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 Simple copy minimization heuristics.
23 * @author Christian Wuerdig
27 * This is the C implementation of the mst algorithm
28 * originally written in Java by Sebastian Hack.
29 * (also known as "heur3" :)
30 * Performs simple copy minimization.
34 #define DISABLE_STATEV
41 #include "raw_bitset.h"
42 #include "irnodemap.h"
59 #include "becopyopt_t.h"
63 #define COL_COST_INFEASIBLE DBL_MAX
64 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
65 #define NEIGHBOUR_CONSTR_COSTS 64.0
70 #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while (0)
71 #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while (0)
73 static firm_dbg_module_t *dbg = NULL;
77 #define DBG_AFF_CHUNK(env, level, chunk)
78 #define DBG_COL_COST(env, level, cost)
83 #define REAL(C) (C ## f)
85 static unsigned last_chunk_id = 0;
86 static int recolor_limit = 7;
87 static double dislike_influence = REAL(0.1);
89 typedef struct col_cost_t {
97 typedef struct aff_chunk_t {
98 const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
99 const ir_node **interfere; /**< An ARR_F containing all inference. */
100 int weight; /**< Weight of this chunk */
101 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
102 unsigned deleted : 1; /**< For debugging: Set if the was deleted. */
103 unsigned id; /**< An id of this chunk. */
106 col_cost_t color_affinity[1];
112 typedef struct aff_edge_t {
113 const ir_node *src; /**< Source node. */
114 const ir_node *tgt; /**< Target node. */
115 int weight; /**< The weight of this edge. */
118 /* main coalescing environment */
119 typedef struct co_mst_env_t {
120 int n_regs; /**< number of regs in class */
121 int k; /**< number of non-ignore registers in class */
122 bitset_t *allocatable_regs; /**< set containing all global ignore registers */
123 ir_nodemap map; /**< phase object holding data for nodes */
125 pqueue_t *chunks; /**< priority queue for chunks */
126 list_head chunklist; /**< list holding all chunks */
127 be_ifg_t *ifg; /**< the interference graph */
128 copy_opt_t *co; /**< the copy opt object */
129 unsigned chunk_visited;
130 col_cost_t **single_cols;
133 /* stores coalescing related information for a node */
134 typedef struct co_mst_irn_t {
135 const ir_node *irn; /**< the irn this information belongs to */
136 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
137 bitset_t *adm_colors; /**< set of admissible colors for this irn */
138 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
139 int n_neighs; /**< length of the interfering neighbours array. */
140 int int_aff_neigh; /**< number of interfering affinity neighbours */
141 int col; /**< color currently assigned */
142 int init_col; /**< the initial color */
143 int tmp_col; /**< a temporary assigned color */
144 unsigned fixed : 1; /**< the color is fixed */
145 struct list_head list; /**< Queue for coloring undo. */
146 real_t constr_factor;
150 * In case there is no phase information for irn, initialize it.
152 static co_mst_irn_t *co_mst_irn_init(co_mst_env_t *env, const ir_node *irn)
154 co_mst_irn_t *res = OALLOC(&env->obst, co_mst_irn_t);
156 const arch_register_req_t *req;
157 neighbours_iter_t nodes_it;
165 res->int_neighs = NULL;
166 res->int_aff_neigh = 0;
167 res->col = arch_register_get_index(arch_get_irn_register(irn));
168 res->init_col = res->col;
169 INIT_LIST_HEAD(&res->list);
171 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
173 /* set admissible registers */
174 res->adm_colors = bitset_obstack_alloc(&env->obst, env->n_regs);
176 /* Exclude colors not assignable to the irn */
177 req = arch_get_irn_register_req(irn);
178 if (arch_register_req_is(req, limited)) {
179 rbitset_copy_to_bitset(req->limited, res->adm_colors);
181 bitset_set_all(res->adm_colors);
184 /* exclude global ignore registers as well */
185 bitset_and(res->adm_colors, env->allocatable_regs);
187 /* compute the constraint factor */
188 res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcount(res->adm_colors)) / env->n_regs;
190 /* set the number of interfering affinity neighbours to -1, they are calculated later */
191 res->int_aff_neigh = -1;
193 /* build list of interfering neighbours */
195 be_ifg_foreach_neighbour(env->ifg, &nodes_it, irn, neigh) {
196 if (!arch_irn_is_ignore(neigh)) {
197 obstack_ptr_grow(&env->obst, neigh);
201 res->int_neighs = (ir_node**)obstack_finish(&env->obst);
206 static co_mst_irn_t *get_co_mst_irn(co_mst_env_t *env, const ir_node *node)
208 co_mst_irn_t *res = (co_mst_irn_t*)ir_nodemap_get(&env->map, node);
210 res = co_mst_irn_init(env, node);
211 ir_nodemap_insert(&env->map, node, res);
216 typedef int decide_func_t(const co_mst_irn_t *node, int col);
221 * Write a chunk to stderr for debugging.
