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 PBQP based register allocation.
23 * @author Thomas Bersch
25 * @version $Id: bechordal.c 26750 2009-11-27 09:37:43Z bersch $
28 /* miscellaneous includes */
37 #include "iredges_t.h"
42 /* libfirm/ir/be includes */
47 #include "bechordal_common.h"
48 #include "bechordal.h"
49 #include "bechordal_t.h"
57 #include "becopyopt.h"
64 #include "heuristical.h"
66 #include "html_dumper.h"
67 #include "pbqp_node_t.h"
68 #include "pbqp_node.h"
73 static bool use_exec_freq = true;
75 typedef struct _be_pbqp_alloc_env_t {
76 pbqp *pbqp_inst; /**< PBQP instance for register allocation */
77 be_irg_t *birg; /**< Back-end IRG session. */
78 ir_graph *irg; /**< The graph under examination. */
79 const arch_register_class_t *cls; /**< Current processed register class */
81 bitset_t *ignored_regs;
82 pbqp_matrix *ife_matrix_template;
83 pbqp_matrix *aff_matrix_template;
85 unsigned *restr_nodes;
86 unsigned *ife_edge_num;
87 be_chordal_env_t *env;
88 } be_pbqp_alloc_env_t;
91 #define is_Reg_Phi(irn) (is_Phi(irn) && mode_is_data(get_irn_mode(irn)))
92 #define get_Perm_src(irn) (get_irn_n(get_Proj_pred(irn), get_Proj_proj(irn)))
93 #define is_Perm_Proj(irn) (is_Proj(irn) && be_is_Perm(get_Proj_pred(irn)))
94 #define insert_edge(pbqp, src_node, trg_node, template_matrix) (add_edge_costs(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node), pbqp_matrix_copy(pbqp, template_matrix)))
95 #define get_free_regs(restr_nodes, cls, irn) (arch_register_class_n_regs(cls) - restr_nodes[get_irn_idx(irn)])
97 static inline int is_2addr_code(const arch_register_req_t *req)
99 return (req->type & arch_register_req_type_should_be_same) != 0;
102 static const lc_opt_table_entry_t options[] = {
103 LC_OPT_ENT_BOOL ("exec_freq", "use exec_freq", &use_exec_freq),
108 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
115 n = strlen(env->birg->main_env->cup_name);
116 tu_name = XMALLOCN(char, n + 1);
117 strcpy(tu_name, env->birg->main_env->cup_name);
118 for (i = 0; i < n; ++i)
119 if (tu_name[i] == '.')
122 ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
124 result = fopen(buf, "wt");
126 panic("Couldn't open '%s' for writing.", buf);
134 static void create_pbqp_node(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *irn) {
135 const arch_register_class_t *cls = pbqp_alloc_env->cls;
136 pbqp *pbqp_inst = pbqp_alloc_env->pbqp_inst;
137 bitset_t *ignored_regs = pbqp_alloc_env->ignored_regs;
138 unsigned colors_n = arch_register_class_n_regs(cls);
139 unsigned cntConstrains = 0;
141 /* create costs vector depending on register constrains */
142 struct vector *costs_vector = vector_alloc(pbqp_inst, colors_n);
144 /* set costs depending on register constrains */
146 for(idx = 0; idx < colors_n; idx++) {
147 if(bitset_is_set(ignored_regs, idx) || !arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, idx))) {
148 vector_set(costs_vector, idx, INF_COSTS);
153 /* add vector to pbqp node */
154 add_node_costs(pbqp_inst, get_irn_idx(irn), costs_vector);
155 pbqp_alloc_env->restr_nodes[get_irn_idx(irn)] = cntConstrains;
158 static void insert_ife_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node)
160 pbqp *pbqp = pbqp_alloc_env->pbqp_inst;
161 const arch_register_class_t *cls = pbqp_alloc_env->cls;
162 pbqp_matrix *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
163 unsigned *restr_nodes = pbqp_alloc_env->restr_nodes;
165 if(get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
167 /* increase ife edge counter */
168 pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
169 pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
171 /* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
172 if(get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
173 unsigned src_idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
174 unsigned trg_idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
175 assert(src_idx != trg_idx && "Interfering nodes could not have the same register!");
178 if(get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
179 if(get_free_regs(restr_nodes, cls, src_node) == 1) {
180 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
