2 * Copyright (C) 1995-2007 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 #endif /* HAVE_CONFIG_H */
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)
80 static int last_chunk_id = 0;
81 static int recolor_limit = 4;
83 typedef struct _col_cost_t {
91 typedef struct _aff_chunk_t {
92 ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
93 bitset_t *nodes; /**< A bitset containing all nodes inside this chunk. */
94 bitset_t *interfere; /**< A bitset containing all interfering neighbours of the nodes in this chunk. */
95 int weight; /**< Weight of this chunk */
96 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
97 unsigned deleted : 1; /**< Set if the was deleted. */
98 int id; /**< For debugging: An id of this chunk. */
104 typedef struct _aff_edge_t {
105 ir_node *src; /**< Source node. */
106 ir_node *tgt; /**< Target node. */
107 double weight; /**< The weight of this edge. */
110 /* main coalescing environment */
111 typedef struct _co_mst_env_t {
112 int n_regs; /**< number of regs in class */
113 int k; /**< number of non-ignore registers in class */
114 bitset_t *ignore_regs; /**< set containing all global ignore registers */
115 ir_phase ph; /**< phase object holding data for nodes */
116 pqueue *chunks; /**< priority queue for chunks */
117 pset *chunkset; /**< set holding all chunks */
118 be_ifg_t *ifg; /**< the interference graph */
119 const arch_env_t *aenv; /**< the arch environment */
120 copy_opt_t *co; /**< the copy opt object */
123 /* stores coalescing related information for a node */
124 typedef struct _co_mst_irn_t {
125 ir_node *irn; /**< the irn this information belongs to */
126 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
127 bitset_t *adm_colors; /**< set of admissible colors for this irn */
128 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
129 int n_neighs; /**< length of the interfering neighbours array. */
130 int int_aff_neigh; /**< number of interfering affinity neighbours */
131 int col; /**< color currently assigned */
132 int init_col; /**< the initial color */
133 int tmp_col; /**< a temporary assigned color */
134 unsigned fixed : 1; /**< the color is fixed */
135 struct list_head list; /**< Queue for coloring undo. */
138 #define get_co_mst_irn(mst_env, irn) (phase_get_or_set_irn_data(&(mst_env)->ph, (irn)))
140 typedef int decide_func_t(const co_mst_irn_t *node, int col);
145 * Write a chunk to stderr for debugging.
147 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) {
149 if (c->weight_consistent)
150 ir_fprintf(stderr, " $%d ", c->weight);
151 ir_fprintf(stderr, "{");
152 bitset_foreach(c->nodes, idx) {
153 ir_node *n = get_idx_irn(env->co->irg, idx);
154 ir_fprintf(stderr, " %+F,", n);
156 ir_fprintf(stderr, "}");
160 * Dump all admissible colors to stderr.
162 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) {
166 if (bitset_popcnt(node->adm_colors) < 1)
167 fprintf(stderr, "no admissible colors?!?");
169 bitset_foreach(node->adm_colors, idx)
170 fprintf(stderr, " %d", idx);
175 * Dump color-cost pairs to stderr.
177 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) {
179 for (i = 0; i < env->n_regs; ++i) {
180 if (cost[i].cost == COL_COST_INFEASIBLE)
181 fprintf(stderr, " (%d, INF)", cost[i].col);
183 fprintf(stderr, " (%d, %.1f)", cost[i].col, cost[i].cost);
187 #endif /* DEBUG_libfirm */
189 static INLINE int get_mst_irn_col(const co_mst_irn_t *node) {
190 return node->tmp_col >= 0 ? node->tmp_col : node->col;
194 * @return 1 if node @p node has color @p col, 0 otherwise.
196 static int decider_has_color(const co_mst_irn_t *node, int col) {
197 return get_mst_irn_col(node) == col;
201 * @return 1 if node @p node has not color @p col, 0 otherwise.
203 static int decider_hasnot_color(const co_mst_irn_t *node, int col) {
204 return get_mst_irn_col(node) != col;
208 * Always returns true.
