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 MIN-algorithm with some twists (aka belady spiller v3)
23 * @author Matthias Braun
28 * - handle phis correctly, decide whether we should spill them ~ok
29 * - merge multiple start worksets of blocks ~ok
30 * - how to and when to do the tentative phase...
44 #include "iredges_t.h"
52 #include "bespilloptions.h"
53 #include "besched_t.h"
56 #define EXPENSIVE_CHECKS
58 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
60 typedef struct loop_edge_t loop_edge_t;
67 typedef struct loop_info_t loop_info_t;
70 loop_edge_t *entry_edges;
71 loop_edge_t *exit_edges;
72 size_t max_register_pressure;
75 typedef struct worklist_entry_t worklist_entry_t;
76 struct worklist_entry_t {
77 struct list_head head;
79 ir_node *reload_point;
80 ir_loop *unused_livethrough_loop;
83 typedef struct worklist_t worklist_t;
85 struct list_head live_values;
87 unsigned long visited;
90 typedef struct block_info_t block_info_t;
92 worklist_t *start_worklist;
93 worklist_t *end_worklist;
96 static const arch_env_t *arch_env;
97 static const arch_register_class_t *cls;
98 static struct obstack obst;
99 static spill_env_t *senv;
100 static size_t n_regs;
101 static size_t max_register_pressure;
102 static bool tentative_mode;
103 static bool should_have_reached_fixpoint;
104 static bool do_push_unused_livethroughs;
105 static ir_exec_freq *exec_freq;
106 static unsigned long worklist_visited;
108 static worklist_t *new_worklist(void)
110 worklist_t *worklist = obstack_alloc(&obst, sizeof(worklist[0]));
111 memset(worklist, 0, sizeof(worklist[0]));
113 INIT_LIST_HEAD(&worklist->live_values);
114 worklist->n_live_values = 0;
119 static void mark_irn_not_visited(ir_node *node)
121 set_irn_visited(node, get_irg_visited(current_ir_graph) - 1);
124 static bool worklist_contains(const ir_node *node)
126 return irn_visited(node);
129 static block_info_t *get_block_info(ir_node *block)
131 block_info_t *info = get_irn_link(block);
135 info = obstack_alloc(&obst, sizeof(info[0]));
136 memset(info, 0, sizeof(info[0]));
137 set_irn_link(block, info);
141 static void deactivate_worklist(const worklist_t *worklist)
143 struct list_head *entry;
145 list_for_each(entry, &worklist->live_values) {
146 worklist_entry_t *wl_entry
147 = list_entry(entry, worklist_entry_t, head);
148 assert(worklist_contains(wl_entry->value));
149 mark_irn_not_visited(wl_entry->value);
150 set_irn_link(wl_entry->value, NULL);
154 static void activate_worklist(const worklist_t *worklist)
156 struct list_head *entry;
158 list_for_each(entry, &worklist->live_values) {
159 worklist_entry_t *wl_entry
160 = list_entry(entry, worklist_entry_t, head);
161 assert(!worklist_contains(wl_entry->value));
162 mark_irn_visited(wl_entry->value);
163 set_irn_link(wl_entry->value, wl_entry);
168 * duplicate a worklist and directly activates it
170 static void fill_and_activate_worklist(worklist_t *new_worklist,
171 const worklist_t *worklist, ir_node *block, ir_node *succ_block,
174 ir_node *reload_point = NULL;
175 size_t n_live_values = 0;
176 struct list_head *entry;
178 if (succ_block != NULL &&
179 (get_Block_n_cfgpreds(succ_block) > 1
180 || get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) > 1)) {
181 reload_point = be_get_end_of_block_insertion_point(block);
184 list_for_each(entry, &worklist->live_values) {
185 worklist_entry_t *wl_entry = list_entry(entry, worklist_entry_t, head);
186 ir_node *value = wl_entry->value;
187 worklist_entry_t *new_entry;
189 if (new_worklist->n_live_values >= n_regs)
192 if (is_Phi(value) && get_nodes_block(value) == succ_block) {
193 value = get_Phi_pred(value, succ_pos);
195 /* can happen for unknown phi preds */
196 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, value))
200 if (irn_visited(value))
203 new_entry = obstack_alloc(&obst, sizeof(new_entry[0]));
204 memset(new_entry, 0, sizeof(new_entry[0]));
206 new_entry->value = value;
207 if (reload_point != NULL) {
