2 * Copyright (C) 1995-2008 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 */
36 #define DISABLE_STATEV
43 #include "raw_bitset.h"
44 #include "irphase_t.h"
60 #include "becopyopt_t.h"
64 #define COL_COST_INFEASIBLE DBL_MAX
65 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
66 #define NEIGHBOUR_CONSTR_COSTS 64.0
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)
74 static firm_dbg_module_t *dbg = NULL;
78 #define DBG_AFF_CHUNK(env, level, chunk)
79 #define DBG_COL_COST(env, level, cost)
84 #define REAL(C) (C ## f)
86 static unsigned last_chunk_id = 0;
87 static int recolor_limit = 7;
88 static real_t dislike_influence = REAL(0.1);
90 typedef struct _col_cost_t {
98 typedef struct _aff_chunk_t {
99 const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
100 const ir_node **interfere; /**< An ARR_F containing all inference. */
101 int weight; /**< Weight of this chunk */
102 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
103 unsigned deleted : 1; /**< For debugging: Set if the was deleted. */
104 unsigned id; /**< An id of this chunk. */
106 col_cost_t color_affinity[1];
112 typedef struct _aff_edge_t {
113 const ir_node *src; /**< Source node. */
114 const ir_node *tgt; /**< Target node. */
115 int weight; /**< The weight of this edge. */
118 /* main coalescing environment */
119 typedef struct _co_mst_env_t {
120 int n_regs; /**< number of regs in class */
121 int k; /**< number of non-ignore registers in class */
122 bitset_t *ignore_regs; /**< set containing all global ignore registers */
123 ir_phase ph; /**< phase object holding data for nodes */
124 pqueue_t *chunks; /**< priority queue for chunks */
125 pset *chunkset; /**< set holding all chunks */
126 be_ifg_t *ifg; /**< the interference graph */
127 const arch_env_t *aenv; /**< the arch environment */
128 copy_opt_t *co; /**< the copy opt object */
129 unsigned chunk_visited;
130 col_cost_t **single_cols;
133 /* stores coalescing related information for a node */
134 typedef struct _co_mst_irn_t {
135 const ir_node *irn; /**< the irn this information belongs to */
136 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
137 bitset_t *adm_colors; /**< set of admissible colors for this irn */
138 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
139 int n_neighs; /**< length of the interfering neighbours array. */
140 int int_aff_neigh; /**< number of interfering affinity neighbours */
141 int col; /**< color currently assigned */
142 int init_col; /**< the initial color */
143 int tmp_col; /**< a temporary assigned color */
144 unsigned fixed : 1; /**< the color is fixed */
145 struct list_head list; /**< Queue for coloring undo. */
146 real_t constr_factor;
149 #define get_co_mst_irn(mst_env, irn) (phase_get_or_set_irn_data(&(mst_env)->ph, (irn)))
151 typedef int decide_func_t(const co_mst_irn_t *node, int col);
156 * Write a chunk to stderr for debugging.
158 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) {
161 if (c->weight_consistent)
162 ir_fprintf(stderr, " $%d ", c->weight);
163 ir_fprintf(stderr, "{");
164 for (i = 0, l = ARR_LEN(c->n); i < l; ++i) {
165 const ir_node *n = c->n[i];
166 ir_fprintf(stderr, " %+F,", n);
168 ir_fprintf(stderr, "}");
172 * Dump all admissible colors to stderr.
174 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) {
178 if (bitset_popcnt(node->adm_colors) < 1)
179 fprintf(stderr, "no admissible colors?!?");
181 bitset_foreach(node->adm_colors, idx) {
182 fprintf(stderr, " %d", idx);
188 * Dump color-cost pairs to stderr.
190 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) {
192 for (i = 0; i < env->n_regs; ++i)
193 fprintf(stderr, " (%d, %.4f)", cost[i].col, cost[i].cost);
196 #endif /* DEBUG_libfirm */
198 static INLINE int get_mst_irn_col(const co_mst_irn_t *node) {
199 return node->tmp_col >= 0 ? node->tmp_col : node->col;
203 * @return 1 if node @p node has color @p col, 0 otherwise.
205 static int decider_has_color(const co_mst_irn_t *node, int col) {
206 return get_mst_irn_col(node) == col;
210 * @return 1 if node @p node has not color @p col, 0 otherwise.
212 static int decider_hasnot_color(const co_mst_irn_t *node, int col) {
213 return get_mst_irn_col(node) != col;
217 * Always returns true.
