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 /* Mach-O does not like labels in the constructor segment, but with ELF
154 * the linker dies horribly if there is no label. */
155 ident *ide = id_unique("constructor_ptr.%u");
156 ir_entity *ptr = new_entity(constructors, ide, ptr_type);
157 ir_graph *irg = get_const_code_irg();
158 ir_node *val = new_rd_SymConst_addr_ent(NULL, irg, mode_P_code, method);
160 set_entity_ld_ident(ptr, new_id_from_chars("", 0));
161 set_entity_compiler_generated(ptr, 1);
162 set_entity_linkage(ptr, IR_LINKAGE_CONSTANT | IR_LINKAGE_HIDDEN_USER);
163 set_entity_visibility(ptr, ir_visibility_private);
164 set_atomic_ent_value(ptr, val);
168 * Returns an entity representing the __init_firmprof function from libfirmprof
169 * This is the equivalent of:
170 * extern void __init_firmprof(char *filename, uint *counters, uint size)
172 static ir_entity *get_init_firmprof_ref(void)
174 ident *init_name = new_id_from_str("__init_firmprof");
175 ir_type *init_type = new_type_method(3, 0);
176 ir_type *uint = new_type_primitive(mode_Iu);
177 ir_type *uintptr = new_type_pointer(uint);
178 ir_type *string = new_type_pointer(new_type_primitive(mode_Bs));
181 set_method_param_type(init_type, 0, string);
182 set_method_param_type(init_type, 1, uintptr);
183 set_method_param_type(init_type, 2, uint);
185 result = new_entity(get_glob_type(), init_name, init_type);
186 set_entity_visibility(result, ir_visibility_external);
192 * Generates a new irg which calls the initializer
195 * static void __firmprof_initializer(void) __attribute__ ((constructor))
197 * __init_firmprof(ent_filename, bblock_counts, n_blocks);
200 static ir_graph *gen_initializer_irg(ir_entity *ent_filename,
201 ir_entity *bblock_counts, int n_blocks)
205 ir_node *bb, *ret, *call, *symconst;
206 ir_type *empty_frame_type;
209 ir_entity *init_ent = get_init_firmprof_ref();
211 ident *name = new_id_from_str("__firmprof_initializer");
212 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(0, 0));
213 set_entity_visibility(ent, ir_visibility_local);
214 set_entity_ld_ident(ent, name);
216 /* create the new ir_graph */
217 irg = new_ir_graph(ent, 0);
218 set_current_ir_graph(irg);
219 empty_frame_type = get_irg_frame_type(irg);
220 set_type_size_bytes(empty_frame_type, 0);
221 set_type_state(empty_frame_type, layout_fixed);
223 bb = get_r_cur_block(irg);
225 sym.entity_p = init_ent;
226 symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
228 sym.entity_p = ent_filename;
229 ins[0] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
230 sym.entity_p = bblock_counts;
231 ins[1] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
232 ins[2] = new_r_Const_long(irg, mode_Iu, n_blocks);
234 call = new_r_Call(bb, get_irg_initial_mem(irg), symconst, 3, ins,
235 get_entity_type(init_ent));
236 ret = new_r_Return(bb, new_r_Proj(call, mode_M, pn_Call_M), 0, NULL);
239 add_immBlock_pred(get_irg_end_block(irg), ret);
240 mature_immBlock(get_irg_end_block(irg));
242 irg_finalize_cons(irg);
244 /* add a pointer to the new function in the constructor section */
245 add_constructor(ent);
251 * Instrument a block with code needed for profiling.
252 * This just inserts the instruction nodes, it doesn't connect the memory
253 * nodes in a meaningful way.
255 static void instrument_block(ir_node *bb, ir_node *address, unsigned int id)
257 ir_graph *irg = get_irn_irg(bb);
258 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown, *cnst;
260 /* We can't instrument the end block */
261 if (bb == get_irg_end_block(irg))
264 unknown = new_r_Unknown(irg, mode_M);
265 cnst = new_r_Const_long(irg, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
266 offset = new_r_Add(bb, address, cnst, get_modeP_data());
267 load = new_r_Load(bb, unknown, offset, mode_Iu, cons_none);
268 projm = new_r_Proj(load, mode_M, pn_Load_M);
269 proji = new_r_Proj(load, mode_Iu, pn_Load_res);
270 cnst = new_r_Const_long(irg, mode_Iu, 1);
271 add = new_r_Add(bb, proji, cnst, mode_Iu);
272 store = new_r_Store(bb, projm, offset, add, cons_none);
273 projm = new_r_Proj(store, mode_M, pn_Store_M);
275 set_irn_link(bb, projm);
276 set_irn_link(projm, load);
280 * SSA Construction for instrumentation code memory.
282 * This introduces a new memory node and connects it to the instrumentation
283 * codes, inserting phiM nodes as necessary. Note that afterwards, the new
284 * memory is not connected to any return nodes and thus still dead.