208 new_entry->reload_point = reload_point;
210 new_entry->reload_point = wl_entry->reload_point;
213 list_add_tail(&new_entry->head, &new_worklist->live_values);
216 mark_irn_visited(value);
217 set_irn_link(value, new_entry);
218 new_worklist->n_live_values++;
222 static worklist_t *duplicate_worklist(const worklist_t *worklist)
224 worklist_t *new_worklist;
225 struct list_head *entry;
227 new_worklist = obstack_alloc(&obst, sizeof(new_worklist[0]));
228 memset(new_worklist, 0, sizeof(new_worklist[0]));
229 INIT_LIST_HEAD(&new_worklist->live_values);
230 new_worklist->n_live_values = worklist->n_live_values;
232 list_for_each(entry, &worklist->live_values) {
233 worklist_entry_t *wl_entry = list_entry(entry, worklist_entry_t, head);
234 worklist_entry_t *new_entry
235 = obstack_alloc(&obst, sizeof(new_entry[0]));
237 memcpy(new_entry, wl_entry, sizeof(new_entry[0]));
238 list_add_tail(&new_entry->head, &new_worklist->live_values);
245 static void print_worklist(const worklist_t *worklist, int level)
247 struct list_head *entry;
249 DB((dbg, level, "%d values: ", worklist->n_live_values));
250 list_for_each(entry, &worklist->live_values) {
251 worklist_entry_t *wl_entry
252 = list_entry(entry, worklist_entry_t, head);
254 DB((dbg, level, "%+F ", wl_entry->value));
261 static void clear_loop_info(ir_loop *loop)
263 int n_elements = get_loop_n_elements(loop);
267 for (i = 0; i < n_elements; ++i) {
268 loop_element element = get_loop_element(loop, i);
269 if (*element.kind != k_ir_loop)
272 clear_loop_info(element.son);
276 static loop_info_t *get_loop_info(ir_loop *loop)
278 loop_info_t *info = loop->link;
282 info = obstack_alloc(&obst, sizeof(info[0]));
283 memset(info, 0, sizeof(info[0]));
290 * constructs loop in+out edges when called in a block walker
292 static void construct_loop_edges(ir_node *block, void *data)
294 int n_cfgpreds = get_Block_n_cfgpreds(block);
295 ir_loop *loop = get_irn_loop(block);
300 for (i = 0; i < n_cfgpreds; ++i) {
301 ir_node *cfgpred_block = get_Block_cfgpred_block(block, i);
302 ir_loop *cfgpred_loop = get_irn_loop(cfgpred_block);
307 if (cfgpred_loop == loop)
310 /* critical edges are splitted, so we can't jump from 1 loop
311 * directly into another without going out of it */
312 assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
314 /* edge out of loop */
315 if (get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
320 is_exit_edge = false;
325 /* this might be a jump out/in multiple loops */
327 loop_info_t *l_info = get_loop_info(l);
329 edge = obstack_alloc(&obst, sizeof(edge[0]));
330 memset(edge, 0, sizeof(edge[0]));
335 edge->next = l_info->exit_edges;
336 l_info->exit_edges = edge;
337 assert(get_loop_depth(loop) < get_loop_depth(l));
339 edge->next = l_info->entry_edges;
340 l_info->entry_edges = edge;
341 assert(get_loop_depth(cfgpred_loop) < get_loop_depth(l));
343 l = get_loop_outer_loop(l);
351 static void place_reload(worklist_entry_t *entry)
356 DB((dbg, LEVEL_1, "reload of %+F before %+F\n", entry->value,
357 entry->reload_point));
358 assert(entry->reload_point != NULL);
359 be_add_reload(senv, entry->value, entry->reload_point, cls, 1);
360 entry->reload_point = NULL;
364 * makes sure the worklist contains not more than n_regs - room_needed entries
366 static void make_room(worklist_t *worklist, size_t room_needed)
370 struct list_head *entry;
372 spills_needed = worklist->n_live_values + room_needed - n_regs;
373 if (spills_needed <= 0)
376 entry = worklist->live_values.next;
377 for(i = spills_needed; i > 0; --i) {
378 struct list_head *next = entry->next;
379 worklist_entry_t *wl_entry
380 = list_entry(entry, worklist_entry_t, head);
382 assert(worklist_contains(wl_entry->value));
383 mark_irn_not_visited(wl_entry->value);
384 place_reload(wl_entry);
389 worklist->n_live_values -= spills_needed;
393 * a value was used, so bring it to the back of the worklist (which might
394 * result in a spill of another value).