219 static int decider_always_yes(const co_mst_irn_t *node, int col) {
225 /** compares two affinity edges by its weight */
226 static int cmp_aff_edge(const void *a, const void *b) {
227 const aff_edge_t *e1 = a;
228 const aff_edge_t *e2 = b;
230 if (e2->weight == e1->weight) {
231 if (e2->src->node_idx == e1->src->node_idx)
232 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
234 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
236 /* sort in descending order */
237 return QSORT_CMP(e2->weight, e1->weight);
240 /** compares to color-cost pairs */
241 static __attribute__((unused)) int cmp_col_cost_lt(const void *a, const void *b) {
242 const col_cost_t *c1 = a;
243 const col_cost_t *c2 = b;
244 real_t diff = c1->cost - c2->cost;
245 return (diff > 0) - (diff < 0);
248 static int cmp_col_cost_gt(const void *a, const void *b) {
249 const col_cost_t *c1 = a;
250 const col_cost_t *c2 = b;
251 real_t diff = c2->cost - c1->cost;
252 return (diff > 0) - (diff < 0);
256 * Creates a new affinity chunk
258 static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) {
259 aff_chunk_t *c = xmalloc(sizeof(*c) + (env->n_regs - 1) * sizeof(c->color_affinity[0]));
260 c->n = NEW_ARR_F(const ir_node *, 0);
261 c->interfere = NEW_ARR_F(const ir_node *, 0);
263 c->weight_consistent = 0;
265 c->id = ++last_chunk_id;
267 pset_insert(env->chunkset, c, c->id);
272 * Frees all memory allocated by an affinity chunk.
274 static INLINE void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
275 pset_remove(env->chunkset, c, c->id);
276 DEL_ARR_F(c->interfere);
283 * binary search of sorted nodes.
285 * @return the position where n is found in the array arr or ~pos
286 * if the nodes is not here.
288 static INLINE int nodes_bsearch(const ir_node **arr, const ir_node *n) {
289 int hi = ARR_LEN(arr);
293 int md = lo + ((hi - lo) >> 1);
306 /** Check if a node n can be found inside arr. */
307 static int node_contains(const ir_node **arr, const ir_node *n) {
308 int i = nodes_bsearch(arr, n);
313 * Insert a node into the sorted nodes list.
315 * @return 1 if the node was inserted, 0 else
317 static int nodes_insert(const ir_node ***arr, const ir_node *irn) {
318 int idx = nodes_bsearch(*arr, irn);
321 int i, n = ARR_LEN(*arr);
324 ARR_APP1(const ir_node *, *arr, irn);
329 for (i = n - 1; i >= idx; --i)
338 * Adds a node to an affinity chunk
340 static INLINE void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) {
343 if (! nodes_insert(&c->n, node->irn))
346 c->weight_consistent = 0;
349 for (i = node->n_neighs - 1; i >= 0; --i) {
350 ir_node *neigh = node->int_neighs[i];
351 nodes_insert(&c->interfere, neigh);
356 * In case there is no phase information for irn, initialize it.
358 static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn, void *old) {
359 co_mst_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0]));
360 co_mst_env_t *env = ph->priv;
363 const arch_register_req_t *req;
364 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
372 res->int_neighs = NULL;
373 res->int_aff_neigh = 0;
374 res->col = arch_register_get_index(arch_get_irn_register(env->aenv, irn));
375 res->init_col = res->col;
376 INIT_LIST_HEAD(&res->list);
378 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
380 /* set admissible registers */
381 res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs);
383 /* Exclude colors not assignable to the irn */
384 req = arch_get_register_req(env->aenv, irn, -1);
385 if (arch_register_req_is(req, limited))
386 rbitset_copy_to_bitset(req->limited, res->adm_colors);
388 bitset_set_all(res->adm_colors);
390 /* exclude global ignore registers as well */
391 bitset_andnot(res->adm_colors, env->ignore_regs);
393 /* compute the constraint factor */
394 res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcnt(res->adm_colors)) / env->n_regs;
396 /* set the number of interfering affinity neighbours to -1, they are calculated later */
397 res->int_aff_neigh = -1;
399 /* build list of interfering neighbours */
401 be_ifg_foreach_neighbour(env->ifg, nodes_it, irn, neigh) {
402 if (! arch_irn_is(env->aenv, neigh, ignore)) {
403 obstack_ptr_grow(phase_obst(ph), neigh);
407 res->int_neighs = obstack_finish(phase_obst(ph));
414 * Check if affinity chunk @p chunk interferes with node @p irn.