223 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c)
227 if (c->weight_consistent)
228 ir_fprintf(stderr, " $%d ", c->weight);
229 ir_fprintf(stderr, "{");
230 for (i = 0, l = ARR_LEN(c->n); i < l; ++i) {
231 const ir_node *n = c->n[i];
232 ir_fprintf(stderr, " %+F,", n);
234 ir_fprintf(stderr, "}");
238 * Dump all admissible colors to stderr.
240 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node)
245 if (bitset_popcount(node->adm_colors) < 1)
246 fprintf(stderr, "no admissible colors?!?");
248 bitset_foreach(node->adm_colors, idx) {
249 ir_fprintf(stderr, " %zu", idx);
255 * Dump color-cost pairs to stderr.
257 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost)
260 for (i = 0; i < env->n_regs; ++i)
261 fprintf(stderr, " (%d, %.4f)", cost[i].col, cost[i].cost);
264 #endif /* DEBUG_libfirm */
266 static inline int get_mst_irn_col(const co_mst_irn_t *node)
268 return node->tmp_col >= 0 ? node->tmp_col : node->col;
272 * @return 1 if node @p node has color @p col, 0 otherwise.
274 static int decider_has_color(const co_mst_irn_t *node, int col)
276 return get_mst_irn_col(node) == col;
280 * @return 1 if node @p node has not color @p col, 0 otherwise.
282 static int decider_hasnot_color(const co_mst_irn_t *node, int col)
284 return get_mst_irn_col(node) != col;
288 * Always returns true.
290 static int decider_always_yes(const co_mst_irn_t *node, int col)
297 /** compares two affinity edges by its weight */
298 static int cmp_aff_edge(const void *a, const void *b)
300 const aff_edge_t *e1 = (const aff_edge_t*)a;
301 const aff_edge_t *e2 = (const aff_edge_t*)b;
303 if (e2->weight == e1->weight) {
304 if (e2->src->node_idx == e1->src->node_idx)
305 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
307 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
309 /* sort in descending order */
310 return QSORT_CMP(e2->weight, e1->weight);
313 /** compares to color-cost pairs */
314 static __attribute__((unused)) int cmp_col_cost_lt(const void *a, const void *b)
316 const col_cost_t *c1 = (const col_cost_t*)a;
317 const col_cost_t *c2 = (const col_cost_t*)b;
318 real_t diff = c1->cost - c2->cost;
319 return (diff > 0) - (diff < 0);
322 static int cmp_col_cost_gt(const void *a, const void *b)
324 const col_cost_t *c1 = (const col_cost_t*)a;
325 const col_cost_t *c2 = (const col_cost_t*)b;
326 real_t diff = c2->cost - c1->cost;
329 return QSORT_CMP(c1->col, c2->col);
331 return (diff > 0) - (diff < 0);
335 * Creates a new affinity chunk
337 static inline aff_chunk_t *new_aff_chunk(co_mst_env_t *env)
339 aff_chunk_t *c = XMALLOCF(aff_chunk_t, color_affinity, env->n_regs);
340 c->n = NEW_ARR_F(const ir_node *, 0);
341 c->interfere = NEW_ARR_F(const ir_node *, 0);
343 c->weight_consistent = 0;
345 c->id = ++last_chunk_id;
347 list_add(&c->list, &env->chunklist);
352 * Frees all memory allocated by an affinity chunk.
354 static inline void delete_aff_chunk(aff_chunk_t *c)
357 DEL_ARR_F(c->interfere);
364 * binary search of sorted nodes.
366 * @return the position where n is found in the array arr or ~pos
367 * if the nodes is not here.
369 static inline int nodes_bsearch(const ir_node **arr, const ir_node *n)
371 int hi = ARR_LEN(arr);
375 int md = lo + ((hi - lo) >> 1);
388 /** Check if a node n can be found inside arr. */
389 static int node_contains(const ir_node **arr, const ir_node *n)
391 int i = nodes_bsearch(arr, n);
396 * Insert a node into the sorted nodes list.
