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 #define DISABLE_STATEV
41 #include "raw_bitset.h"
42 #include "irphase_t.h"
58 #include "becopyopt_t.h"
62 #define COL_COST_INFEASIBLE DBL_MAX
63 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
64 #define NEIGHBOUR_CONSTR_COSTS 64.0
69 #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while (0)
70 #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while (0)
72 static firm_dbg_module_t *dbg = NULL;
76 #define DBG_AFF_CHUNK(env, level, chunk)
77 #define DBG_COL_COST(env, level, cost)
82 #define REAL(C) (C ## f)
84 static unsigned last_chunk_id = 0;
85 static int recolor_limit = 7;
86 static real_t dislike_influence = REAL(0.1);
88 typedef struct _col_cost_t {
96 typedef struct _aff_chunk_t {
97 const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
98 const ir_node **interfere; /**< An ARR_F containing all inference. */
99 int weight; /**< Weight of this chunk */
100 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
101 unsigned deleted : 1; /**< For debugging: Set if the was deleted. */
102 unsigned id; /**< An id of this chunk. */
104 col_cost_t color_affinity[1];
110 typedef struct _aff_edge_t {
111 const ir_node *src; /**< Source node. */
112 const ir_node *tgt; /**< Target node. */
113 int weight; /**< The weight of this edge. */
116 /* main coalescing environment */
117 typedef struct _co_mst_env_t {
118 int n_regs; /**< number of regs in class */
119 int k; /**< number of non-ignore registers in class */
120 bitset_t *ignore_regs; /**< set containing all global ignore registers */
121 ir_phase ph; /**< phase object holding data for nodes */
122 pqueue_t *chunks; /**< priority queue for chunks */
123 pset *chunkset; /**< set holding all chunks */
124 be_ifg_t *ifg; /**< the interference graph */
125 copy_opt_t *co; /**< the copy opt object */
126 unsigned chunk_visited;
127 col_cost_t **single_cols;
130 /* stores coalescing related information for a node */
131 typedef struct _co_mst_irn_t {
132 const ir_node *irn; /**< the irn this information belongs to */
133 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
134 bitset_t *adm_colors; /**< set of admissible colors for this irn */
135 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
136 int n_neighs; /**< length of the interfering neighbours array. */
137 int int_aff_neigh; /**< number of interfering affinity neighbours */
138 int col; /**< color currently assigned */
139 int init_col; /**< the initial color */
140 int tmp_col; /**< a temporary assigned color */
141 unsigned fixed : 1; /**< the color is fixed */
142 struct list_head list; /**< Queue for coloring undo. */
143 real_t constr_factor;
146 #define get_co_mst_irn(mst_env, irn) (phase_get_or_set_irn_data(&(mst_env)->ph, (irn)))
148 typedef int decide_func_t(const co_mst_irn_t *node, int col);
153 * Write a chunk to stderr for debugging.
155 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c)
159 if (c->weight_consistent)
160 ir_fprintf(stderr, " $%d ", c->weight);
161 ir_fprintf(stderr, "{");
162 for (i = 0, l = ARR_LEN(c->n); i < l; ++i) {
163 const ir_node *n = c->n[i];
164 ir_fprintf(stderr, " %+F,", n);
166 ir_fprintf(stderr, "}");
170 * Dump all admissible colors to stderr.
172 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node)
177 if (bitset_popcount(node->adm_colors) < 1)
178 fprintf(stderr, "no admissible colors?!?");
180 bitset_foreach(node->adm_colors, idx) {
181 fprintf(stderr, " %d", idx);
187 * Dump color-cost pairs to stderr.
189 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)
200 return node->tmp_col >= 0 ? node->tmp_col : node->col;
204 * @return 1 if node @p node has color @p col, 0 otherwise.
206 static int decider_has_color(const co_mst_irn_t *node, int col)
208 return get_mst_irn_col(node) == col;
212 * @return 1 if node @p node has not color @p col, 0 otherwise.
214 static int decider_hasnot_color(const co_mst_irn_t *node, int col)
216 return get_mst_irn_col(node) != col;
220 * Always returns true.