416 static INLINE int aff_chunk_interferes(const aff_chunk_t *chunk, const ir_node *irn) {
417 return node_contains(chunk->interfere, irn);
421 * Check if there are interference edges from c1 to c2.
423 * @param c2 Another chunk
424 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
426 static INLINE int aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2) {
432 /* check if there is a node in c2 having an interfering neighbor in c1 */
433 for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) {
434 const ir_node *irn = c2->n[i];
436 if (node_contains(c1->interfere, irn))
443 * Returns the affinity chunk of @p irn or creates a new
444 * one with @p irn as element if there is none assigned.
446 static INLINE aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn) {
447 co_mst_irn_t *node = get_co_mst_irn(env, irn);
452 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
453 * are no interference edges from chunk(src) to chunk(tgt)).
454 * @return 1 if successful, 0 if not possible
456 static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node *tgt) {
457 aff_chunk_t *c1 = get_aff_chunk(env, src);
458 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
461 DB((dbg, LEVEL_4, "Attempt to let c1 (id %u): ", c1 ? c1->id : 0));
463 DBG_AFF_CHUNK(env, LEVEL_4, c1);
465 DB((dbg, LEVEL_4, "{%+F}", src));
467 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %u): ", c2 ? c2->id : 0));
469 DBG_AFF_CHUNK(env, LEVEL_4, c2);
471 DB((dbg, LEVEL_4, "{%+F}", tgt));
473 DB((dbg, LEVEL_4, "\n"));
478 /* no chunk exists */
479 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
482 for (i = mirn->n_neighs - 1; i >= 0; --i) {
483 if (mirn->int_neighs[i] == tgt)
487 /* create one containing both nodes */
488 c1 = new_aff_chunk(env);
489 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
490 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
494 /* c2 already exists */
495 if (! aff_chunk_interferes(c2, src)) {
496 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
500 } else if (c2 == NULL) {
501 /* c1 already exists */
502 if (! aff_chunk_interferes(c1, tgt)) {
503 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
506 } else if (c1 != c2 && ! aff_chunks_interfere(c1, c2)) {
509 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx)
510 aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx]));
512 for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) {
513 const ir_node *irn = c2->interfere[idx];
514 nodes_insert(&c1->interfere, irn);
517 c1->weight_consistent = 0;
519 delete_aff_chunk(env, c2);
522 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
526 DB((dbg, LEVEL_4, " ... absorbed\n"));
531 * Assures that the weight of the given chunk is consistent.
533 static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) {
534 if (! c->weight_consistent) {
538 for (i = 0; i < env->n_regs; ++i) {
539 c->color_affinity[i].col = i;
540 c->color_affinity[i].cost = REAL(0.0);
543 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
544 const ir_node *n = c->n[idx];
545 const affinity_node_t *an = get_affinity_info(env->co, n);
546 co_mst_irn_t *node = get_co_mst_irn(env, n);
549 if (node->constr_factor > REAL(0.0)) {
551 bitset_foreach (node->adm_colors, col)
552 c->color_affinity[col].cost += node->constr_factor;
557 co_gs_foreach_neighb(an, neigh) {
558 const ir_node *m = neigh->irn;
560 /* skip ignore nodes */
561 if (arch_irn_is(env->aenv, m, ignore))
564 w += node_contains(c->n, m) ? neigh->costs : 0;
569 for (i = 0; i < env->n_regs; ++i)
570 c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n));
573 // c->weight = bitset_popcnt(c->nodes);
574 c->weight_consistent = 1;
579 * Count the number of interfering affinity neighbours
581 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an) {
582 const neighb_t *neigh;
583 const ir_node *irn = an->irn;
584 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
587 co_gs_foreach_neighb(an, neigh) {
588 const ir_node *n = neigh->irn;
591 /* skip ignore nodes */
592 if (arch_irn_is(env->aenv, n, ignore))
595 /* check if the affinity neighbour interfere */
596 for (i = 0; i < node->n_neighs; ++i) {
597 if (node->int_neighs[i] == n) {
608 * Build chunks of nodes connected by affinity edges.
609 * We start at the heaviest affinity edge.
610 * The chunks of the two edge-defining nodes will be
611 * merged if there are no interference edges from one
612 * chunk to the other.