398 * @return 1 if the node was inserted, 0 else
400 static int nodes_insert(const ir_node ***arr, const ir_node *irn)
402 int idx = nodes_bsearch(*arr, irn);
405 int i, n = ARR_LEN(*arr);
408 ARR_APP1(const ir_node *, *arr, irn);
413 for (i = n - 1; i >= idx; --i)
422 * Adds a node to an affinity chunk
424 static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node)
428 if (! nodes_insert(&c->n, node->irn))
431 c->weight_consistent = 0;
434 for (i = node->n_neighs - 1; i >= 0; --i) {
435 ir_node *neigh = node->int_neighs[i];
436 nodes_insert(&c->interfere, neigh);
441 * Check if affinity chunk @p chunk interferes with node @p irn.
443 static inline int aff_chunk_interferes(const aff_chunk_t *chunk, const ir_node *irn)
445 return node_contains(chunk->interfere, irn);
449 * Check if there are interference edges from c1 to c2.
451 * @param c2 Another chunk
452 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
454 static inline int aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2)
461 /* check if there is a node in c2 having an interfering neighbor in c1 */
462 for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) {
463 const ir_node *irn = c2->n[i];
465 if (node_contains(c1->interfere, irn))
472 * Returns the affinity chunk of @p irn or creates a new
473 * one with @p irn as element if there is none assigned.
475 static inline aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn)
477 co_mst_irn_t *node = get_co_mst_irn(env, irn);
482 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
483 * are no interference edges from chunk(src) to chunk(tgt)).
484 * @return 1 if successful, 0 if not possible
486 static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node *tgt)
488 aff_chunk_t *c1 = get_aff_chunk(env, src);
489 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
492 DB((dbg, LEVEL_4, "Attempt to let c1 (id %u): ", c1 ? c1->id : 0));
494 DBG_AFF_CHUNK(env, LEVEL_4, c1);
496 DB((dbg, LEVEL_4, "{%+F}", src));
498 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %u): ", c2 ? c2->id : 0));
500 DBG_AFF_CHUNK(env, LEVEL_4, c2);
502 DB((dbg, LEVEL_4, "{%+F}", tgt));
504 DB((dbg, LEVEL_4, "\n"));
509 /* no chunk exists */
510 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
513 for (i = mirn->n_neighs - 1; i >= 0; --i) {
514 if (mirn->int_neighs[i] == tgt)
518 /* create one containing both nodes */
519 c1 = new_aff_chunk(env);
520 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
521 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
525 /* c2 already exists */
526 if (! aff_chunk_interferes(c2, src)) {
527 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
531 } else if (c2 == NULL) {
532 /* c1 already exists */
533 if (! aff_chunk_interferes(c1, tgt)) {
534 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
537 } else if (c1 != c2 && ! aff_chunks_interfere(c1, c2)) {
540 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx)
541 aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx]));
543 for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) {
544 const ir_node *irn = c2->interfere[idx];
545 nodes_insert(&c1->interfere, irn);
548 c1->weight_consistent = 0;
550 delete_aff_chunk(c2);
553 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
557 DB((dbg, LEVEL_4, " ... absorbed\n"));
562 * Assures that the weight of the given chunk is consistent.
564 static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c)
566 if (! c->weight_consistent) {
570 for (i = 0; i < env->n_regs; ++i) {
571 c->color_affinity[i].col = i;
572 c->color_affinity[i].cost = REAL(0.0);
575 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
576 const ir_node *n = c->n[idx];
577 const affinity_node_t *an = get_affinity_info(env->co, n);
578 co_mst_irn_t *node = get_co_mst_irn(env, n);
581 if (node->constr_factor > REAL(0.0)) {
583 bitset_foreach (node->adm_colors, col)
584 c->color_affinity[col].cost += node->constr_factor;
589 co_gs_foreach_neighb(an, neigh) {
590 const ir_node *m = neigh->irn;
592 if (arch_irn_is_ignore(m))
595 w += node_contains(c->n, m) ? neigh->costs : 0;
600 for (i = 0; i < env->n_regs; ++i)
601 c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n));
604 // c->weight = bitset_popcount(c->nodes);
605 c->weight_consistent = 1;
610 * Count the number of interfering affinity neighbours
612 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an)
614 const neighb_t *neigh;
615 const ir_node *irn = an->irn;
616 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
619 co_gs_foreach_neighb(an, neigh) {
620 const ir_node *n = neigh->irn;
623 if (arch_irn_is_ignore(n))
626 /* check if the affinity neighbour interfere */
627 for (i = 0; i < node->n_neighs; ++i) {
628 if (node->int_neighs[i] == n) {
639 * Build chunks of nodes connected by affinity edges.