181 vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
184 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
185 vector_set(get_node(pbqp, get_irn_idx(src_node))->costs, idx, INF_COSTS);
190 /* insert interference edge */
191 insert_edge(pbqp, src_node, trg_node, ife_matrix_template);
195 static void inser_afe_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node, int pos)
197 pbqp *pbqp = pbqp_alloc_env->pbqp_inst;
198 const arch_register_class_t *cls = pbqp_alloc_env->cls;
199 unsigned *restr_nodes = pbqp_alloc_env->restr_nodes;
200 pbqp_matrix *afe_matrix = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
201 unsigned colors_n = arch_register_class_n_regs(cls);
203 if(get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
205 /* get exec_freq for copy_block */
206 ir_node *root_bl = get_nodes_block(src_node);
207 ir_node *copy_bl = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
208 unsigned long res = get_block_execfreq_ulong(pbqp_alloc_env->birg->exec_freq, copy_bl);
210 /* create afe-matrix */
212 for(row = 0; row < colors_n; row++) {
213 for(col = 0; col < colors_n; col++) {
215 pbqp_matrix_set(afe_matrix, row, col, (num)res);
220 afe_matrix = pbqp_alloc_env->aff_matrix_template;
223 /* do useful optimization to speed up pbqp solving */
224 if(get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
227 if(get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
228 if(get_free_regs(restr_nodes, cls, src_node) == 1) {
229 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
230 vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
233 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
234 vector_add_matrix_col(get_node(pbqp, get_irn_idx(src_node))->costs, afe_matrix, regIdx);
239 /* insert interference edge */
240 insert_edge(pbqp, src_node, trg_node, afe_matrix);
244 static void create_affinity_edges(ir_node *irn, void *env)
246 be_pbqp_alloc_env_t *pbqp_alloc_env = env;
247 const arch_register_class_t *cls = pbqp_alloc_env->cls;
248 const arch_register_req_t *req = arch_get_register_req_out(irn);
251 if (is_Reg_Phi(irn)) { /* Phis */
252 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
253 ir_node *arg = get_irn_n(irn, pos);
255 if (!arch_irn_consider_in_reg_alloc(cls, arg))
258 /* no edges to itself */
263 inser_afe_edge(pbqp_alloc_env, irn, arg, pos);
266 else if (is_Perm_Proj(irn)) { /* Perms */
267 ir_node *arg = get_Perm_src(irn);
268 if (!arch_irn_consider_in_reg_alloc(cls, arg))
271 inser_afe_edge(pbqp_alloc_env, irn, arg, -1);
273 else { /* 2-address code */
274 if (is_2addr_code(req)) {
275 const unsigned other = req->other_same;
278 for (i = 0; 1U << i <= other; ++i) {
279 if (other & (1U << i)) {
280 ir_node *other = get_irn_n(skip_Proj(irn), i);
281 if (!arch_irn_consider_in_reg_alloc(cls, other))
284 /* no edges to itself */
289 inser_afe_edge(pbqp_alloc_env, irn, other, i);
296 static void create_pbqp_coloring_instance(ir_node *block, void *data)
298 be_pbqp_alloc_env_t *pbqp_alloc_env = data;
299 be_lv_t *lv = pbqp_alloc_env->lv;
300 const arch_register_class_t *cls = pbqp_alloc_env->cls;
301 plist_t *rpeo = pbqp_alloc_env->rpeo;
302 pbqp *pbqp_inst = pbqp_alloc_env->pbqp_inst;
303 unsigned *restr_nodes = pbqp_alloc_env->restr_nodes;
304 pqueue_t *queue = new_pqueue();
305 pqueue_t *restr_nodes_queue = new_pqueue();
306 plist_t *temp_list = plist_new();
308 ir_nodeset_t live_nodes;
310 /* first, determine the pressure */
311 /* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
312 create_borders(block, pbqp_alloc_env->env);
314 /* calculate living nodes for the first step */
315 ir_nodeset_init(&live_nodes);
316 be_liveness_end_of_block(lv, cls, block, &live_nodes);
318 /* create pbqp nodes, interference edges and reverse perfect elimination order */
319 sched_foreach_reverse(block, irn) {
321 ir_nodeset_iterator_t iter;
323 if (get_irn_mode(irn) == mode_T) {
324 const ir_edge_t *edge;
325 foreach_out_edge(irn, edge) {
326 ir_node *proj = get_edge_src_irn(edge);
327 if (!arch_irn_consider_in_reg_alloc(cls, proj))
330 /* create pbqp source node if it dosn't exist */