222 static int decider_always_yes(const co_mst_irn_t *node, int col)
229 /** compares two affinity edges by its weight */
230 static int cmp_aff_edge(const void *a, const void *b)
232 const aff_edge_t *e1 = a;
233 const aff_edge_t *e2 = b;
235 if (e2->weight == e1->weight) {
236 if (e2->src->node_idx == e1->src->node_idx)
237 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
239 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
241 /* sort in descending order */
242 return QSORT_CMP(e2->weight, e1->weight);
245 /** compares to color-cost pairs */
246 static __attribute__((unused)) int cmp_col_cost_lt(const void *a, const void *b)
248 const col_cost_t *c1 = a;
249 const col_cost_t *c2 = b;
250 real_t diff = c1->cost - c2->cost;
251 return (diff > 0) - (diff < 0);
254 static int cmp_col_cost_gt(const void *a, const void *b)
256 const col_cost_t *c1 = a;
257 const col_cost_t *c2 = b;
258 real_t diff = c2->cost - c1->cost;
259 return (diff > 0) - (diff < 0);
263 * Creates a new affinity chunk
265 static inline aff_chunk_t *new_aff_chunk(co_mst_env_t *env)
267 aff_chunk_t *c = XMALLOCF(aff_chunk_t, color_affinity, env->n_regs);
268 c->n = NEW_ARR_F(const ir_node *, 0);
269 c->interfere = NEW_ARR_F(const ir_node *, 0);
271 c->weight_consistent = 0;
273 c->id = ++last_chunk_id;
275 pset_insert(env->chunkset, c, c->id);
280 * Frees all memory allocated by an affinity chunk.
282 static inline void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c)
284 pset_remove(env->chunkset, c, c->id);
285 DEL_ARR_F(c->interfere);
292 * binary search of sorted nodes.
294 * @return the position where n is found in the array arr or ~pos
295 * if the nodes is not here.
297 static inline int nodes_bsearch(const ir_node **arr, const ir_node *n)
299 int hi = ARR_LEN(arr);
303 int md = lo + ((hi - lo) >> 1);
316 /** Check if a node n can be found inside arr. */
317 static int node_contains(const ir_node **arr, const ir_node *n)
319 int i = nodes_bsearch(arr, n);
324 * Insert a node into the sorted nodes list.
326 * @return 1 if the node was inserted, 0 else
328 static int nodes_insert(const ir_node ***arr, const ir_node *irn)
330 int idx = nodes_bsearch(*arr, irn);
333 int i, n = ARR_LEN(*arr);
336 ARR_APP1(const ir_node *, *arr, irn);
341 for (i = n - 1; i >= idx; --i)
350 * Adds a node to an affinity chunk
352 static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node)
356 if (! nodes_insert(&c->n, node->irn))
359 c->weight_consistent = 0;
362 for (i = node->n_neighs - 1; i >= 0; --i) {
363 ir_node *neigh = node->int_neighs[i];
364 nodes_insert(&c->interfere, neigh);
369 * In case there is no phase information for irn, initialize it.
371 static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn)
373 co_mst_irn_t *res = phase_alloc(ph, sizeof(res[0]));
374 co_mst_env_t *env = ph->priv;
376 const arch_register_req_t *req;
377 neighbours_iter_t nodes_it;
385 res->int_neighs = NULL;
386 res->int_aff_neigh = 0;
387 res->col = arch_register_get_index(arch_get_irn_register(irn));
388 res->init_col = res->col;
389 INIT_LIST_HEAD(&res->list);
391 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
393 /* set admissible registers */
394 res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs);
396 /* Exclude colors not assignable to the irn */
397 req = arch_get_register_req_out(irn);
398 if (arch_register_req_is(req, limited))
399 rbitset_copy_to_bitset(req->limited, res->adm_colors);
401 bitset_set_all(res->adm_colors);
403 /* exclude global ignore registers as well */
404 bitset_andnot(res->adm_colors, env->ignore_regs);
406 /* compute the constraint factor */
407 res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcount(res->adm_colors)) / env->n_regs;
409 /* set the number of interfering affinity neighbours to -1, they are calculated later */
410 res->int_aff_neigh = -1;
412 /* build list of interfering neighbours */
414 be_ifg_foreach_neighbour(env->ifg, &nodes_it, irn, neigh) {
415 if (!arch_irn_is_ignore(neigh)) {
416 obstack_ptr_grow(phase_obst(ph), neigh);
420 res->int_neighs = obstack_finish(phase_obst(ph));
426 * Check if affinity chunk @p chunk interferes with node @p irn.