640 * We start at the heaviest affinity edge.
641 * The chunks of the two edge-defining nodes will be
642 * merged if there are no interference edges from one
643 * chunk to the other.
645 static void build_affinity_chunks(co_mst_env_t *env)
647 nodes_iter_t nodes_it;
648 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
651 aff_chunk_t *curr_chunk;
654 /* at first we create the affinity edge objects */
655 be_ifg_foreach_node(env->ifg, &nodes_it, n) {
656 int n_idx = get_irn_idx(n);
660 if (arch_irn_is_ignore(n))
663 n1 = get_co_mst_irn(env, n);
664 an = get_affinity_info(env->co, n);
669 if (n1->int_aff_neigh < 0)
670 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
672 /* build the affinity edges */
673 co_gs_foreach_neighb(an, neigh) {
674 const ir_node *m = neigh->irn;
675 int m_idx = get_irn_idx(m);
677 /* record the edge in only one direction */
682 /* skip ignore nodes */
683 if (arch_irn_is_ignore(m))
689 n2 = get_co_mst_irn(env, m);
690 if (n2->int_aff_neigh < 0) {
691 affinity_node_t *am = get_affinity_info(env->co, m);
692 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
695 * these weights are pure hackery ;-).
696 * It's not chriswue's fault but mine.
698 edge.weight = neigh->costs;
699 ARR_APP1(aff_edge_t, edges, edge);
705 /* now: sort edges and build the affinity chunks */
706 len = ARR_LEN(edges);
707 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
708 for (i = 0; i < len; ++i) {
709 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
711 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
714 /* now insert all chunks into a priority queue */
715 list_for_each_entry(aff_chunk_t, curr_chunk, &env->chunklist, list) {
716 aff_chunk_assure_weight(env, curr_chunk);
718 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
719 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
720 DBG((dbg, LEVEL_1, "\n"));
722 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
725 for (pn = 0; pn < ARR_LEN(env->map.data); ++pn) {
726 co_mst_irn_t *mirn = env->map.data[pn];
729 if (mirn->chunk != NULL)
732 /* no chunk is allocated so far, do it now */
733 aff_chunk_t *curr_chunk = new_aff_chunk(env);
734 aff_chunk_add_node(curr_chunk, mirn);
736 aff_chunk_assure_weight(env, curr_chunk);
738 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
739 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
740 DBG((dbg, LEVEL_1, "\n"));
742 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
748 static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk)
750 pqueue_t *grow = new_pqueue();
751 ir_node const *max_node = NULL;
755 for (i = ARR_LEN(chunk->n); i != 0;) {
756 const ir_node *irn = chunk->n[--i];
757 affinity_node_t *an = get_affinity_info(env->co, irn);
761 if (arch_irn_is_ignore(irn))
765 co_gs_foreach_neighb(an, neigh)
768 if (w > max_weight) {
776 bitset_t *visited = bitset_irg_malloc(env->co->irg);
778 for (i = ARR_LEN(chunk->n); i != 0;)
779 bitset_add_irn(visited, chunk->n[--i]);
781 pqueue_put(grow, (void *) max_node, max_weight);
782 bitset_remv_irn(visited, max_node);
784 while (!pqueue_empty(grow)) {
785 ir_node *irn = (ir_node*)pqueue_pop_front(grow);
786 affinity_node_t *an = get_affinity_info(env->co, irn);
789 if (arch_irn_is_ignore(irn))
792 assert(i <= ARR_LEN(chunk->n));
797 /* build the affinity edges */
798 co_gs_foreach_neighb(an, neigh) {
799 co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn);
801 if (bitset_contains_irn(visited, node->irn)) {
802 pqueue_put(grow, (void *) neigh->irn, neigh->costs);
803 bitset_remv_irn(visited, node->irn);
809 bitset_free(visited);
814 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
816 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
817 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
819 waitq *nodes = new_waitq();
821 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col));
823 /* init queue and chunk */
824 waitq_put(nodes, node);
825 bitset_set(visited, get_irn_idx(node->irn));
826 aff_chunk_add_node(chunk, node);
827 DB((dbg, LEVEL_1, " %+F", node->irn));
829 /* as long as there are nodes in the queue */
830 while (! waitq_empty(nodes)) {
831 co_mst_irn_t *n = (co_mst_irn_t*)waitq_get(nodes);
832 affinity_node_t *an = get_affinity_info(env->co, n->irn);
834 /* check all affinity neighbors */
837 co_gs_foreach_neighb(an, neigh) {
838 const ir_node *m = neigh->irn;
839 int m_idx = get_irn_idx(m);
842 if (arch_irn_is_ignore(m))
845 n2 = get_co_mst_irn(env, m);
847 if (! bitset_is_set(visited, m_idx) &&
850 ! aff_chunk_interferes(chunk, m) &&
851 node_contains(orig_chunk->n, m))
854 following conditions are met:
855 - neighbour is not visited
856 - neighbour likes the color
857 - neighbour has not yet a fixed color
858 - the new chunk doesn't interfere with the neighbour
859 - neighbour belongs or belonged once to the original chunk
861 bitset_set(visited, m_idx);
862 aff_chunk_add_node(chunk, n2);
863 DB((dbg, LEVEL_1, " %+F", n2->irn));
864 /* enqueue for further search */
865 waitq_put(nodes, n2);
871 DB((dbg, LEVEL_1, "\n"));
877 * Fragment the given chunk into chunks having given color and not having given color.
