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"
54 #include "becopyopt_t.h"
58 #define COL_COST_INFEASIBLE DBL_MAX
59 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
60 #define NEIGHBOUR_CONSTR_COSTS 64.0
64 #define DBG_AFF_CHUNK(env, level, chunk)
65 #define DBG_COL_COST(env, level, cost)
69 static firm_dbg_module_t *dbg = NULL;
70 #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while(0)
71 #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while(0)
75 static int last_chunk_id = 0;
77 typedef struct _col_cost_t {
85 typedef struct _aff_chunk_t {
86 ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
87 bitset_t *nodes; /**< A bitset containing all nodes inside this chunk. */
88 bitset_t *interfere; /**< A bitset containing all interfering neighbours of the nodes in this chunk. */
89 int weight; /**< Weight of this chunk */
90 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
91 unsigned deleted : 1; /**< Set if the was deleted. */
92 int id; /**< For debugging: An id of this chunk. */
98 typedef struct _aff_edge_t {
99 ir_node *src; /**< Source node. */
100 ir_node *tgt; /**< Target node. */
101 double weight; /**< The weight of this edge. */
104 /* main coalescing environment */
105 typedef struct _co_mst_env_t {
106 int n_regs; /**< number of regs in class */
107 int k; /**< number of non-ignore registers in class */
108 bitset_t *ignore_regs; /**< set containing all global ignore registers */
109 ir_phase ph; /**< phase object holding data for nodes */
110 pqueue *chunks; /**< priority queue for chunks */
111 pset *chunkset; /**< set holding all chunks */
112 be_ifg_t *ifg; /**< the interference graph */
113 const arch_env_t *aenv; /**< the arch environment */
114 copy_opt_t *co; /**< the copy opt object */
117 /* stores coalescing related information for a node */
118 typedef struct _co_mst_irn_t {
119 ir_node *irn; /**< the irn this information belongs to */
120 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
121 bitset_t *adm_colors; /**< set of admissible colors for this irn */
122 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
123 int n_neighs; /**< length of the interfering neighbours array. */
124 int int_aff_neigh; /**< number of interfering affinity neighbours */
125 int col; /**< color currently assigned */
126 int init_col; /**< the initial color */
127 int tmp_col; /**< a temporary assigned color */
128 unsigned fixed : 1; /**< the color is fixed */
129 unsigned tmp_fixed : 1; /**< the color is temporary fixed */
132 #define get_co_mst_irn(mst_env, irn) (phase_get_or_set_irn_data(&(mst_env)->ph, (irn)))
134 typedef int decide_func_t(const co_mst_irn_t *node, int col);
139 * Write a chunk to stderr for debugging.
141 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) {
143 if (c->weight_consistent)
144 ir_fprintf(stderr, " $%d ", c->weight);
145 ir_fprintf(stderr, "{");
146 bitset_foreach(c->nodes, idx) {
147 ir_node *n = get_idx_irn(env->co->irg, idx);
148 ir_fprintf(stderr, " %+F,", n);
150 ir_fprintf(stderr, "}");
154 * Dump all admissible colors to stderr.
156 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) {
160 if (bitset_popcnt(node->adm_colors) < 1)
161 fprintf(stderr, "no admissible colors?!?");
163 bitset_foreach(node->adm_colors, idx)
164 fprintf(stderr, " %d", idx);
169 * Dump color-cost pairs to stderr.
171 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) {
173 for (i = 0; i < env->n_regs; ++i) {
174 if (cost[i].cost == COL_COST_INFEASIBLE)
175 fprintf(stderr, " (%d, INF)", cost[i].col);
177 fprintf(stderr, " (%d, %.1f)", cost[i].col, cost[i].cost);
181 #endif /* DEBUG_libfirm */
183 static INLINE int get_mst_irn_col(const co_mst_irn_t *node) {
184 return node->tmp_fixed ? node->tmp_col : node->col;
188 * @return 1 if node @p node has color @p col, 0 otherwise.
190 static int decider_has_color(const co_mst_irn_t *node, int col) {
191 return get_mst_irn_col(node) == col;
195 * @return 1 if node @p node has not color @p col, 0 otherwise.