210 static int decider_always_yes(const co_mst_irn_t *node, int col) {
216 /** compares two affinity edges by its weight */
217 static int cmp_aff_edge(const void *a, const void *b) {
218 const aff_edge_t *e1 = a;
219 const aff_edge_t *e2 = b;
221 if (e2->weight == e1->weight) {
222 if (e2->src->node_idx == e1->src->node_idx)
223 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
225 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
227 /* sort in descending order */
228 return QSORT_CMP(e2->weight, e1->weight);
231 /** compares to color-cost pairs */
232 static int cmp_col_cost(const void *a, const void *b) {
233 const col_cost_t *c1 = a;
234 const col_cost_t *c2 = b;
235 double diff = c1->cost - c2->cost;
236 return (diff > 0) - (diff < 0);
240 * Creates a new affinity chunk
242 static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) {
243 aff_chunk_t *c = xmalloc(sizeof(*c));
245 c->weight_consistent = 0;
246 c->n = NEW_ARR_F(ir_node *, 0);
247 c->nodes = bitset_irg_malloc(env->co->irg);
248 c->interfere = bitset_irg_malloc(env->co->irg);
249 c->id = last_chunk_id++;
250 pset_insert(env->chunkset, c, c->id);
255 * Frees all memory allocated by an affinity chunk.
257 static INLINE void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
258 pset_remove(env->chunkset, c, c->id);
259 bitset_free(c->nodes);
260 bitset_free(c->interfere);
267 * Adds a node to an affinity chunk
269 static INLINE void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) {
272 if (bitset_is_set(c->nodes, get_irn_idx(node->irn)))
275 c->weight_consistent = 0;
277 bitset_set(c->nodes, get_irn_idx(node->irn));
279 ARR_APP1(ir_node *, c->n, node->irn);
281 for (i = node->n_neighs - 1; i >= 0; --i) {
282 ir_node *neigh = node->int_neighs[i];
283 bitset_set(c->interfere, get_irn_idx(neigh));
288 * In case there is no phase information for irn, initialize it.
290 static void *co_mst_irn_init(ir_phase *ph, ir_node *irn, void *old) {
291 co_mst_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
292 co_mst_env_t *env = ph->priv;
295 const arch_register_req_t *req;
296 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
304 res->int_neighs = NULL;
305 res->int_aff_neigh = 0;
306 res->col = arch_register_get_index(arch_get_irn_register(env->aenv, irn));
307 res->init_col = res->col;
308 INIT_LIST_HEAD(&res->list);
310 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
312 /* set admissible registers */
313 res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs);
315 /* Exclude colors not assignable to the irn */
316 req = arch_get_register_req(env->aenv, irn, -1);
317 if (arch_register_req_is(req, limited))
318 rbitset_copy_to_bitset(req->limited, res->adm_colors);
320 bitset_set_all(res->adm_colors);
322 /* exclude global ignore registers as well */
323 bitset_andnot(res->adm_colors, env->ignore_regs);
325 /* set the number of interfering affinity neighbours to -1, they are calculated later */
326 res->int_aff_neigh = -1;
328 /* build list of interfering neighbours */
330 be_ifg_foreach_neighbour(env->ifg, nodes_it, irn, neigh) {
331 if (! arch_irn_is(env->aenv, neigh, ignore)) {
332 obstack_ptr_grow(phase_obst(ph), neigh);
336 res->int_neighs = obstack_finish(phase_obst(ph));
343 * Check if affinity chunk @p chunk interferes with node @p irn.
345 static INLINE int aff_chunk_interferes(co_mst_env_t *env, const aff_chunk_t *chunk, ir_node *irn) {
347 return bitset_is_set(chunk->interfere, get_irn_idx(irn));
351 * Check if there are interference edges from c1 to c2.
352 * @param env The global co_mst environment
354 * @param c2 Another chunk
355 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
357 static INLINE int aff_chunks_interfere(co_mst_env_t *env, const aff_chunk_t *c1, const aff_chunk_t *c2) {
362 /* check if there is a node in c2 having an interfering neighbor in c1 */
363 return bitset_intersect(c1->interfere, c2->nodes);
367 * Returns the affinity chunk of @p irn or creates a new
368 * one with @p irn as element if there is none assigned.
370 static INLINE aff_chunk_t *get_aff_chunk(co_mst_env_t *env, ir_node *irn) {
371 co_mst_irn_t *node = get_co_mst_irn(env, irn);
376 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
377 * are no interference edges from chunk(src) to chunk(tgt)).