396 static void val_used(worklist_t *worklist, ir_node *value, ir_node *sched_point)
398 /* already in the worklist? move around, otherwise add at back */
399 worklist_entry_t *entry = get_irn_link(value);
401 assert(arch_irn_consider_in_reg_alloc(arch_env, cls, value));
403 if (worklist_contains(value)) {
404 assert(entry != NULL);
406 list_del(&entry->head);
409 entry = obstack_alloc(&obst, sizeof(entry[0]));
410 memset(entry, 0, sizeof(entry[0]));
412 entry->value = value;
413 set_irn_link(value, entry);
416 ++worklist->n_live_values;
417 mark_irn_visited(value);
420 entry->reload_point = sched_point;
421 list_add_tail(&entry->head, &worklist->live_values);
424 static void worklist_remove(worklist_t *worklist, ir_node *value)
426 worklist_entry_t *entry = get_irn_link(value);
427 assert(entry != NULL);
428 list_del(&entry->head);
429 --worklist->n_live_values;
431 assert(worklist_contains(value));
432 mark_irn_not_visited(value);
435 static void update_max_pressure(worklist_t *worklist)
437 if (worklist->n_live_values > max_register_pressure)
438 max_register_pressure = worklist->n_live_values;
441 static void do_spilling(ir_node *block, worklist_t *worklist)
445 assert(worklist != NULL);
447 sched_foreach_reverse(block, node) {
451 DB((dbg, LEVEL_2, "\t%+F... ", node));
452 update_max_pressure(worklist);
456 /* TODO: if we have some free registers, then we could decide to
457 * not spill some phis (but not for phis where at least 1 input is
460 /* we have to spill all phis that are not live */
461 sched_foreach_reverse_from(node, node2) {
462 assert(is_Phi(node2));
464 if (worklist_contains(node2))
466 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node2))
470 be_spill_phi(senv, node2);
472 DB((dbg, LEVEL_2, "\n"));
476 /* remove values defined by this instruction from the workset. Values
477 * defined but not in the workset need free registers */
478 if (get_irn_mode(node) == mode_T) {
479 const ir_edge_t *edge;
481 foreach_out_edge(node, edge) {
482 ir_node *proj = get_edge_src_irn(edge);
483 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
485 if (worklist_contains(proj)) {
486 worklist_remove(worklist, proj);
491 } else if (arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
492 if (worklist_contains(node)) {
493 worklist_remove(worklist, node);
499 /* make sure we have enough free registers for the spills */
500 make_room(worklist, n_defs);
502 /* put all values used by the instruction into the workset */
503 arity = get_irn_arity(node);
504 for(i = 0; i < arity; ++i) {
505 ir_node *use = get_irn_n(node, i);
507 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, use))
510 val_used(worklist, use, node);
513 /* we might have too many values in the worklist now and need to spill
515 make_room(worklist, 0);
518 print_worklist(worklist, LEVEL_2);
519 DB((dbg, LEVEL_2, "\n"));
523 update_max_pressure(worklist);
526 DB((dbg, LEVEL_1, "worklist at begin of %+F:", block));
527 print_worklist(worklist, LEVEL_1);
528 DB((dbg, LEVEL_1, "\n"));
532 static bool worklists_equal(const worklist_t *wl1, const worklist_t *wl2)
534 const struct list_head *i1 = &wl1->live_values;
535 const struct list_head *i2 = &wl2->live_values;
537 for ( ; i1 != &wl1->live_values && i2 != &wl2->live_values;
538 i1 = i1->next, i2 = i2->next) {
