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"
55 #include "becopyopt_t.h"
59 #define COL_COST_INFEASIBLE DBL_MAX
60 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
61 #define NEIGHBOUR_CONSTR_COSTS 64.0
65 #define DBG_AFF_CHUNK(env, level, chunk)
66 #define DBG_COL_COST(env, level, cost)
70 static firm_dbg_module_t *dbg = NULL;
71 #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while(0)
72 #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while(0)
76 static int last_chunk_id = 0;
78 typedef struct _col_cost_t {
86 typedef struct _aff_chunk_t {
87 ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
88 bitset_t *nodes; /**< A bitset containing all nodes inside this chunk. */
89 bitset_t *interfere; /**< A bitset containing all interfering neighbours of the nodes in this chunk. */
90 int weight; /**< Weight of this chunk */
91 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
92 unsigned deleted : 1; /**< Set if the was deleted. */
93 int id; /**< For debugging: An id of this chunk. */
99 typedef struct _aff_edge_t {
100 ir_node *src; /**< Source node. */
101 ir_node *tgt; /**< Target node. */
102 double weight; /**< The weight of this edge. */
105 /* main coalescing environment */
106 typedef struct _co_mst_env_t {
107 int n_regs; /**< number of regs in class */
108 int k; /**< number of non-ignore registers in class */
109 bitset_t *ignore_regs; /**< set containing all global ignore registers */
110 ir_phase ph; /**< phase object holding data for nodes */
111 pqueue *chunks; /**< priority queue for chunks */
112 pset *chunkset; /**< set holding all chunks */
113 be_ifg_t *ifg; /**< the interference graph */
114 const arch_env_t *aenv; /**< the arch environment */
115 copy_opt_t *co; /**< the copy opt object */
118 /* stores coalescing related information for a node */
119 typedef struct _co_mst_irn_t {
120 ir_node *irn; /**< the irn this information belongs to */
121 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
122 bitset_t *adm_colors; /**< set of admissible colors for this irn */
123 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
124 int n_neighs; /**< length of the interfering neighbours array. */
125 int int_aff_neigh; /**< number of interfering affinity neighbours */
126 int col; /**< color currently assigned */
127 int init_col; /**< the initial color */
128 int tmp_col; /**< a temporary assigned color */
129 unsigned fixed : 1; /**< the color is fixed */
130 struct list_head list; /**< Queue for coloring undo. */
133 #define get_co_mst_irn(mst_env, irn) (phase_get_or_set_irn_data(&(mst_env)->ph, (irn)))
135 typedef int decide_func_t(const co_mst_irn_t *node, int col);
140 * Write a chunk to stderr for debugging.
142 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) {
144 if (c->weight_consistent)
145 ir_fprintf(stderr, " $%d ", c->weight);
146 ir_fprintf(stderr, "{");
147 bitset_foreach(c->nodes, idx) {
148 ir_node *n = get_idx_irn(env->co->irg, idx);
149 ir_fprintf(stderr, " %+F,", n);
151 ir_fprintf(stderr, "}");
155 * Dump all admissible colors to stderr.
157 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) {
161 if (bitset_popcnt(node->adm_colors) < 1)
162 fprintf(stderr, "no admissible colors?!?");
164 bitset_foreach(node->adm_colors, idx)
165 fprintf(stderr, " %d", idx);
170 * Dump color-cost pairs to stderr.
172 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) {
174 for (i = 0; i < env->n_regs; ++i) {
175 if (cost[i].cost == COL_COST_INFEASIBLE)
176 fprintf(stderr, " (%d, INF)", cost[i].col);
178 fprintf(stderr, " (%d, %.1f)", cost[i].col, cost[i].cost);
182 #endif /* DEBUG_libfirm */
184 static INLINE int get_mst_irn_col(const co_mst_irn_t *node) {
185 return node->tmp_col >= 0 ? node->tmp_col : node->col;
189 * @return 1 if node @p node has color @p col, 0 otherwise.
191 static int decider_has_color(const co_mst_irn_t *node, int col) {
192 return get_mst_irn_col(node) == col;
196 * @return 1 if node @p node has not color @p col, 0 otherwise.
198 static int decider_hasnot_color(const co_mst_irn_t *node, int col) {
199 return get_mst_irn_col(node) != col;
203 * Always returns true.