378 * @return 1 if successful, 0 if not possible
380 static int aff_chunk_absorb(co_mst_env_t *env, ir_node *src, ir_node *tgt) {
381 aff_chunk_t *c1 = get_aff_chunk(env, src);
382 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
385 DB((dbg, LEVEL_4, "Attempt to let c1 (id %d): ", c1 ? c1->id : -1));
387 DBG_AFF_CHUNK(env, LEVEL_4, c1);
389 DB((dbg, LEVEL_4, "{%+F}", src));
391 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %d): ", c2 ? c2->id : -1));
393 DBG_AFF_CHUNK(env, LEVEL_4, c2);
395 DB((dbg, LEVEL_4, "{%+F}", tgt));
397 DB((dbg, LEVEL_4, "\n"));
402 /* no chunk exists */
403 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
406 for (i = mirn->n_neighs - 1; i >= 0; --i) {
407 if (mirn->int_neighs[i] == tgt)
411 /* create one containing both nodes */
412 c1 = new_aff_chunk(env);
413 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
414 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
418 /* c2 already exists */
419 if (! aff_chunk_interferes(env, c2, src)) {
420 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
424 } else if (c2 == NULL) {
425 /* c1 already exists */
426 if (! aff_chunk_interferes(env, c1, tgt)) {
427 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
430 } else if (c1 != c2 && ! aff_chunks_interfere(env, c1, c2)) {
433 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx) {
434 ir_node *n = c2->n[idx];
435 co_mst_irn_t *mn = get_co_mst_irn(env, n);
439 if (! bitset_is_set(c1->nodes, get_irn_idx(n)))
440 ARR_APP1(ir_node *, c1->n, n);
443 bitset_or(c1->nodes, c2->nodes);
444 bitset_or(c1->interfere, c2->interfere);
445 c1->weight_consistent = 0;
447 delete_aff_chunk(env, c2);
450 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
454 DB((dbg, LEVEL_4, " ... absorbed\n"));
459 * Assures that the weight of the given chunk is consistent.
461 static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) {
462 if (! c->weight_consistent) {
468 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
469 ir_node *n = c->n[idx];
470 const affinity_node_t *an = get_affinity_info(env->co, n);
471 co_mst_irn_t *node = get_co_mst_irn(env, n);
472 int col_cnt = bitset_popcnt(node->adm_colors);
474 if (col_cnt < env->k) {
475 /* calculate costs for constrained interfering neighbors */
476 ratio += 1.0 - (double) col_cnt / env->k;
482 co_gs_foreach_neighb(an, neigh) {
483 const ir_node *m = neigh->irn;
484 const int m_idx = get_irn_idx(m);
486 /* skip ignore nodes */
487 if (arch_irn_is(env->aenv, m, ignore))
490 w += bitset_is_set(c->nodes, m_idx) ? neigh->costs : 0;
495 w *= 1.0 + (double) n_constr / ARR_LEN(c->n) * ratio;
498 c->weight_consistent = 1;
503 * Count the number of interfering affinity neighbours
505 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an) {
506 const neighb_t *neigh;
507 ir_node *irn = an->irn;
508 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
511 co_gs_foreach_neighb(an, neigh) {
512 const ir_node *n = neigh->irn;
515 /* skip ignore nodes */
516 if (arch_irn_is(env->aenv, n, ignore))
519 /* check if the affinity neighbour interfere */
520 for (i = 0; i < node->n_neighs; ++i) {
521 if (node->int_neighs[i] == n) {
532 * Build chunks of nodes connected by affinity edges.
533 * We start at the heaviest affinity edge.
534 * The chunks of the two edge-defining nodes will be
535 * merged if there are no interference edges from one
536 * chunk to the other.
538 static void build_affinity_chunks(co_mst_env_t *env) {
539 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
540 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
543 aff_chunk_t *curr_chunk;
545 /* at first we create the affinity edge objects */
546 be_ifg_foreach_node(env->ifg, nodes_it, n) {
547 int n_idx = get_irn_idx(n);
551 /* skip ignore nodes */
552 if (arch_irn_is(env->aenv, n, ignore))
555 n1 = get_co_mst_irn(env, n);
556 an = get_affinity_info(env->co, n);
561 if (n1->int_aff_neigh < 0)
562 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
564 /* build the affinity edges */
565 co_gs_foreach_neighb(an, neigh) {
566 ir_node *m = neigh->irn;
567 int m_idx = get_irn_idx(m);
569 /* record the edge in only one direction */
574 /* skip ignore nodes */
575 if (arch_irn_is(env->aenv, m, ignore))
581 n2 = get_co_mst_irn(env, m);
582 if (n2->int_aff_neigh < 0) {
583 affinity_node_t *am = get_affinity_info(env->co, m);
584 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
587 * these weights are pure hackery ;-).