879 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp)
881 bitset_t *visited = bitset_irg_malloc(env->co->irg);
883 aff_chunk_t *best = NULL;
885 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
888 aff_chunk_t *tmp_chunk;
889 decide_func_t *decider;
893 if (bitset_is_set(visited, get_irn_idx(irn)))
896 node = get_co_mst_irn(env, irn);
898 if (get_mst_irn_col(node) == col) {
899 decider = decider_has_color;
901 DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col));
904 decider = decider_hasnot_color;
906 DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col));
909 /* create a new chunk starting at current node */
910 tmp_chunk = new_aff_chunk(env);
911 waitq_put(tmp, tmp_chunk);
912 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
913 assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk");
915 /* remember the local best */
916 aff_chunk_assure_weight(env, tmp_chunk);
917 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
921 assert(best && "No chunk found?");
922 bitset_free(visited);
927 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
928 * ATTENTION: the queue is empty after calling this function!
930 static inline void reject_coloring(struct list_head *nodes)
932 co_mst_irn_t *n, *temp;
933 DB((dbg, LEVEL_4, "\treject coloring for"));
934 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
935 DB((dbg, LEVEL_4, " %+F", n->irn));
936 assert(n->tmp_col >= 0);
938 list_del_init(&n->list);
940 DB((dbg, LEVEL_4, "\n"));
943 static inline void materialize_coloring(struct list_head *nodes)
945 co_mst_irn_t *n, *temp;
946 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
947 assert(n->tmp_col >= 0);
950 list_del_init(&n->list);
954 static inline void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
957 assert(!node->fixed);
958 assert(node->tmp_col < 0);
959 assert(node->list.next == &node->list && node->list.prev == &node->list);
960 assert(bitset_is_set(node->adm_colors, col));
962 list_add_tail(&node->list, changed);
966 static inline int is_loose(co_mst_irn_t *node)
968 return !node->fixed && node->tmp_col < 0;
972 * Determines the costs for each color if it would be assigned to node @p node.
974 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs)
976 int *neigh_cols = ALLOCAN(int, env->n_regs);
981 for (i = 0; i < env->n_regs; ++i) {
984 costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0);
987 for (i = 0; i < node->n_neighs; ++i) {
988 co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]);
989 int col = get_mst_irn_col(n);
994 costs[col].cost = REAL(0.0);
998 coeff = REAL(1.0) / n_loose;
999 for (i = 0; i < env->n_regs; ++i)
1000 costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i];
1004 /* need forward declaration due to recursive call */
1005 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones, int depth, int *max_depth, int *trip);
1008 * Tries to change node to a color but @p explude_col.
1009 * @return 1 if succeeded, 0 otherwise.
1011 static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, struct list_head *changed, int depth, int *max_depth, int *trip)
1013 int col = get_mst_irn_col(node);
1016 /* neighbours has already a different color -> good, temporary fix it */
1017 if (col != exclude_col) {
1019 set_temp_color(node, col, changed);
1023 /* The node has the color it should not have _and_ has not been visited yet. */
1024 if (is_loose(node)) {
1025 col_cost_t *costs = ALLOCAN(col_cost_t, env->n_regs);
1027 /* Get the costs for giving the node a specific color. */
1028 determine_color_costs(env, node, costs);
1030 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
1031 costs[exclude_col].cost = REAL(0.0);
1033 /* sort the colors according costs, cheapest first. */
1034 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt);
1036 /* Try recoloring the node using the color list. */
1037 res = recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip);
1044 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
1045 * ATTENTION: Expect @p costs already sorted by increasing costs.