205 static int decider_always_yes(const co_mst_irn_t *node, int col) {
211 /** compares two affinity edges by its weight */
212 static int cmp_aff_edge(const void *a, const void *b) {
213 const aff_edge_t *e1 = a;
214 const aff_edge_t *e2 = b;
216 if (e2->weight == e1->weight) {
217 if (e2->src->node_idx == e1->src->node_idx)
218 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
220 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
222 /* sort in descending order */
223 return QSORT_CMP(e2->weight, e1->weight);
226 /** compares to color-cost pairs */
227 static int cmp_col_cost(const void *a, const void *b) {
228 const col_cost_t *c1 = a;
229 const col_cost_t *c2 = b;
231 return c1->cost < c2->cost ? -1 : 1;
235 * Creates a new affinity chunk
237 static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) {
238 aff_chunk_t *c = xmalloc(sizeof(*c));
240 c->weight_consistent = 0;
241 c->n = NEW_ARR_F(ir_node *, 0);
242 c->nodes = bitset_irg_malloc(env->co->irg);
243 c->interfere = bitset_irg_malloc(env->co->irg);
244 c->id = last_chunk_id++;
245 pset_insert(env->chunkset, c, c->id);
250 * Frees all memory allocated by an affinity chunk.
252 static INLINE void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
253 pset_remove(env->chunkset, c, c->id);
254 bitset_free(c->nodes);
255 bitset_free(c->interfere);
262 * Adds a node to an affinity chunk
264 static INLINE void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) {
267 if (bitset_is_set(c->nodes, get_irn_idx(node->irn)))
270 c->weight_consistent = 0;
272 bitset_set(c->nodes, get_irn_idx(node->irn));
274 ARR_APP1(ir_node *, c->n, node->irn);
276 for (i = node->n_neighs - 1; i >= 0; --i) {
277 ir_node *neigh = node->int_neighs[i];
278 bitset_set(c->interfere, get_irn_idx(neigh));
283 * In case there is no phase information for irn, initialize it.
285 static void *co_mst_irn_init(ir_phase *ph, ir_node *irn, void *old) {
286 co_mst_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
287 co_mst_env_t *env = ph->priv;
290 const arch_register_req_t *req;
291 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
299 res->int_neighs = NULL;
300 res->int_aff_neigh = 0;
301 res->col = arch_register_get_index(arch_get_irn_register(env->aenv, irn));
302 res->init_col = res->col;
303 INIT_LIST_HEAD(&res->list);
305 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
307 /* set admissible registers */
308 res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs);
310 /* Exclude colors not assignable to the irn */
311 req = arch_get_register_req(env->aenv, irn, -1);
312 if (arch_register_req_is(req, limited))
313 rbitset_copy_to_bitset(req->limited, res->adm_colors);
315 bitset_set_all(res->adm_colors);
317 /* exclude global ignore registers as well */
318 bitset_andnot(res->adm_colors, env->ignore_regs);
320 /* set the number of interfering affinity neighbours to -1, they are calculated later */
321 res->int_aff_neigh = -1;
323 /* build list of interfering neighbours */
325 be_ifg_foreach_neighbour(env->ifg, nodes_it, irn, neigh) {
326 if (! arch_irn_is(env->aenv, neigh, ignore)) {
327 obstack_ptr_grow(phase_obst(ph), neigh);
331 res->int_neighs = obstack_finish(phase_obst(ph));
338 * Check if affinity chunk @p chunk interferes with node @p irn.
340 static INLINE int aff_chunk_interferes(co_mst_env_t *env, const aff_chunk_t *chunk, ir_node *irn) {
342 return bitset_is_set(chunk->interfere, get_irn_idx(irn));
346 * Check if there are interference edges from c1 to c2.
