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
42 #include "execfreq_t.h"
43 #include "irprofile.h"
46 /* Instrument blocks walker. */
47 typedef struct block_id_walker_data_t {
48 unsigned int id; /**< current block id number */
49 ir_node *symconst; /**< the SymConst representing the counter array */
50 } block_id_walker_data_t;
52 /* Associate counters with blocks. */
53 typedef struct block_assoc_t {
54 unsigned int i; /**< current block id number */
55 unsigned int *counters; /**< block execution counts */
58 /* minimal execution frequency (an execfreq of 0 confuses algos) */
59 #define MIN_EXECFREQ 0.00001
61 /* keep the execcounts here because they are only read once per compiler run */
62 static set *profile = NULL;
64 /* Hook for vcg output. */
67 /* The debug module handle. */
68 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
70 /* Since the backend creates a new firm graph we cannot associate counts with
71 * blocks directly. Instead we associate them with the block ids, which are
74 typedef struct execcount_t {
75 unsigned long block; /**< block id */
76 uint32_t count; /**< execution count */
80 * Compare two execcount_t entries.
82 static int cmp_execcount(const void *a, const void *b, size_t size)
84 const execcount_t *ea = (const execcount_t*)a;
85 const execcount_t *eb = (const execcount_t*)b;
87 return ea->block != eb->block;
90 uint32_t ir_profile_get_block_execcount(const ir_node *block)
92 execcount_t *ec, query;
94 query.block = get_irn_node_nr(block);
95 ec = set_find(execcount_t, profile, &query, sizeof(query), query.block);
101 "Warning: Profile contains no data for %+F\n", block));
107 * Block walker, count number of blocks.
109 static void block_counter(ir_node *bb, void *data)
111 unsigned *count = (unsigned*) data;
117 * Returns the number of blocks the given graph.
119 static unsigned int get_irg_n_blocks(ir_graph *irg)
121 unsigned int count = 0;
122 irg_block_walk_graph(irg, block_counter, NULL, &count);
127 * Returns the number of basic blocks in the current ir program.
129 static unsigned int get_irp_n_blocks(void)
131 int i, n = get_irp_n_irgs();
132 unsigned int count = 0;
134 for (i = 0; i < n; i++) {
135 ir_graph *irg = get_irp_irg(i);
136 count += get_irg_n_blocks(irg);
143 static void dump_profile_node_info(void *ctx, FILE *f, const ir_node *irn)
147 unsigned int execcount = ir_profile_get_block_execcount(irn);
148 fprintf(f, "profiled execution count: %u\n", execcount);
153 * Add the given method entity as a constructor.
155 static void add_constructor(ir_entity *method)
157 ir_type *method_type = get_entity_type(method);
158 ir_type *ptr_type = new_type_pointer(method_type);
160 ir_type *constructors = get_segment_type(IR_SEGMENT_CONSTRUCTORS);
161 ident *ide = id_unique("constructor_ptr.%u");
162 ir_entity *ptr = new_entity(constructors, ide, ptr_type);
163 ir_graph *irg = get_const_code_irg();
164 ir_node *val = new_rd_SymConst_addr_ent(NULL, irg, mode_P_code, method);
166 set_entity_ld_ident(ptr, new_id_from_chars("", 0));
167 set_entity_compiler_generated(ptr, 1);
168 set_entity_linkage(ptr, IR_LINKAGE_CONSTANT | IR_LINKAGE_HIDDEN_USER);
169 set_entity_visibility(ptr, ir_visibility_private);
170 set_atomic_ent_value(ptr, val);
174 * Returns an entity representing the __init_firmprof function from libfirmprof
175 * This is the equivalent of:
176 * extern void __init_firmprof(char *filename, uint *counters, uint size)
178 static ir_entity *get_init_firmprof_ref(void)
180 ident *init_name = new_id_from_str("__init_firmprof");
181 ir_type *init_type = new_type_method(3, 0);
182 ir_type *uint = new_type_primitive(mode_Iu);
183 ir_type *uintptr = new_type_pointer(uint);
184 ir_type *string = new_type_pointer(new_type_primitive(mode_Bs));
187 set_method_param_type(init_type, 0, string);
188 set_method_param_type(init_type, 1, uintptr);
189 set_method_param_type(init_type, 2, uint);
191 result = new_entity(get_glob_type(), init_name, init_type);
192 set_entity_visibility(result, ir_visibility_external);
