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 "irphase_t.h"
57 #include "becopyopt_t.h"
61 #define COL_COST_INFEASIBLE DBL_MAX
62 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
63 #define NEIGHBOUR_CONSTR_COSTS 64.0
68 #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while (0)
69 #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while (0)
71 static firm_dbg_module_t *dbg = NULL;
75 #define DBG_AFF_CHUNK(env, level, chunk)
76 #define DBG_COL_COST(env, level, cost)
81 #define REAL(C) (C ## f)
83 static unsigned last_chunk_id = 0;
84 static int recolor_limit = 7;
85 static real_t dislike_influence = REAL(0.1);
87 typedef struct col_cost_t {
95 typedef struct aff_chunk_t {
96 const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
97 const ir_node **interfere; /**< An ARR_F containing all inference. */
98 int weight; /**< Weight of this chunk */
99 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
100 unsigned deleted : 1; /**< For debugging: Set if the was deleted. */
101 unsigned id; /**< An id of this chunk. */
104 col_cost_t color_affinity[1];
110 typedef struct aff_edge_t {
111 const ir_node *src; /**< Source node. */
112 const ir_node *tgt; /**< Target node. */
113 int weight; /**< The weight of this edge. */
116 /* main coalescing environment */
117 typedef struct co_mst_env_t {
118 int n_regs; /**< number of regs in class */
119 int k; /**< number of non-ignore registers in class */
120 bitset_t *allocatable_regs; /**< set containing all global ignore registers */
121 ir_phase ph; /**< phase object holding data for nodes */
122 pqueue_t *chunks; /**< priority queue for chunks */
123 list_head chunklist; /**< list holding all chunks */
124 be_ifg_t *ifg; /**< the interference graph */
125 copy_opt_t *co; /**< the copy opt object */
126 unsigned chunk_visited;
127 col_cost_t **single_cols;
130 /* stores coalescing related information for a node */
131 typedef struct co_mst_irn_t {
132 const ir_node *irn; /**< the irn this information belongs to */
133 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
134 bitset_t *adm_colors; /**< set of admissible colors for this irn */
135 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
136 int n_neighs; /**< length of the interfering neighbours array. */
137 int int_aff_neigh; /**< number of interfering affinity neighbours */
138 int col; /**< color currently assigned */
139 int init_col; /**< the initial color */
140 int tmp_col; /**< a temporary assigned color */
141 unsigned fixed : 1; /**< the color is fixed */
142 struct list_head list; /**< Queue for coloring undo. */
143 real_t constr_factor;
146 static co_mst_irn_t *get_co_mst_irn(co_mst_env_t *env, const ir_node *node)
148 return (co_mst_irn_t*)phase_get_or_set_irn_data(&env->ph, node);
151 typedef int decide_func_t(const co_mst_irn_t *node, int col);
156 * Write a chunk to stderr for debugging.
158 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c)
162 if (c->weight_consistent)
163 ir_fprintf(stderr, " $%d ", c->weight);
164 ir_fprintf(stderr, "{");
165 for (i = 0, l = ARR_LEN(c->n); i < l; ++i) {
166 const ir_node *n = c->n[i];
167 ir_fprintf(stderr, " %+F,", n);
169 ir_fprintf(stderr, "}");
173 * Dump all admissible colors to stderr.
175 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node)
180 if (bitset_popcount(node->adm_colors) < 1)
181 fprintf(stderr, "no admissible colors?!?");
183 bitset_foreach(node->adm_colors, idx) {
184 ir_fprintf(stderr, " %zu", idx);
190 * Dump color-cost pairs to stderr.
192 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost)
195 for (i = 0; i < env->n_regs; ++i)
196 fprintf(stderr, " (%d, %.4f)", cost[i].col, cost[i].cost);
199 #endif /* DEBUG_libfirm */
201 static inline int get_mst_irn_col(const co_mst_irn_t *node)
203 return node->tmp_col >= 0 ? node->tmp_col : node->col;
207 * @return 1 if node @p node has color @p col, 0 otherwise.
209 static int decider_has_color(const co_mst_irn_t *node, int col)
211 return get_mst_irn_col(node) == col;
215 * @return 1 if node @p node has not color @p col, 0 otherwise.
217 static int decider_hasnot_color(const co_mst_irn_t *node, int col)
219 return get_mst_irn_col(node) != col;
223 * Always returns true.