539 worklist_entry_t *entry1 = list_entry(i1, worklist_entry_t, head);
540 worklist_entry_t *entry2 = list_entry(i2, worklist_entry_t, head);
542 if (entry1->value != entry2->value)
545 /* both lists at the end */
546 if (i1 != &wl1->live_values || i2 != &wl2->live_values)
552 static bool fill_start_worklist(worklist_t *new_worklist, ir_node *block)
554 double best_execfreq = -1;
555 worklist_t *best_worklist = NULL;
556 ir_node *best_succ_block = NULL;
558 const ir_edge_t *edge;
560 /* construct worklist */
561 foreach_block_succ(block, edge) {
562 ir_node *succ_block = get_edge_src_irn(edge);
563 double execfreq = get_block_execfreq(exec_freq, succ_block);
564 block_info_t *block_info;
565 worklist_t *succ_worklist;
567 if (execfreq < best_execfreq)
570 block_info = get_block_info(succ_block);
571 succ_worklist = block_info->start_worklist;
573 if (succ_worklist == NULL || succ_worklist->visited >= worklist_visited)
576 best_execfreq = execfreq;
577 best_worklist = succ_worklist;
578 best_succ_block = succ_block;
579 best_pos = get_edge_src_pos(edge);
582 if (best_worklist == NULL)
584 best_worklist->visited = worklist_visited;
586 fill_and_activate_worklist(new_worklist, best_worklist, block,
587 best_succ_block, best_pos);
591 static worklist_t *construct_start_worklist(ir_node *block)
593 worklist_t *worklist = new_worklist();
597 while(fill_start_worklist(worklist, block)) {
598 if (worklist->n_live_values >= n_regs)
605 static void process_block(ir_node *block, void *env)
607 block_info_t *block_info;
608 worklist_t *worklist;
612 DB((dbg, LEVEL_1, "Processing %+F\n", block));
614 worklist = construct_start_worklist(block);
615 block_info = get_block_info(block);
618 DB((dbg, LEVEL_1, "worklist at end of %+F:", block));
619 print_worklist(worklist, LEVEL_1);
620 DB((dbg, LEVEL_1, "\n"));
622 if (should_have_reached_fixpoint) {
623 assert(worklists_equal(block_info->end_worklist, worklist));
626 block_info->end_worklist = duplicate_worklist(worklist);
628 do_spilling(block, worklist);
629 deactivate_worklist(worklist);
632 if (should_have_reached_fixpoint) {
633 assert(worklists_equal(block_info->start_worklist, worklist));
636 block_info->start_worklist = worklist;
638 /* we shouldn't have any live values left at the start block */
639 n_preds = get_Block_n_cfgpreds(block);
640 //assert(n_preds != 0 || worklist->n_live_values == 0);
643 typedef struct block_or_loop_t block_or_loop_t;
644 struct block_or_loop_t {
652 static block_or_loop_t *loop_blocks;
653 static ir_loop *current_loop;
655 static void find_blocks(ir_node *block);
657 static void find_in_loop(ir_loop *loop, ir_node *entry)
659 loop_info_t *loop_info = get_loop_info(loop);
660 block_or_loop_t block_or_loop;
663 /* simply mark 1 block in the loop to indicate that the loop was already
665 ir_node *some_block = loop_info->entry_edges->block;
666 if (Block_block_visited(some_block))
669 block_or_loop.v.loop = loop;
670 block_or_loop.is_loop = true;
671 ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
675 /* we should be 1 of the entry blocks */
676 loop_edge_t *edge = loop_info->entry_edges;
678 for ( ; edge != NULL; edge = edge->next) {
679 if (edge->block == entry)
685 /* check all loop successors */
686 for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
687 ir_node *succ = edge->block;