347 * @param env The global co_mst environment
349 * @param c2 Another chunk
350 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
352 static INLINE int aff_chunks_interfere(co_mst_env_t *env, const aff_chunk_t *c1, const aff_chunk_t *c2) {
358 /* check if there is a node in c2 having an interfering neighbor in c1 */
359 tmp = bitset_alloca(get_irg_last_idx(env->co->irg));
360 tmp = bitset_copy(tmp, c1->interfere);
361 tmp = bitset_and(tmp, c2->nodes);
363 return bitset_popcnt(tmp) > 0;
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(const co_mst_env_t *env, aff_chunk_t *c) {
462 if (! c->weight_consistent) {
466 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
467 ir_node *n = c->n[idx];
468 const affinity_node_t *an = get_affinity_info(env->co, n);
472 co_gs_foreach_neighb(an, neigh) {
473 const ir_node *m = neigh->irn;
474 const int m_idx = get_irn_idx(m);
476 /* skip ignore nodes */
477 if (arch_irn_is(env->aenv, m, ignore))
480 w += bitset_is_set(c->nodes, m_idx) ? neigh->costs : 0;
486 c->weight_consistent = 1;
491 * Count the number of interfering affinity neighbours
493 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an) {
494 const neighb_t *neigh;
495 ir_node *irn = an->irn;
496 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
499 co_gs_foreach_neighb(an, neigh) {
500 const ir_node *n = neigh->irn;
503 /* skip ignore nodes */
504 if (arch_irn_is(env->aenv, n, ignore))
507 /* check if the affinity neighbour interfere */
508 for (i = 0; i < node->n_neighs; ++i) {
509 if (node->int_neighs[i] == n) {
520 * Build chunks of nodes connected by affinity edges.
521 * We start at the heaviest affinity edge.
522 * The chunks of the two edge-defining nodes will be
523 * merged if there are no interference edges from one
524 * chunk to the other.
526 static void build_affinity_chunks(co_mst_env_t *env) {
527 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
528 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
531 aff_chunk_t *curr_chunk;
533 /* at first we create the affinity edge objects */
534 be_ifg_foreach_node(env->ifg, nodes_it, n) {
535 int n_idx = get_irn_idx(n);
539 /* skip ignore nodes */
540 if (arch_irn_is(env->aenv, n, ignore))
543 n1 = get_co_mst_irn(env, n);
544 an = get_affinity_info(env->co, n);
549 if (n1->int_aff_neigh < 0)
550 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
552 /* build the affinity edges */
553 co_gs_foreach_neighb(an, neigh) {
554 ir_node *m = neigh->irn;
555 int m_idx = get_irn_idx(m);
557 /* record the edge in only one direction */
562 /* skip ignore nodes */
563 if (arch_irn_is(env->aenv, m, ignore))
569 n2 = get_co_mst_irn(env, m);
570 if (n2->int_aff_neigh < 0) {
571 affinity_node_t *am = get_affinity_info(env->co, m);
572 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
575 * these weights are pure hackery ;-).
576 * It's not chriswue's fault but mine.
578 edge.weight = (double)neigh->costs / (double)(1 + n1->int_aff_neigh + n2->int_aff_neigh);
579 ARR_APP1(aff_edge_t, edges, edge);
585 /* now: sort edges and build the affinity chunks */
586 len = ARR_LEN(edges);
587 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
588 for (i = 0; i < len; ++i) {
589 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
591 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
594 /* now insert all chunks into a priority queue */
595 foreach_pset(env->chunkset, curr_chunk) {
596 aff_chunk_assure_weight(env, curr_chunk);
598 DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id));
599 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
600 DBG((dbg, LEVEL_1, "\n"));
602 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
604 foreach_phase_irn(&env->ph, n) {
605 co_mst_irn_t *mirn = get_co_mst_irn(env, n);
607 if (mirn->chunk == NULL) {
608 /* no chunk is allocated so far, do it now */
609 aff_chunk_t *curr_chunk = new_aff_chunk(env);
610 aff_chunk_add_node(curr_chunk, mirn);
612 aff_chunk_assure_weight(env, curr_chunk);
614 DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id));
615 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
616 DBG((dbg, LEVEL_1, "\n"));
618 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
626 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
628 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
629 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
631 waitq *nodes = new_waitq();
633 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%d) from %+F, color %d:", chunk->id, node->irn, col));
635 /* init queue and chunk */
636 waitq_put(nodes, node);
637 bitset_set(visited, get_irn_idx(node->irn));
638 aff_chunk_add_node(chunk, node);
639 DB((dbg, LEVEL_1, " %+F", node->irn));
641 /* as long as there are nodes in the queue */
642 while (! waitq_empty(nodes)) {
643 co_mst_irn_t *n = waitq_get(nodes);
644 affinity_node_t *an = get_affinity_info(env->co, n->irn);
646 /* check all affinity neighbors */
649 co_gs_foreach_neighb(an, neigh) {
650 ir_node *m = neigh->irn;
651 int m_idx = get_irn_idx(m);
654 /* skip ignore nodes */
655 if (arch_irn_is(env->aenv, m, ignore))
658 n2 = get_co_mst_irn(env, m);
660 if (! bitset_is_set(visited, m_idx) &&
663 ! aff_chunk_interferes(env, chunk, m) &&
664 bitset_is_set(orig_chunk->nodes, m_idx))
667 following conditions are met:
668 - neighbour is not visited
669 - neighbour likes the color
670 - neighbour has not yet a fixed color
671 - the new chunk doesn't interfere with the neighbour
672 - neighbour belongs or belonged once to the original chunk
674 bitset_set(visited, m_idx);
675 aff_chunk_add_node(chunk, n2);
676 DB((dbg, LEVEL_1, " %+F", n2->irn));
677 /* enqueue for further search */
678 waitq_put(nodes, n2);
684 DB((dbg, LEVEL_1, "\n"));
690 * Fragment the given chunk into chunks having given color and not having given color.