225 static int decider_always_yes(const co_mst_irn_t *node, int col)
232 /** compares two affinity edges by its weight */
233 static int cmp_aff_edge(const void *a, const void *b)
235 const aff_edge_t *e1 = (const aff_edge_t*)a;
236 const aff_edge_t *e2 = (const aff_edge_t*)b;
238 if (e2->weight == e1->weight) {
239 if (e2->src->node_idx == e1->src->node_idx)
240 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
242 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
244 /* sort in descending order */
245 return QSORT_CMP(e2->weight, e1->weight);
248 /** compares to color-cost pairs */
249 static __attribute__((unused)) int cmp_col_cost_lt(const void *a, const void *b)
251 const col_cost_t *c1 = (const col_cost_t*)a;
252 const col_cost_t *c2 = (const col_cost_t*)b;
253 real_t diff = c1->cost - c2->cost;
254 return (diff > 0) - (diff < 0);
257 static int cmp_col_cost_gt(const void *a, const void *b)
259 const col_cost_t *c1 = (const col_cost_t*)a;
260 const col_cost_t *c2 = (const col_cost_t*)b;
261 real_t diff = c2->cost - c1->cost;
264 return QSORT_CMP(c1->col, c2->col);
266 return (diff > 0) - (diff < 0);
270 * Creates a new affinity chunk
272 static inline aff_chunk_t *new_aff_chunk(co_mst_env_t *env)
274 aff_chunk_t *c = XMALLOCF(aff_chunk_t, color_affinity, env->n_regs);
275 c->n = NEW_ARR_F(const ir_node *, 0);
276 c->interfere = NEW_ARR_F(const ir_node *, 0);
278 c->weight_consistent = 0;
280 c->id = ++last_chunk_id;
282 list_add(&c->list, &env->chunklist);
287 * Frees all memory allocated by an affinity chunk.
289 static inline void delete_aff_chunk(aff_chunk_t *c)
292 DEL_ARR_F(c->interfere);
299 * binary search of sorted nodes.
301 * @return the position where n is found in the array arr or ~pos
302 * if the nodes is not here.
304 static inline int nodes_bsearch(const ir_node **arr, const ir_node *n)
306 int hi = ARR_LEN(arr);
310 int md = lo + ((hi - lo) >> 1);
323 /** Check if a node n can be found inside arr. */
324 static int node_contains(const ir_node **arr, const ir_node *n)
326 int i = nodes_bsearch(arr, n);
331 * Insert a node into the sorted nodes list.
333 * @return 1 if the node was inserted, 0 else
335 static int nodes_insert(const ir_node ***arr, const ir_node *irn)
337 int idx = nodes_bsearch(*arr, irn);
340 int i, n = ARR_LEN(*arr);
343 ARR_APP1(const ir_node *, *arr, irn);
348 for (i = n - 1; i >= idx; --i)
357 * Adds a node to an affinity chunk
359 static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node)
363 if (! nodes_insert(&c->n, node->irn))
366 c->weight_consistent = 0;
369 for (i = node->n_neighs - 1; i >= 0; --i) {
370 ir_node *neigh = node->int_neighs[i];
371 nodes_insert(&c->interfere, neigh);
376 * In case there is no phase information for irn, initialize it.
378 static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn)
380 co_mst_irn_t *res = (co_mst_irn_t*)phase_alloc(ph, sizeof(res[0]));
381 co_mst_env_t *env = (co_mst_env_t*)ph->priv;
383 const arch_register_req_t *req;
384 neighbours_iter_t nodes_it;
392 res->int_neighs = NULL;
393 res->int_aff_neigh = 0;
394 res->col = arch_register_get_index(arch_get_irn_register(irn));
395 res->init_col = res->col;
396 INIT_LIST_HEAD(&res->list);
398 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
400 /* set admissible registers */
401 res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs);
403 /* Exclude colors not assignable to the irn */
404 req = arch_get_register_req_out(irn);
405 if (arch_register_req_is(req, limited))
406 rbitset_copy_to_bitset(req->limited, res->adm_colors);
408 bitset_set_all(res->adm_colors);
410 /* exclude global ignore registers as well */
411 bitset_and(res->adm_colors, env->allocatable_regs);
413 /* compute the constraint factor */
414 res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcount(res->adm_colors)) / env->n_regs;
416 /* set the number of interfering affinity neighbours to -1, they are calculated later */
417 res->int_aff_neigh = -1;
419 /* build list of interfering neighbours */
421 be_ifg_foreach_neighbour(env->ifg, &nodes_it, irn, neigh) {
422 if (!arch_irn_is_ignore(neigh)) {
423 obstack_ptr_grow(phase_obst(ph), neigh);
427 res->int_neighs = (ir_node**)obstack_finish(phase_obst(ph));
433 * Check if affinity chunk @p chunk interferes with node @p irn.