428 static inline int aff_chunk_interferes(const aff_chunk_t *chunk, const ir_node *irn)
430 return node_contains(chunk->interfere, irn);
434 * Check if there are interference edges from c1 to c2.
436 * @param c2 Another chunk
437 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
439 static inline int aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2)
446 /* check if there is a node in c2 having an interfering neighbor in c1 */
447 for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) {
448 const ir_node *irn = c2->n[i];
450 if (node_contains(c1->interfere, irn))
457 * Returns the affinity chunk of @p irn or creates a new
458 * one with @p irn as element if there is none assigned.
460 static inline aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn)
462 co_mst_irn_t *node = get_co_mst_irn(env, irn);
467 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
468 * are no interference edges from chunk(src) to chunk(tgt)).
469 * @return 1 if successful, 0 if not possible
471 static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node *tgt)
473 aff_chunk_t *c1 = get_aff_chunk(env, src);
474 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
477 DB((dbg, LEVEL_4, "Attempt to let c1 (id %u): ", c1 ? c1->id : 0));
479 DBG_AFF_CHUNK(env, LEVEL_4, c1);
481 DB((dbg, LEVEL_4, "{%+F}", src));
483 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %u): ", c2 ? c2->id : 0));
485 DBG_AFF_CHUNK(env, LEVEL_4, c2);
487 DB((dbg, LEVEL_4, "{%+F}", tgt));
489 DB((dbg, LEVEL_4, "\n"));
494 /* no chunk exists */
495 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
498 for (i = mirn->n_neighs - 1; i >= 0; --i) {
499 if (mirn->int_neighs[i] == tgt)
503 /* create one containing both nodes */
504 c1 = new_aff_chunk(env);
505 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
506 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
510 /* c2 already exists */
511 if (! aff_chunk_interferes(c2, src)) {
512 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
516 } else if (c2 == NULL) {
517 /* c1 already exists */
518 if (! aff_chunk_interferes(c1, tgt)) {
519 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
522 } else if (c1 != c2 && ! aff_chunks_interfere(c1, c2)) {
525 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx)
526 aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx]));
528 for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) {
529 const ir_node *irn = c2->interfere[idx];
530 nodes_insert(&c1->interfere, irn);
533 c1->weight_consistent = 0;
535 delete_aff_chunk(env, c2);
538 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
542 DB((dbg, LEVEL_4, " ... absorbed\n"));
547 * Assures that the weight of the given chunk is consistent.
549 static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c)
551 if (! c->weight_consistent) {
555 for (i = 0; i < env->n_regs; ++i) {
556 c->color_affinity[i].col = i;
557 c->color_affinity[i].cost = REAL(0.0);
560 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
561 const ir_node *n = c->n[idx];
562 const affinity_node_t *an = get_affinity_info(env->co, n);
563 co_mst_irn_t *node = get_co_mst_irn(env, n);
566 if (node->constr_factor > REAL(0.0)) {
568 bitset_foreach (node->adm_colors, col)
569 c->color_affinity[col].cost += node->constr_factor;
574 co_gs_foreach_neighb(an, neigh) {
575 const ir_node *m = neigh->irn;
577 if (arch_irn_is_ignore(m))
580 w += node_contains(c->n, m) ? neigh->costs : 0;
585 for (i = 0; i < env->n_regs; ++i)
586 c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n));
589 // c->weight = bitset_popcount(c->nodes);
590 c->weight_consistent = 1;
595 * Count the number of interfering affinity neighbours
597 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an)
599 const neighb_t *neigh;
600 const ir_node *irn = an->irn;
601 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
604 co_gs_foreach_neighb(an, neigh) {
605 const ir_node *n = neigh->irn;
608 if (arch_irn_is_ignore(n))
611 /* check if the affinity neighbour interfere */
612 for (i = 0; i < node->n_neighs; ++i) {
613 if (node->int_neighs[i] == n) {
624 * Build chunks of nodes connected by affinity edges.
625 * We start at the heaviest affinity edge.
626 * The chunks of the two edge-defining nodes will be
627 * merged if there are no interference edges from one
628 * chunk to the other.