1046 * @return 1 if coloring could be applied, 0 otherwise.
1048 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed, int depth, int *max_depth, int *trip)
1051 struct list_head local_changed;
1054 if (depth > *max_depth)
1057 DBG((dbg, LEVEL_4, "\tRecoloring %+F with color-costs", node->irn));
1058 DBG_COL_COST(env, LEVEL_4, costs);
1059 DB((dbg, LEVEL_4, "\n"));
1061 if (depth >= recolor_limit) {
1062 DBG((dbg, LEVEL_4, "\tHit recolor limit\n"));
1066 for (i = 0; i < env->n_regs; ++i) {
1067 int tgt_col = costs[i].col;
1071 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
1072 if (costs[i].cost == REAL(0.0)) {
1073 DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n"));
1077 /* Set the new color of the node and mark the node as temporarily fixed. */
1078 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
1079 INIT_LIST_HEAD(&local_changed);
1080 set_temp_color(node, tgt_col, &local_changed);
1081 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
1083 /* try to color all interfering neighbours with current color forbidden */
1084 for (j = 0; j < node->n_neighs; ++j) {
1088 neigh = node->int_neighs[j];
1090 if (arch_irn_is_ignore(neigh))
1093 nn = get_co_mst_irn(env, neigh);
1094 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
1095 neigh, j, nn->fixed, nn->tmp_col, nn->col));
1098 Try to change the color of the neighbor and record all nodes which
1099 get changed in the tmp list. Add this list to the "changed" list for
1100 that color. If we did not succeed to change the color of the neighbor,
1101 we bail out and try the next color.
1103 if (get_mst_irn_col(nn) == tgt_col) {
1104 /* try to color neighbour with tgt_col forbidden */
1105 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip);
1113 We managed to assign the target color to all neighbors, so from the perspective
1114 of the current node, every thing was ok and we can return safely.
1117 /* append the local_changed ones to global ones */
1118 list_splice(&local_changed, changed);
1122 /* coloring of neighbours failed, so we try next color */
1123 reject_coloring(&local_changed);
1127 DBG((dbg, LEVEL_4, "\tAll colors failed\n"));
1132 * Tries to bring node @p node and all its neighbours to color @p tgt_col.
1133 * @return 1 if color @p col could be applied, 0 otherwise
1135 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed)
1137 int col = get_mst_irn_col(node);
1139 /* if node already has the target color -> good, temporary fix it */
1140 if (col == tgt_col) {
1141 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1143 set_temp_color(node, tgt_col, changed);
1148 Node has not yet a fixed color and target color is admissible
1149 -> try to recolor node and its affinity neighbours
1151 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1152 col_cost_t *costs = env->single_cols[tgt_col];
1153 int res, max_depth, trip;
1158 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1159 res = recolor_nodes(env, node, costs, changed, 0, &max_depth, &trip);
1160 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1161 stat_ev_int("heur4_recolor_depth_max", max_depth);
1162 stat_ev_int("heur4_recolor_trip", trip);
1168 #ifdef DEBUG_libfirm
1169 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1170 if (!is_loose(node))
1171 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1173 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1174 dbg_admissible_colors(env, node);
1175 DB((dbg, LEVEL_4, ")\n"));
1184 * Tries to color an affinity chunk (or at least a part of it).
1185 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1187 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c)
1189 aff_chunk_t *best_chunk = NULL;
1190 int n_nodes = ARR_LEN(c->n);
1191 int best_color = -1;
1192 int n_int_chunks = 0;
1193 waitq *tmp_chunks = new_waitq();
1194 waitq *best_starts = NULL;
1195 col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs);
1202 struct list_head changed;
1204 DB((dbg, LEVEL_2, "fragmentizing chunk #%u", c->id));
1205 DBG_AFF_CHUNK(env, LEVEL_2, c);
1206 DB((dbg, LEVEL_2, "\n"));
1208 stat_ev_ctx_push_fmt("heur4_color_chunk", "%u", c->id);
1210 ++env->chunk_visited;
1212 /* compute color preference */
1213 for (pos = 0, len = ARR_LEN(c->interfere); pos < len; ++pos) {
1214 const ir_node *n = c->interfere[pos];
1215 co_mst_irn_t *node = get_co_mst_irn(env, n);
1216 aff_chunk_t *chunk = node->chunk;
1218 if (is_loose(node) && chunk && chunk->visited < env->chunk_visited) {
1219 assert(!chunk->deleted);
1220 chunk->visited = env->chunk_visited;
1223 aff_chunk_assure_weight(env, chunk);
1224 for (i = 0; i < env->n_regs; ++i)
1225 order[i].cost += chunk->color_affinity[i].cost;
1229 for (i = 0; i < env->n_regs; ++i) {
1230 real_t dislike = n_int_chunks > 0 ? REAL(1.0) - order[i].cost / n_int_chunks : REAL(0.0);
1232 order[i].cost = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike;
1235 qsort(order, env->n_regs, sizeof(order[0]), cmp_col_cost_gt);
1237 DBG_COL_COST(env, LEVEL_2, order);
1238 DB((dbg, LEVEL_2, "\n"));
1240 /* check which color is the "best" for the given chunk.
