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...
42 #include "iredges_t.h"
50 #include "bespilloptions.h"
51 #include "besched_t.h"
55 #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;
90 typedef struct block_info_t block_info_t;
92 worklist_t *start_worklist;
93 worklist_t *end_worklist;
96 static const arch_register_class_t *cls;
97 static struct obstack obst;
98 static spill_env_t *senv;
100 static size_t max_register_pressure;
101 static bool tentative_mode;
102 static bool should_have_reached_fixpoint;
103 static bool do_push_unused_livethroughs;
104 static ir_exec_freq *exec_freq;
105 static ir_visited_t worklist_visited;
107 static worklist_t *new_worklist(void)
109 worklist_t *worklist = obstack_alloc(&obst, sizeof(worklist[0]));
110 memset(worklist, 0, sizeof(worklist[0]));
112 INIT_LIST_HEAD(&worklist->live_values);
113 worklist->n_live_values = 0;
118 static void mark_irn_not_visited(ir_node *node)
120 set_irn_visited(node, get_irg_visited(current_ir_graph) - 1);
123 static bool worklist_contains(const ir_node *node)
125 return irn_visited(node);
128 static block_info_t *get_block_info(ir_node *block)
130 block_info_t *info = get_irn_link(block);
134 info = obstack_alloc(&obst, sizeof(info[0]));
135 memset(info, 0, sizeof(info[0]));
136 set_irn_link(block, info);
140 static void deactivate_worklist(const worklist_t *worklist)
142 struct list_head *entry;
144 list_for_each(entry, &worklist->live_values) {
145 worklist_entry_t *wl_entry
146 = list_entry(entry, worklist_entry_t, head);
147 assert(worklist_contains(wl_entry->value));
148 mark_irn_not_visited(wl_entry->value);
149 set_irn_link(wl_entry->value, NULL);
153 static void activate_worklist(const worklist_t *worklist)
155 struct list_head *entry;
157 list_for_each(entry, &worklist->live_values) {
158 worklist_entry_t *wl_entry
159 = list_entry(entry, worklist_entry_t, head);
160 assert(!worklist_contains(wl_entry->value));
161 mark_irn_visited(wl_entry->value);
162 set_irn_link(wl_entry->value, wl_entry);
167 * duplicate a worklist and directly activates it
169 static void fill_and_activate_worklist(worklist_t *new_worklist,
170 const worklist_t *worklist, ir_node *block, ir_node *succ_block,
173 ir_node *reload_point = NULL;
174 size_t n_live_values = 0;
175 struct list_head *entry;
177 if (succ_block != NULL &&
178 (get_Block_n_cfgpreds(succ_block) > 1
179 || get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) > 1)) {
180 reload_point = be_get_end_of_block_insertion_point(block);
183 list_for_each(entry, &worklist->live_values) {
184 worklist_entry_t *wl_entry = list_entry(entry, worklist_entry_t, head);
185 ir_node *value = wl_entry->value;
186 worklist_entry_t *new_entry;
188 if (new_worklist->n_live_values >= n_regs)
191 if (is_Phi(value) && get_nodes_block(value) == succ_block) {
192 value = get_Phi_pred(value, succ_pos);
194 /* can happen for unknown phi preds */
195 if (!arch_irn_consider_in_reg_alloc(cls, value))
199 if (irn_visited_else_mark(value))
202 new_entry = obstack_alloc(&obst, sizeof(new_entry[0]));
203 memset(new_entry, 0, sizeof(new_entry[0]));
205 new_entry->value = value;
206 if (reload_point != NULL) {
207 new_entry->reload_point = reload_point;
209 new_entry->reload_point = wl_entry->reload_point;
212 list_add_tail(&new_entry->head, &new_worklist->live_values);
215 set_irn_link(value, new_entry);
216 new_worklist->n_live_values++;
220 static worklist_t *duplicate_worklist(const worklist_t *worklist)
222 worklist_t *new_worklist;
223 struct list_head *entry;