630 static void build_affinity_chunks(co_mst_env_t *env)
632 nodes_iter_t nodes_it;
633 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
636 aff_chunk_t *curr_chunk;
638 /* at first we create the affinity edge objects */
639 be_ifg_foreach_node(env->ifg, &nodes_it, n) {
640 int n_idx = get_irn_idx(n);
644 if (arch_irn_is_ignore(n))
647 n1 = get_co_mst_irn(env, n);
648 an = get_affinity_info(env->co, n);
653 if (n1->int_aff_neigh < 0)
654 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
656 /* build the affinity edges */
657 co_gs_foreach_neighb(an, neigh) {
658 const ir_node *m = neigh->irn;
659 int m_idx = get_irn_idx(m);
661 /* record the edge in only one direction */
666 /* skip ignore nodes */
667 if (arch_irn_is_ignore(m))
673 n2 = get_co_mst_irn(env, m);
674 if (n2->int_aff_neigh < 0) {
675 affinity_node_t *am = get_affinity_info(env->co, m);
676 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
679 * these weights are pure hackery ;-).
680 * It's not chriswue's fault but mine.
682 edge.weight = neigh->costs;
683 ARR_APP1(aff_edge_t, edges, edge);
689 /* now: sort edges and build the affinity chunks */
690 len = ARR_LEN(edges);
691 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
692 for (i = 0; i < len; ++i) {
693 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
695 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
698 /* now insert all chunks into a priority queue */
699 foreach_pset(env->chunkset, curr_chunk) {
700 aff_chunk_assure_weight(env, curr_chunk);
702 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
703 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
704 DBG((dbg, LEVEL_1, "\n"));
706 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
709 foreach_phase_irn(&env->ph, n) {
710 co_mst_irn_t *mirn = get_co_mst_irn(env, n);
712 if (mirn->chunk == NULL) {
713 /* no chunk is allocated so far, do it now */
714 aff_chunk_t *curr_chunk = new_aff_chunk(env);
715 aff_chunk_add_node(curr_chunk, mirn);
717 aff_chunk_assure_weight(env, curr_chunk);
719 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
720 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
721 DBG((dbg, LEVEL_1, "\n"));
723 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
730 static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk)
732 pqueue_t *grow = new_pqueue();
733 const ir_node *max_node = NULL;
737 for (i = ARR_LEN(chunk->n) - 1; i >= 0; i--) {
738 const ir_node *irn = chunk->n[i];
739 affinity_node_t *an = get_affinity_info(env->co, irn);
743 if (arch_irn_is_ignore(irn))
747 co_gs_foreach_neighb(an, neigh)
750 if (w > max_weight) {
758 bitset_t *visited = bitset_irg_malloc(env->co->irg);
760 for (i = ARR_LEN(chunk->n) - 1; i >= 0; --i)
761 bitset_add_irn(visited, chunk->n[i]);
763 pqueue_put(grow, (void *) max_node, max_weight);
764 bitset_remv_irn(visited, max_node);
766 while (!pqueue_empty(grow)) {
767 ir_node *irn = pqueue_pop_front(grow);
768 affinity_node_t *an = get_affinity_info(env->co, irn);
771 if (arch_irn_is_ignore(irn))
774 assert(i <= ARR_LEN(chunk->n));
779 /* build the affinity edges */
780 co_gs_foreach_neighb(an, neigh) {
781 co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn);
783 if (bitset_contains_irn(visited, node->irn)) {
784 pqueue_put(grow, (void *) neigh->irn, neigh->costs);
785 bitset_remv_irn(visited, node->irn);
791 bitset_free(visited);
796 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
798 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
799 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
801 waitq *nodes = new_waitq();
803 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col));
805 /* init queue and chunk */
806 waitq_put(nodes, node);
807 bitset_set(visited, get_irn_idx(node->irn));
808 aff_chunk_add_node(chunk, node);
809 DB((dbg, LEVEL_1, " %+F", node->irn));
811 /* as long as there are nodes in the queue */
812 while (! waitq_empty(nodes)) {
813 co_mst_irn_t *n = waitq_get(nodes);
814 affinity_node_t *an = get_affinity_info(env->co, n->irn);
816 /* check all affinity neighbors */
819 co_gs_foreach_neighb(an, neigh) {
820 const ir_node *m = neigh->irn;
821 int m_idx = get_irn_idx(m);
824 if (arch_irn_is_ignore(m))
827 n2 = get_co_mst_irn(env, m);
829 if (! bitset_is_set(visited, m_idx) &&
832 ! aff_chunk_interferes(chunk, m) &&
833 node_contains(orig_chunk->n, m))
836 following conditions are met:
837 - neighbour is not visited
838 - neighbour likes the color
839 - neighbour has not yet a fixed color
840 - the new chunk doesn't interfere with the neighbour
841 - neighbour belongs or belonged once to the original chunk
843 bitset_set(visited, m_idx);
844 aff_chunk_add_node(chunk, n2);
845 DB((dbg, LEVEL_1, " %+F", n2->irn));
846 /* enqueue for further search */
847 waitq_put(nodes, n2);
853 DB((dbg, LEVEL_1, "\n"));
859 * Fragment the given chunk into chunks having given color and not having given color.