197 static int decider_hasnot_color(const co_mst_irn_t *node, int col) {
198 return get_mst_irn_col(node) != col;
202 * Always returns true.
204 static int decider_always_yes(const co_mst_irn_t *node, int col) {
210 /** compares two affinity edges by its weight */
211 static int cmp_aff_edge(const void *a, const void *b) {
212 const aff_edge_t *e1 = a;
213 const aff_edge_t *e2 = b;
215 if (e2->weight == e1->weight) {
216 if (e2->src->node_idx == e1->src->node_idx)
217 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
219 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
221 /* sort in descending order */
222 return QSORT_CMP(e2->weight, e1->weight);
225 /** compares to color-cost pairs */
226 static int cmp_col_cost(const void *a, const void *b) {
227 const col_cost_t *c1 = a;
228 const col_cost_t *c2 = b;
230 return c1->cost < c2->cost ? -1 : 1;
234 * Creates a new affinity chunk
236 static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) {
237 aff_chunk_t *c = xmalloc(sizeof(*c));
239 c->weight_consistent = 0;
240 c->n = NEW_ARR_F(ir_node *, 0);
241 c->nodes = bitset_irg_malloc(env->co->irg);
242 c->interfere = bitset_irg_malloc(env->co->irg);
243 c->id = last_chunk_id++;
244 pset_insert(env->chunkset, c, c->id);
249 * Frees all memory allocated by an affinity chunk.
251 static INLINE void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
252 pset_remove(env->chunkset, c, c->id);
253 bitset_free(c->nodes);
254 bitset_free(c->interfere);
261 * Adds a node to an affinity chunk
263 static INLINE void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) {
266 if (bitset_is_set(c->nodes, get_irn_idx(node->irn)))
269 c->weight_consistent = 0;
271 bitset_set(c->nodes, get_irn_idx(node->irn));
273 ARR_APP1(ir_node *, c->n, node->irn);
275 for (i = node->n_neighs - 1; i >= 0; --i) {
276 ir_node *neigh = node->int_neighs[i];
277 bitset_set(c->interfere, get_irn_idx(neigh));
282 * In case there is no phase information for irn, initialize it.
284 static void *co_mst_irn_init(ir_phase *ph, ir_node *irn, void *old) {
285 co_mst_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
286 co_mst_env_t *env = ph->priv;
289 const arch_register_req_t *req;
290 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;
304 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
306 /* set admissible registers */
307 res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs);
309 /* Exclude colors not assignable to the irn */
310 req = arch_get_register_req(env->aenv, irn, -1);
311 if (arch_register_req_is(req, limited))
312 rbitset_copy_to_bitset(req->limited, res->adm_colors);
314 bitset_set_all(res->adm_colors);
316 /* exclude global ignore registers as well */
317 bitset_andnot(res->adm_colors, env->ignore_regs);
319 /* set the number of interfering affinity neighbours to -1, they are calculated later */
320 res->int_aff_neigh = -1;
322 /* build list of interfering neighbours */
324 be_ifg_foreach_neighbour(env->ifg, nodes_it, irn, neigh) {
325 if (! arch_irn_is(env->aenv, neigh, ignore)) {
326 obstack_ptr_grow(phase_obst(ph), neigh);
330 res->int_neighs = obstack_finish(phase_obst(ph));
337 * Check if affinity chunk @p chunk interferes with node @p irn.
339 static INLINE int aff_chunk_interferes(co_mst_env_t *env, const aff_chunk_t *chunk, ir_node *irn) {
341 return bitset_is_set(chunk->interfere, get_irn_idx(irn));
345 * Check if there are interference edges from c1 to c2.
346 * @param env The global co_mst environment
348 * @param c2 Another chunk
349 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
351 static INLINE int aff_chunks_interfere(co_mst_env_t *env, const aff_chunk_t *c1, const aff_chunk_t *c2) {
357 /* check if there is a node in c2 having an interfering neighbor in c1 */
358 tmp = bitset_alloca(get_irg_last_idx(env->co->irg));
359 tmp = bitset_copy(tmp, c1->interfere);
360 tmp = bitset_and(tmp, c2->nodes);
362 return bitset_popcnt(tmp) > 0;
366 * Returns the affinity chunk of @p irn or creates a new
367 * one with @p irn as element if there is none assigned.