692 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp) {
693 bitset_t *visited = bitset_irg_malloc(env->co->irg);
695 aff_chunk_t *best = NULL;
697 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
700 aff_chunk_t *tmp_chunk;
701 decide_func_t *decider;
705 if (bitset_is_set(visited, get_irn_idx(irn)))
708 node = get_co_mst_irn(env, irn);
710 if (get_mst_irn_col(node) == col) {
711 decider = decider_has_color;
713 DBG((dbg, LEVEL_4, "\tcolor %d wanted", col));
716 decider = decider_hasnot_color;
718 DBG((dbg, LEVEL_4, "\tcolor %d forbidden", col));
721 /* create a new chunk starting at current node */
722 tmp_chunk = new_aff_chunk(env);
723 waitq_put(tmp, tmp_chunk);
724 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
725 assert(bitset_popcnt(tmp_chunk->nodes) > 0 && "No nodes added to chunk");
727 /* remember the local best */
728 aff_chunk_assure_weight(env, tmp_chunk);
729 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
733 assert(best && "No chunk found?");
734 bitset_free(visited);
739 * Initializes an array of color-cost pairs.
740 * Sets forbidden colors to costs COL_COST_INFEASIBLE and all others to @p c.
742 static INLINE void col_cost_init(co_mst_env_t *env, col_cost_t *cost, double c) {
745 for (i = 0; i < env->n_regs; ++i) {
747 if (bitset_is_set(env->ignore_regs, i))
748 cost[i].cost = COL_COST_INFEASIBLE;
755 * Initializes an array of color-cost pairs.
756 * Sets all colors except color @p col to COL_COST_INFEASIBLE and @p col to 0.0
758 static INLINE void col_cost_init_single(co_mst_env_t *env, col_cost_t *cost, int col) {
759 assert(! bitset_is_set(env->ignore_regs, col) && "Attempt to use forbidden color.");
760 col_cost_init(env, cost, COL_COST_INFEASIBLE);
767 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
768 * ATTENTION: the queue is empty after calling this function!
770 static INLINE void reject_coloring(struct list_head *nodes) {
771 co_mst_irn_t *n, *temp;
772 DB((dbg, LEVEL_4, "\treject coloring for"));
773 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
774 DB((dbg, LEVEL_4, " %+F", n->irn));
775 assert(n->tmp_col >= 0);
777 list_del_init(&n->list);
779 DB((dbg, LEVEL_4, "\n"));
782 static INLINE void materialize_coloring(struct list_head *nodes) {
783 co_mst_irn_t *n, *temp;
784 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
785 assert(n->tmp_col >= 0);
788 list_del_init(&n->list);
792 static INLINE void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
795 assert(!node->fixed);
796 assert(node->tmp_col < 0);
797 assert(node->list.next == &node->list && node->list.prev == &node->list);
799 list_add_tail(&node->list, changed);
803 static INLINE int is_loose(co_mst_irn_t *node)
805 return !node->fixed && node->tmp_col < 0;
809 * Determines the costs for each color if it would be assigned to node @p node.