435 static inline int aff_chunk_interferes(const aff_chunk_t *chunk, const ir_node *irn)
437 return node_contains(chunk->interfere, irn);
441 * Check if there are interference edges from c1 to c2.
443 * @param c2 Another chunk
444 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
446 static inline int aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2)
453 /* check if there is a node in c2 having an interfering neighbor in c1 */
454 for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) {
455 const ir_node *irn = c2->n[i];
457 if (node_contains(c1->interfere, irn))
464 * Returns the affinity chunk of @p irn or creates a new
465 * one with @p irn as element if there is none assigned.
467 static inline aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn)
469 co_mst_irn_t *node = get_co_mst_irn(env, irn);
474 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
475 * are no interference edges from chunk(src) to chunk(tgt)).
476 * @return 1 if successful, 0 if not possible
478 static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node *tgt)
480 aff_chunk_t *c1 = get_aff_chunk(env, src);
481 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
484 DB((dbg, LEVEL_4, "Attempt to let c1 (id %u): ", c1 ? c1->id : 0));
486 DBG_AFF_CHUNK(env, LEVEL_4, c1);
488 DB((dbg, LEVEL_4, "{%+F}", src));
490 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %u): ", c2 ? c2->id : 0));
492 DBG_AFF_CHUNK(env, LEVEL_4, c2);
494 DB((dbg, LEVEL_4, "{%+F}", tgt));
496 DB((dbg, LEVEL_4, "\n"));
501 /* no chunk exists */
502 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
505 for (i = mirn->n_neighs - 1; i >= 0; --i) {
506 if (mirn->int_neighs[i] == tgt)
510 /* create one containing both nodes */
511 c1 = new_aff_chunk(env);
512 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
513 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
517 /* c2 already exists */
518 if (! aff_chunk_interferes(c2, src)) {
519 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
523 } else if (c2 == NULL) {
524 /* c1 already exists */
525 if (! aff_chunk_interferes(c1, tgt)) {
526 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
529 } else if (c1 != c2 && ! aff_chunks_interfere(c1, c2)) {
532 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx)
533 aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx]));
535 for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) {
536 const ir_node *irn = c2->interfere[idx];
537 nodes_insert(&c1->interfere, irn);
540 c1->weight_consistent = 0;
542 delete_aff_chunk(c2);
545 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
549 DB((dbg, LEVEL_4, " ... absorbed\n"));
554 * Assures that the weight of the given chunk is consistent.
556 static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c)
558 if (! c->weight_consistent) {
562 for (i = 0; i < env->n_regs; ++i) {
563 c->color_affinity[i].col = i;
564 c->color_affinity[i].cost = REAL(0.0);
567 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
568 const ir_node *n = c->n[idx];
569 const affinity_node_t *an = get_affinity_info(env->co, n);
570 co_mst_irn_t *node = get_co_mst_irn(env, n);
573 if (node->constr_factor > REAL(0.0)) {
575 bitset_foreach (node->adm_colors, col)
576 c->color_affinity[col].cost += node->constr_factor;
581 co_gs_foreach_neighb(an, neigh) {
582 const ir_node *m = neigh->irn;
584 if (arch_irn_is_ignore(m))
587 w += node_contains(c->n, m) ? neigh->costs : 0;
592 for (i = 0; i < env->n_regs; ++i)
593 c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n));
596 // c->weight = bitset_popcount(c->nodes);
597 c->weight_consistent = 1;
602 * Count the number of interfering affinity neighbours
604 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an)
606 const neighb_t *neigh;
607 const ir_node *irn = an->irn;
608 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
611 co_gs_foreach_neighb(an, neigh) {
612 const ir_node *n = neigh->irn;
615 if (arch_irn_is_ignore(n))
618 /* check if the affinity neighbour interfere */
619 for (i = 0; i < node->n_neighs; ++i) {
620 if (node->int_neighs[i] == n) {
631 * Build chunks of nodes connected by affinity edges.
632 * We start at the heaviest affinity edge.
633 * The chunks of the two edge-defining nodes will be
634 * merged if there are no interference edges from one
635 * chunk to the other.