614 static void build_affinity_chunks(co_mst_env_t *env) {
615 void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg);
616 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
619 aff_chunk_t *curr_chunk;
621 /* at first we create the affinity edge objects */
622 be_ifg_foreach_node(env->ifg, nodes_it, n) {
623 int n_idx = get_irn_idx(n);
627 /* skip ignore nodes */
628 if (arch_irn_is(env->aenv, n, ignore))
631 n1 = get_co_mst_irn(env, n);
632 an = get_affinity_info(env->co, n);
637 if (n1->int_aff_neigh < 0)
638 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
640 /* build the affinity edges */
641 co_gs_foreach_neighb(an, neigh) {
642 const ir_node *m = neigh->irn;
643 int m_idx = get_irn_idx(m);
645 /* record the edge in only one direction */
650 /* skip ignore nodes */
651 if (arch_irn_is(env->aenv, m, ignore))
657 n2 = get_co_mst_irn(env, m);
658 if (n2->int_aff_neigh < 0) {
659 affinity_node_t *am = get_affinity_info(env->co, m);
660 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
663 * these weights are pure hackery ;-).
664 * It's not chriswue's fault but mine.
666 edge.weight = neigh->costs;
667 ARR_APP1(aff_edge_t, edges, edge);
673 /* now: sort edges and build the affinity chunks */
674 len = ARR_LEN(edges);
675 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
676 for (i = 0; i < len; ++i) {
677 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
679 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
682 /* now insert all chunks into a priority queue */
683 foreach_pset(env->chunkset, curr_chunk) {
684 aff_chunk_assure_weight(env, curr_chunk);
686 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
687 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
688 DBG((dbg, LEVEL_1, "\n"));
690 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
693 foreach_phase_irn(&env->ph, n) {
694 co_mst_irn_t *mirn = get_co_mst_irn(env, n);
696 if (mirn->chunk == NULL) {
697 /* no chunk is allocated so far, do it now */
698 aff_chunk_t *curr_chunk = new_aff_chunk(env);
699 aff_chunk_add_node(curr_chunk, mirn);
701 aff_chunk_assure_weight(env, curr_chunk);
703 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
704 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
705 DBG((dbg, LEVEL_1, "\n"));
707 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
714 static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk)
716 pqueue_t *grow = new_pqueue();
717 const ir_node *max_node = NULL;
721 for (i = ARR_LEN(chunk->n) - 1; i >= 0; i--) {
722 const ir_node *irn = chunk->n[i];
723 affinity_node_t *an = get_affinity_info(env->co, irn);
727 if (arch_irn_is(env->aenv, irn, ignore))
731 co_gs_foreach_neighb(an, neigh)
734 if (w > max_weight) {
742 bitset_t *visited = bitset_irg_malloc(env->co->irg);
744 for (i = ARR_LEN(chunk->n) - 1; i >= 0; --i)
745 bitset_add_irn(visited, chunk->n[i]);
747 pqueue_put(grow, (void *) max_node, max_weight);
748 bitset_remv_irn(visited, max_node);
750 while (!pqueue_empty(grow)) {
751 ir_node *irn = pqueue_pop_front(grow);
752 affinity_node_t *an = get_affinity_info(env->co, irn);
755 if (arch_irn_is(env->aenv, irn, ignore))
758 assert(i <= ARR_LEN(chunk->n));
763 /* build the affinity edges */
764 co_gs_foreach_neighb(an, neigh) {
765 co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn);
767 if (bitset_contains_irn(visited, node->irn)) {
768 pqueue_put(grow, (void *) neigh->irn, neigh->costs);
769 bitset_remv_irn(visited, node->irn);
775 bitset_free(visited);
780 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
782 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
783 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
785 waitq *nodes = new_waitq();
787 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col));
789 /* init queue and chunk */
790 waitq_put(nodes, node);
791 bitset_set(visited, get_irn_idx(node->irn));
792 aff_chunk_add_node(chunk, node);
793 DB((dbg, LEVEL_1, " %+F", node->irn));
795 /* as long as there are nodes in the queue */
796 while (! waitq_empty(nodes)) {
797 co_mst_irn_t *n = waitq_get(nodes);
798 affinity_node_t *an = get_affinity_info(env->co, n->irn);
800 /* check all affinity neighbors */
803 co_gs_foreach_neighb(an, neigh) {
804 const ir_node *m = neigh->irn;
805 int m_idx = get_irn_idx(m);
808 /* skip ignore nodes */
809 if (arch_irn_is(env->aenv, m, ignore))
812 n2 = get_co_mst_irn(env, m);
814 if (! bitset_is_set(visited, m_idx) &&
817 ! aff_chunk_interferes(chunk, m) &&
818 node_contains(orig_chunk->n, m))
821 following conditions are met:
822 - neighbour is not visited
823 - neighbour likes the color
824 - neighbour has not yet a fixed color
825 - the new chunk doesn't interfere with the neighbour
826 - neighbour belongs or belonged once to the original chunk
828 bitset_set(visited, m_idx);
829 aff_chunk_add_node(chunk, n2);
830 DB((dbg, LEVEL_1, " %+F", n2->irn));
831 /* enqueue for further search */
832 waitq_put(nodes, n2);
838 DB((dbg, LEVEL_1, "\n"));
844 * Fragment the given chunk into chunks having given color and not having given color.