198 * Generates a new irg which calls the initializer
201 * static void __firmprof_initializer(void) __attribute__ ((constructor))
203 * __init_firmprof(ent_filename, bblock_counts, n_blocks);
206 static ir_graph *gen_initializer_irg(ir_entity *ent_filename,
207 ir_entity *bblock_counts, int n_blocks)
211 ir_node *bb, *ret, *call, *symconst;
212 ir_type *empty_frame_type;
215 ir_entity *init_ent = get_init_firmprof_ref();
217 ident *name = new_id_from_str("__firmprof_initializer");
218 ir_entity *ent = new_entity(get_glob_type(), name, new_type_method(0, 0));
219 set_entity_visibility(ent, ir_visibility_local);
220 set_entity_ld_ident(ent, name);
222 /* create the new ir_graph */
223 irg = new_ir_graph(ent, 0);
224 set_current_ir_graph(irg);
225 empty_frame_type = get_irg_frame_type(irg);
226 set_type_size_bytes(empty_frame_type, 0);
227 set_type_state(empty_frame_type, layout_fixed);
229 bb = get_r_cur_block(irg);
231 sym.entity_p = init_ent;
232 symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
234 sym.entity_p = ent_filename;
235 ins[0] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
236 sym.entity_p = bblock_counts;
237 ins[1] = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
238 ins[2] = new_r_Const_long(irg, mode_Iu, n_blocks);
240 call = new_r_Call(bb, get_irg_initial_mem(irg), symconst, 3, ins,
241 get_entity_type(init_ent));
242 ret = new_r_Return(bb, new_r_Proj(call, mode_M, pn_Call_M), 0, NULL);
245 add_immBlock_pred(get_irg_end_block(irg), ret);
246 mature_immBlock(get_irg_end_block(irg));
248 irg_finalize_cons(irg);
250 /* add a pointer to the new function in the constructor section */
251 add_constructor(ent);
257 * Instrument a block with code needed for profiling.
258 * This just inserts the instruction nodes, it doesn't connect the memory
259 * nodes in a meaningful way.
261 static void instrument_block(ir_node *bb, ir_node *address, unsigned int id)
263 ir_graph *irg = get_irn_irg(bb);
264 ir_node *load, *store, *offset, *add, *projm, *proji, *unknown, *cnst;
266 /* We can't instrument the end block */
267 if (bb == get_irg_end_block(irg))
270 unknown = new_r_Unknown(irg, mode_M);
271 cnst = new_r_Const_long(irg, mode_Iu, get_mode_size_bytes(mode_Iu) * id);
272 offset = new_r_Add(bb, address, cnst, get_modeP_data());
273 load = new_r_Load(bb, unknown, offset, mode_Iu, cons_none);
274 projm = new_r_Proj(load, mode_M, pn_Load_M);
275 proji = new_r_Proj(load, mode_Iu, pn_Load_res);
276 cnst = new_r_Const(irg, get_mode_one(mode_Iu));
277 add = new_r_Add(bb, proji, cnst, mode_Iu);
278 store = new_r_Store(bb, projm, offset, add, cons_none);
279 projm = new_r_Proj(store, mode_M, pn_Store_M);
281 set_irn_link(bb, projm);
282 set_irn_link(projm, load);
286 * SSA Construction for instrumentation code memory.
288 * This introduces a new memory node and connects it to the instrumentation
289 * codes, inserting phiM nodes as necessary. Note that afterwards, the new
290 * memory is not connected to any return nodes and thus still dead.
292 static void fix_ssa(ir_node *bb, void *data)
294 ir_graph *irg = get_irn_irg(bb);
295 ir_node *mem, *proj, *load;
296 int n, arity = get_Block_n_cfgpreds(bb);
300 /* end blocks are not instrumented, skip! */
301 if (bb == get_irg_end_block(irg))
304 if (bb == get_irg_start_block(irg)) {
305 mem = get_irg_initial_mem(irg);
306 } else if (arity == 1) {
307 ir_node *pred = get_Block_cfgpred_block(bb, 0);
309 mem = (ir_node*) get_irn_link(pred);
311 mem = new_r_NoMem(irg);
313 ir_node **ins = ALLOCAN(ir_node*, arity);
314 for (n = arity - 1; n >= 0; --n) {
315 ir_node *pred = get_Block_cfgpred_block(bb, n);
317 ins[n] = (ir_node*) get_irn_link(pred);
319 ins[n] = new_r_NoMem(irg);
321 mem = new_r_Phi(bb, arity, ins, mode_M);
324 /* The block link fields point to the projm from the instrumentation code,
325 * the projm in turn links to the initial load which lacks a memory
326 * argument at this point. */
327 proj = (ir_node*) get_irn_link(bb);
328 load = (ir_node*) get_irn_link(proj);
329 set_Load_mem(load, mem);
333 * Instrument a single block.