369 static INLINE aff_chunk_t *get_aff_chunk(co_mst_env_t *env, ir_node *irn) {
370 co_mst_irn_t *node = get_co_mst_irn(env, irn);
375 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
376 * are no interference edges from chunk(src) to chunk(tgt)).
377 * @return 1 if successful, 0 if not possible
379 static int aff_chunk_absorb(co_mst_env_t *env, ir_node *src, ir_node *tgt) {
380 aff_chunk_t *c1 = get_aff_chunk(env, src);
381 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
384 DB((dbg, LEVEL_4, "Attempt to let c1 (id %d): ", c1 ? c1->id : -1));
386 DBG_AFF_CHUNK(env, LEVEL_4, c1);
388 DB((dbg, LEVEL_4, "{%+F}", src));
390 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %d): ", c2 ? c2->id : -1));
392 DBG_AFF_CHUNK(env, LEVEL_4, c2);
394 DB((dbg, LEVEL_4, "{%+F}", tgt));
396 DB((dbg, LEVEL_4, "\n"));
401 /* no chunk exists */
402 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
405 for (i = mirn->n_neighs - 1; i >= 0; --i) {
406 if (mirn->int_neighs[i] == tgt)
410 /* create one containing both nodes */
411 c1 = new_aff_chunk(env);
412 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
413 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
417 /* c2 already exists */
418 if (! aff_chunk_interferes(env, c2, src)) {
419 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
423 } else if (c2 == NULL) {
424 /* c1 already exists */
425 if (! aff_chunk_interferes(env, c1, tgt)) {
426 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
429 } else if (c1 != c2 && ! aff_chunks_interfere(env, c1, c2)) {
432 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx) {
433 ir_node *n = c2->n[idx];
434 co_mst_irn_t *mn = get_co_mst_irn(env, n);
438 if (! bitset_is_set(c1->nodes, get_irn_idx(n)))
439 ARR_APP1(ir_node *, c1->n, n);
442 bitset_or(c1->nodes, c2->nodes);
443 bitset_or(c1->interfere, c2->interfere);
444 c1->weight_consistent = 0;
446 delete_aff_chunk(env, c2);
449 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
453 DB((dbg, LEVEL_4, " ... absorbed\n"));
458 * Assures that the weight of the given chunk is consistent.
460 static void aff_chunk_assure_weight(const co_mst_env_t *env, aff_chunk_t *c) {
461 if (! c->weight_consistent) {
465 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
466 ir_node *n = c->n[idx];
467 const affinity_node_t *an = get_affinity_info(env->co, n);
471 co_gs_foreach_neighb(an, neigh) {
472 const ir_node *m = neigh->irn;
473 const int m_idx = get_irn_idx(m);
475 /* skip ignore nodes */
476 if (arch_irn_is(env->aenv, m, ignore))
479 w += bitset_is_set(c->nodes, m_idx) ? neigh->costs : 0;
485 c->weight_consistent = 1;
490 * Count the number of interfering affinity neighbours
492 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an) {
493 const neighb_t *neigh;
494 ir_node *irn = an->irn;
495 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
498 co_gs_foreach_neighb(an, neigh) {
499 const ir_node *n = neigh->irn;
502 /* skip ignore nodes */
503 if (arch_irn_is(env->aenv, n, ignore))
506 /* check if the affinity neighbour interfere */
507 for (i = 0; i < node->n_neighs; ++i) {
508 if (node->int_neighs[i] == n) {
519 * Build chunks of nodes connected by affinity edges.
520 * We start at the heaviest affinity edge.
521 * The chunks of the two edge-defining nodes will be
522 * merged if there are no interference edges from one
523 * chunk to the other.