846 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp) {
847 bitset_t *visited = bitset_irg_malloc(env->co->irg);
849 aff_chunk_t *best = NULL;
851 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
854 aff_chunk_t *tmp_chunk;
855 decide_func_t *decider;
859 if (bitset_is_set(visited, get_irn_idx(irn)))
862 node = get_co_mst_irn(env, irn);
864 if (get_mst_irn_col(node) == col) {
865 decider = decider_has_color;
867 DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col));
870 decider = decider_hasnot_color;
872 DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col));
875 /* create a new chunk starting at current node */
876 tmp_chunk = new_aff_chunk(env);
877 waitq_put(tmp, tmp_chunk);
878 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
879 assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk");
881 /* remember the local best */
882 aff_chunk_assure_weight(env, tmp_chunk);
883 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
887 assert(best && "No chunk found?");
888 bitset_free(visited);
893 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
894 * ATTENTION: the queue is empty after calling this function!
896 static INLINE void reject_coloring(struct list_head *nodes) {
897 co_mst_irn_t *n, *temp;
898 DB((dbg, LEVEL_4, "\treject coloring for"));
899 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
900 DB((dbg, LEVEL_4, " %+F", n->irn));
901 assert(n->tmp_col >= 0);
903 list_del_init(&n->list);
905 DB((dbg, LEVEL_4, "\n"));
908 static INLINE void materialize_coloring(struct list_head *nodes) {
909 co_mst_irn_t *n, *temp;
910 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
911 assert(n->tmp_col >= 0);
914 list_del_init(&n->list);
918 static INLINE void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
921 assert(!node->fixed);
922 assert(node->tmp_col < 0);
923 assert(node->list.next == &node->list && node->list.prev == &node->list);
924 assert(bitset_is_set(node->adm_colors, col));
926 list_add_tail(&node->list, changed);
930 static INLINE int is_loose(co_mst_irn_t *node)
932 return !node->fixed && node->tmp_col < 0;
936 * Determines the costs for each color if it would be assigned to node @p node.
938 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) {
939 int *neigh_cols = alloca(env->n_regs * sizeof(*neigh_cols));
944 for (i = 0; i < env->n_regs; ++i) {
947 costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0);
950 for (i = 0; i < node->n_neighs; ++i) {
951 co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]);
952 int col = get_mst_irn_col(n);
957 costs[col].cost = REAL(0.0);
961 coeff = REAL(1.0) / n_loose;
962 for (i = 0; i < env->n_regs; ++i)
963 costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i];
967 /* need forward declaration due to recursive call */
968 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones, int depth, int *max_depth, int *trip);
971 * Tries to change node to a color but @p explude_col.
972 * @return 1 if succeeded, 0 otherwise.
974 static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, struct list_head *changed, int depth, int *max_depth, int *trip) {
975 int col = get_mst_irn_col(node);
978 /* neighbours has already a different color -> good, temporary fix it */
979 if (col != exclude_col) {
981 set_temp_color(node, col, changed);
985 /* The node has the color it should not have _and_ has not been visited yet. */
986 if (is_loose(node)) {
987 col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0]));
989 /* Get the costs for giving the node a specific color. */
990 determine_color_costs(env, node, costs);
992 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
993 costs[exclude_col].cost = REAL(0.0);
995 /* sort the colors according costs, cheapest first. */
996 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt);
998 /* Try recoloring the node using the color list. */
999 res = recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip);
1006 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
1007 * ATTENTION: Expect @p costs already sorted by increasing costs.
1008 * @return 1 if coloring could be applied, 0 otherwise.