335 static void block_instrument_walker(ir_node *bb, void *data)
337 block_id_walker_data_t *wd = (block_id_walker_data_t*)data;
338 instrument_block(bb, wd->symconst, wd->id);
343 * Synchronize the original memory input of node with the additional operand
344 * from the profiling code.
346 static ir_node *sync_mem(ir_node *bb, ir_node *mem)
349 ins[0] = (ir_node*) get_irn_link(bb);
351 return new_r_Sync(bb, 2, ins);
355 * Instrument a single ir_graph, counters should point to the bblock
358 static void instrument_irg(ir_graph *irg, ir_entity *counters,
359 block_id_walker_data_t *wd)
361 ir_node *end = get_irg_end(irg);
362 ir_node *endbb = get_irg_end_block(irg);
365 /* generate a symbolic constant pointing to the count array */
367 sym.entity_p = counters;
368 wd->symconst = new_r_SymConst(irg, mode_P_data, sym, symconst_addr_ent);
370 /* instrument each block in the current irg */
371 irg_block_walk_graph(irg, block_instrument_walker, NULL, wd);
372 irg_block_walk_graph(irg, fix_ssa, NULL, NULL);
374 /* connect the new memory nodes to the return nodes */
375 for (i = get_Block_n_cfgpreds(endbb) - 1; i >= 0; --i) {
376 ir_node *node = skip_Proj(get_Block_cfgpred(endbb, i));
377 ir_node *bb = get_Block_cfgpred_block(endbb, i);
380 switch (get_irn_opcode(node)) {
382 mem = get_Return_mem(node);
383 set_Return_mem(node, sync_mem(bb, mem));
386 mem = get_Raise_mem(node);
387 set_Raise_mem(node, sync_mem(bb, mem));
392 /* A fragile's op exception. There should be another path to End,
395 assert(is_fragile_op(node) && \
396 "unexpected End control flow predecessor");
400 /* as well as calls with attribute noreturn */
401 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
402 ir_node *node = get_End_keepalive(end, i);
404 ir_node *bb = get_nodes_block(node);
405 ir_node *mem = get_Call_mem(node);
406 set_Call_mem(node, sync_mem(bb, mem));
412 * Creates a new entity representing the equivalent of
413 * static unsigned int name[size]
415 static ir_entity *new_array_entity(ident *name, int size)
418 ir_type *uint_type, *array_type;
420 uint_type = new_type_primitive(mode_Iu);
421 set_type_alignment_bytes(uint_type, get_type_size_bytes(uint_type));
423 array_type = new_type_array(1, uint_type);
424 set_array_bounds_int(array_type, 0, 0, size);
425 set_type_size_bytes(array_type, size * get_mode_size_bytes(mode_Iu));
426 set_type_alignment_bytes(array_type, get_mode_size_bytes(mode_Iu));
427 set_type_state(array_type, layout_fixed);
429 result = new_entity(get_glob_type(), name, array_type);
430 set_entity_visibility(result, ir_visibility_local);
431 set_entity_compiler_generated(result, 1);
437 * Creates a new entity representing the equivalent of
438 * static const char name[strlen(string)+1] = string
440 static ir_entity *new_static_string_entity(ident *name, const char *string)
444 ir_type *char_type = new_type_primitive(mode_Bs);
445 ir_type *string_type = new_type_array(1, char_type);
447 ir_initializer_t *contents;
449 size_t i, length = strlen(string)+1;
451 /* Create the type for a fixed-length string */
452 set_array_bounds_int(string_type, 0, 0, length);
453 set_type_size_bytes(string_type, length);
454 set_type_alignment_bytes(string_type, 1);
455 set_type_state(string_type, layout_fixed);
457 result = new_entity(get_glob_type(), name, string_type);
458 set_entity_visibility(result, ir_visibility_local);
459 set_entity_linkage(result, IR_LINKAGE_CONSTANT);
460 set_entity_compiler_generated(result, 1);