811 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) {
812 affinity_node_t *an = get_affinity_info(env->co, node->irn);
817 col_cost_init(env, costs, 0.0);
819 /* calculate (negative) costs for affinity neighbours */
821 co_gs_foreach_neighb(an, aff_neigh) {
822 ir_node *m = aff_neigh->irn;
826 /* skip ignore nodes */
827 if (arch_irn_is(env->aenv, m, ignore))
830 neigh = get_co_mst_irn(env, m);
831 c = (double)aff_neigh->costs;
833 /* calculate costs for fixed affinity neighbours */
834 if (!is_loose(neigh)) {
835 int col = get_mst_irn_col(neigh);
836 costs[col].cost -= c * AFF_NEIGHBOUR_FIX_BENEFIT;
841 /* calculate (positive) costs for interfering neighbours */
842 for (i = 0; i < node->n_neighs; ++i) {
847 int_neigh = node->int_neighs[i];
849 /* skip ignore nodes */
850 if (arch_irn_is(env->aenv, int_neigh, ignore))
853 neigh = get_co_mst_irn(env, int_neigh);
854 col = get_mst_irn_col(neigh);
855 col_cnt = bitset_popcnt(neigh->adm_colors);
857 if (!is_loose(neigh)) {
858 /* colors of fixed interfering neighbours are infeasible */
859 costs[col].cost = COL_COST_INFEASIBLE;
861 else if (col_cnt < env->k) {
862 /* calculate costs for constrained interfering neighbours */
863 double ratio = 1.0 - ((double)col_cnt / (double)env->k);
865 bitset_foreach_clear(neigh->adm_colors, idx) {
866 /* check only explicitly forbidden colors (skip global forbidden ones) */
867 if (! bitset_is_set(env->ignore_regs, idx)) {
868 costs[col].cost += ratio * NEIGHBOUR_CONSTR_COSTS;
873 DB((dbg, LEVEL_4, "\tneigh %+F, loose: %d, color: %d\n", int_neigh, is_loose(neigh), col));
876 /* set all not admissible colors to COL_COST_INFEASIBLE */
877 bitset_foreach_clear(node->adm_colors, idx)
878 costs[idx].cost = COL_COST_INFEASIBLE;
881 /* need forward declaration due to recursive call */
882 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones);
885 * Tries to change node to a color but @p explude_col.
886 * @return 1 if succeeded, 0 otherwise.
888 static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, struct list_head *changed_ones) {
889 int col = get_mst_irn_col(node);
892 /* neighbours has already a different color -> good, temporary fix it */
893 if (col != exclude_col) {
895 set_temp_color(node, col, changed_ones);
899 /* The node has the color it should not have _and_ has not been visited yet. */
900 if (is_loose(node)) {
901 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
903 /* Get the costs for giving the node a specific color. */
904 determine_color_costs(env, node, costs);
906 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
907 costs[exclude_col].cost = COL_COST_INFEASIBLE;
909 /* sort the colors according costs, cheapest first. */
910 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost);
912 /* Try recoloring the node using the color list. */
913 res = recolor_nodes(env, node, costs, changed_ones);
920 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
921 * ATTENTION: Expect @p costs already sorted by increasing costs.
922 * @return 1 if coloring could be applied, 0 otherwise.
924 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones) {
926 struct list_head local_changed;
928 DBG((dbg, LEVEL_1, "\tRecoloring %+F with color-costs", node->irn));
929 DBG_COL_COST(env, LEVEL_1, costs);
930 DB((dbg, LEVEL_1, "\n"));
932 for (i = 0; i < env->n_regs; ++i) {
933 int tgt_col = costs[i].col;
937 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
938 if (costs[i].cost == COL_COST_INFEASIBLE) {
942 /* Set the new color of the node and mark the node as temporarily fixed. */
943 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
944 INIT_LIST_HEAD(&local_changed);
945 set_temp_color(node, tgt_col, &local_changed);
946 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
948 /* try to color all interfering neighbours with current color forbidden */
949 for (j = 0; j < node->n_neighs; ++j) {
953 neigh = node->int_neighs[j];
955 /* skip ignore nodes */
956 if (arch_irn_is(env->aenv, neigh, ignore))
959 nn = get_co_mst_irn(env, neigh);
960 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
961 neigh, j, nn->fixed, nn->tmp_col, nn->col));
964 Try to change the color of the neighbor and record all nodes which
965 get changed in the tmp list. Add this list to the "changed" list for
966 that color. If we did not succeed to change the color of the neighbor,
967 we bail out and try the next color.