688 ir_loop *succ_loop = get_irn_loop(succ);
690 if (succ_loop == current_loop) {
693 assert(get_loop_depth(succ_loop) < get_loop_depth(current_loop));
698 static void find_blocks(ir_node *block)
700 const ir_edge_t *edge;
701 block_or_loop_t block_or_loop;
703 if (Block_block_visited(block))
706 block_or_loop.v.block = block;
707 block_or_loop.is_loop = false;
709 ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
710 mark_Block_block_visited(block);
712 foreach_block_succ(block, edge) {
713 ir_node *succ = get_edge_src_irn(edge);
715 /* is it part of the current loop? */
716 ir_loop *loop = get_irn_loop(succ);
717 if (loop != current_loop) {
719 if (get_loop_depth(loop) > get_loop_depth(current_loop)) {
720 find_in_loop(loop, succ);
722 /* parent loop: we're not interested in the block */
731 * append an entry to a worklist. WARNING: The entry must not already be in the
734 static void worklist_append(worklist_t *worklist, ir_node *value,
735 ir_node *reload_point,
736 ir_loop *unused_livethrough_loop)
738 worklist_entry_t *entry = obstack_alloc(&obst, sizeof(entry[0]));
739 memset(entry, 0, sizeof(entry[0]));
741 #ifdef EXPENSIVE_CHECKS
743 struct list_head *entry;
744 list_for_each(entry, &worklist->live_values) {
745 worklist_entry_t *wl_entry
746 = list_entry(entry, worklist_entry_t, head);
747 assert(wl_entry->value != value);
752 entry->value = value;
753 entry->reload_point = reload_point;
754 entry->unused_livethrough_loop = unused_livethrough_loop;
755 list_add_tail(&entry->head, &worklist->live_values);
756 ++worklist->n_live_values;
757 assert(worklist->n_live_values <= n_regs);
760 static void push_unused_livethrough(loop_info_t *loop_info, ir_node *value)
765 /* add the value to all loop exit and entry blocks */
766 for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
768 = get_Block_cfgpred_block(edge->block, edge->pos);
769 const block_info_t *info = get_block_info(block);
770 worklist_t *worklist = info->end_worklist;
771 ir_node *reload_point = NULL;
773 if (worklist->visited >= worklist_visited)
775 worklist->visited = worklist_visited;
777 /* TODO: we need a smarter mechanism here, that makes the reloader place
778 * reload nodes on all loop exits... */
780 worklist_append(worklist, value, reload_point, loop_info->loop);
782 edge = loop_info->entry_edges;
783 for ( ; edge != NULL; edge = edge->next) {
784 ir_node *entry_block = edge->block;
785 const block_info_t *info = get_block_info(entry_block);
786 worklist_t *worklist = info->start_worklist;
788 ir_node *reload_point;
790 if (worklist->visited >= worklist_visited)
792 worklist->visited = worklist_visited;
794 pred_block = get_Block_cfgpred_block(entry_block, edge->pos);
795 reload_point = be_get_end_of_block_insertion_point(pred_block);
797 worklist_append(worklist, value, reload_point, loop_info->loop);
800 set_irn_link(value, NULL);
801 ++loop_info->max_register_pressure;
804 static void push_unused_livethroughs(loop_info_t *loop_info)
808 /* we can only push unused livethroughs if register pressure inside the loop
810 if (loop_info->max_register_pressure >= n_regs)
813 /* find unused livethroughs: register pressure in the loop was low enough
814 * which means that we had no spills which implies that at every point in
816 for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