225 new_worklist = obstack_alloc(&obst, sizeof(new_worklist[0]));
226 memset(new_worklist, 0, sizeof(new_worklist[0]));
227 INIT_LIST_HEAD(&new_worklist->live_values);
228 new_worklist->n_live_values = worklist->n_live_values;
230 list_for_each(entry, &worklist->live_values) {
231 worklist_entry_t *wl_entry = list_entry(entry, worklist_entry_t, head);
232 worklist_entry_t *new_entry
233 = obstack_alloc(&obst, sizeof(new_entry[0]));
235 memcpy(new_entry, wl_entry, sizeof(new_entry[0]));
236 list_add_tail(&new_entry->head, &new_worklist->live_values);
243 static void print_worklist(const worklist_t *worklist, int level)
245 struct list_head *entry;
247 DB((dbg, level, "%d values: ", worklist->n_live_values));
248 list_for_each(entry, &worklist->live_values) {
249 worklist_entry_t *wl_entry
250 = list_entry(entry, worklist_entry_t, head);
252 DB((dbg, level, "%+F ", wl_entry->value));
259 static void clear_loop_info(ir_loop *loop)
261 int n_elements = get_loop_n_elements(loop);
265 for (i = 0; i < n_elements; ++i) {
266 loop_element element = get_loop_element(loop, i);
267 if (*element.kind != k_ir_loop)
270 clear_loop_info(element.son);
274 static loop_info_t *get_loop_info(ir_loop *loop)
276 loop_info_t *info = loop->link;
280 info = obstack_alloc(&obst, sizeof(info[0]));
281 memset(info, 0, sizeof(info[0]));
288 * constructs loop in+out edges when called in a block walker
290 static void construct_loop_edges(ir_node *block, void *data)
292 int n_cfgpreds = get_Block_n_cfgpreds(block);
293 ir_loop *loop = get_irn_loop(block);
298 for (i = 0; i < n_cfgpreds; ++i) {
299 ir_node *cfgpred_block = get_Block_cfgpred_block(block, i);
300 ir_loop *cfgpred_loop = get_irn_loop(cfgpred_block);
305 if (cfgpred_loop == loop)
308 /* critical edges are splitted, so we can't jump from 1 loop
309 * directly into another without going out of it */
310 assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
312 /* edge out of loop */
313 if (get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
318 is_exit_edge = false;
323 /* this might be a jump out/in multiple loops */
325 loop_info_t *l_info = get_loop_info(l);
327 edge = obstack_alloc(&obst, sizeof(edge[0]));
328 memset(edge, 0, sizeof(edge[0]));
333 edge->next = l_info->exit_edges;
334 l_info->exit_edges = edge;
335 assert(get_loop_depth(loop) < get_loop_depth(l));
337 edge->next = l_info->entry_edges;
338 l_info->entry_edges = edge;
339 assert(get_loop_depth(cfgpred_loop) < get_loop_depth(l));
341 l = get_loop_outer_loop(l);
349 static void place_reload(worklist_entry_t *entry)
354 DB((dbg, LEVEL_1, "reload of %+F before %+F\n", entry->value,
355 entry->reload_point));
356 assert(entry->reload_point != NULL);
357 be_add_reload(senv, entry->value, entry->reload_point, cls, 1);
358 entry->reload_point = NULL;
362 * makes sure the worklist contains not more than n_regs - room_needed entries
364 static void make_room(worklist_t *worklist, size_t room_needed)
368 struct list_head *entry;
370 spills_needed = worklist->n_live_values + room_needed - n_regs;
371 if (spills_needed <= 0)
374 entry = worklist->live_values.next;
375 for(i = spills_needed; i > 0; --i) {
376 struct list_head *next = entry->next;
377 worklist_entry_t *wl_entry
378 = list_entry(entry, worklist_entry_t, head);
380 assert(worklist_contains(wl_entry->value));
381 mark_irn_not_visited(wl_entry->value);
382 place_reload(wl_entry);
387 worklist->n_live_values -= spills_needed;
391 * a value was used, so bring it to the back of the worklist (which might
392 * result in a spill of another value).