969 if (get_mst_irn_col(nn) == tgt_col) {
970 /* try to color neighbour with tgt_col forbidden */
971 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed);
979 We managed to assign the target color to all neighbors, so from the perspective
980 of the current node, every thing was ok and we can return safely.
983 /* append the local_changed ones to global ones */
984 list_splice(&local_changed, changed_ones);
988 /* coloring of neighbours failed, so we try next color */
989 reject_coloring(&local_changed);
997 * Tries to bring node @p node and all it's neighbours to color @p tgt_col.
998 * @return 1 if color @p col could be applied, 0 otherwise
1000 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed_ones) {
1001 int col = get_mst_irn_col(node);
1003 /* if node already has the target color -> good, temporary fix it */
1004 if (col == tgt_col) {
1005 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1007 set_temp_color(node, tgt_col, changed_ones);
1012 Node has not yet a fixed color and target color is admissible
1013 -> try to recolor node and it's affinity neighbours
1015 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1016 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
1019 col_cost_init_single(env, costs, tgt_col);
1021 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1022 res = recolor_nodes(env, node, costs, changed_ones);
1023 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1029 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1030 if (!is_loose(node))
1031 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1033 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1034 dbg_admissible_colors(env, node);
1035 DB((dbg, LEVEL_4, ")\n"));
1044 * Tries to color an affinity chunk (or at least a part of it).
1045 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1047 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
1048 aff_chunk_t *best_chunk = NULL;
1049 int best_color = -1;
1051 waitq *tmp_chunks = new_waitq();
1052 waitq *best_starts = NULL;
1055 struct list_head changed_ones;
1057 DB((dbg, LEVEL_2, "fragmentizing chunk #%d", c->id));
1058 DBG_AFF_CHUNK(env, LEVEL_2, c);
1059 DB((dbg, LEVEL_2, "\n"));
1062 /* check which color is the "best" for the given chunk.
1063 * if we found a color which was ok for all nodes, we take it
1064 * and do not look further. (see did_all flag usage below.)
1065 * If we have many colors which fit all nodes it is hard to decide
1066 * which one to take anyway.
1067 * TODO Sebastian: Perhaps we should at all nodes and figure out
1068 * a suitable color using costs as done above (determine_color_costs).
1070 for (col = 0; col < env->n_regs && !did_all; ++col) {
1072 waitq *good_starts = new_waitq();
1073 aff_chunk_t *local_best;
1075 /* skip ignore colors */
1076 if (bitset_is_set(env->ignore_regs, col))
1079 DB((dbg, LEVEL_3, "\ttrying color %d\n", col));
1081 /* suppose we can color all nodes to the same color */
1084 INIT_LIST_HEAD(&changed_ones);
1086 /* try to bring all nodes of given chunk to the current color. */
1087 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1088 ir_node *irn = c->n[idx];
1089 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1092 assert(! node->fixed && "Node must not have a fixed color.");
1093 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1096 The order of the colored nodes is important, so we record the successfully
1097 colored ones in the order they appeared.