637 static void build_affinity_chunks(co_mst_env_t *env)
639 nodes_iter_t nodes_it;
640 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
643 aff_chunk_t *curr_chunk;
645 /* at first we create the affinity edge objects */
646 be_ifg_foreach_node(env->ifg, &nodes_it, n) {
647 int n_idx = get_irn_idx(n);
651 if (arch_irn_is_ignore(n))
654 n1 = get_co_mst_irn(env, n);
655 an = get_affinity_info(env->co, n);
660 if (n1->int_aff_neigh < 0)
661 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
663 /* build the affinity edges */
664 co_gs_foreach_neighb(an, neigh) {
665 const ir_node *m = neigh->irn;
666 int m_idx = get_irn_idx(m);
668 /* record the edge in only one direction */
673 /* skip ignore nodes */
674 if (arch_irn_is_ignore(m))
680 n2 = get_co_mst_irn(env, m);
681 if (n2->int_aff_neigh < 0) {
682 affinity_node_t *am = get_affinity_info(env->co, m);
683 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
686 * these weights are pure hackery ;-).
687 * It's not chriswue's fault but mine.
689 edge.weight = neigh->costs;
690 ARR_APP1(aff_edge_t, edges, edge);
696 /* now: sort edges and build the affinity chunks */
697 len = ARR_LEN(edges);
698 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
699 for (i = 0; i < len; ++i) {
700 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
702 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
705 /* now insert all chunks into a priority queue */
706 list_for_each_entry(aff_chunk_t, curr_chunk, &env->chunklist, list) {
707 aff_chunk_assure_weight(env, curr_chunk);
709 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
710 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
711 DBG((dbg, LEVEL_1, "\n"));
713 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
716 foreach_phase_irn(&env->ph, n) {
717 co_mst_irn_t *mirn = get_co_mst_irn(env, n);
719 if (mirn->chunk == NULL) {
720 /* no chunk is allocated so far, do it now */
721 aff_chunk_t *curr_chunk = new_aff_chunk(env);
722 aff_chunk_add_node(curr_chunk, mirn);
724 aff_chunk_assure_weight(env, curr_chunk);
726 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
727 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
728 DBG((dbg, LEVEL_1, "\n"));
730 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
737 static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk)
739 pqueue_t *grow = new_pqueue();
740 const ir_node *max_node = NULL;
744 for (i = ARR_LEN(chunk->n) - 1; i >= 0; i--) {
745 const ir_node *irn = chunk->n[i];
746 affinity_node_t *an = get_affinity_info(env->co, irn);
750 if (arch_irn_is_ignore(irn))
754 co_gs_foreach_neighb(an, neigh)
757 if (w > max_weight) {
765 bitset_t *visited = bitset_irg_malloc(env->co->irg);
767 for (i = ARR_LEN(chunk->n) - 1; i >= 0; --i)
768 bitset_add_irn(visited, chunk->n[i]);
770 pqueue_put(grow, (void *) max_node, max_weight);
771 bitset_remv_irn(visited, max_node);
773 while (!pqueue_empty(grow)) {
774 ir_node *irn = (ir_node*)pqueue_pop_front(grow);
775 affinity_node_t *an = get_affinity_info(env->co, irn);
778 if (arch_irn_is_ignore(irn))
781 assert(i <= ARR_LEN(chunk->n));
786 /* build the affinity edges */
787 co_gs_foreach_neighb(an, neigh) {
788 co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn);
790 if (bitset_contains_irn(visited, node->irn)) {
791 pqueue_put(grow, (void *) neigh->irn, neigh->costs);
792 bitset_remv_irn(visited, node->irn);
798 bitset_free(visited);
803 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
805 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
806 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
808 waitq *nodes = new_waitq();
810 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col));
812 /* init queue and chunk */
813 waitq_put(nodes, node);
814 bitset_set(visited, get_irn_idx(node->irn));
815 aff_chunk_add_node(chunk, node);
816 DB((dbg, LEVEL_1, " %+F", node->irn));
818 /* as long as there are nodes in the queue */
819 while (! waitq_empty(nodes)) {
820 co_mst_irn_t *n = (co_mst_irn_t*)waitq_get(nodes);
821 affinity_node_t *an = get_affinity_info(env->co, n->irn);
823 /* check all affinity neighbors */
826 co_gs_foreach_neighb(an, neigh) {
827 const ir_node *m = neigh->irn;
828 int m_idx = get_irn_idx(m);
831 if (arch_irn_is_ignore(m))
834 n2 = get_co_mst_irn(env, m);
836 if (! bitset_is_set(visited, m_idx) &&
839 ! aff_chunk_interferes(chunk, m) &&
840 node_contains(orig_chunk->n, m))
843 following conditions are met:
844 - neighbour is not visited
845 - neighbour likes the color
846 - neighbour has not yet a fixed color
847 - the new chunk doesn't interfere with the neighbour
848 - neighbour belongs or belonged once to the original chunk
850 bitset_set(visited, m_idx);
851 aff_chunk_add_node(chunk, n2);
852 DB((dbg, LEVEL_1, " %+F", n2->irn));
853 /* enqueue for further search */
854 waitq_put(nodes, n2);
860 DB((dbg, LEVEL_1, "\n"));
866 * Fragment the given chunk into chunks having given color and not having given color.