331 if(get_node(pbqp_inst, get_irn_idx(proj)) == NULL) {
332 create_pbqp_node(pbqp_alloc_env, proj);
335 /* create nodes and interference edges */
336 foreach_ir_nodeset(&live_nodes, live, iter) {
337 /* create pbqp source node if it dosn't exist */
338 if(get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
339 create_pbqp_node(pbqp_alloc_env, live);
342 /* no edges to itself */
347 insert_ife_edge(pbqp_alloc_env, proj, live);
352 if (arch_irn_consider_in_reg_alloc(cls, irn)) {
353 /* create pbqp source node if it dosn't exist */
354 if(get_node(pbqp_inst, get_irn_idx(irn)) == NULL) {
355 create_pbqp_node(pbqp_alloc_env, irn);
358 /* create nodes and interference edges */
359 foreach_ir_nodeset(&live_nodes, live, iter) {
360 /* create pbqp source node if it dosn't exist */
361 if(get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
362 create_pbqp_node(pbqp_alloc_env, live);
365 /* no edges to itself */
370 /* insert interference edge */
371 insert_ife_edge(pbqp_alloc_env, irn, live);
376 /* get living nodes for next step */
378 be_liveness_transfer(cls, irn, &live_nodes);
381 /* order nodes for perfect elimination order */
382 if (get_irn_mode(irn) == mode_T) {
383 plist_element_t *first = plist_first(temp_list);
384 const ir_edge_t *edge;
386 foreach_out_edge(irn, edge) {
387 ir_node *proj = get_edge_src_irn(edge);
388 if (!arch_irn_consider_in_reg_alloc(cls, proj))
391 // insert proj node into priority queue (descending by the number of interference edges)
392 if(get_free_regs(restr_nodes, cls, proj) <= 4/*bitset_is_set(restr_nodes, get_irn_idx(proj))*/) {
393 pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
396 pqueue_put(queue,proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
400 /* first insert all restricted nodes */
401 while(!pqueue_empty(restr_nodes_queue)) {
403 plist_insert_back(temp_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(restr_nodes_queue))));
404 first = plist_first(temp_list);
406 plist_insert_before(temp_list, first, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(restr_nodes_queue))));
410 /* insert proj nodes descending by their number of interference edges */
411 while(!pqueue_empty(queue)) {
413 plist_insert_back(temp_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(queue))));
414 first = plist_first(temp_list);
416 plist_insert_before(temp_list, first, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(queue))));
421 if (arch_irn_consider_in_reg_alloc(cls, irn)) {
422 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
427 /* insert nodes into reverse perfect elimination order */
429 foreach_plist(temp_list, el) {
430 plist_insert_back(rpeo, el->data);
433 /* free reserved memory */
434 ir_nodeset_destroy(&live_nodes);
435 plist_free(temp_list);
437 del_pqueue(restr_nodes_queue);
440 static void insert_perms(ir_node *block, void *data)
443 * Start silent in the start block.
444 * The silence remains until the first barrier is seen.
445 * Each other block is begun loud.
447 be_chordal_env_t *env = data;
449 int silent = block == get_irg_start_block(get_irn_irg(block));
452 * If the block is the start block search the barrier and
453 * start handling constraints from there.
455 for (irn = sched_first(block); !sched_is_end(irn);) {
456 int silent_old = silent; /* store old silent value */
457 if (be_is_Barrier(irn))
458 silent = !silent; /* toggle silent flag */
460 be_insn_t *insn = chordal_scan_insn(env, irn);
461 irn = insn->next_insn;
466 if (!insn->has_constraints)
469 pre_process_constraints(env, &insn);
473 void be_pbqp_coloring(be_chordal_env_t *env)
475 ir_graph *irg = env->irg;
476 be_irg_t *birg = env->birg;
477 const arch_register_class_t *cls = env->cls;
478 unsigned colors_n = arch_register_class_n_regs(cls);
479 be_pbqp_alloc_env_t pbqp_alloc_env;
480 unsigned idx, row, col;
484 ir_timer_t *t_ra_pbqp_alloc_create = ir_timer_new();
485 ir_timer_t *t_ra_pbqp_alloc_solve = ir_timer_new();
486 ir_timer_t *t_ra_pbqp_alloc_create_aff = ir_timer_new();
488 printf("#### ----- === Allocating registers of %s (%s) ===\n", cls->name, get_entity_name(get_irg_entity(irg)));
490 lv = be_assure_liveness(birg);
491 be_liveness_assure_sets(lv);