286 static void fix_ssa(ir_node *bb, void *data)
288 ir_graph *irg = get_irn_irg(bb);
289 ir_node *mem, *proj, *load;
290 int n, arity = get_Block_n_cfgpreds(bb);
294 /* end blocks are not instrumented, skip! */
295 if (bb == get_irg_end_block(irg))
298 if (bb == get_irg_start_block(irg)) {
299 mem = get_irg_initial_mem(irg);
300 } else if (arity == 1) {
301 ir_node *pred = get_Block_cfgpred_block(bb, 0);
303 mem = (ir_node*) get_irn_link(pred);
305 mem = new_r_NoMem(irg);
307 ir_node **ins = ALLOCAN(ir_node*, arity);
308 for (n = arity - 1; n >= 0; --n) {
309 ir_node *pred = get_Block_cfgpred_block(bb, n);
311 ins[n] = (ir_node*) get_irn_link(pred);
313 ins[n] = new_r_NoMem(irg);
315 mem = new_r_Phi(bb, arity, ins, mode_M);
318 /* The block link fields point to the projm from the instrumentation code,
319 * the projm in turn links to the initial load which lacks a memory
320 * argument at this point. */
321 proj = (ir_node*) get_irn_link(bb);
322 load = (ir_node*) get_irn_link(proj);
323 set_Load_mem(load, mem);
327 * Instrument a single block.
329 static void block_instrument_walker(ir_node *bb, void *data)
331 block_id_walker_data_t *wd = (block_id_walker_data_t*)data;
332 instrument_block(bb, wd->symconst, wd->id);
337 * Synchronize the original memory input of node with the additional operand
338 * from the profiling code.
340 static ir_node *sync_mem(ir_node *bb, ir_node *mem)
343 ins[0] = (ir_node*) get_irn_link(bb);
345 return new_r_Sync(bb, 2, ins);
349 * Instrument a single ir_graph, counters should point to the bblock
352 static void instrument_irg(ir_graph *irg, ir_entity *counters,
353 block_id_walker_data_t *wd)
355 ir_node *end = get_irg_end(irg);
356 ir_node *endbb = get_irg_end_block(irg);
359 /* generate a symbolic constant pointing to the count array */
361 sym.entity_p = counters;
362 wd->symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
364 /* instrument each block in the current irg */
365 irg_block_walk_graph(irg, block_instrument_walker, NULL, wd);
366 irg_block_walk_graph(irg, fix_ssa, NULL, NULL);
368 /* connect the new memory nodes to the return nodes */
369 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
370 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
371 ir_node *bb = get_Block_cfgpred_block(endbb, i);
374 switch (get_irn_opcode(node)) {
376 mem = get_Return_mem(node);
377 set_Return_mem(node, sync_mem(bb, mem));
380 mem = get_Raise_mem(node);
381 set_Raise_mem(node, sync_mem(bb, mem));
386 /* A fragile's op exception. There should be another path to End,
389 assert(is_fragile_op(node) && \
390 "unexpected End control flow predecessor");
394 /* as well as calls with attribute noreturn */
395 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
396 ir_node *node = get_End_keepalive(end, i);
398 ir_node *bb = get_nodes_block(node);
399 ir_node *mem = get_Call_mem(node);
400 set_Call_mem(node, sync_mem(bb, mem));
406 * Creates a new entity representing the equivalent of
407 * static unsigned int name[size]
409 static ir_entity *new_array_entity(ident *name, int size)
412 ir_type *uint_type, *array_type;
414 uint_type = new_type_primitive(mode_Iu);
415 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
417 array_type = new_type_array(1, uint_type);
418 set_array_bounds_int(array_type, 0, 0, size);
419 set_type_size_bytes(array_type, size * get_mode_size_bytes(mode_Iu));
420 set_type_alignment_bytes(array_type, get_mode_size_bytes(mode_Iu));
421 set_type_state(array_type, layout_fixed);
423 result = new_entity(get_glob_type(), name, array_type);
424 set_entity_visibility(result, ir_visibility_local);
425 set_entity_compiler_generated(result, 1);
431 * Creates a new entity representing the equivalent of
432 * static const char name[strlen(string)+1] = string
434 static ir_entity *new_static_string_entity(ident *name, const char *string)
438 ir_type *char_type = new_type_primitive(mode_Bs);
439 ir_type *string_type = new_type_array(1, char_type);
441 ir_initializer_t *contents;
443 size_t i, length = strlen(string)+1;
445 /* Create the type for a fixed-length string */
446 set_array_bounds_int(string_type, 0, 0, length);
447 set_type_size_bytes(string_type, length);
448 set_type_alignment_bytes(string_type, 1);
449 set_type_state(string_type, layout_fixed);
451 result = new_entity(get_glob_type(), name, string_type);
452 set_entity_visibility(result, ir_visibility_local);
453 set_entity_linkage(result, IR_LINKAGE_CONSTANT);
454 set_entity_compiler_generated(result, 1);
456 /* There seems to be no simpler way to do this. Or at least, cparser
457 * does exactly the same thing... */
458 contents = create_initializer_compound(length);
459 for (i = 0; i < length; i++) {
460 ir_tarval *c = new_tarval_from_long(string[i], mode_Bs);
461 ir_initializer_t *init = create_initializer_tarval(c);
462 set_initializer_compound_value(contents, i, init);
464 set_entity_initializer(result, contents);
470 * Instrument all ir_graphs in the current ir_program. Currently this only
471 * works for graphs in the backend. Additionally, the resulting program
472 * has to be linked with libfirmprof.