868 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp)
870 bitset_t *visited = bitset_irg_malloc(env->co->irg);
872 aff_chunk_t *best = NULL;
874 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
877 aff_chunk_t *tmp_chunk;
878 decide_func_t *decider;
882 if (bitset_is_set(visited, get_irn_idx(irn)))
885 node = get_co_mst_irn(env, irn);
887 if (get_mst_irn_col(node) == col) {
888 decider = decider_has_color;
890 DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col));
893 decider = decider_hasnot_color;
895 DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col));
898 /* create a new chunk starting at current node */
899 tmp_chunk = new_aff_chunk(env);
900 waitq_put(tmp, tmp_chunk);
901 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
902 assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk");
904 /* remember the local best */
905 aff_chunk_assure_weight(env, tmp_chunk);
906 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
910 assert(best && "No chunk found?");
911 bitset_free(visited);
916 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
917 * ATTENTION: the queue is empty after calling this function!
919 static inline void reject_coloring(struct list_head *nodes)
921 co_mst_irn_t *n, *temp;
922 DB((dbg, LEVEL_4, "\treject coloring for"));
923 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
924 DB((dbg, LEVEL_4, " %+F", n->irn));
925 assert(n->tmp_col >= 0);
927 list_del_init(&n->list);
929 DB((dbg, LEVEL_4, "\n"));
932 static inline void materialize_coloring(struct list_head *nodes)
934 co_mst_irn_t *n, *temp;
935 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
936 assert(n->tmp_col >= 0);
939 list_del_init(&n->list);
943 static inline void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
946 assert(!node->fixed);
947 assert(node->tmp_col < 0);
948 assert(node->list.next == &node->list && node->list.prev == &node->list);
949 assert(bitset_is_set(node->adm_colors, col));
951 list_add_tail(&node->list, changed);
955 static inline int is_loose(co_mst_irn_t *node)
957 return !node->fixed && node->tmp_col < 0;
961 * Determines the costs for each color if it would be assigned to node @p node.
963 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs)
965 int *neigh_cols = ALLOCAN(int, env->n_regs);
970 for (i = 0; i < env->n_regs; ++i) {
973 costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0);
976 for (i = 0; i < node->n_neighs; ++i) {
977 co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]);
978 int col = get_mst_irn_col(n);
983 costs[col].cost = REAL(0.0);
987 coeff = REAL(1.0) / n_loose;
988 for (i = 0; i < env->n_regs; ++i)
989 costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i];
993 /* need forward declaration due to recursive call */
994 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);
997 * Tries to change node to a color but @p explude_col.
998 * @return 1 if succeeded, 0 otherwise.
1000 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)
1002 int col = get_mst_irn_col(node);
1005 /* neighbours has already a different color -> good, temporary fix it */
1006 if (col != exclude_col) {
1008 set_temp_color(node, col, changed);
1012 /* The node has the color it should not have _and_ has not been visited yet. */
1013 if (is_loose(node)) {
1014 col_cost_t *costs = ALLOCAN(col_cost_t, env->n_regs);
1016 /* Get the costs for giving the node a specific color. */
1017 determine_color_costs(env, node, costs);
1019 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
1020 costs[exclude_col].cost = REAL(0.0);
1022 /* sort the colors according costs, cheapest first. */
1023 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt);
1025 /* Try recoloring the node using the color list. */
1026 res = recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip);
1033 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
1034 * ATTENTION: Expect @p costs already sorted by increasing costs.
1035 * @return 1 if coloring could be applied, 0 otherwise.