525 static void build_affinity_chunks(co_mst_env_t *env) {
526 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
527 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
530 aff_chunk_t *curr_chunk;
532 /* at first we create the affinity edge objects */
533 be_ifg_foreach_node(env->ifg, nodes_it, n) {
534 int n_idx = get_irn_idx(n);
538 /* skip ignore nodes */
539 if (arch_irn_is(env->aenv, n, ignore))
542 n1 = get_co_mst_irn(env, n);
543 an = get_affinity_info(env->co, n);
548 if (n1->int_aff_neigh < 0)
549 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
551 /* build the affinity edges */
552 co_gs_foreach_neighb(an, neigh) {
553 ir_node *m = neigh->irn;
554 int m_idx = get_irn_idx(m);
556 /* record the edge in only one direction */
561 /* skip ignore nodes */
562 if (arch_irn_is(env->aenv, m, ignore))
568 n2 = get_co_mst_irn(env, m);
569 if (n2->int_aff_neigh < 0) {
570 affinity_node_t *am = get_affinity_info(env->co, m);
571 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
574 * these weights are pure hackery ;-).
575 * It's not chriswue's fault but mine.
577 edge.weight = (double)neigh->costs / (double)(1 + n1->int_aff_neigh + n2->int_aff_neigh);
578 ARR_APP1(aff_edge_t, edges, edge);
584 /* now: sort edges and build the affinity chunks */
585 len = ARR_LEN(edges);
586 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
587 for (i = 0; i < len; ++i) {
588 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
590 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
593 /* now insert all chunks into a priority queue */
594 foreach_pset(env->chunkset, curr_chunk) {
595 aff_chunk_assure_weight(env, curr_chunk);
597 DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id));
598 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
599 DBG((dbg, LEVEL_1, "\n"));
601 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
603 foreach_phase_irn(&env->ph, n) {
604 co_mst_irn_t *mirn = get_co_mst_irn(env, n);
606 if (mirn->chunk == NULL) {
607 /* no chunk is allocated so far, do it now */
608 aff_chunk_t *curr_chunk = new_aff_chunk(env);
609 aff_chunk_add_node(curr_chunk, mirn);
611 aff_chunk_assure_weight(env, curr_chunk);
613 DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id));
614 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
615 DBG((dbg, LEVEL_1, "\n"));
617 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
625 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
627 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
628 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
630 waitq *nodes = new_waitq();
632 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%d) from %+F, color %d:", chunk->id, node->irn, col));
634 /* init queue and chunk */
635 waitq_put(nodes, node);
636 bitset_set(visited, get_irn_idx(node->irn));
637 aff_chunk_add_node(chunk, node);
638 DB((dbg, LEVEL_1, " %+F", node->irn));
640 /* as long as there are nodes in the queue */
641 while (! waitq_empty(nodes)) {
642 co_mst_irn_t *n = waitq_get(nodes);
643 affinity_node_t *an = get_affinity_info(env->co, n->irn);
645 /* check all affinity neighbors */
648 co_gs_foreach_neighb(an, neigh) {
649 ir_node *m = neigh->irn;
650 int m_idx = get_irn_idx(m);
653 /* skip ignore nodes */
654 if (arch_irn_is(env->aenv, m, ignore))
657 n2 = get_co_mst_irn(env, m);
659 if (! bitset_is_set(visited, m_idx) &&
662 ! aff_chunk_interferes(env, chunk, m) &&
663 bitset_is_set(orig_chunk->nodes, m_idx))
666 following conditions are met:
667 - neighbour is not visited
668 - neighbour likes the color
669 - neighbour has not yet a fixed color
670 - the new chunk doesn't interfere with the neighbour
671 - neighbour belongs or belonged once to the original chunk
673 bitset_set(visited, m_idx);
674 aff_chunk_add_node(chunk, n2);
675 DB((dbg, LEVEL_1, " %+F", n2->irn));
676 /* enqueue for further search */
677 waitq_put(nodes, n2);
683 DB((dbg, LEVEL_1, "\n"));
689 * Fragment the given chunk into chunks having given color and not having given color.