861 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp)
863 bitset_t *visited = bitset_irg_malloc(env->co->irg);
865 aff_chunk_t *best = NULL;
867 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
870 aff_chunk_t *tmp_chunk;
871 decide_func_t *decider;
875 if (bitset_is_set(visited, get_irn_idx(irn)))
878 node = get_co_mst_irn(env, irn);
880 if (get_mst_irn_col(node) == col) {
881 decider = decider_has_color;
883 DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col));
886 decider = decider_hasnot_color;
888 DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col));
891 /* create a new chunk starting at current node */
892 tmp_chunk = new_aff_chunk(env);
893 waitq_put(tmp, tmp_chunk);
894 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
895 assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk");
897 /* remember the local best */
898 aff_chunk_assure_weight(env, tmp_chunk);
899 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
903 assert(best && "No chunk found?");
904 bitset_free(visited);
909 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
910 * ATTENTION: the queue is empty after calling this function!
912 static inline void reject_coloring(struct list_head *nodes)
914 co_mst_irn_t *n, *temp;
915 DB((dbg, LEVEL_4, "\treject coloring for"));
916 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
917 DB((dbg, LEVEL_4, " %+F", n->irn));
918 assert(n->tmp_col >= 0);
920 list_del_init(&n->list);
922 DB((dbg, LEVEL_4, "\n"));
925 static inline void materialize_coloring(struct list_head *nodes)
927 co_mst_irn_t *n, *temp;
928 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
929 assert(n->tmp_col >= 0);
932 list_del_init(&n->list);
936 static inline void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
939 assert(!node->fixed);
940 assert(node->tmp_col < 0);
941 assert(node->list.next == &node->list && node->list.prev == &node->list);
942 assert(bitset_is_set(node->adm_colors, col));
944 list_add_tail(&node->list, changed);
948 static inline int is_loose(co_mst_irn_t *node)
950 return !node->fixed && node->tmp_col < 0;
954 * Determines the costs for each color if it would be assigned to node @p node.
956 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs)
958 int *neigh_cols = ALLOCAN(int, env->n_regs);
963 for (i = 0; i < env->n_regs; ++i) {
966 costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0);
969 for (i = 0; i < node->n_neighs; ++i) {
970 co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]);
971 int col = get_mst_irn_col(n);
976 costs[col].cost = REAL(0.0);
980 coeff = REAL(1.0) / n_loose;
981 for (i = 0; i < env->n_regs; ++i)
982 costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i];
986 /* need forward declaration due to recursive call */
987 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);
990 * Tries to change node to a color but @p explude_col.
991 * @return 1 if succeeded, 0 otherwise.
993 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)
995 int col = get_mst_irn_col(node);
998 /* neighbours has already a different color -> good, temporary fix it */
999 if (col != exclude_col) {
1001 set_temp_color(node, col, changed);
1005 /* The node has the color it should not have _and_ has not been visited yet. */
1006 if (is_loose(node)) {
1007 col_cost_t *costs = ALLOCAN(col_cost_t, env->n_regs);
1009 /* Get the costs for giving the node a specific color. */
1010 determine_color_costs(env, node, costs);
1012 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
1013 costs[exclude_col].cost = REAL(0.0);
1015 /* sort the colors according costs, cheapest first. */
1016 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt);
1018 /* Try recoloring the node using the color list. */
1019 res = recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip);
1026 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
1027 * ATTENTION: Expect @p costs already sorted by increasing costs.
1028 * @return 1 if coloring could be applied, 0 otherwise.