1037 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)
1040 struct list_head local_changed;
1043 if (depth > *max_depth)
1046 DBG((dbg, LEVEL_4, "\tRecoloring %+F with color-costs", node->irn));
1047 DBG_COL_COST(env, LEVEL_4, costs);
1048 DB((dbg, LEVEL_4, "\n"));
1050 if (depth >= recolor_limit) {
1051 DBG((dbg, LEVEL_4, "\tHit recolor limit\n"));
1055 for (i = 0; i < env->n_regs; ++i) {
1056 int tgt_col = costs[i].col;
1060 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
1061 if (costs[i].cost == REAL(0.0)) {
1062 DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n"));
1066 /* Set the new color of the node and mark the node as temporarily fixed. */
1067 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
1068 INIT_LIST_HEAD(&local_changed);
1069 set_temp_color(node, tgt_col, &local_changed);
1070 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
1072 /* try to color all interfering neighbours with current color forbidden */
1073 for (j = 0; j < node->n_neighs; ++j) {
1077 neigh = node->int_neighs[j];
1079 if (arch_irn_is_ignore(neigh))
1082 nn = get_co_mst_irn(env, neigh);
1083 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
1084 neigh, j, nn->fixed, nn->tmp_col, nn->col));
1087 Try to change the color of the neighbor and record all nodes which
1088 get changed in the tmp list. Add this list to the "changed" list for
1089 that color. If we did not succeed to change the color of the neighbor,
1090 we bail out and try the next color.
1092 if (get_mst_irn_col(nn) == tgt_col) {
1093 /* try to color neighbour with tgt_col forbidden */
1094 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip);
1102 We managed to assign the target color to all neighbors, so from the perspective
1103 of the current node, every thing was ok and we can return safely.
1106 /* append the local_changed ones to global ones */
1107 list_splice(&local_changed, changed);
1111 /* coloring of neighbours failed, so we try next color */
1112 reject_coloring(&local_changed);
1116 DBG((dbg, LEVEL_4, "\tAll colors failed\n"));
1121 * Tries to bring node @p node and all it's neighbours to color @p tgt_col.
1122 * @return 1 if color @p col could be applied, 0 otherwise
1124 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed)
1126 int col = get_mst_irn_col(node);
1128 /* if node already has the target color -> good, temporary fix it */
1129 if (col == tgt_col) {
1130 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1132 set_temp_color(node, tgt_col, changed);
1137 Node has not yet a fixed color and target color is admissible
1138 -> try to recolor node and it's affinity neighbours
1140 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1141 col_cost_t *costs = env->single_cols[tgt_col];
1142 int res, max_depth, trip;
1147 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1148 res = recolor_nodes(env, node, costs, changed, 0, &max_depth, &trip);
1149 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1150 stat_ev_int("heur4_recolor_depth_max", max_depth);
1151 stat_ev_int("heur4_recolor_trip", trip);
1157 #ifdef DEBUG_libfirm
1158 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1159 if (!is_loose(node))
1160 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1162 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1163 dbg_admissible_colors(env, node);
1164 DB((dbg, LEVEL_4, ")\n"));
1173 * Tries to color an affinity chunk (or at least a part of it).
1174 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1176 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c)
1178 aff_chunk_t *best_chunk = NULL;
1179 int n_nodes = ARR_LEN(c->n);
1180 int best_color = -1;
1181 int n_int_chunks = 0;
1182 waitq *tmp_chunks = new_waitq();
1183 waitq *best_starts = NULL;
1184 col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs);
1186 int idx, len, i, nidx, pos;
1187 struct list_head changed;
1189 DB((dbg, LEVEL_2, "fragmentizing chunk #%u", c->id));
1190 DBG_AFF_CHUNK(env, LEVEL_2, c);
1191 DB((dbg, LEVEL_2, "\n"));
1193 stat_ev_ctx_push_fmt("heur4_color_chunk", "%u", c->id);
1195 ++env->chunk_visited;
1197 /* compute color preference */
1198 for (pos = 0, len = ARR_LEN(c->interfere); pos < len; ++pos) {
1199 const ir_node *n = c->interfere[pos];
1200 co_mst_irn_t *node = get_co_mst_irn(env, n);
1201 aff_chunk_t *chunk = node->chunk;
1203 if (is_loose(node) && chunk && chunk->visited < env->chunk_visited) {
1204 assert(!chunk->deleted);
1205 chunk->visited = env->chunk_visited;
1208 aff_chunk_assure_weight(env, chunk);
1209 for (i = 0; i < env->n_regs; ++i)
1210 order[i].cost += chunk->color_affinity[i].cost;
1214 for (i = 0; i < env->n_regs; ++i) {
1215 real_t dislike = n_int_chunks > 0 ? REAL(1.0) - order[i].cost / n_int_chunks : REAL(0.0);
1217 order[i].cost = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike;
1220 qsort(order, env->n_regs, sizeof(order[0]), cmp_col_cost_gt);
1222 DBG_COL_COST(env, LEVEL_2, order);
1223 DB((dbg, LEVEL_2, "\n"));
1225 /* check which color is the "best" for the given chunk.