691 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp) {
692 bitset_t *visited = bitset_irg_malloc(env->co->irg);
694 aff_chunk_t *best = NULL;
696 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
699 aff_chunk_t *tmp_chunk;
700 decide_func_t *decider;
704 if (bitset_is_set(visited, get_irn_idx(irn)))
707 node = get_co_mst_irn(env, irn);
709 if (get_mst_irn_col(node) == col) {
710 decider = decider_has_color;
712 DBG((dbg, LEVEL_4, "\tcolor %d wanted", col));
715 decider = decider_hasnot_color;
717 DBG((dbg, LEVEL_4, "\tcolor %d forbidden", col));
720 /* create a new chunk starting at current node */
721 tmp_chunk = new_aff_chunk(env);
722 waitq_put(tmp, tmp_chunk);
723 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
724 assert(bitset_popcnt(tmp_chunk->nodes) > 0 && "No nodes added to chunk");
726 /* remember the local best */
727 aff_chunk_assure_weight(env, tmp_chunk);
728 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
732 assert(best && "No chunk found?");
733 bitset_free(visited);
738 * Initializes an array of color-cost pairs.
739 * Sets forbidden colors to costs COL_COST_INFEASIBLE and all others to @p c.
741 static INLINE void col_cost_init(co_mst_env_t *env, col_cost_t *cost, double c) {
744 for (i = 0; i < env->n_regs; ++i) {
746 if (bitset_is_set(env->ignore_regs, i))
747 cost[i].cost = COL_COST_INFEASIBLE;
754 * Initializes an array of color-cost pairs.
755 * Sets all colors except color @p col to COL_COST_INFEASIBLE and @p col to 0.0
757 static INLINE void col_cost_init_single(co_mst_env_t *env, col_cost_t *cost, int col) {
758 assert(! bitset_is_set(env->ignore_regs, col) && "Attempt to use forbidden color.");
759 col_cost_init(env, cost, COL_COST_INFEASIBLE);
766 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
767 * ATTENTION: the queue is empty after calling this function!
769 static INLINE void reject_coloring(waitq *nodes) {
770 DB((dbg, LEVEL_4, "\treject coloring for"));
771 while (! waitq_empty(nodes)) {
772 co_mst_irn_t *n = waitq_get(nodes);
773 DB((dbg, LEVEL_4, " %+F", n->irn));
776 DB((dbg, LEVEL_4, "\n"));
780 * Determines the costs for each color if it would be assigned to node @p node.
782 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) {
783 affinity_node_t *an = get_affinity_info(env->co, node->irn);
788 col_cost_init(env, costs, 0.0);
790 /* calculate (negative) costs for affinity neighbours */
792 co_gs_foreach_neighb(an, aff_neigh) {
793 ir_node *m = aff_neigh->irn;
797 /* skip ignore nodes */
798 if (arch_irn_is(env->aenv, m, ignore))
801 neigh = get_co_mst_irn(env, m);
802 c = (double)aff_neigh->costs;
804 /* calculate costs for fixed affinity neighbours */
805 if (neigh->tmp_fixed || neigh->fixed) {
806 int col = get_mst_irn_col(neigh);
807 costs[col].cost -= c * AFF_NEIGHBOUR_FIX_BENEFIT;
812 /* calculate (positive) costs for interfering neighbours */
813 for (i = 0; i < node->n_neighs; ++i) {
818 int_neigh = node->int_neighs[i];
820 /* skip ignore nodes */
821 if (arch_irn_is(env->aenv, int_neigh, ignore))
824 neigh = get_co_mst_irn(env, int_neigh);
825 col = get_mst_irn_col(neigh);
826 col_cnt = bitset_popcnt(neigh->adm_colors);
828 if (neigh->tmp_fixed || neigh->fixed) {
829 /* colors of fixed interfering neighbours are infeasible */
830 costs[col].cost = COL_COST_INFEASIBLE;
832 else if (col_cnt < env->k) {
833 /* calculate costs for constrained interfering neighbours */
834 double ratio = 1.0 - ((double)col_cnt / (double)env->k);
836 bitset_foreach_clear(neigh->adm_colors, idx) {
837 /* check only explicitly forbidden colors (skip global forbidden ones) */
838 if (! bitset_is_set(env->ignore_regs, idx)) {
839 costs[col].cost += ratio * NEIGHBOUR_CONSTR_COSTS;
845 /* set all not admissible colors to COL_COST_INFEASIBLE */
846 bitset_foreach_clear(node->adm_colors, idx)
847 costs[idx].cost = COL_COST_INFEASIBLE;
850 /* need forward declaration due to recursive call */
851 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, waitq *changed_ones);
854 * Tries to change node to a color but @p explude_col.