1030 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)
1033 struct list_head local_changed;
1036 if (depth > *max_depth)
1039 DBG((dbg, LEVEL_4, "\tRecoloring %+F with color-costs", node->irn));
1040 DBG_COL_COST(env, LEVEL_4, costs);
1041 DB((dbg, LEVEL_4, "\n"));
1043 if (depth >= recolor_limit) {
1044 DBG((dbg, LEVEL_4, "\tHit recolor limit\n"));
1048 for (i = 0; i < env->n_regs; ++i) {
1049 int tgt_col = costs[i].col;
1053 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
1054 if (costs[i].cost == REAL(0.0)) {
1055 DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n"));
1059 /* Set the new color of the node and mark the node as temporarily fixed. */
1060 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
1061 INIT_LIST_HEAD(&local_changed);
1062 set_temp_color(node, tgt_col, &local_changed);
1063 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
1065 /* try to color all interfering neighbours with current color forbidden */
1066 for (j = 0; j < node->n_neighs; ++j) {
1070 neigh = node->int_neighs[j];
1072 if (arch_irn_is_ignore(neigh))
1075 nn = get_co_mst_irn(env, neigh);
1076 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
1077 neigh, j, nn->fixed, nn->tmp_col, nn->col));
1080 Try to change the color of the neighbor and record all nodes which
1081 get changed in the tmp list. Add this list to the "changed" list for
1082 that color. If we did not succeed to change the color of the neighbor,
1083 we bail out and try the next color.
1085 if (get_mst_irn_col(nn) == tgt_col) {
1086 /* try to color neighbour with tgt_col forbidden */
1087 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip);
1095 We managed to assign the target color to all neighbors, so from the perspective
1096 of the current node, every thing was ok and we can return safely.
1099 /* append the local_changed ones to global ones */
1100 list_splice(&local_changed, changed);
1104 /* coloring of neighbours failed, so we try next color */
1105 reject_coloring(&local_changed);
1109 DBG((dbg, LEVEL_4, "\tAll colors failed\n"));
1114 * Tries to bring node @p node and all it's neighbours to color @p tgt_col.
1115 * @return 1 if color @p col could be applied, 0 otherwise
1117 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed)
1119 int col = get_mst_irn_col(node);
1121 /* if node already has the target color -> good, temporary fix it */
1122 if (col == tgt_col) {
1123 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1125 set_temp_color(node, tgt_col, changed);
1130 Node has not yet a fixed color and target color is admissible
1131 -> try to recolor node and it's affinity neighbours
1133 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1134 col_cost_t *costs = env->single_cols[tgt_col];
1135 int res, max_depth, trip;
1140 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1141 res = recolor_nodes(env, node, costs, changed, 0, &max_depth, &trip);
1142 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1143 stat_ev_int("heur4_recolor_depth_max", max_depth);
1144 stat_ev_int("heur4_recolor_trip", trip);
1150 #ifdef DEBUG_libfirm
1151 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1152 if (!is_loose(node))
1153 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1155 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1156 dbg_admissible_colors(env, node);
1157 DB((dbg, LEVEL_4, ")\n"));
1166 * Tries to color an affinity chunk (or at least a part of it).
1167 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1169 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c)
1171 aff_chunk_t *best_chunk = NULL;
1172 int n_nodes = ARR_LEN(c->n);
1173 int best_color = -1;
1174 int n_int_chunks = 0;
1175 waitq *tmp_chunks = new_waitq();
1176 waitq *best_starts = NULL;
1177 col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs);
1179 int idx, len, i, nidx, pos;
1180 struct list_head changed;
1182 DB((dbg, LEVEL_2, "fragmentizing chunk #%u", c->id));
1183 DBG_AFF_CHUNK(env, LEVEL_2, c);
1184 DB((dbg, LEVEL_2, "\n"));
1186 stat_ev_ctx_push_fmt("heur4_color_chunk", "%u", c->id);
1188 ++env->chunk_visited;
1190 /* compute color preference */
1191 for (pos = 0, len = ARR_LEN(c->interfere); pos < len; ++pos) {
1192 const ir_node *n = c->interfere[pos];
1193 co_mst_irn_t *node = get_co_mst_irn(env, n);
1194 aff_chunk_t *chunk = node->chunk;
1196 if (is_loose(node) && chunk && chunk->visited < env->chunk_visited) {
1197 assert(!chunk->deleted);
1198 chunk->visited = env->chunk_visited;
1201 aff_chunk_assure_weight(env, chunk);
1202 for (i = 0; i < env->n_regs; ++i)
1203 order[i].cost += chunk->color_affinity[i].cost;
1207 for (i = 0; i < env->n_regs; ++i) {
1208 real_t dislike = n_int_chunks > 0 ? REAL(1.0) - order[i].cost / n_int_chunks : REAL(0.0);
1210 order[i].cost = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike;
1213 qsort(order, env->n_regs, sizeof(order[0]), cmp_col_cost_gt);
1215 DBG_COL_COST(env, LEVEL_2, order);
1216 DB((dbg, LEVEL_2, "\n"));
1218 /* check which color is the "best" for the given chunk.