1226 * if we found a color which was ok for all nodes, we take it
1227 * and do not look further. (see did_all flag usage below.)
1228 * If we have many colors which fit all nodes it is hard to decide
1229 * which one to take anyway.
1230 * TODO Sebastian: Perhaps we should at all nodes and figure out
1231 * a suitable color using costs as done above (determine_color_costs).
1233 for (i = 0; i < env->k; ++i) {
1234 int col = order[i].col;
1235 waitq *good_starts = new_waitq();
1236 aff_chunk_t *local_best;
1239 /* skip ignore colors */
1240 if (!bitset_is_set(env->allocatable_regs, col))
1243 DB((dbg, LEVEL_2, "\ttrying color %d\n", col));
1247 /* try to bring all nodes of given chunk to the current color. */
1248 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1249 const ir_node *irn = c->n[idx];
1250 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1253 assert(! node->fixed && "Node must not have a fixed color.");
1254 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1257 The order of the colored nodes is important, so we record the successfully
1258 colored ones in the order they appeared.
1260 INIT_LIST_HEAD(&changed);
1262 good = change_node_color(env, node, col, &changed);
1263 stat_ev_tim_pop("heur4_recolor");
1265 waitq_put(good_starts, node);
1266 materialize_coloring(&changed);
1271 reject_coloring(&changed);
1273 n_succeeded += good;
1274 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, good ? "succeeded" : "failed"));
1277 /* unfix all nodes */
1278 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1279 co_mst_irn_t *node = get_co_mst_irn(env, c->n[idx]);
1283 /* try next color when failed */
1284 if (n_succeeded == 0)
1287 /* fragment the chunk according to the coloring */
1288 local_best = fragment_chunk(env, col, c, tmp_chunks);
1290 /* search the best of the good list
1291 and make it the new best if it is better than the current */
1293 aff_chunk_assure_weight(env, local_best);
1295 DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %u) for color %d: ", local_best->id, col));
1296 DBG_AFF_CHUNK(env, LEVEL_3, local_best);
1298 if (! best_chunk || best_chunk->weight < local_best->weight) {
1299 best_chunk = local_best;
1302 del_waitq(best_starts);
1303 best_starts = good_starts;
1304 DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %u), color %d\n", best_chunk->id, best_color));
1306 DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n"));
1307 del_waitq(good_starts);
1311 del_waitq(good_starts);
1314 /* if all nodes were recolored, bail out */
1315 if (n_succeeded == n_nodes)
1319 stat_ev_int("heur4_colors_tried", i);
1321 /* free all intermediate created chunks except best one */
1322 while (! waitq_empty(tmp_chunks)) {
1323 aff_chunk_t *tmp = (aff_chunk_t*)waitq_get(tmp_chunks);
1324 if (tmp != best_chunk)
1325 delete_aff_chunk(tmp);
1327 del_waitq(tmp_chunks);
1329 /* return if coloring failed */
1332 del_waitq(best_starts);
1336 DB((dbg, LEVEL_2, "\tbest chunk #%u ", best_chunk->id));
1337 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1338 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1340 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1341 const ir_node *irn = best_chunk->n[idx];
1342 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1345 /* bring the node to the color. */
1346 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id));
1347 INIT_LIST_HEAD(&changed);
1349 res = change_node_color(env, node, best_color, &changed);
1350 stat_ev_tim_pop("heur4_recolor");
1352 materialize_coloring(&changed);
1355 assert(list_empty(&changed));
1358 /* remove the nodes in best chunk from original chunk */
1359 len = ARR_LEN(best_chunk->n);
1360 for (idx = 0; idx < len; ++idx) {
1361 const ir_node *irn = best_chunk->n[idx];
1362 int pos = nodes_bsearch(c->n, irn);
1367 len = ARR_LEN(c->n);
1368 for (idx = nidx = 0; idx < len; ++idx) {
1369 const ir_node *irn = c->n[idx];
1375 ARR_SHRINKLEN(c->n, nidx);
1378 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1379 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1380 const ir_node *n = c->n[idx];
1381 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1385 /* fragment the remaining chunk */
1386 visited = bitset_irg_malloc(env->co->irg);
1387 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx)
1388 bitset_set(visited, get_irn_idx(best_chunk->n[idx]));
1390 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1391 const ir_node *irn = c->n[idx];
1392 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1393 aff_chunk_t *new_chunk = new_aff_chunk(env);
1394 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1396 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1397 aff_chunk_assure_weight(env, new_chunk);
1398 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1402 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1403 const ir_node *n = best_chunk->n[idx];
1404 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1408 /* clear obsolete chunks and free some memory */
1409 delete_aff_chunk(best_chunk);
1410 bitset_free(visited);
1412 del_waitq(best_starts);
1414 stat_ev_ctx_pop("heur4_color_chunk");
1418 * Main driver for mst safe coalescing algorithm.