394 static void val_used(worklist_t *worklist, ir_node *value, ir_node *sched_point)
396 /* already in the worklist? move around, otherwise add at back */
397 worklist_entry_t *entry = get_irn_link(value);
399 assert(arch_irn_consider_in_reg_alloc(cls, value));
401 if (worklist_contains(value)) {
402 assert(entry != NULL);
404 list_del(&entry->head);
407 entry = obstack_alloc(&obst, sizeof(entry[0]));
408 memset(entry, 0, sizeof(entry[0]));
410 entry->value = value;
411 set_irn_link(value, entry);
414 ++worklist->n_live_values;
415 mark_irn_visited(value);
418 entry->reload_point = sched_point;
419 list_add_tail(&entry->head, &worklist->live_values);
422 static void worklist_remove(worklist_t *worklist, ir_node *value)
424 worklist_entry_t *entry = get_irn_link(value);
425 assert(entry != NULL);
426 list_del(&entry->head);
427 --worklist->n_live_values;
429 assert(worklist_contains(value));
430 mark_irn_not_visited(value);
433 static void update_max_pressure(worklist_t *worklist)
435 if (worklist->n_live_values > max_register_pressure)
436 max_register_pressure = worklist->n_live_values;
439 static void do_spilling(ir_node *block, worklist_t *worklist)
443 assert(worklist != NULL);
445 sched_foreach_reverse(block, node) {
449 DB((dbg, LEVEL_2, "\t%+F... ", node));
450 update_max_pressure(worklist);
454 /* TODO: if we have some free registers, then we could decide to
455 * not spill some phis (but not for phis where at least 1 input is
458 /* we have to spill all phis that are not live */
459 sched_foreach_reverse_from(node, node2) {
460 assert(is_Phi(node2));
462 if (worklist_contains(node2))
464 if (!arch_irn_consider_in_reg_alloc(cls, node2))
468 be_spill_phi(senv, node2);
470 DB((dbg, LEVEL_2, "\n"));
474 /* remove values defined by this instruction from the workset. Values
475 * defined but not in the workset need free registers */
476 if (get_irn_mode(node) == mode_T) {
477 const ir_edge_t *edge;
479 foreach_out_edge(node, edge) {
480 ir_node *proj = get_edge_src_irn(edge);
481 if (!arch_irn_consider_in_reg_alloc(cls, proj))
483 if (worklist_contains(proj)) {
484 worklist_remove(worklist, proj);
489 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
490 if (worklist_contains(node)) {
491 worklist_remove(worklist, node);
497 /* make sure we have enough free registers for the spills */
498 make_room(worklist, n_defs);
500 /* put all values used by the instruction into the workset */
501 arity = get_irn_arity(node);
502 for(i = 0; i < arity; ++i) {
503 ir_node *use = get_irn_n(node, i);
505 if (!arch_irn_consider_in_reg_alloc(cls, use))
508 val_used(worklist, use, node);
511 /* we might have too many values in the worklist now and need to spill
513 make_room(worklist, 0);
516 print_worklist(worklist, LEVEL_2);
517 DB((dbg, LEVEL_2, "\n"));
521 update_max_pressure(worklist);
524 DB((dbg, LEVEL_1, "worklist at begin of %+F:", block));
525 print_worklist(worklist, LEVEL_1);
526 DB((dbg, LEVEL_1, "\n"));
530 static bool worklists_equal(const worklist_t *wl1, const worklist_t *wl2)
532 const struct list_head *i1 = &wl1->live_values;
533 const struct list_head *i2 = &wl2->live_values;
535 for ( ; i1 != &wl1->live_values && i2 != &wl2->live_values;
536 i1 = i1->next, i2 = i2->next) {
537 worklist_entry_t *entry1 = list_entry(i1, worklist_entry_t, head);
538 worklist_entry_t *entry2 = list_entry(i2, worklist_entry_t, head);
540 if (entry1->value != entry2->value)
543 /* both lists at the end */
544 if (i1 != &wl1->live_values || i2 != &wl2->live_values)
550 static bool fill_start_worklist(worklist_t *new_worklist, ir_node *block)
552 double best_execfreq = -1;
553 worklist_t *best_worklist = NULL;
554 ir_node *best_succ_block = NULL;
556 const ir_edge_t *edge;
558 /* construct worklist */
559 foreach_block_succ(block, edge) {
560 ir_node *succ_block = get_edge_src_irn(edge);
561 double execfreq = get_block_execfreq(exec_freq, succ_block);