1219 * if we found a color which was ok for all nodes, we take it
1220 * and do not look further. (see did_all flag usage below.)
1221 * If we have many colors which fit all nodes it is hard to decide
1222 * which one to take anyway.
1223 * TODO Sebastian: Perhaps we should at all nodes and figure out
1224 * a suitable color using costs as done above (determine_color_costs).
1226 for (i = 0; i < env->k; ++i) {
1227 int col = order[i].col;
1228 waitq *good_starts = new_waitq();
1229 aff_chunk_t *local_best;
1232 /* skip ignore colors */
1233 if (bitset_is_set(env->ignore_regs, col))
1236 DB((dbg, LEVEL_2, "\ttrying color %d\n", col));
1240 /* try to bring all nodes of given chunk to the current color. */
1241 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1242 const ir_node *irn = c->n[idx];
1243 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1246 assert(! node->fixed && "Node must not have a fixed color.");
1247 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1250 The order of the colored nodes is important, so we record the successfully
1251 colored ones in the order they appeared.
1253 INIT_LIST_HEAD(&changed);
1255 good = change_node_color(env, node, col, &changed);
1256 stat_ev_tim_pop("heur4_recolor");
1258 waitq_put(good_starts, node);
1259 materialize_coloring(&changed);
1264 reject_coloring(&changed);
1266 n_succeeded += good;
1267 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, good ? "succeeded" : "failed"));
1270 /* unfix all nodes */
1271 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1272 co_mst_irn_t *node = get_co_mst_irn(env, c->n[idx]);
1276 /* try next color when failed */
1277 if (n_succeeded == 0)
1280 /* fragment the chunk according to the coloring */
1281 local_best = fragment_chunk(env, col, c, tmp_chunks);
1283 /* search the best of the good list
1284 and make it the new best if it is better than the current */
1286 aff_chunk_assure_weight(env, local_best);
1288 DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %u) for color %d: ", local_best->id, col));
1289 DBG_AFF_CHUNK(env, LEVEL_3, local_best);
1291 if (! best_chunk || best_chunk->weight < local_best->weight) {
1292 best_chunk = local_best;
1295 del_waitq(best_starts);
1296 best_starts = good_starts;
1297 DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %u), color %d\n", best_chunk->id, best_color));
1299 DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n"));
1300 del_waitq(good_starts);
1304 del_waitq(good_starts);
1307 /* if all nodes were recolored, bail out */
1308 if (n_succeeded == n_nodes)
1312 stat_ev_int("heur4_colors_tried", i);
1314 /* free all intermediate created chunks except best one */
1315 while (! waitq_empty(tmp_chunks)) {
1316 aff_chunk_t *tmp = waitq_get(tmp_chunks);
1317 if (tmp != best_chunk)
1318 delete_aff_chunk(env, tmp);
1320 del_waitq(tmp_chunks);
1322 /* return if coloring failed */
1325 del_waitq(best_starts);
1329 DB((dbg, LEVEL_2, "\tbest chunk #%u ", best_chunk->id));
1330 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1331 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1333 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1334 const ir_node *irn = best_chunk->n[idx];
1335 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1338 /* bring the node to the color. */
1339 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id));
1340 INIT_LIST_HEAD(&changed);
1342 res = change_node_color(env, node, best_color, &changed);
1343 stat_ev_tim_pop("heur4_recolor");
1345 materialize_coloring(&changed);
1348 assert(list_empty(&changed));
1351 /* remove the nodes in best chunk from original chunk */
1352 len = ARR_LEN(best_chunk->n);
1353 for (idx = 0; idx < len; ++idx) {
1354 const ir_node *irn = best_chunk->n[idx];
1355 int pos = nodes_bsearch(c->n, irn);
1360 len = ARR_LEN(c->n);
1361 for (idx = nidx = 0; idx < len; ++idx) {
1362 const ir_node *irn = c->n[idx];
1368 ARR_SHRINKLEN(c->n, nidx);
1371 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1372 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1373 const ir_node *n = c->n[idx];
1374 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1378 /* fragment the remaining chunk */
1379 visited = bitset_irg_malloc(env->co->irg);
1380 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx)
1381 bitset_set(visited, get_irn_idx(best_chunk->n[idx]));
1383 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1384 const ir_node *irn = c->n[idx];
1385 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1386 aff_chunk_t *new_chunk = new_aff_chunk(env);
1387 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1389 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1390 aff_chunk_assure_weight(env, new_chunk);
1391 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1395 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1396 const ir_node *n = best_chunk->n[idx];
1397 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1401 /* clear obsolete chunks and free some memory */
1402 delete_aff_chunk(env, best_chunk);
1403 bitset_free(visited);
1405 del_waitq(best_starts);
1407 stat_ev_ctx_pop("heur4_color_chunk");
1411 * Main driver for mst safe coalescing algorithm.