817 ir_node *block = edge->block;
818 const block_info_t *info = get_block_info(block);
819 worklist_t *start_worklist = info->start_worklist;
821 const block_info_t *exit_info;
822 worklist_t *end_worklist;
823 struct list_head *entry;
825 if (start_worklist == NULL)
828 exit_block = get_Block_cfgpred_block(edge->block, edge->pos);
829 exit_info = get_block_info(exit_block);
830 end_worklist = exit_info->end_worklist;
832 activate_worklist(end_worklist);
833 /* all values contained in the start_worklist, which are not available
834 * in the end_worklist, must be unused livethroughs */
836 list_for_each(entry, &start_worklist->live_values) {
837 worklist_entry_t *wl_entry
838 = list_entry(entry, worklist_entry_t, head);
839 ir_node *value = wl_entry->value;
841 if (loop_info->max_register_pressure >= n_regs)
845 if (!worklist_contains(value)) {
846 /* add it to all loop exits */
847 DB((dbg, LEVEL_2, "Found unused livethrough: %+F (regpressure in loop %d)\n", value, loop_info->max_register_pressure));
848 push_unused_livethrough(loop_info, value);
849 /* the value should now be part of end_worklist, so mark it */
850 mark_irn_visited(value);
854 deactivate_worklist(end_worklist);
858 static void process_block_or_loop(const block_or_loop_t *block_or_loop)
860 if (block_or_loop->is_loop) {
861 loop_info_t *loop_info = get_loop_info(block_or_loop->v.loop);
863 if (do_push_unused_livethroughs)
864 push_unused_livethroughs(loop_info);
866 if (loop_info->max_register_pressure > max_register_pressure)
867 max_register_pressure = loop_info->max_register_pressure;
871 process_block(block_or_loop->v.block, NULL);
874 static void process_loop(ir_loop *loop)
876 int n_elements = get_loop_n_elements(loop);
878 loop_info_t *loop_info;
882 /* first handle all sub-loops */
883 for (i = 0; i < n_elements; ++i) {
884 loop_element element = get_loop_element(loop, i);
885 if (*element.kind != k_ir_loop)
888 process_loop(element.son);
891 /* create a postorder of the blocks */
892 loop_info = get_loop_info(loop);
893 edge = loop_info->entry_edges;
895 some_block = edge->block;
897 assert(loop == get_irg_loop(current_ir_graph));
898 some_block = get_irg_start_block(current_ir_graph);
901 loop_blocks = NEW_ARR_F(block_or_loop_t,0);
904 ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
905 inc_irg_block_visited(current_ir_graph);
906 find_blocks(some_block);
907 /* for endless loops the end-block might be unreachable */
908 if (loop == get_irg_loop(current_ir_graph)) {
909 find_blocks(get_irg_end_block(current_ir_graph));
911 ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
913 DB((dbg, LEVEL_3, "Block List for loop %p\n", loop));
914 len = ARR_LEN(loop_blocks);
915 for (i = 0; i < len; ++i) {
916 block_or_loop_t *block_or_loop = &loop_blocks[i];
917 if (block_or_loop->is_loop) {
918 DB((dbg, LEVEL_3, " L-%p", block_or_loop->v.loop));
920 DB((dbg, LEVEL_3, " B-%+F", block_or_loop->v.block));
923 DB((dbg, LEVEL_3, "\n"));
925 max_register_pressure = 0;
927 /* run1: tentative phase */
928 tentative_mode = true;
929 for (i = len-1; i >= 0; --i) {
930 process_block_or_loop(&loop_blocks[i]);
933 /* run2: tentative phase - should reach fixpoint */
934 tentative_mode = true;
935 for (i = len-1; i >= 0; --i) {
936 process_block_or_loop(&loop_blocks[i]);
940 /* run3: tentative phase - check fixpoint */