855 * @return 1 if succeeded, 0 otherwise.
857 static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, waitq *changed_ones) {
858 int col = get_mst_irn_col(node);
861 /* neighbours has already a different color -> good, temporary fix it */
862 if (col != exclude_col) {
865 waitq_put(changed_ones, node);
869 /* The node has the color it should not have _and_ has not been visited yet. */
870 if (! (node->tmp_fixed || node->fixed)) {
871 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
873 /* Get the costs for giving the node a specific color. */
874 determine_color_costs(env, node, costs);
876 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
877 costs[exclude_col].cost = COL_COST_INFEASIBLE;
879 /* sort the colors according costs, cheapest first. */
880 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost);
882 /* Try recoloring the node using the color list. */
883 res = recolor_nodes(env, node, costs, changed_ones);
890 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
891 * ATTENTION: Expect @p costs already sorted by increasing costs.
892 * @return 1 if coloring could be applied, 0 otherwise.
894 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, waitq *changed_ones) {
896 waitq *local_changed = new_waitq();
897 waitq *tmp = new_waitq();
899 DBG((dbg, LEVEL_1, "\tRecoloring %+F with color-costs", node->irn));
900 DBG_COL_COST(env, LEVEL_1, costs);
901 DB((dbg, LEVEL_1, "\n"));
903 for (i = 0; i < env->n_regs; ++i) {
904 int tgt_col = costs[i].col;
908 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
909 if (costs[i].cost == COL_COST_INFEASIBLE) {
911 del_waitq(local_changed);
916 /* Set the new color of the node and mark the node as temporarily fixed. */
917 assert(! node->tmp_fixed && "Node must not have been temporary fixed.");
919 node->tmp_col = tgt_col;
920 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
922 assert(waitq_empty(local_changed) && "Node queue should be empty here.");
923 waitq_put(local_changed, node);
925 /* try to color all interfering neighbours with current color forbidden */
926 for (j = 0; j < node->n_neighs; ++j) {
930 neigh = node->int_neighs[j];
932 /* skip ignore nodes */
933 if (arch_irn_is(env->aenv, neigh, ignore))
936 nn = get_co_mst_irn(env, neigh);
937 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_fixed: %d, tmp_col: %d, col: %d)\n",
938 neigh, j, nn->fixed, nn->tmp_fixed, nn->tmp_col, nn->col));
941 Try to change the color of the neighbor and record all nodes which
942 get changed in the tmp list. Add this list to the "changed" list for
943 that color. If we did not succeed to change the color of the neighbor,
944 we bail out and try the next color.
946 if (get_mst_irn_col(nn) == tgt_col) {
947 /* try to color neighbour with tgt_col forbidden */
948 neigh_ok = change_node_color_excluded(env, nn, tgt_col, tmp);
950 /* join lists of changed nodes */
951 while (! waitq_empty(tmp))
952 waitq_put(local_changed, waitq_get(tmp));
960 We managed to assign the target color to all neighbors, so from the perspective
961 of the current node, every thing was ok and we can return safely.
964 /* append the local_changed ones to global ones */
965 while (! waitq_empty(local_changed))
966 waitq_put(changed_ones, waitq_get(local_changed));
967 del_waitq(local_changed);
972 /* coloring of neighbours failed, so we try next color */
973 reject_coloring(local_changed);
977 del_waitq(local_changed);
983 * Tries to bring node @p node and all it's neighbours to color @p tgt_col.
984 * @return 1 if color @p col could be applied, 0 otherwise
986 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, waitq *changed_ones) {
987 int col = get_mst_irn_col(node);
989 /* if node already has the target color -> good, temporary fix it */
990 if (col == tgt_col) {
991 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
992 if (! node->tmp_fixed) {
994 node->tmp_col = tgt_col;
995 waitq_put(changed_ones, node);
1001 Node has not yet a fixed color and target color is admissible
1002 -> try to recolor node and it's affinity neighbours
1004 if (! (node->fixed || node->tmp_fixed) && bitset_is_set(node->adm_colors, tgt_col)) {
1005 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
1008 col_cost_init_single(env, costs, tgt_col);
1010 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1011 res = recolor_nodes(env, node, costs, changed_ones);
1012 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1018 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1019 if (node->fixed || node->tmp_fixed)
1020 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1022 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1023 dbg_admissible_colors(env, node);
1024 DB((dbg, LEVEL_4, ")\n"));
1033 * Tries to color an affinity chunk (or at least a part of it).