1099 good = change_node_color(env, node, col, &changed_ones);
1101 waitq_put(good_starts, node);
1107 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, one_good ? "succeeded" : "failed"));
1110 /* try next color when failed */
1112 reject_coloring(&changed_ones);
1116 /* fragment the chunk according to the coloring */
1117 local_best = fragment_chunk(env, col, c, tmp_chunks);
1119 /* search the best of the good list
1120 and make it the new best if it is better than the current */
1122 aff_chunk_assure_weight(env, local_best);
1124 DB((dbg, LEVEL_4, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, col));
1125 DBG_AFF_CHUNK(env, LEVEL_4, local_best);
1127 if (! best_chunk || best_chunk->weight < local_best->weight) {
1128 best_chunk = local_best;
1131 del_waitq(best_starts);
1132 best_starts = good_starts;
1133 DB((dbg, LEVEL_4, "\n\t\t... setting global best chunk (id %d), color %d\n", best_chunk->id, best_color));
1135 DB((dbg, LEVEL_4, "\n\t\t... omitting, global best is better\n"));
1136 del_waitq(good_starts);
1140 del_waitq(good_starts);
1143 reject_coloring(&changed_ones);
1146 /* free all intermediate created chunks except best one */
1147 while (! waitq_empty(tmp_chunks)) {
1148 aff_chunk_t *tmp = waitq_get(tmp_chunks);
1149 if (tmp != best_chunk)
1150 delete_aff_chunk(env, tmp);
1152 del_waitq(tmp_chunks);
1154 /* return if coloring failed */
1157 del_waitq(best_starts);
1161 DB((dbg, LEVEL_2, "\tbest chunk #%d ", best_chunk->id));
1162 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1163 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1165 INIT_LIST_HEAD(&changed_ones);
1166 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1167 ir_node *irn = best_chunk->n[idx];
1168 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1171 /* bring the node to the color. */
1172 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%d\n", best_color, node->irn, best_chunk->id));
1173 INIT_LIST_HEAD(&changed_ones);
1174 res = change_node_color(env, node, best_color, &changed_ones);
1176 materialize_coloring(&changed_ones);
1181 /* remove the nodes in best chunk from original chunk */
1182 bitset_andnot(c->nodes, best_chunk->nodes);
1183 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1184 ir_node *irn = c->n[idx];
1186 if (bitset_is_set(best_chunk->nodes, get_irn_idx(irn))) {
1187 int last = ARR_LEN(c->n) - 1;
1189 c->n[idx] = c->n[last];
1190 ARR_SHRINKLEN(c->n, last);
1195 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1196 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1197 ir_node *n = c->n[idx];
1198 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1202 /* fragment the remaining chunk */
1203 visited = bitset_irg_malloc(env->co->irg);
1204 bitset_or(visited, best_chunk->nodes);
1205 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1206 ir_node *irn = c->n[idx];
1207 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1208 aff_chunk_t *new_chunk = new_aff_chunk(env);
1209 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1211 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1212 aff_chunk_assure_weight(env, new_chunk);
1213 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1217 /* clear obsolete chunks and free some memory */
1218 delete_aff_chunk(env, best_chunk);
1219 bitset_free(visited);
1221 del_waitq(best_starts);
1225 * Main driver for mst safe coalescing algorithm.
1227 int co_solve_heuristic_mst(copy_opt_t *co) {
1228 unsigned n_regs = co->cls->n_regs;
1229 bitset_t *ignore_regs = bitset_alloca(n_regs);
1232 co_mst_env_t mst_env;
1235 phase_init(&mst_env.ph, "co_mst", co->irg, PHASE_DEFAULT_GROWTH, co_mst_irn_init, &mst_env);
1237 k = be_put_ignore_regs(co->cenv->birg, co->cls, ignore_regs);
1240 mst_env.n_regs = n_regs;
1242 mst_env.chunks = new_pqueue();
1244 mst_env.ignore_regs = ignore_regs;
1245 mst_env.ifg = co->cenv->ifg;
1246 mst_env.aenv = co->aenv;
1247 mst_env.chunkset = pset_new_ptr(512);
1249 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1251 /* build affinity chunks */
1252 build_affinity_chunks(&mst_env);
1254 /* color chunks as long as there are some */
1255 while (! pqueue_empty(mst_env.chunks)) {
1256 aff_chunk_t *chunk = pqueue_get(mst_env.chunks);
1258 color_aff_chunk(&mst_env, chunk);
1259 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%d) done\n", chunk->id));
1260 delete_aff_chunk(&mst_env, chunk);
1263 /* apply coloring */
1264 foreach_phase_irn(&mst_env.ph, irn) {
1265 co_mst_irn_t *mirn = get_co_mst_irn(&mst_env, irn);
1266 const arch_register_t *reg;
1268 if (arch_irn_is(mst_env.aenv, irn, ignore))
1271 // assert(mirn->fixed && "Node should have fixed color");
1273 /* skip nodes where color hasn't changed */
1274 if (mirn->init_col == mirn->col)
1277 reg = arch_register_for_index(co->cls, mirn->col);
1278 arch_set_irn_register(co->aenv, irn, reg);
1279 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1282 /* free allocated memory */
1283 del_pqueue(mst_env.chunks);
1284 phase_free(&mst_env.ph);
1285 del_pset(mst_env.chunkset);
1290 void be_init_copyheur4(void) {
1291 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");
1294 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);