941 tentative_mode = true;
942 should_have_reached_fixpoint = true;
943 for (i = len-1; i >= 0; --i) {
944 process_block_or_loop(&loop_blocks[i]);
946 should_have_reached_fixpoint = false;
949 /* run4: add spills/reloads */
950 tentative_mode = false;
951 do_push_unused_livethroughs = true;
952 for (i = len-1; i >= 0; --i) {
953 process_block_or_loop(&loop_blocks[i]);
955 do_push_unused_livethroughs = false;
957 loop_info->max_register_pressure = max_register_pressure;
958 DB((dbg, LEVEL_2, "Regpressure in loop %p: %u\n", loop,
959 (unsigned) max_register_pressure));
961 DEL_ARR_F(loop_blocks);
964 static void fix_block_borders(ir_node *block, void *data)
966 block_info_t *block_info = get_block_info(block);
967 worklist_t *start_worklist = block_info->start_worklist;
968 int n_cfgpreds = get_Block_n_cfgpreds(block);
969 ir_loop *loop = get_irn_loop(block);
973 assert(start_worklist != NULL);
975 for (i = 0; i < n_cfgpreds; ++i) {
976 ir_node *pred_block = get_Block_cfgpred_block(block, i);
977 block_info_t *pred_block_info = get_block_info(pred_block);
978 worklist_t *end_worklist = pred_block_info->end_worklist;
979 ir_loop *pred_loop = get_irn_loop(pred_block);
980 bool is_loop_entry = false;
981 struct list_head *entry;
983 assert(end_worklist != NULL);
985 if (get_loop_depth(pred_loop) < get_loop_depth(loop)) {
986 is_loop_entry = true;
989 /* reload missing values */
990 activate_worklist(end_worklist);
992 list_for_each(entry, &start_worklist->live_values) {
993 worklist_entry_t *wl_entry
994 = list_entry(entry, worklist_entry_t, head);
995 ir_node *value = wl_entry->value;
997 if (is_Phi(value) && get_nodes_block(value) == block) {
998 value = get_irn_n(value, i);
1000 /* we might have unknowns as argument for the phi */
1001 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, value))
1005 if (worklist_contains(value))
1007 if (wl_entry->unused_livethrough_loop != NULL && !is_loop_entry)
1010 be_add_reload_on_edge(senv, value, block, i, cls, 1);
1013 deactivate_worklist(end_worklist);
1017 static void be_spill_belady3(be_irg_t *birg, const arch_register_class_t *ncls)
1019 ir_graph *irg = be_get_birg_irg(birg);
1022 n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1024 /* shortcut for register classes with ignore regs only */
1028 worklist_visited = 0;
1029 arch_env = be_get_birg_arch_env(birg);
1030 exec_freq = be_get_birg_exec_freq(birg);
1032 be_clear_links(irg);
1033 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
1034 | IR_RESOURCE_LOOP_LINK);
1035 inc_irg_visited(irg);
1037 obstack_init(&obst);
1038 senv = be_new_spill_env(birg);
1040 assure_cf_loop(irg);
1041 clear_loop_info(get_irg_loop(irg));
1042 irg_block_walk_graph(irg, construct_loop_edges, NULL, NULL);
1044 process_loop(get_irg_loop(current_ir_graph));
1046 irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1048 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
1049 | IR_RESOURCE_LOOP_LINK);
1051 be_insert_spills_reloads(senv);
1053 obstack_free(&obst, NULL);
1056 be_delete_spill_env(senv);
1059 void be_init_spillbelady3(void)
1061 static be_spiller_t belady3_spiller = {
1065 be_register_spiller("belady3", &belady3_spiller);
1066 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady3");
1069 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady3);