1241 * if we found a color which was ok for all nodes, we take it
1242 * and do not look further. (see did_all flag usage below.)
1243 * If we have many colors which fit all nodes it is hard to decide
1244 * which one to take anyway.
1245 * TODO Sebastian: Perhaps we should at all nodes and figure out
1246 * a suitable color using costs as done above (determine_color_costs).
1248 for (i = 0; i < env->k; ++i) {
1249 int col = order[i].col;
1250 waitq *good_starts = new_waitq();
1251 aff_chunk_t *local_best;
1254 /* skip ignore colors */
1255 if (!bitset_is_set(env->allocatable_regs, col))
1258 DB((dbg, LEVEL_2, "\ttrying color %d\n", col));
1262 /* try to bring all nodes of given chunk to the current color. */
1263 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1264 const ir_node *irn = c->n[idx];
1265 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1268 assert(! node->fixed && "Node must not have a fixed color.");
1269 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1272 The order of the colored nodes is important, so we record the successfully
1273 colored ones in the order they appeared.
1275 INIT_LIST_HEAD(&changed);
1277 good = change_node_color(env, node, col, &changed);
1278 stat_ev_tim_pop("heur4_recolor");
1280 waitq_put(good_starts, node);
1281 materialize_coloring(&changed);
1286 reject_coloring(&changed);
1288 n_succeeded += good;
1289 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, good ? "succeeded" : "failed"));
1292 /* unfix all nodes */
1293 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1294 co_mst_irn_t *node = get_co_mst_irn(env, c->n[idx]);
1298 /* try next color when failed */
1299 if (n_succeeded == 0)
1302 /* fragment the chunk according to the coloring */
1303 local_best = fragment_chunk(env, col, c, tmp_chunks);
1305 /* search the best of the good list
1306 and make it the new best if it is better than the current */
1308 aff_chunk_assure_weight(env, local_best);
1310 DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %u) for color %d: ", local_best->id, col));
1311 DBG_AFF_CHUNK(env, LEVEL_3, local_best);
1313 if (! best_chunk || best_chunk->weight < local_best->weight) {
1314 best_chunk = local_best;
1317 del_waitq(best_starts);
1318 best_starts = good_starts;
1319 DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %u), color %d\n", best_chunk->id, best_color));
1321 DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n"));
1322 del_waitq(good_starts);
1326 del_waitq(good_starts);
1329 /* if all nodes were recolored, bail out */
1330 if (n_succeeded == n_nodes)
1334 stat_ev_int("heur4_colors_tried", i);
1336 /* free all intermediate created chunks except best one */
1337 while (! waitq_empty(tmp_chunks)) {
1338 aff_chunk_t *tmp = (aff_chunk_t*)waitq_get(tmp_chunks);
1339 if (tmp != best_chunk)
1340 delete_aff_chunk(tmp);
1342 del_waitq(tmp_chunks);
1344 /* return if coloring failed */
1347 del_waitq(best_starts);
1351 DB((dbg, LEVEL_2, "\tbest chunk #%u ", best_chunk->id));
1352 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1353 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1355 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1356 const ir_node *irn = best_chunk->n[idx];
1357 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1360 /* bring the node to the color. */
1361 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id));
1362 INIT_LIST_HEAD(&changed);
1364 res = change_node_color(env, node, best_color, &changed);
1365 stat_ev_tim_pop("heur4_recolor");
1367 materialize_coloring(&changed);
1370 assert(list_empty(&changed));
1373 /* remove the nodes in best chunk from original chunk */
1374 len = ARR_LEN(best_chunk->n);
1375 for (idx = 0; idx < len; ++idx) {
1376 const ir_node *irn = best_chunk->n[idx];
1377 int pos = nodes_bsearch(c->n, irn);
1382 len = ARR_LEN(c->n);
1383 for (idx = nidx = 0; idx < len; ++idx) {
1384 const ir_node *irn = c->n[idx];
1390 ARR_SHRINKLEN(c->n, nidx);
1393 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1394 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1395 const ir_node *n = c->n[idx];
1396 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1400 /* fragment the remaining chunk */
1401 visited = bitset_irg_malloc(env->co->irg);
1402 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx)
1403 bitset_set(visited, get_irn_idx(best_chunk->n[idx]));
1405 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1406 const ir_node *irn = c->n[idx];
1407 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1408 aff_chunk_t *new_chunk = new_aff_chunk(env);
1409 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1411 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1412 aff_chunk_assure_weight(env, new_chunk);
1413 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1417 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1418 const ir_node *n = best_chunk->n[idx];
1419 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1423 /* clear obsolete chunks and free some memory */
1424 delete_aff_chunk(best_chunk);
1425 bitset_free(visited);
1427 del_waitq(best_starts);
1429 stat_ev_ctx_pop("heur4_color_chunk");
1433 * Main driver for mst safe coalescing algorithm.