588 * It's not chriswue's fault but mine.
590 edge.weight = (double)neigh->costs / (double)(1 + n1->int_aff_neigh + n2->int_aff_neigh);
591 ARR_APP1(aff_edge_t, edges, edge);
597 /* now: sort edges and build the affinity chunks */
598 len = ARR_LEN(edges);
599 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
600 for (i = 0; i < len; ++i) {
601 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
603 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
606 /* now insert all chunks into a priority queue */
607 foreach_pset(env->chunkset, curr_chunk) {
608 aff_chunk_assure_weight(env, curr_chunk);
610 DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id));
611 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
612 DBG((dbg, LEVEL_1, "\n"));
614 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
616 foreach_phase_irn(&env->ph, n) {
617 co_mst_irn_t *mirn = get_co_mst_irn(env, n);
619 if (mirn->chunk == NULL) {
620 /* no chunk is allocated so far, do it now */
621 aff_chunk_t *curr_chunk = new_aff_chunk(env);
622 aff_chunk_add_node(curr_chunk, mirn);
624 aff_chunk_assure_weight(env, curr_chunk);
626 DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id));
627 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
628 DBG((dbg, LEVEL_1, "\n"));
630 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
637 static void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk)
639 pqueue *grow = new_pqueue();
642 ir_node *max_node = NULL;
644 for (i = ARR_LEN(chunk->n) - 1; i >= 0; i--) {
645 ir_node *irn = chunk->n[i];
646 affinity_node_t *an = get_affinity_info(env->co, irn);
650 if (arch_irn_is(env->aenv, irn, ignore))
654 co_gs_foreach_neighb(an, neigh)
657 if (w > max_weight) {
665 bitset_t *visited = bitset_irg_malloc(env->co->irg);
667 for (i = ARR_LEN(chunk->n) - 1; i >= 0; --i)
668 bitset_add_irn(visited, chunk->n[i]);
670 pqueue_put(grow, max_node, max_weight);
671 bitset_remv_irn(visited, max_node);
673 while (!pqueue_empty(grow)) {
674 ir_node *irn = pqueue_get(grow);
675 affinity_node_t *an = get_affinity_info(env->co, irn);
678 if (arch_irn_is(env->aenv, irn, ignore))
681 assert(i <= ARR_LEN(chunk->n));
686 /* build the affinity edges */
687 co_gs_foreach_neighb(an, neigh) {
688 co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn);
690 if (bitset_contains_irn(visited, node->irn)) {
691 pqueue_put(grow, neigh->irn, neigh->costs);
692 bitset_remv_irn(visited, node->irn);
698 bitset_free(visited);
703 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
705 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
706 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
708 waitq *nodes = new_waitq();
710 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%d) from %+F, color %d:", chunk->id, node->irn, col));
712 /* init queue and chunk */
713 waitq_put(nodes, node);
714 bitset_set(visited, get_irn_idx(node->irn));
715 aff_chunk_add_node(chunk, node);
716 DB((dbg, LEVEL_1, " %+F", node->irn));
718 /* as long as there are nodes in the queue */
719 while (! waitq_empty(nodes)) {
720 co_mst_irn_t *n = waitq_get(nodes);
721 affinity_node_t *an = get_affinity_info(env->co, n->irn);
723 /* check all affinity neighbors */
726 co_gs_foreach_neighb(an, neigh) {
727 ir_node *m = neigh->irn;
728 int m_idx = get_irn_idx(m);
731 /* skip ignore nodes */
732 if (arch_irn_is(env->aenv, m, ignore))
735 n2 = get_co_mst_irn(env, m);
737 if (! bitset_is_set(visited, m_idx) &&
740 ! aff_chunk_interferes(env, chunk, m) &&
741 bitset_is_set(orig_chunk->nodes, m_idx))
744 following conditions are met:
745 - neighbour is not visited
746 - neighbour likes the color
747 - neighbour has not yet a fixed color
748 - the new chunk doesn't interfere with the neighbour
749 - neighbour belongs or belonged once to the original chunk
751 bitset_set(visited, m_idx);
752 aff_chunk_add_node(chunk, n2);
753 DB((dbg, LEVEL_1, " %+F", n2->irn));
754 /* enqueue for further search */
755 waitq_put(nodes, n2);
761 DB((dbg, LEVEL_1, "\n"));
767 * Fragment the given chunk into chunks having given color and not having given color.