1010 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed, int depth, int *max_depth, int *trip) {
1012 struct list_head local_changed;
1015 if (depth > *max_depth)
1018 DBG((dbg, LEVEL_4, "\tRecoloring %+F with color-costs", node->irn));
1019 DBG_COL_COST(env, LEVEL_4, costs);
1020 DB((dbg, LEVEL_4, "\n"));
1022 if (depth >= recolor_limit) {
1023 DBG((dbg, LEVEL_4, "\tHit recolor limit\n"));
1027 for (i = 0; i < env->n_regs; ++i) {
1028 int tgt_col = costs[i].col;
1032 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
1033 if (costs[i].cost == REAL(0.0)) {
1034 DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n"));
1038 /* Set the new color of the node and mark the node as temporarily fixed. */
1039 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
1040 INIT_LIST_HEAD(&local_changed);
1041 set_temp_color(node, tgt_col, &local_changed);
1042 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
1044 /* try to color all interfering neighbours with current color forbidden */
1045 for (j = 0; j < node->n_neighs; ++j) {
1049 neigh = node->int_neighs[j];
1051 /* skip ignore nodes */
1052 if (arch_irn_is(env->aenv, neigh, ignore))
1055 nn = get_co_mst_irn(env, neigh);
1056 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
1057 neigh, j, nn->fixed, nn->tmp_col, nn->col));
1060 Try to change the color of the neighbor and record all nodes which
1061 get changed in the tmp list. Add this list to the "changed" list for
1062 that color. If we did not succeed to change the color of the neighbor,
1063 we bail out and try the next color.
1065 if (get_mst_irn_col(nn) == tgt_col) {
1066 /* try to color neighbour with tgt_col forbidden */
1067 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip);
1075 We managed to assign the target color to all neighbors, so from the perspective
1076 of the current node, every thing was ok and we can return safely.
1079 /* append the local_changed ones to global ones */
1080 list_splice(&local_changed, changed);
1084 /* coloring of neighbours failed, so we try next color */
1085 reject_coloring(&local_changed);
1089 DBG((dbg, LEVEL_4, "\tAll colors failed\n"));
1094 * Tries to bring node @p node and all it's neighbours to color @p tgt_col.
1095 * @return 1 if color @p col could be applied, 0 otherwise
1097 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed) {
1098 int col = get_mst_irn_col(node);
1100 /* if node already has the target color -> good, temporary fix it */
1101 if (col == tgt_col) {
1102 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1104 set_temp_color(node, tgt_col, changed);
1109 Node has not yet a fixed color and target color is admissible
1110 -> try to recolor node and it's affinity neighbours
1112 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1113 col_cost_t *costs = env->single_cols[tgt_col];
1114 int res, max_depth, trip;
1119 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1120 res = recolor_nodes(env, node, costs, changed, 0, &max_depth, &trip);
1121 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1122 stat_ev_int("heur4_recolor_depth_max", max_depth);
1123 stat_ev_int("heur4_recolor_trip", trip);
1129 #ifdef DEBUG_libfirm
1130 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1131 if (!is_loose(node))
1132 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1134 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1135 dbg_admissible_colors(env, node);
1136 DB((dbg, LEVEL_4, ")\n"));
1145 * Tries to color an affinity chunk (or at least a part of it).
1146 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1148 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) {
1149 aff_chunk_t *best_chunk = NULL;
1150 int n_nodes = ARR_LEN(c->n);
1151 int best_color = -1;
1152 int n_int_chunks = 0;
1153 waitq *tmp_chunks = new_waitq();
1154 waitq *best_starts = NULL;
1155 col_cost_t *order = alloca(env->n_regs * sizeof(order[0]));
1157 int idx, len, i, nidx, pos;
1158 struct list_head changed;
1160 DB((dbg, LEVEL_2, "fragmentizing chunk #%u", c->id));
1161 DBG_AFF_CHUNK(env, LEVEL_2, c);
1162 DB((dbg, LEVEL_2, "\n"));
1164 stat_ev_ctx_push_fmt("heur4_color_chunk", "%u", c->id);
1166 ++env->chunk_visited;
1168 /* compute color preference */
1169 memset(order, 0, env->n_regs * sizeof(order[0]));
1171 for (pos = 0, len = ARR_LEN(c->interfere); pos < len; ++pos) {
1172 const ir_node *n = c->interfere[pos];
1173 co_mst_irn_t *node = get_co_mst_irn(env, n);
1174 aff_chunk_t *chunk = node->chunk;
1176 if (is_loose(node) && chunk && chunk->visited < env->chunk_visited) {
1177 assert(!chunk->deleted);
1178 chunk->visited = env->chunk_visited;
1181 aff_chunk_assure_weight(env, chunk);
1182 for (i = 0; i < env->n_regs; ++i)
1183 order[i].cost += chunk->color_affinity[i].cost;
1187 for (i = 0; i < env->n_regs; ++i) {
1188 real_t dislike = n_int_chunks > 0 ? REAL(1.0) - order[i].cost / n_int_chunks : REAL(0.0);
1190 order[i].cost = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike;
1193 qsort(order, env->n_regs, sizeof(order[0]), cmp_col_cost_gt);
1195 DBG_COL_COST(env, LEVEL_2, order);
1196 DB((dbg, LEVEL_2, "\n"));
1198 /* check which color is the "best" for the given chunk.