1435 static int co_solve_heuristic_mst(copy_opt_t *co)
1437 unsigned n_regs = co->cls->n_regs;
1438 bitset_t *allocatable_regs = bitset_alloca(n_regs);
1443 co_mst_env_t mst_env;
1450 ir_nodemap_init(&mst_env.map, co->irg);
1451 obstack_init(&mst_env.obst);
1453 be_put_allocatable_regs(co->cenv->irg, co->cls, allocatable_regs);
1454 k = bitset_popcount(allocatable_regs);
1456 mst_env.n_regs = n_regs;
1458 mst_env.chunks = new_pqueue();
1460 mst_env.allocatable_regs = allocatable_regs;
1461 mst_env.ifg = co->cenv->ifg;
1462 INIT_LIST_HEAD(&mst_env.chunklist);
1463 mst_env.chunk_visited = 0;
1464 mst_env.single_cols = OALLOCN(&mst_env.obst, col_cost_t*, n_regs);
1466 for (i = 0; i < n_regs; ++i) {
1467 col_cost_t *vec = OALLOCN(&mst_env.obst, col_cost_t, n_regs);
1469 mst_env.single_cols[i] = vec;
1470 for (j = 0; j < n_regs; ++j) {
1472 vec[j].cost = REAL(0.0);
1476 vec[0].cost = REAL(1.0);
1479 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1481 /* build affinity chunks */
1483 build_affinity_chunks(&mst_env);
1484 stat_ev_tim_pop("heur4_initial_chunk");
1486 /* color chunks as long as there are some */
1487 while (! pqueue_empty(mst_env.chunks)) {
1488 aff_chunk_t *chunk = (aff_chunk_t*)pqueue_pop_front(mst_env.chunks);
1490 color_aff_chunk(&mst_env, chunk);
1491 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id));
1492 delete_aff_chunk(chunk);
1495 /* apply coloring */
1496 for (pn = 0; pn < ARR_LEN(mst_env.map.data); ++pn) {
1497 co_mst_irn_t *mirn = mst_env.map.data[pn];
1498 const arch_register_t *reg;
1501 irn = get_idx_irn(co->irg, pn);
1502 if (arch_irn_is_ignore(irn))
1505 /* skip nodes where color hasn't changed */
1506 if (mirn->init_col == mirn->col)
1509 reg = arch_register_for_index(co->cls, mirn->col);
1510 arch_set_irn_register(irn, reg);
1511 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1514 /* free allocated memory */
1515 del_pqueue(mst_env.chunks);
1516 obstack_free(&mst_env.obst, NULL);
1517 ir_nodemap_destroy(&mst_env.map);
1519 stat_ev_tim_pop("heur4_total");
1524 static const lc_opt_table_entry_t options[] = {
1525 LC_OPT_ENT_INT ("limit", "limit recoloring", &recolor_limit),
1526 LC_OPT_ENT_DBL ("di", "dislike influence", &dislike_influence),
1530 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4)
1531 void be_init_copyheur4(void)
1533 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1534 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
1535 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
1536 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
1537 lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4");
1539 static co_algo_info copyheur = {
1540 co_solve_heuristic_mst, 0
1543 lc_opt_add_table(heur4_grp, options);
1544 be_register_copyopt("heur4", ©heur);
1546 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");