492 be_liveness_assure_chk(lv);
496 dom_tree_walk_irg(irg, insert_perms, NULL, env);
498 /* dump graph after inserting perms */
499 if (env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
501 snprintf(buf, sizeof(buf), "-%s-constr", cls->name);
502 be_dump(irg, buf, dump_ir_block_graph_sched);
506 /* initialize pbqp allocation data structure */
507 pbqp_alloc_env.pbqp_inst = alloc_pbqp(get_irg_last_idx(irg)); /* initialize pbqp instance */
508 pbqp_alloc_env.birg = birg;
509 pbqp_alloc_env.cls = cls;
510 pbqp_alloc_env.irg = irg;
511 pbqp_alloc_env.lv = lv;
512 pbqp_alloc_env.ignored_regs = bitset_malloc(colors_n);
513 pbqp_alloc_env.rpeo = plist_new();
514 pbqp_alloc_env.restr_nodes = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
515 pbqp_alloc_env.ife_edge_num = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
516 pbqp_alloc_env.env = env;
517 be_put_ignore_regs(birg, cls, pbqp_alloc_env.ignored_regs); /* get ignored registers */
520 /* create costs matrix template for interference edges */
521 struct pbqp_matrix *ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
523 for(row = 0, col=0; row < colors_n; row++, col++)
524 pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
526 pbqp_alloc_env.ife_matrix_template = ife_matrix;
530 /* create costs matrix template for affinity edges */
531 struct pbqp_matrix *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
533 for(row = 0; row < colors_n; row++) {
534 for(col = 0; col < colors_n; col++) {
536 pbqp_matrix_set(afe_matrix, row, col, 2);
539 pbqp_alloc_env.aff_matrix_template = afe_matrix;
543 /* create pbqp instance */
545 ir_timer_reset_and_start(t_ra_pbqp_alloc_create);
548 dom_tree_walk_irg(irg, create_pbqp_coloring_instance , NULL, &pbqp_alloc_env);
550 ir_timer_stop(t_ra_pbqp_alloc_create);
554 /* set up affinity edges */
556 ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
559 foreach_plist(pbqp_alloc_env.rpeo, el) {
560 pbqp_node *node = el->data;
562 ir_node *irn = get_idx_irn(irg, idx);
563 create_affinity_edges(irn, &pbqp_alloc_env);
566 ir_timer_stop(t_ra_pbqp_alloc_create_aff);
571 // dump graph before solving pbqp
572 FILE *file_before = my_open(env, "", "-pbqp_coloring.html");
573 set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
577 /* solve pbqp instance */
579 ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
581 solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
583 ir_timer_stop(t_ra_pbqp_alloc_solve);
585 num solution = get_solution(pbqp_alloc_env.pbqp_inst);
586 assert(solution != INF_COSTS && "No PBQP solution found");
590 plist_element_t *element;
591 foreach_plist(pbqp_alloc_env.rpeo, element) {
592 pbqp_node *node = element->data;
594 ir_node *irn = get_idx_irn(irg, idx);
595 num color = get_node_solution(pbqp_alloc_env.pbqp_inst, idx);
596 const arch_register_t *reg = arch_register_for_index(cls, color);
598 arch_set_irn_register(irn, reg);
603 printf("%-20s: %8.3lf msec\n", "pbqp alloc create",
604 (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create) / 1000.0);
605 printf("%-20s: %8.3lf msec\n", "pbqp alloc solve",
606 (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_solve) / 1000.0);
607 printf("%-20s: %8.3lf msec\n", "pbqp alloc create aff",
608 (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create_aff) / 1000.0);
612 /* free reserved memory */
616 bitset_free(pbqp_alloc_env.ignored_regs);
617 free_pbqp(pbqp_alloc_env.pbqp_inst);
618 plist_free(pbqp_alloc_env.rpeo);
619 xfree(pbqp_alloc_env.restr_nodes);
620 xfree(pbqp_alloc_env.ife_edge_num);
625 * Initializes this module.
627 void be_init_pbqp_coloring(void)
629 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
630 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
631 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
632 lc_opt_entry_t *coloring_grp = lc_opt_get_grp(chordal_grp, "coloring");
633 lc_opt_entry_t *pbqp_grp = lc_opt_get_grp(coloring_grp, "pbqp");
635 static be_ra_chordal_coloring_t coloring = {
639 lc_opt_add_table(pbqp_grp, options);
640 be_register_chordal_coloring("pbqp", &coloring);
643 BE_REGISTER_MODULE_CONSTRUCTOR(be_pbqp_alloc);