1199 * if we found a color which was ok for all nodes, we take it
1200 * and do not look further. (see did_all flag usage below.)
1201 * If we have many colors which fit all nodes it is hard to decide
1202 * which one to take anyway.
1203 * TODO Sebastian: Perhaps we should at all nodes and figure out
1204 * a suitable color using costs as done above (determine_color_costs).
1206 for (i = 0; i < env->k; ++i) {
1207 int col = order[i].col;
1208 waitq *good_starts = new_waitq();
1209 aff_chunk_t *local_best;
1212 /* skip ignore colors */
1213 if (bitset_is_set(env->ignore_regs, col))
1216 DB((dbg, LEVEL_2, "\ttrying color %d\n", col));
1220 /* try to bring all nodes of given chunk to the current color. */
1221 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1222 const ir_node *irn = c->n[idx];
1223 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1226 assert(! node->fixed && "Node must not have a fixed color.");
1227 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1230 The order of the colored nodes is important, so we record the successfully
1231 colored ones in the order they appeared.
1233 INIT_LIST_HEAD(&changed);
1235 good = change_node_color(env, node, col, &changed);
1236 stat_ev_tim_pop("heur4_recolor");
1238 waitq_put(good_starts, node);
1239 materialize_coloring(&changed);
1244 reject_coloring(&changed);
1246 n_succeeded += good;
1247 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, good ? "succeeded" : "failed"));
1250 /* unfix all nodes */
1251 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1252 co_mst_irn_t *node = get_co_mst_irn(env, c->n[idx]);
1256 /* try next color when failed */
1257 if (n_succeeded == 0)
1260 /* fragment the chunk according to the coloring */
1261 local_best = fragment_chunk(env, col, c, tmp_chunks);
1263 /* search the best of the good list
1264 and make it the new best if it is better than the current */
1266 aff_chunk_assure_weight(env, local_best);
1268 DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %u) for color %d: ", local_best->id, col));
1269 DBG_AFF_CHUNK(env, LEVEL_3, local_best);
1271 if (! best_chunk || best_chunk->weight < local_best->weight) {
1272 best_chunk = local_best;
1275 del_waitq(best_starts);
1276 best_starts = good_starts;
1277 DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %u), color %d\n", best_chunk->id, best_color));
1279 DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n"));
1280 del_waitq(good_starts);
1284 del_waitq(good_starts);
1287 /* if all nodes were recolored, bail out */
1288 if (n_succeeded == n_nodes)
1292 stat_ev_int("heur4_colors_tried", i);
1294 /* free all intermediate created chunks except best one */
1295 while (! waitq_empty(tmp_chunks)) {
1296 aff_chunk_t *tmp = waitq_get(tmp_chunks);
1297 if (tmp != best_chunk)
1298 delete_aff_chunk(env, tmp);
1300 del_waitq(tmp_chunks);
1302 /* return if coloring failed */
1305 del_waitq(best_starts);
1309 DB((dbg, LEVEL_2, "\tbest chunk #%u ", best_chunk->id));
1310 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1311 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1313 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1314 const ir_node *irn = best_chunk->n[idx];
1315 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1318 /* bring the node to the color. */
1319 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id));
1320 INIT_LIST_HEAD(&changed);
1322 res = change_node_color(env, node, best_color, &changed);
1323 stat_ev_tim_pop("heur4_recolor");
1325 materialize_coloring(&changed);
1328 assert(list_empty(&changed));
1331 /* remove the nodes in best chunk from original chunk */
1332 len = ARR_LEN(best_chunk->n);
1333 for (idx = 0; idx < len; ++idx) {
1334 const ir_node *irn = best_chunk->n[idx];
1335 int pos = nodes_bsearch(c->n, irn);
1340 len = ARR_LEN(c->n);
1341 for (idx = nidx = 0; idx < len; ++idx) {
1342 const ir_node *irn = c->n[idx];
1348 ARR_SHRINKLEN(c->n, nidx);
1351 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1352 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1353 const ir_node *n = c->n[idx];
1354 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1358 /* fragment the remaining chunk */
1359 visited = bitset_irg_malloc(env->co->irg);
1360 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx)
1361 bitset_set(visited, get_irn_idx(best_chunk->n[idx]));
1363 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1364 const ir_node *irn = c->n[idx];
1365 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1366 aff_chunk_t *new_chunk = new_aff_chunk(env);
1367 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1369 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1370 aff_chunk_assure_weight(env, new_chunk);
1371 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1375 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1376 const ir_node *n = best_chunk->n[idx];
1377 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1381 /* clear obsolete chunks and free some memory */
1382 delete_aff_chunk(env, best_chunk);
1383 bitset_free(visited);
1385 del_waitq(best_starts);
1387 stat_ev_ctx_pop("heur4_color_chunk");
1391 * Main driver for mst safe coalescing algorithm.