1420 static int co_solve_heuristic_mst(copy_opt_t *co)
1422 unsigned n_regs = co->cls->n_regs;
1423 bitset_t *allocatable_regs = bitset_alloca(n_regs);
1427 co_mst_env_t mst_env;
1434 phase_init(&mst_env.ph, co->irg, co_mst_irn_init);
1435 phase_set_private(&mst_env.ph, &mst_env);
1437 be_put_allocatable_regs(co->cenv->irg, co->cls, allocatable_regs);
1438 k = bitset_popcount(allocatable_regs);
1440 mst_env.n_regs = n_regs;
1442 mst_env.chunks = new_pqueue();
1444 mst_env.allocatable_regs = allocatable_regs;
1445 mst_env.ifg = co->cenv->ifg;
1446 INIT_LIST_HEAD(&mst_env.chunklist);
1447 mst_env.chunk_visited = 0;
1448 mst_env.single_cols = (col_cost_t**)phase_alloc(&mst_env.ph, sizeof(*mst_env.single_cols) * n_regs);
1450 for (i = 0; i < n_regs; ++i) {
1451 col_cost_t *vec = (col_cost_t*)phase_alloc(&mst_env.ph, sizeof(*vec) * n_regs);
1453 mst_env.single_cols[i] = vec;
1454 for (j = 0; j < n_regs; ++j) {
1456 vec[j].cost = REAL(0.0);
1460 vec[0].cost = REAL(1.0);
1463 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1465 /* build affinity chunks */
1467 build_affinity_chunks(&mst_env);
1468 stat_ev_tim_pop("heur4_initial_chunk");
1470 /* color chunks as long as there are some */
1471 while (! pqueue_empty(mst_env.chunks)) {
1472 aff_chunk_t *chunk = (aff_chunk_t*)pqueue_pop_front(mst_env.chunks);
1474 color_aff_chunk(&mst_env, chunk);
1475 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id));
1476 delete_aff_chunk(chunk);
1479 /* apply coloring */
1480 foreach_phase_irn(&mst_env.ph, irn) {
1482 const arch_register_t *reg;
1484 if (arch_irn_is_ignore(irn))
1487 mirn = get_co_mst_irn(&mst_env, irn);
1488 // assert(mirn->fixed && "Node should have fixed color");
1490 /* skip nodes where color hasn't changed */
1491 if (mirn->init_col == mirn->col)
1494 reg = arch_register_for_index(co->cls, mirn->col);
1495 arch_set_irn_register(irn, reg);
1496 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1499 /* free allocated memory */
1500 del_pqueue(mst_env.chunks);
1501 phase_deinit(&mst_env.ph);
1503 stat_ev_tim_pop("heur4_total");
1508 static const lc_opt_table_entry_t options[] = {
1509 LC_OPT_ENT_INT ("limit", "limit recoloring", &recolor_limit),
1510 LC_OPT_ENT_DBL ("di", "dislike influence", &dislike_influence),
1514 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);
1515 void be_init_copyheur4(void)
1517 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1518 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
1519 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
1520 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
1521 lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4");
1523 static co_algo_info copyheur = {
1524 co_solve_heuristic_mst, 0
1527 lc_opt_add_table(heur4_grp, options);
1528 be_register_copyopt("heur4", ©heur);
1530 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");