1034 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1036 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
1037 aff_chunk_t *best_chunk = NULL;
1038 int best_color = -1;
1040 waitq *changed_ones = new_waitq();
1041 waitq *tmp_chunks = new_waitq();
1042 waitq *best_starts = NULL;
1046 DB((dbg, LEVEL_2, "fragmentizing chunk #%d", c->id));
1047 DBG_AFF_CHUNK(env, LEVEL_2, c);
1048 DB((dbg, LEVEL_2, "\n"));
1051 /* check which color is the "best" for the given chunk.
1052 * if we found a color which was ok for all nodes, we take it
1053 * and do not look further. (see did_all flag usage below.)
1054 * If we have many colors which fit all nodes it is hard to decide
1055 * which one to take anyway.
1056 * TODO Sebastian: Perhaps we should at all nodes and figure out
1057 * a suitable color using costs as done above (determine_color_costs).
1059 for (col = 0; col < env->n_regs && !did_all; ++col) {
1061 waitq *good_starts = new_waitq();
1062 aff_chunk_t *local_best;
1064 /* skip ignore colors */
1065 if (bitset_is_set(env->ignore_regs, col))
1068 DB((dbg, LEVEL_3, "\ttrying color %d\n", col));
1070 /* suppose we can color all nodes to the same color */
1073 /* try to bring all nodes of given chunk to the current color. */
1074 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1075 ir_node *irn = c->n[idx];
1076 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1079 assert(! node->fixed && "Node must not have a fixed color.");
1080 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1083 The order of the colored nodes is important, so we record the successfully
1084 colored ones in the order they appeared.
1086 good = change_node_color(env, node, col, changed_ones);
1088 waitq_put(good_starts, node);
1092 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, one_good ? "succeeded" : "failed"));
1095 /* try next color when failed */
1099 /* fragment the chunk according to the coloring */
1100 local_best = fragment_chunk(env, col, c, tmp_chunks);
1102 /* search the best of the good list
1103 and make it the new best if it is better than the current */
1105 aff_chunk_assure_weight(env, local_best);
1107 DB((dbg, LEVEL_4, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, col));
1108 DBG_AFF_CHUNK(env, LEVEL_4, local_best);
1110 if (! best_chunk || best_chunk->weight < local_best->weight) {
1111 best_chunk = local_best;
1114 del_waitq(best_starts);
1115 best_starts = good_starts;
1116 DB((dbg, LEVEL_4, "\n\t\t... setting global best chunk (id %d), color %d\n", best_chunk->id, best_color));
1118 DB((dbg, LEVEL_4, "\n\t\t... omitting, global best is better\n"));
1119 del_waitq(good_starts);
1123 del_waitq(good_starts);
1126 /* reject the coloring and bring the coloring to the initial state */
1127 reject_coloring(changed_ones);
1130 /* free all intermediate created chunks except best one */
1131 while (! waitq_empty(tmp_chunks)) {
1132 aff_chunk_t *tmp = waitq_get(tmp_chunks);
1133 if (tmp != best_chunk)
1134 delete_aff_chunk(env, tmp);
1136 del_waitq(tmp_chunks);
1138 /* return if coloring failed */
1140 del_waitq(changed_ones);
1142 del_waitq(best_starts);
1146 DB((dbg, LEVEL_2, "\tbest chunk #%d ", best_chunk->id));
1147 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1148 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1150 /* get the best fragment from the best list and color it */
1151 while (! waitq_empty(best_starts)) {
1152 co_mst_irn_t *node = waitq_get(best_starts);
1155 if (! bitset_is_set(best_chunk->nodes, get_irn_idx(node->irn)))
1158 res = change_node_color(env, node, best_color, changed_ones);
1160 panic("Color manifesting failed for %+F, color %d in chunk %d\n", node->irn, best_color, best_chunk->id);
1162 node->chunk = best_chunk;
1164 /* we colored the successful start nodes, now color the rest of the chunk */