462 /* There seems to be no simpler way to do this. Or at least, cparser
463 * does exactly the same thing... */
464 contents = create_initializer_compound(length);
465 for (i = 0; i < length; i++) {
466 ir_tarval *c = new_tarval_from_long(string[i], mode_Bs);
467 ir_initializer_t *init = create_initializer_tarval(c);
468 set_initializer_compound_value(contents, i, init);
470 set_entity_initializer(result, contents);
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 *parse_profile(const char *filename, unsigned int num_blocks)
512 FILE *f = fopen(filename, "rb");
514 DBG((dbg, LEVEL_2, "Failed to open profile file (%s)\n", filename));
519 uint32_t *result = NULL;
521 size_t ret = fread(buf, 8, 1, f);
522 if (ret == 0 || strncmp(buf, "firmprof", 8) != 0) {
523 DBG((dbg, LEVEL_2, "Broken fileheader in profile\n"));
527 result = XMALLOCN(unsigned int, num_blocks);
529 /* The profiling output format is defined to be a sequence of integer
530 * values stored little endian format. */
531 for (unsigned i = 0; i < num_blocks; ++i) {
532 unsigned char bytes[4];
534 if ((ret = fread(bytes, 1, 4, f)) < 1)
537 result[i] = (bytes[0] << 0) | (bytes[1] << 8)
538 | (bytes[2] << 16) | (bytes[3] << 24);
542 DBG((dbg, LEVEL_4, "Failed to read counters... (size: %u)\n",
543 sizeof(unsigned int) * num_blocks));
554 * Reads the corresponding profile info file if it exists.
556 static void block_associate_walker(ir_node *bb, void *env)
558 block_assoc_t *b = (block_assoc_t*) env;
561 query.block = get_irn_node_nr(bb);
562 query.count = b->counters[(b->i)++];
563 DBG((dbg, LEVEL_4, "execcount(%+F, %u): %u\n", bb, query.block,
565 (void)set_insert(execcount_t, profile, &query, sizeof(query), query.block);
568 static void irp_associate_blocks(block_assoc_t *env)
570 for (int n = get_irp_n_irgs() - 1; n >= 0; --n) {
571 ir_graph *irg = get_irp_irg(n);
572 irg_block_walk_graph(irg, block_associate_walker, NULL, env);
576 void ir_profile_free(void)
584 dump_remove_node_info_callback(hook);
589 bool ir_profile_read(const char *filename)
592 FIRM_DBG_REGISTER(dbg, "firm.ir.profile");
594 unsigned n_blocks = get_irp_n_blocks();
596 env.counters = parse_profile(filename, n_blocks);
601 profile = new_set(cmp_execcount, 16);
603 irp_associate_blocks(&env);
606 /* register the vcg hook */
607 hook = dump_add_node_info_callback(dump_profile_node_info, NULL);
611 typedef struct initialize_execfreq_env_t {
613 } initialize_execfreq_env_t;
615 static void initialize_execfreq(ir_node *block, void *data)
617 const initialize_execfreq_env_t *env
618 = (const initialize_execfreq_env_t*) data;
619 ir_graph *irg = get_irn_irg(block);
622 if (block == get_irg_start_block(irg) || block == get_irg_end_block(irg)) {
625 freq = ir_profile_get_block_execcount(block);
626 freq *= env->freq_factor;
627 if (freq < MIN_EXECFREQ)
631 set_block_execfreq(block, freq);
634 static void ir_set_execfreqs_from_profile(ir_graph *irg)
636 /* Find the first block containing instructions */
637 ir_node *start_block = get_irg_start_block(irg);
638 unsigned count = ir_profile_get_block_execcount(start_block);
640 /* the function was never executed, so fallback to estimated freqs */
641 ir_estimate_execfreq(irg);
645 initialize_execfreq_env_t env;
646 env.freq_factor = 1.0 / count;
647 irg_block_walk_graph(irg, initialize_execfreq, NULL, &env);
650 void ir_create_execfreqs_from_profile(void)
652 for (int n = get_irp_n_irgs() - 1; n >= 0; --n) {
653 ir_graph *irg = get_irp_irg(n);
654 ir_set_execfreqs_from_profile(irg);