562 block_info_t *block_info;
563 worklist_t *succ_worklist;
565 if (execfreq < best_execfreq)
568 block_info = get_block_info(succ_block);
569 succ_worklist = block_info->start_worklist;
571 if (succ_worklist == NULL || succ_worklist->visited >= worklist_visited)
574 best_execfreq = execfreq;
575 best_worklist = succ_worklist;
576 best_succ_block = succ_block;
577 best_pos = get_edge_src_pos(edge);
580 if (best_worklist == NULL)
582 best_worklist->visited = worklist_visited;
584 fill_and_activate_worklist(new_worklist, best_worklist, block,
585 best_succ_block, best_pos);
589 static worklist_t *construct_start_worklist(ir_node *block)
591 worklist_t *worklist = new_worklist();
595 while(fill_start_worklist(worklist, block)) {
596 if (worklist->n_live_values >= n_regs)
603 static void process_block(ir_node *block, void *env)
605 block_info_t *block_info;
606 worklist_t *worklist;
610 DB((dbg, LEVEL_1, "Processing %+F\n", block));
612 worklist = construct_start_worklist(block);
613 block_info = get_block_info(block);
616 DB((dbg, LEVEL_1, "worklist at end of %+F:", block));
617 print_worklist(worklist, LEVEL_1);
618 DB((dbg, LEVEL_1, "\n"));
620 if (should_have_reached_fixpoint) {
621 assert(worklists_equal(block_info->end_worklist, worklist));
624 block_info->end_worklist = duplicate_worklist(worklist);
626 do_spilling(block, worklist);
627 deactivate_worklist(worklist);
630 if (should_have_reached_fixpoint) {
631 assert(worklists_equal(block_info->start_worklist, worklist));
634 block_info->start_worklist = worklist;
636 /* we shouldn't have any live values left at the start block */
637 n_preds = get_Block_n_cfgpreds(block);
638 //assert(n_preds != 0 || worklist->n_live_values == 0);
641 typedef struct block_or_loop_t block_or_loop_t;
642 struct block_or_loop_t {
650 static block_or_loop_t *loop_blocks;
651 static ir_loop *current_loop;
653 static void find_blocks(ir_node *block);
655 static void find_in_loop(ir_loop *loop, ir_node *entry)
657 loop_info_t *loop_info = get_loop_info(loop);
658 block_or_loop_t block_or_loop;
661 /* simply mark 1 block in the loop to indicate that the loop was already
663 ir_node *some_block = loop_info->entry_edges->block;
664 if (Block_block_visited(some_block))
667 block_or_loop.v.loop = loop;
668 block_or_loop.is_loop = true;
669 ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
673 /* we should be 1 of the entry blocks */
674 loop_edge_t *edge = loop_info->entry_edges;
676 for ( ; edge != NULL; edge = edge->next) {
677 if (edge->block == entry)
686 /* check all loop successors */
687 for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
688 ir_node *succ = edge->block;
689 ir_loop *succ_loop = get_irn_loop(succ);
691 if (succ_loop == current_loop) {
694 assert(get_loop_depth(succ_loop) < get_loop_depth(current_loop));
699 static void find_blocks(ir_node *block)
701 const ir_edge_t *edge;
702 block_or_loop_t block_or_loop;
704 if (Block_block_visited(block))
707 block_or_loop.v.block = block;
708 block_or_loop.is_loop = false;
710 ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
711 mark_Block_block_visited(block);
713 foreach_block_succ(block, edge) {
714 ir_node *succ = get_edge_src_irn(edge);
716 /* is it part of the current loop? */
717 ir_loop *loop = get_irn_loop(succ);
718 if (loop != current_loop) {
720 if (get_loop_depth(loop) > get_loop_depth(current_loop)) {
721 find_in_loop(loop, succ);
723 /* parent loop: we're not interested in the block */
732 * append an entry to a worklist. WARNING: The entry must not already be in the
735 static void worklist_append(worklist_t *worklist, ir_node *value,
736 ir_node *reload_point,
737 ir_loop *unused_livethrough_loop)
739 worklist_entry_t *entry = obstack_alloc(&obst, sizeof(entry[0]));
740 memset(entry, 0, sizeof(entry[0]));
742 #ifdef EXPENSIVE_CHECKS
744 struct list_head *entry;
745 list_for_each(entry, &worklist->live_values) {
