_node*) get_irn_link(proj);
321 set_Load_mem(load, mem);
325 * Instrument a single block.
327 static void block_instrument_walker(ir_node *bb, void *data)
329 block_id_walker_data_t *wd = (block_id_walker_data_t*)data;
330 instrument_block(bb, wd->symconst, wd->id);
335 * Synchronize the original memory input of node with the additional operand
336 * from the profiling code.
338 static ir_node *sync_mem(ir_node *bb, ir_node *mem)
341 ins[0] = (ir_node*) get_irn_link(bb);
343 return new_r_Sync(bb, 2, ins);
347 * Instrument a single ir_graph, counters should point to the bblock
350 static void instrument_irg(ir_graph *irg, ir_entity *counters,
351 block_id_walker_data_t *wd)
353 ir_node *end = get_irg_end(irg);
354 ir_node *endbb = get_irg_end_block(irg);
357 /* generate a symbolic constant pointing to the count array */
359 sym.entity_p = counters;
360 wd->symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
362 /* instrument each block in the current irg */
363 irg_block_walk_graph(irg, block_instrument_walker, NULL, wd);
364 irg_block_walk_graph(irg, fix_ssa, NULL, NULL);
366 /* connect the new memory nodes to the return nodes */
367 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
368 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
369 ir_node *bb = get_Block_cfgpred_block(endbb, i);
372 switch (get_irn_opcode(node)) {
374 mem = get_Return_mem(node);
375 set_Return_mem(node, sync_mem(bb, mem));
378 mem = get_Raise_mem(node);
379 set_Raise_mem(node, sync_mem(bb, mem));
384 /* A fragile's op exception. There should be another path to End,
387 assert(is_fragile_op(node) && \
388 "unexpected End control flow predecessor");
392 /* as well as calls with attribute noreturn */
393 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
394 ir_node *node = get_End_keepalive(end, i);
396 ir_node *bb = get_nodes_block(node);
397 ir_node *mem = get_Call_mem(node);
398 set_Call_mem(node, sync_mem(bb, mem));
404 * Creates a new entity representing the equivalent of
405 * static unsigned int name[size]
407 static ir_entity *new_array_entity(ident *name, int size)
410 ir_type *uint_type, *array_type;
412 uint_type = new_type_primitive(mode_Iu);
413 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
415 array_type = new_type_array(1, uint_type);
416 set_array_bounds_int(array_type, 0, 0, size);
417 set_type_size_bytes(array_type, size * get_mode_size_bytes(mode_Iu));
418 set_type_alignment_bytes(array_type, get_mode_size_bytes(mode_Iu));
419 set_type_state(array_type, layout_fixed);
421 result = new_entity(get_glob_type(), name, array_type);
422 set_entity_visibility(result, ir_visibility_local);
423 set_entity_compiler_generated(result, 1);
429 * Creates a new entity representing the equivalent of
430 * static const char name[strlen(string)+1] = string
432 static ir_entity *new_static_string_entity(ident *name, const char *string)
436 ir_type *char_type = new_type_primitive(mode_Bs);
437 ir_type *string_type = new_type_array(1, char_type);
439 ir_initializer_t *contents;
441 size_t i, length = strlen(string)+1;
443 /* Create the type for a fixed-length string */
444 set_array_bounds_int(string_type, 0, 0, length);
445 set_type_size_bytes(string_type, length);
446 set_type_alignment_bytes(string_type, 1);
447 set_type_state(string_type, layout_fixed);
449 result = new_entity(get_glob_type(), name, string_type);
450 set_entity_visibility(result, ir_visibility_local);
451 set_entity_linkage(result, IR_LINKAGE_CONSTANT);
452 set_entity_compiler_generated(result, 1);
454 /* There seems to be no simpler way to do this. Or at least, cparser
455 * does exactly the same thing... */
456 contents = create_initializer_compound(length);
457 for (i = 0; i < length; i++) {
458 ir_tarval *c = new_tarval_from_long(string[i], mode_Bs);
459 ir_initializer_t *init = create_initializer_tarval(c);
460 set_initializer_compound_value(contents, i, init);
462 set_entity_initializer(result, contents);
468 * Instrument all ir_graphs in the current ir_program. Currently this only
469 * works for graphs in the backend. Additionally, the resulting program
470 * has to be linked with libfirmprof.