1393 int co_solve_heuristic_mst(copy_opt_t *co) {
1394 unsigned n_regs = co->cls->n_regs;
1395 bitset_t *ignore_regs = bitset_alloca(n_regs);
1398 co_mst_env_t mst_env;
1405 phase_init(&mst_env.ph, "co_mst", co->irg, PHASE_DEFAULT_GROWTH, co_mst_irn_init, &mst_env);
1407 k = be_put_ignore_regs(co->cenv->birg, co->cls, ignore_regs);
1410 mst_env.n_regs = n_regs;
1412 mst_env.chunks = new_pqueue();
1414 mst_env.ignore_regs = ignore_regs;
1415 mst_env.ifg = co->cenv->ifg;
1416 mst_env.aenv = co->aenv;
1417 mst_env.chunkset = pset_new_ptr(512);
1418 mst_env.chunk_visited = 0;
1419 mst_env.single_cols = phase_alloc(&mst_env.ph, sizeof(*mst_env.single_cols) * n_regs);
1421 for (i = 0; i < n_regs; ++i) {
1422 col_cost_t *vec = phase_alloc(&mst_env.ph, sizeof(*vec) * n_regs);
1424 mst_env.single_cols[i] = vec;
1425 for (j = 0; j < n_regs; ++j) {
1427 vec[j].cost = REAL(0.0);
1431 vec[0].cost = REAL(1.0);
1434 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1436 /* build affinity chunks */
1438 build_affinity_chunks(&mst_env);
1439 stat_ev_tim_pop("heur4_initial_chunk");
1441 /* color chunks as long as there are some */
1442 while (! pqueue_empty(mst_env.chunks)) {
1443 aff_chunk_t *chunk = pqueue_pop_front(mst_env.chunks);
1445 color_aff_chunk(&mst_env, chunk);
1446 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id));
1447 delete_aff_chunk(&mst_env, chunk);
1450 /* apply coloring */
1451 foreach_phase_irn(&mst_env.ph, irn) {
1453 const arch_register_t *reg;
1455 if (arch_irn_is(mst_env.aenv, irn, ignore))
1458 mirn = get_co_mst_irn(&mst_env, irn);
1459 // assert(mirn->fixed && "Node should have fixed color");
1461 /* skip nodes where color hasn't changed */
1462 if (mirn->init_col == mirn->col)
1465 reg = arch_register_for_index(co->cls, mirn->col);
1466 arch_set_irn_register(co->aenv, irn, reg);
1467 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1470 /* free allocated memory */
1471 del_pqueue(mst_env.chunks);
1472 phase_free(&mst_env.ph);
1473 del_pset(mst_env.chunkset);
1475 stat_ev_tim_pop("heur4_total");
1480 static const lc_opt_table_entry_t options[] = {
1481 LC_OPT_ENT_INT ("limit", "limit recoloring", &recolor_limit),
1482 LC_OPT_ENT_DBL ("di", "dislike influence", &dislike_influence),
1487 void be_init_copyheur4(void) {
1488 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1489 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
1490 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
1491 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
1492 lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4");
1494 lc_opt_add_table(heur4_grp, options);
1495 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");
1499 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);