474 * @param filename the name of the profile file (usually module_name.prof)
475 * @returns the module initializer, may be NULL
477 ir_graph *ir_profile_instrument(const char *filename)
480 ident *counter_id, *filename_id;
481 ir_entity *bblock_counts, *ent_filename;
482 block_id_walker_data_t wd;
483 FIRM_DBG_REGISTER(dbg, "firm.ir.profile");
485 /* Don't do anything for modules without code. Else the linker will
487 if (get_irp_n_irgs() == 0)
490 /* count the number of block first */
491 n_blocks = get_irp_n_blocks();
493 /* create all the necessary types and entities. Note that the
494 * types must have a fixed layout, because we are already running in the
496 counter_id = new_id_from_str("__FIRMPROF__BLOCK_COUNTS");
497 bblock_counts = new_array_entity(counter_id, n_blocks);
499 filename_id = new_id_from_str("__FIRMPROF__FILE_NAME");
500 ent_filename = new_static_string_entity(filename_id, filename);
502 /* initialize block id array and instrument blocks */
504 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
505 ir_graph *irg = get_irp_irg(n);
506 instrument_irg(irg, bblock_counts, &wd);
509 return gen_initializer_irg(ent_filename, bblock_counts, n_blocks);
512 static unsigned int *
513 parse_profile(const char *filename, unsigned int num_blocks)
515 unsigned int *result = NULL;
519 FILE *f = fopen(filename, "rb");
521 DBG((dbg, LEVEL_2, "Failed to open profile file (%s)\n", filename));
526 ret = fread(buf, 8, 1, f);
527 if (ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
528 DBG((dbg, LEVEL_2, "Broken fileheader in profile\n"));
532 result = XMALLOCN(unsigned int, num_blocks);
533 if (fread(result, sizeof(unsigned int) * num_blocks, 1, f) < 1) {
534 DBG((dbg, LEVEL_4, "Failed to read counters... (size: %u)\n",
535 sizeof(unsigned int) * num_blocks));
546 * Reads the corresponding profile info file if it exists.
549 static void block_associate_walker(ir_node *bb, void *env)
551 block_assoc_t *b = (block_assoc_t*) env;
554 query.block = get_irn_node_nr(bb);
555 query.count = b->counters[(b->i)++];
556 DBG((dbg, LEVEL_4, "execcount(%+F, %u): %u\n", bb, query.block,
558 set_insert(profile, &query, sizeof(query), query.block);
561 static void irp_associate_blocks(block_assoc_t *env)
564 for (n = get_irp_n_irgs() - 1; n >= 0; --n) {
565 ir_graph *irg = get_irp_irg(n);
566 irg_block_walk_graph(irg, block_associate_walker, NULL, env);
570 bool ir_profile_read(const char *filename)
573 FIRM_DBG_REGISTER(dbg, "firm.ir.profile");
576 env.counters = parse_profile(filename, get_irp_n_blocks());
582 profile = new_set(cmp_execcount, 16);
584 irp_associate_blocks(&env);
587 /* register the vcg hook */
588 hook = dump_add_node_info_callback(dump_profile_node_info, NULL);
593 * Frees the profile info
595 void ir_profile_free(void)
603 dump_remove_node_info_callback(hook);
609 * Tells whether profile module has acquired data
611 bool ir_profile_has_data(void)
613 return profile != NULL;
617 * Get block execution count as determined by profiling
619 unsigned int ir_profile_get_block_execcount(const ir_node *block)
621 execcount_t *ec, query;
623 if (!ir_profile_has_data())
626 query.block = get_irn_node_nr(block);
627 ec = (execcount_t*)set_find(profile, &query, sizeof(query), query.block);
633 "Warning: Profile contains no data for %+F\n", block));
638 static void initialize_execfreq(ir_node *block, void *data)
640 initialize_execfreq_env_t *env = (initialize_execfreq_env_t*)data;
643 if (block == get_irg_start_block(env->irg)
644 || block == get_irg_end_block(env->irg)) {
647 freq = ir_profile_get_block_execcount(block);
648 freq *= env->freq_factor;
649 if (freq < MIN_EXECFREQ)
653 set_execfreq(env->execfreqs, block, freq);
656 ir_exec_freq *ir_create_execfreqs_from_profile(ir_graph *irg)
658 ir_node *start_block;
659 initialize_execfreq_env_t env;
663 env.execfreqs = create_execfreq(irg);
665 /* Find the first block containing instructions */
666 start_block = get_irg_start_block(irg);
667 count = ir_profile_get_block_execcount(start_block);
669 /* the function was never executed, so fallback to estimated freqs */
670 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;