472 * @param filename the name of the profile file (usually module_name.prof)
473 * @returns the module initializer, may be NULL
475 ir_graph *ir_profile_instrument(const char *filename)
478 ident *counter_id, *filename_id;
479 ir_entity *bblock_counts, *ent_filename;
480 block_id_walker_data_t wd;
481 FIRM_DBG_REGISTER(dbg, "firm.ir.profile");
483 /* Don't do anything for modules without code. Else the linker will
485 if (get_irp_n_irgs() == 0)
488 /* count the number of block first */
489 n_blocks = get_irp_n_blocks();
491 /* create all the necessary types and entities. Note that the
492 * types must have a fixed layout, because we are already running in the
494 counter_id = new_id_from_str("__FIRMPROF__BLOCK_COUNTS");
495 bblock_counts = new_array_entity(counter_id, n_blocks);
497 filename_id = new_id_from_str("__FIRMPROF__FILE_NAME");
498 ent_filename = new_static_string_entity(filename_id, filename);
500 /* initialize block id array and instrument blocks */
502 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
503 ir_graph *irg = get_irp_irg(n);
504 instrument_irg(irg, bblock_counts, &wd);
507 return gen_initializer_irg(ent_filename, bblock_counts, n_blocks);
510 static unsigned int *
511 parse_profile(const char *filename, unsigned int num_blocks)
513 unsigned int *result = NULL;
517 FILE *f = fopen(filename, "rb");
519 DBG((dbg, LEVEL_2, "Failed to open profile file (%s)\n", filename));
524 ret = fread(buf, 8, 1, f);
525 if (ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
526 DBG((dbg, LEVEL_2, "Broken fileheader in profile\n"));
530 result = XMALLOCN(unsigned int, num_blocks);
531 if (fread(result, sizeof(unsigned int) * num_blocks, 1, f) < 1) {
532 DBG((dbg, LEVEL_4, "Failed to read counters... (size: %u)\n",
533 sizeof(unsigned int) * num_blocks));
544 * Reads the corresponding profile info file if it exists.
547 static void block_associate_walker(ir_node *bb, void *env)
549 block_assoc_t *b = (block_assoc_t*) env;
552 query.block = get_irn_node_nr(bb);
553 query.count = b->counters[(b->i)++];
554 DBG((dbg, LEVEL_4, "execcount(%+F, %u): %u\n", bb, query.block,
556 set_insert(profile, &query, sizeof(query), query.block);
559 static void irp_associate_blocks(block_assoc_t *env)
562 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
563 ir_graph *irg = get_irp_irg(n);
564 irg_block_walk_graph(irg, block_associate_walker, NULL, env);
568 bool ir_profile_read(const char *filename)
571 FIRM_DBG_REGISTER(dbg, "firm.ir.profile");
574 env.counters = parse_profile(filename, get_irp_n_blocks());
580 profile = new_set(cmp_execcount, 16);
582 irp_associate_blocks(&env);
585 /* register the vcg hook */
586 hook = dump_add_node_info_callback(dump_profile_node_info, NULL);
591 * Frees the profile info
593 void ir_profile_free(void)
601 dump_remove_node_info_callback(hook);
607 * Tells whether profile module has acquired data
609 bool ir_profile_has_data(void)
611 return profile != NULL;
615 * Get block execution count as determined by profiling
617 unsigned int ir_profile_get_block_execcount(const ir_node *block)
619 execcount_t *ec, query;
621 if (!ir_profile_has_data())
624 query.block = get_irn_node_nr(block);
625 ec = (execcount_t*)set_find(profile, &query, sizeof(query), query.block);
631 "Warning: Profile contains no data for %+F\n", block));
636 static void initialize_execfreq(ir_node *block, void *data)
638 initialize_execfreq_env_t *env = (initialize_execfreq_env_t*)data;
641 if (block == get_irg_start_block(env->irg)
642 || block == get_irg_end_block(env->irg)) {
645 freq = ir_profile_get_block_execcount(block);
646 freq *= env->freq_factor;
647 if (freq < MIN_EXECFREQ)
651 set_execfreq(env->execfreqs, block, freq);
654 ir_exec_freq *ir_create_execfreqs_from_profile(ir_graph *irg)
656 ir_node *start_block;
657 initialize_execfreq_env_t env;
661 env.execfreqs = create_execfreq(irg);
663 /* Find the first block containing instructions */
664 start_block = get_irg_start_block(irg);
665 count = ir_profile_get_block_execcount(start_block);
667 /* the function was never executed, so fallback to estimated freqs */
668 free_execfreq(env.execfreqs);
669 return compute_execfreq(irg, 10);
672 env.freq_factor = 1.0 / count;
673 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
675 return env.execfreqs;