1165 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1166 ir_node *irn = best_chunk->n[idx];
1167 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1170 res = change_node_color(env, node, best_color, changed_ones);
1172 panic("Color manifesting failed for %+F, color %d in chunk %d\n", irn, best_color, best_chunk->id);
1173 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%d\n", best_color, irn, best_chunk->id));
1175 node->chunk = best_chunk;
1178 /* materialize colors on changed nodes */
1179 while (! waitq_empty(changed_ones)) {
1180 co_mst_irn_t *n = waitq_get(changed_ones);
1182 n->col = n->tmp_col;
1185 /* remove the nodes in best chunk from original chunk */
1186 bitset_andnot(c->nodes, best_chunk->nodes);
1187 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1188 ir_node *irn = c->n[idx];
1190 if (bitset_is_set(best_chunk->nodes, get_irn_idx(irn))) {
1191 int last = ARR_LEN(c->n) - 1;
1193 c->n[idx] = c->n[last];
1194 ARR_SHRINKLEN(c->n, last);
1199 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1200 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1201 ir_node *n = c->n[idx];
1202 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1206 /* fragment the remaining chunk */
1207 visited = bitset_irg_malloc(env->co->irg);
1208 bitset_or(visited, best_chunk->nodes);
1209 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1210 ir_node *irn = c->n[idx];
1211 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1212 aff_chunk_t *new_chunk = new_aff_chunk(env);
1213 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1215 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1216 aff_chunk_assure_weight(env, new_chunk);
1217 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1221 /* clear obsolete chunks and free some memory */
1222 delete_aff_chunk(env, best_chunk);
1223 bitset_free(visited);
1224 del_waitq(changed_ones);
1226 del_waitq(best_starts);
1230 * Main driver for mst safe coalescing algorithm.
1232 int co_solve_heuristic_mst(copy_opt_t *co) {
1233 unsigned n_regs = co->cls->n_regs;
1234 bitset_t *ignore_regs = bitset_alloca(n_regs);
1237 co_mst_env_t mst_env;
1240 phase_init(&mst_env.ph, "co_mst", co->irg, PHASE_DEFAULT_GROWTH, co_mst_irn_init, &mst_env);
1242 k = be_put_ignore_regs(co->cenv->birg, co->cls, ignore_regs);
1245 mst_env.n_regs = n_regs;
1247 mst_env.chunks = new_pqueue();
1249 mst_env.ignore_regs = ignore_regs;
1250 mst_env.ifg = co->cenv->ifg;
1251 mst_env.aenv = co->aenv;
1252 mst_env.chunkset = pset_new_ptr(512);
1254 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1256 /* build affinity chunks */
1257 build_affinity_chunks(&mst_env);
1259 /* color chunks as long as there are some */
1260 while (! pqueue_empty(mst_env.chunks)) {
1261 aff_chunk_t *chunk = pqueue_get(mst_env.chunks);
1263 color_aff_chunk(&mst_env, chunk);
1264 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%d) done\n", chunk->id));
1265 delete_aff_chunk(&mst_env, chunk);
1268 /* apply coloring */
1269 foreach_phase_irn(&mst_env.ph, irn) {
1270 co_mst_irn_t *mirn = get_co_mst_irn(&mst_env, irn);
1271 const arch_register_t *reg;
1273 if (arch_irn_is(mst_env.aenv, irn, ignore))
1276 assert(mirn->fixed && "Node should have fixed color");
1278 /* skip nodes where color hasn't changed */
1279 if (mirn->init_col == mirn->col)
1282 reg = arch_register_for_index(co->cls, mirn->col);
1283 arch_set_irn_register(co->aenv, irn, reg);
1284 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1287 /* free allocated memory */
1288 del_pqueue(mst_env.chunks);
1289 phase_free(&mst_env.ph);
1290 del_pset(mst_env.chunkset);
1295 void be_init_copyheur4(void) {
1296 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");
1299 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);