746 worklist_entry_t *wl_entry
747 = list_entry(entry, worklist_entry_t, head);
748 assert(wl_entry->value != value);
753 entry->value = value;
754 entry->reload_point = reload_point;
755 entry->unused_livethrough_loop = unused_livethrough_loop;
756 list_add_tail(&entry->head, &worklist->live_values);
757 ++worklist->n_live_values;
758 assert(worklist->n_live_values <= n_regs);
761 static void push_unused_livethrough(loop_info_t *loop_info, ir_node *value)
766 /* add the value to all loop exit and entry blocks */
767 for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
769 = get_Block_cfgpred_block(edge->block, edge->pos);
770 const block_info_t *info = get_block_info(block);
771 worklist_t *worklist = info->end_worklist;
772 ir_node *reload_point = NULL;
774 if (worklist->visited >= worklist_visited)
776 worklist->visited = worklist_visited;
778 /* TODO: we need a smarter mechanism here, that makes the reloader place
779 * reload nodes on all loop exits... */
781 worklist_append(worklist, value, reload_point, loop_info->loop);
783 edge = loop_info->entry_edges;
784 for ( ; edge != NULL; edge = edge->next) {
785 ir_node *entry_block = edge->block;
786 const block_info_t *info = get_block_info(entry_block);
787 worklist_t *worklist = info->start_worklist;
789 ir_node *reload_point;
791 if (worklist->visited >= worklist_visited)
793 worklist->visited = worklist_visited;
795 pred_block = get_Block_cfgpred_block(entry_block, edge->pos);
796 reload_point = be_get_end_of_block_insertion_point(pred_block);
798 worklist_append(worklist, value, reload_point, loop_info->loop);
801 set_irn_link(value, NULL);
802 ++loop_info->max_register_pressure;
805 static void push_unused_livethroughs(loop_info_t *loop_info)
809 /* we can only push unused livethroughs if register pressure inside the loop
811 if (loop_info->max_register_pressure >= n_regs)
814 /* find unused livethroughs: register pressure in the loop was low enough
815 * which means that we had no spills which implies that at every point in
817 for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
818 ir_node *block = edge->block;
819 const block_info_t *info = get_block_info(block);
820 worklist_t *start_worklist = info->start_worklist;
822 const block_info_t *exit_info;
823 worklist_t *end_worklist;
824 struct list_head *entry;
826 if (start_worklist == NULL)
829 exit_block = get_Block_cfgpred_block(edge->block, edge->pos);
830 exit_info = get_block_info(exit_block);
831 end_worklist = exit_info->end_worklist;
833 activate_worklist(end_worklist);
834 /* all values contained in the start_worklist, which are not available
835 * in the end_worklist, must be unused livethroughs */
837 list_for_each(entry, &start_worklist->live_values) {
838 worklist_entry_t *wl_entry
839 = list_entry(entry, worklist_entry_t, head);
840 ir_node *value = wl_entry->value;
842 if (loop_info->max_register_pressure >= n_regs)
846 if (!worklist_contains(value)) {
847 /* add it to all loop exits */
848 DB((dbg, LEVEL_2, "Found unused livethrough: %+F (regpressure in loop %d)\n", value, loop_info->max_register_pressure));
849 push_unused_livethrough(loop_info, value);
850 /* the value should now be part of end_worklist, so mark it */
851 mark_irn_visited(value);
855 deactivate_worklist(end_worklist);
859 static void process_block_or_loop(const block_or_loop_t *block_or_loop)
861 if (block_or_loop->is_loop) {
862 loop_info_t *loop_info = get_loop_info(block_or_loop->v.loop);
864 if (do_push_unused_livethroughs)
865 push_unused_livethroughs(loop_info);
867 if (loop_info->max_register_pressure > max_register_pressure)
868 max_register_pressure = loop_info->max_register_pressure;
872 process_block(block_or_loop->v.block, NULL);
875 static void process_loop(ir_loop *loop)
877 int n_elements = get_loop_n_elements(loop);
879 loop_info_t *loop_info;
883 /* first handle all sub-loops */