769 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp) {
770 bitset_t *visited = bitset_irg_malloc(env->co->irg);
772 aff_chunk_t *best = NULL;
774 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
777 aff_chunk_t *tmp_chunk;
778 decide_func_t *decider;
782 if (bitset_is_set(visited, get_irn_idx(irn)))
785 node = get_co_mst_irn(env, irn);
787 if (get_mst_irn_col(node) == col) {
788 decider = decider_has_color;
790 DBG((dbg, LEVEL_4, "\tcolor %d wanted", col));
793 decider = decider_hasnot_color;
795 DBG((dbg, LEVEL_4, "\tcolor %d forbidden", col));
798 /* create a new chunk starting at current node */
799 tmp_chunk = new_aff_chunk(env);
800 waitq_put(tmp, tmp_chunk);
801 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
802 assert(bitset_popcnt(tmp_chunk->nodes) > 0 && "No nodes added to chunk");
804 /* remember the local best */
805 aff_chunk_assure_weight(env, tmp_chunk);
806 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
810 assert(best && "No chunk found?");
811 bitset_free(visited);
816 * Initializes an array of color-cost pairs.
817 * Sets forbidden colors to costs COL_COST_INFEASIBLE and all others to @p c.
819 static INLINE void col_cost_init(co_mst_env_t *env, col_cost_t *cost, double c) {
822 for (i = 0; i < env->n_regs; ++i) {
824 if (bitset_is_set(env->ignore_regs, i))
825 cost[i].cost = COL_COST_INFEASIBLE;
832 * Initializes an array of color-cost pairs.
833 * Sets all colors except color @p col to COL_COST_INFEASIBLE and @p col to 0.0
835 static INLINE void col_cost_init_single(co_mst_env_t *env, col_cost_t *cost, int col) {
836 assert(! bitset_is_set(env->ignore_regs, col) && "Attempt to use forbidden color.");
837 col_cost_init(env, cost, COL_COST_INFEASIBLE);
844 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
845 * ATTENTION: the queue is empty after calling this function!
847 static INLINE void reject_coloring(struct list_head *nodes) {
848 co_mst_irn_t *n, *temp;
849 DB((dbg, LEVEL_4, "\treject coloring for"));
850 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
851 DB((dbg, LEVEL_4, " %+F", n->irn));
852 assert(n->tmp_col >= 0);
854 list_del_init(&n->list);
856 DB((dbg, LEVEL_4, "\n"));
859 static INLINE void materialize_coloring(struct list_head *nodes) {
860 co_mst_irn_t *n, *temp;
861 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
862 assert(n->tmp_col >= 0);
865 list_del_init(&n->list);
869 static INLINE void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
872 assert(!node->fixed);
873 assert(node->tmp_col < 0);
874 assert(node->list.next == &node->list && node->list.prev == &node->list);
876 list_add_tail(&node->list, changed);
880 static INLINE int is_loose(co_mst_irn_t *node)
882 return !node->fixed && node->tmp_col < 0;
886 * Determines the costs for each color if it would be assigned to node @p node.