1413 static int co_solve_heuristic_mst(copy_opt_t *co)
1415 unsigned n_regs = co->cls->n_regs;
1416 bitset_t *ignore_regs = bitset_alloca(n_regs);
1419 co_mst_env_t mst_env;
1426 phase_init(&mst_env.ph, co->irg, co_mst_irn_init);
1427 phase_set_private(&mst_env.ph, &mst_env);
1429 k = be_put_ignore_regs(co->cenv->irg, co->cls, ignore_regs);
1432 mst_env.n_regs = n_regs;
1434 mst_env.chunks = new_pqueue();
1436 mst_env.ignore_regs = ignore_regs;
1437 mst_env.ifg = co->cenv->ifg;
1438 mst_env.chunkset = pset_new_ptr(512);
1439 mst_env.chunk_visited = 0;
1440 mst_env.single_cols = phase_alloc(&mst_env.ph, sizeof(*mst_env.single_cols) * n_regs);
1442 for (i = 0; i < n_regs; ++i) {
1443 col_cost_t *vec = phase_alloc(&mst_env.ph, sizeof(*vec) * n_regs);
1445 mst_env.single_cols[i] = vec;
1446 for (j = 0; j < n_regs; ++j) {
1448 vec[j].cost = REAL(0.0);
1452 vec[0].cost = REAL(1.0);
1455 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1457 /* build affinity chunks */
1459 build_affinity_chunks(&mst_env);
1460 stat_ev_tim_pop("heur4_initial_chunk");
1462 /* color chunks as long as there are some */
1463 while (! pqueue_empty(mst_env.chunks)) {
1464 aff_chunk_t *chunk = pqueue_pop_front(mst_env.chunks);
1466 color_aff_chunk(&mst_env, chunk);
1467 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id));
1468 delete_aff_chunk(&mst_env, chunk);
1471 /* apply coloring */
1472 foreach_phase_irn(&mst_env.ph, irn) {
1474 const arch_register_t *reg;
1476 if (arch_irn_is_ignore(irn))
1479 mirn = get_co_mst_irn(&mst_env, irn);
1480 // assert(mirn->fixed && "Node should have fixed color");
1482 /* skip nodes where color hasn't changed */
1483 if (mirn->init_col == mirn->col)
1486 reg = arch_register_for_index(co->cls, mirn->col);
1487 arch_set_irn_register(irn, reg);
1488 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1491 /* free allocated memory */
1492 del_pqueue(mst_env.chunks);
1493 phase_deinit(&mst_env.ph);
1494 del_pset(mst_env.chunkset);
1496 stat_ev_tim_pop("heur4_total");
1501 static const lc_opt_table_entry_t options[] = {
1502 LC_OPT_ENT_INT ("limit", "limit recoloring", &recolor_limit),
1503 LC_OPT_ENT_DBL ("di", "dislike influence", &dislike_influence),
1507 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);
1508 void be_init_copyheur4(void)
1510 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1511 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
1512 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
1513 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
1514 lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4");
1516 static co_algo_info copyheur = {
1517 co_solve_heuristic_mst, 0
1520 lc_opt_add_table(heur4_grp, options);
1521 be_register_copyopt("heur4", ©heur);
1523 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");