884 for (i = 0; i < n_elements; ++i) {
885 loop_element element = get_loop_element(loop, i);
886 if (*element.kind != k_ir_loop)
889 process_loop(element.son);
892 /* create a postorder of the blocks */
893 loop_info = get_loop_info(loop);
894 edge = loop_info->entry_edges;
896 some_block = edge->block;
898 assert(loop == get_irg_loop(current_ir_graph));
899 some_block = get_irg_start_block(current_ir_graph);
902 loop_blocks = NEW_ARR_F(block_or_loop_t,0);
905 ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
906 inc_irg_block_visited(current_ir_graph);
907 find_blocks(some_block);
908 /* for endless loops the end-block might be unreachable */
909 if (loop == get_irg_loop(current_ir_graph)) {
910 find_blocks(get_irg_end_block(current_ir_graph));
912 ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
914 DB((dbg, LEVEL_3, "Block List for loop %p\n", loop));
915 len = ARR_LEN(loop_blocks);
916 for (i = 0; i < len; ++i) {
917 block_or_loop_t *block_or_loop = &loop_blocks[i];
918 if (block_or_loop->is_loop) {
919 DB((dbg, LEVEL_3, " L-%p", block_or_loop->v.loop));
921 DB((dbg, LEVEL_3, " B-%+F", block_or_loop->v.block));
924 DB((dbg, LEVEL_3, "\n"));
926 max_register_pressure = 0;
928 /* run1: tentative phase */
929 tentative_mode = true;
930 for (i = len-1; i >= 0; --i) {
931 process_block_or_loop(&loop_blocks[i]);
934 /* run2: tentative phase - should reach fixpoint */
935 tentative_mode = true;
936 for (i = len-1; i >= 0; --i) {
937 process_block_or_loop(&loop_blocks[i]);
941 /* run3: tentative phase - check fixpoint */
942 tentative_mode = true;
943 should_have_reached_fixpoint = true;
944 for (i = len-1; i >= 0; --i) {
945 process_block_or_loop(&loop_blocks[i]);
947 should_have_reached_fixpoint = false;
950 /* run4: add spills/reloads */
951 tentative_mode = false;
952 do_push_unused_livethroughs = true;
953 for (i = len-1; i >= 0; --i) {
954 process_block_or_loop(&loop_blocks[i]);
956 do_push_unused_livethroughs = false;
958 loop_info->max_register_pressure = max_register_pressure;
959 DB((dbg, LEVEL_2, "Regpressure in loop %p: %u\n", loop,
960 (unsigned) max_register_pressure));
962 DEL_ARR_F(loop_blocks);
965 static void fix_block_borders(ir_node *block, void *data)
967 block_info_t *block_info = get_block_info(block);
968 worklist_t *start_worklist = block_info->start_worklist;
969 int n_cfgpreds = get_Block_n_cfgpreds(block);
970 ir_loop *loop = get_irn_loop(block);
974 assert(start_worklist != NULL);
976 for (i = 0; i < n_cfgpreds; ++i) {
977 ir_node *pred_block = get_Block_cfgpred_block(block, i);
978 block_info_t *pred_block_info = get_block_info(pred_block);
979 worklist_t *end_worklist = pred_block_info->end_worklist;
980 ir_loop *pred_loop = get_irn_loop(pred_block);
981 bool is_loop_entry = false;
982 struct list_head *entry;
984 assert(end_worklist != NULL);
986 if (get_loop_depth(pred_loop) < get_loop_depth(loop)) {
987 is_loop_entry = true;
990 /* reload missing values */
991 activate_worklist(end_worklist);
993 list_for_each(entry, &start_worklist->live_values) {
994 worklist_entry_t *wl_entry
995 = list_entry(entry, worklist_entry_t, head);
996 ir_node *value = wl_entry->value;
998 if (is_Phi(value) && get_nodes_block(value) == block) {
999 value = get_irn_n(value, i);
1001 /* we might have unknowns as argument for the phi */
1002 if (!arch_irn_consider_in_reg_alloc(cls, value))
1006 if (worklist_contains(value))
1008 if (wl_entry->unused_livethrough_loop != NULL && !is_loop_entry)
1011 be_add_reload_on_edge(senv, value, block, i, cls, 1);
1014 deactivate_worklist(end_worklist);
1018 static void be_spill_belady3(be_irg_t *birg, const arch_register_class_t *ncls)
1020 ir_graph *irg = be_get_birg_irg(birg);
1023 n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1025 /* shortcut for register classes with ignore regs only */
1029 worklist_visited = 0;
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);