888 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) {
889 affinity_node_t *an = get_affinity_info(env->co, node->irn);
894 col_cost_init(env, costs, 0.0);
896 /* calculate (negative) costs for affinity neighbours */
898 co_gs_foreach_neighb(an, aff_neigh) {
899 ir_node *m = aff_neigh->irn;
903 /* skip ignore nodes */
904 if (arch_irn_is(env->aenv, m, ignore))
907 neigh = get_co_mst_irn(env, m);
908 c = (double)aff_neigh->costs;
910 /* calculate costs for fixed affinity neighbours */
911 if (!is_loose(neigh)) {
912 int col = get_mst_irn_col(neigh);
913 costs[col].cost -= c * AFF_NEIGHBOUR_FIX_BENEFIT;
918 /* calculate (positive) costs for interfering neighbours */
919 for (i = 0; i < node->n_neighs; ++i) {
924 int_neigh = node->int_neighs[i];
926 assert(!arch_irn_is(env->aenv, int_neigh, ignore));
928 neigh = get_co_mst_irn(env, int_neigh);
929 col = get_mst_irn_col(neigh);
930 col_cnt = bitset_popcnt(neigh->adm_colors);
932 if (!is_loose(neigh)) {
933 /* colors of fixed interfering neighbours are infeasible */
934 costs[col].cost = COL_COST_INFEASIBLE;
936 else if (col_cnt < env->k) {
937 /* calculate costs for constrained interfering neighbours */
938 double ratio = 1.0 - ((double)col_cnt / (double)env->k);
940 bitset_foreach_clear(neigh->adm_colors, idx) {
941 /* check only explicitly forbidden colors (skip global forbidden ones) */
942 if (! bitset_is_set(env->ignore_regs, idx)) {
943 costs[col].cost += ratio * NEIGHBOUR_CONSTR_COSTS;
948 DB((dbg, LEVEL_4, "\tneigh %+F, loose: %d, color: %d\n", int_neigh, is_loose(neigh), col));
951 /* set all not admissible colors to COL_COST_INFEASIBLE */
952 bitset_foreach_clear(node->adm_colors, idx)
953 costs[idx].cost = COL_COST_INFEASIBLE;
956 /* need forward declaration due to recursive call */
957 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);
960 * Tries to change node to a color but @p explude_col.
961 * @return 1 if succeeded, 0 otherwise.
963 static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, struct list_head *changed_ones, int depth) {
964 int col = get_mst_irn_col(node);
967 /* neighbours has already a different color -> good, temporary fix it */
968 if (col != exclude_col) {
970 set_temp_color(node, col, changed_ones);
974 /* The node has the color it should not have _and_ has not been visited yet. */
975 if (is_loose(node)) {
976 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
978 /* Get the costs for giving the node a specific color. */
979 determine_color_costs(env, node, costs);
981 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
982 costs[exclude_col].cost = COL_COST_INFEASIBLE;
984 /* sort the colors according costs, cheapest first. */
985 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost);
987 /* Try recoloring the node using the color list. */
988 res = recolor_nodes(env, node, costs, changed_ones, depth + 1);
995 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
996 * ATTENTION: Expect @p costs already sorted by increasing costs.
997 * @return 1 if coloring could be applied, 0 otherwise.
999 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) {
1001 struct list_head local_changed;
1003 if (depth >= recolor_limit)
1006 DBG((dbg, LEVEL_1, "\tRecoloring %+F with color-costs", node->irn));
1007 DBG_COL_COST(env, LEVEL_1, costs);
1008 DB((dbg, LEVEL_1, "\n"));
1010 for (i = 0; i < env->n_regs; ++i) {
1011 int tgt_col = costs[i].col;
1015 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
1016 if (costs[i].cost == COL_COST_INFEASIBLE) {
1020 /* Set the new color of the node and mark the node as temporarily fixed. */
1021 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
1022 INIT_LIST_HEAD(&local_changed);
1023 set_temp_color(node, tgt_col, &local_changed);
1024 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
1026 /* try to color all interfering neighbours with current color forbidden */
1027 for (j = 0; j < node->n_neighs; ++j) {
1031 neigh = node->int_neighs[j];
1033 /* skip ignore nodes */
1034 if (arch_irn_is(env->aenv, neigh, ignore))
1037 nn = get_co_mst_irn(env, neigh);
1038 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
1039 neigh, j, nn->fixed, nn->tmp_col, nn->col));
1042 Try to change the color of the neighbor and record all nodes which
1043 get changed in the tmp list. Add this list to the "changed" list for
1044 that color. If we did not succeed to change the color of the neighbor,
1045 we bail out and try the next color.
1047 if (get_mst_irn_col(nn) == tgt_col) {
1048 /* try to color neighbour with tgt_col forbidden */
1049 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1);
1057 We managed to assign the target color to all neighbors, so from the perspective
1058 of the current node, every thing was ok and we can return safely.
1061 /* append the local_changed ones to global ones */
1062 list_splice(&local_changed, changed_ones);
1066 /* coloring of neighbours failed, so we try next color */
1067 reject_coloring(&local_changed);
1075 * Tries to bring node @p node and all it's neighbours to color @p tgt_col.
1076 * @return 1 if color @p col could be applied, 0 otherwise
1078 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed_ones, int depth) {
1079 int col = get_mst_irn_col(node);
1081 /* if node already has the target color -> good, temporary fix it */
1082 if (col == tgt_col) {
1083 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1085 set_temp_color(node, tgt_col, changed_ones);
1090 Node has not yet a fixed color and target color is admissible
1091 -> try to recolor node and it's affinity neighbours
1093 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1094 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
1097 col_cost_init_single(env, costs, tgt_col);
1099 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1100 res = recolor_nodes(env, node, costs, changed_ones, depth);
1101 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1106 #ifdef DEBUG_libfirm
1107 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1108 if (!is_loose(node))
1109 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1111 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1112 dbg_admissible_colors(env, node);
1113 DB((dbg, LEVEL_4, ")\n"));
1122 * Tries to color an affinity chunk (or at least a part of it).
1123 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1125 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
1126 aff_chunk_t *best_chunk = NULL;
1127 int best_color = -1;
1129 waitq *tmp_chunks = new_waitq();
1130 waitq *best_starts = NULL;
1133 struct list_head changed_ones;
1135 DB((dbg, LEVEL_2, "fragmentizing chunk #%d", c->id));
1136 DBG_AFF_CHUNK(env, LEVEL_2, c);
1137 DB((dbg, LEVEL_2, "\n"));
1139 chunk_order_nodes(env, c);
1141 /* check which color is the "best" for the given chunk.
1142 * if we found a color which was ok for all nodes, we take it
1143 * and do not look further. (see did_all flag usage below.)
1144 * If we have many colors which fit all nodes it is hard to decide
1145 * which one to take anyway.
1146 * TODO Sebastian: Perhaps we should at all nodes and figure out
1147 * a suitable color using costs as done above (determine_color_costs).
1149 for (col = 0; col < env->n_regs && !did_all; ++col) {
1151 waitq *good_starts = new_waitq();
1152 aff_chunk_t *local_best;
1154 /* skip ignore colors */
1155 if (bitset_is_set(env->ignore_regs, col))
1158 DB((dbg, LEVEL_3, "\ttrying color %d\n", col));
1160 /* suppose we can color all nodes to the same color */
1163 INIT_LIST_HEAD(&changed_ones);
1165 /* try to bring all nodes of given chunk to the current color. */
1166 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1167 ir_node *irn = c->n[idx];
1168 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1171 assert(! node->fixed && "Node must not have a fixed color.");
1172 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1175 The order of the colored nodes is important, so we record the successfully
1176 colored ones in the order they appeared.
1178 good = change_node_color(env, node, col, &changed_ones, 0);
1180 waitq_put(good_starts, node);
1186 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, one_good ? "succeeded" : "failed"));
1189 /* try next color when failed */
1191 reject_coloring(&changed_ones);
1195 /* fragment the chunk according to the coloring */
1196 local_best = fragment_chunk(env, col, c, tmp_chunks);
1198 /* search the best of the good list
1199 and make it the new best if it is better than the current */
1201 aff_chunk_assure_weight(env, local_best);
1203 DB((dbg, LEVEL_4, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, col));
1204 DBG_AFF_CHUNK(env, LEVEL_4, local_best);
1206 if (! best_chunk || best_chunk->weight < local_best->weight) {
1207 best_chunk = local_best;
1210 del_waitq(best_starts);
1211 best_starts = good_starts;
1212 DB((dbg, LEVEL_4, "\n\t\t... setting global best chunk (id %d), color %d\n", best_chunk->id, best_color));
1214 DB((dbg, LEVEL_4, "\n\t\t... omitting, global best is better\n"));
1215 del_waitq(good_starts);
1219 del_waitq(good_starts);
1222 reject_coloring(&changed_ones);
1225 /* free all intermediate created chunks except best one */
1226 while (! waitq_empty(tmp_chunks)) {
1227 aff_chunk_t *tmp = waitq_get(tmp_chunks);
1228 if (tmp != best_chunk)
1229 delete_aff_chunk(env, tmp);
1231 del_waitq(tmp_chunks);
1233 /* return if coloring failed */
1236 del_waitq(best_starts);
1240 DB((dbg, LEVEL_2, "\tbest chunk #%d ", best_chunk->id));
1241 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1242 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1244 INIT_LIST_HEAD(&changed_ones);
1245 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1246 ir_node *irn = best_chunk->n[idx];
1247 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1250 /* bring the node to the color. */
1251 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%d\n", best_color, node->irn, best_chunk->id));
1252 INIT_LIST_HEAD(&changed_ones);
1253 res = change_node_color(env, node, best_color, &changed_ones, 0);
1255 materialize_coloring(&changed_ones);
1260 /* remove the nodes in best chunk from original chunk */
1261 bitset_andnot(c->nodes, best_chunk->nodes);
1262 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1263 ir_node *irn = c->n[idx];
1265 if (bitset_is_set(best_chunk->nodes, get_irn_idx(irn))) {
1266 int last = ARR_LEN(c->n) - 1;
1268 c->n[idx] = c->n[last];
1269 ARR_SHRINKLEN(c->n, last);
1274 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1275 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1276 ir_node *n = c->n[idx];
1277 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1281 /* fragment the remaining chunk */
1282 visited = bitset_irg_malloc(env->co->irg);
1283 bitset_or(visited, best_chunk->nodes);
1284 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1285 ir_node *irn = c->n[idx];
1286 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1287 aff_chunk_t *new_chunk = new_aff_chunk(env);
1288 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1290 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1291 aff_chunk_assure_weight(env, new_chunk);
1292 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1296 /* clear obsolete chunks and free some memory */
1297 delete_aff_chunk(env, best_chunk);
1298 bitset_free(visited);
1300 del_waitq(best_starts);
1304 * Main driver for mst safe coalescing algorithm.
1306 int co_solve_heuristic_mst(copy_opt_t *co) {
1307 unsigned n_regs = co->cls->n_regs;
1308 bitset_t *ignore_regs = bitset_alloca(n_regs);
1311 co_mst_env_t mst_env;
1314 phase_init(&mst_env.ph, "co_mst", co->irg, PHASE_DEFAULT_GROWTH, co_mst_irn_init, &mst_env);
1316 k = be_put_ignore_regs(co->cenv->birg, co->cls, ignore_regs);
1319 mst_env.n_regs = n_regs;
1321 mst_env.chunks = new_pqueue();
1323 mst_env.ignore_regs = ignore_regs;
1324 mst_env.ifg = co->cenv->ifg;
1325 mst_env.aenv = co->aenv;
1326 mst_env.chunkset = pset_new_ptr(512);
1328 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1330 /* build affinity chunks */
1331 build_affinity_chunks(&mst_env);
1333 /* color chunks as long as there are some */
1334 while (! pqueue_empty(mst_env.chunks)) {
1335 aff_chunk_t *chunk = pqueue_get(mst_env.chunks);
1337 color_aff_chunk(&mst_env, chunk);
1338 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%d) done\n", chunk->id));
1339 delete_aff_chunk(&mst_env, chunk);
1342 /* apply coloring */
1343 foreach_phase_irn(&mst_env.ph, irn) {
1344 co_mst_irn_t *mirn = get_co_mst_irn(&mst_env, irn);
1345 const arch_register_t *reg;
1347 if (arch_irn_is(mst_env.aenv, irn, ignore))
1350 // assert(mirn->fixed && "Node should have fixed color");
1352 /* skip nodes where color hasn't changed */
1353 if (mirn->init_col == mirn->col)
1356 reg = arch_register_for_index(co->cls, mirn->col);
1357 arch_set_irn_register(co->aenv, irn, reg);
1358 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1361 /* free allocated memory */
1362 del_pqueue(mst_env.chunks);
1363 phase_free(&mst_env.ph);
1364 del_pset(mst_env.chunkset);
1369 static const lc_opt_table_entry_t options[] = {
1370 LC_OPT_ENT_INT ("limit", "limit recoloring", &recolor_limit),
1375 void be_init_copyheur4(void) {
1376 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1377 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
1378 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
1379 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
1380 lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4");
1382 lc_opt_add_table(heur4_grp, options);
1383 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");
1386 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);