4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
7 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
8 * examined. A qnode has a 'target color', nodes out of the opt unit and
9 * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges'
10 * A 'max indep set' is determined form these. We try to color this mis using a
11 * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges'
12 * and the qnode is reinserted in the queue. The first qnode colored without
13 * conflicts is the best one.
28 #include "becopyopt_t.h"
29 #include "becopystat.h"
32 static firm_dbg_module_t *dbg = NULL;
34 #define SEARCH_FREE_COLORS
36 #define SLOTS_PINNED_GLOBAL 64
37 #define SLOTS_CONFLICTS 8
38 #define SLOTS_CHANGED_NODES 32
40 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
41 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
44 * Modeling additional conflicts between nodes. NOT live range interference
46 typedef struct _conflict_t {
47 const ir_node *n1, *n2;
51 * If an irn is changed, the changes first get stored in a node_stat_t,
52 * to allow undo of changes (=drop new data) in case of conflicts.
54 typedef struct _node_stat_t {
61 * Represents a node in the optimization queue.
63 typedef struct _qnode_t {
64 struct list_head queue; /**< chaining of unit_t->queue */
65 const unit_t *ou; /**< the opt unit this qnode belongs to */
66 int color; /**< target color */
67 set *conflicts; /**< contains conflict_t's. All internal conflicts */
68 int mis_costs; /**< costs of nodes/copies in the mis. */
69 int mis_size; /**< size of the array below */
70 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
71 set *changed_nodes; /**< contains node_stat_t's. */
74 static pset *pinned_global; /**< optimized nodes should not be altered any more */
76 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
77 const conflict_t *xx = x;
78 const conflict_t *yy = y;
79 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
83 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
84 * The next maximum independent set build, will regard it.
86 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
88 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
90 if ((int)n1 < (int)n2) {
97 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
101 * Checks if two nodes are in a conflict.
103 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
105 /* search for live range interference */
106 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
108 /* search for recoloring conflicts */
109 if ((int)n1 < (int)n2) {
116 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
119 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
120 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
124 * Finds a node status entry of a node if existent. Otherwise return NULL
126 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
129 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
133 * Finds a node status entry of a node if existent. Otherwise it will return
134 * an initialized new entry for this node.
136 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
139 find.new_color = NO_COLOR;
140 find.pinned_local = 0;
141 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
145 * Returns the virtual color of a node if set before, else returns the real color.
147 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
148 node_stat_t *found = qnode_find_node(qn, irn);
150 return found->new_color;
152 return get_irn_col(qn->ou->co, irn);
156 * Sets the virtual color of a node.
158 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
159 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
160 found->new_color = color;
161 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
165 * Checks if a node is local pinned. A node is local pinned, iff it belongs
166 * to the same optimization unit and has been optimized before the current
169 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
170 node_stat_t *found = qnode_find_node(qn, irn);
172 return found->pinned_local;
178 * Local-pins a node, so optimizations of further nodes of the same opt unit
179 * can handle situations in which a color change would undo prior optimizations.
181 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
182 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
183 found->pinned_local = 1;
184 if (found->new_color == NO_COLOR)
185 found->new_color = get_irn_col(qn->ou->co, irn);
190 * Possible return values of qnode_color_irn()
192 #define CHANGE_SAVE NULL
193 #define CHANGE_IMPOSSIBLE (ir_node *)1
194 #define is_conflicting_node(n) (((int)n) > 1)
197 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
198 * other nodes are changed too, as required to preserve correctness. Function is
199 * aware of local and global pinning. Recursive.
201 * If irn == trigger the color @p col must be used. (the first recoloring)
202 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
204 * @param irn The node to set the color for
205 * @param col The color to set
206 * @param trigger The irn that caused the wish to change the color of the irn
207 * External callers must call with trigger = irn
209 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
210 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
211 * Else the first conflicting ir_node encountered is returned.
214 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
215 copy_opt_t *co = qn->ou->co;
216 const be_chordal_env_t *chordal_env = co->cenv;
217 const arch_register_class_t *cls = co->cls;
218 const arch_env_t *arch_env = co->aenv;
219 int irn_col = qnode_get_new_color(qn, irn);
220 ir_node *sub_res, *curr;
221 be_ifg_t *ifg = chordal_env->ifg;
222 void *iter = be_ifg_neighbours_iter_alloca(ifg);
225 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
227 /* If the target color is already set do nothing */
228 if (irn_col == col) {
229 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
233 /* If the irn is pinned, changing color is impossible */
234 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
235 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
239 #ifdef SEARCH_FREE_COLORS
240 /* If we resolve conflicts (recursive calls) we can use any unused color.
241 * In case of the first call @p col must be used.
243 if (irn != trigger) {
244 bitset_t *free_cols = bitset_alloca(cls->n_regs);
245 arch_register_req_t req;
249 /* Get all possible colors */
250 arch_put_non_ignore_regs(arch_env, cls, free_cols);
252 /* Exclude colors not assignable to the irn */
253 arch_get_register_req(arch_env, &req, irn, -1);
254 if (arch_register_req_is(&req, limited)) {
255 bitset_t *limited = bitset_alloca(cls->n_regs);
256 req.limited(req.limited_env, limited);
257 bitset_and(free_cols, limited);
260 /* Exclude the color of the irn, because it must _change_ its color */
261 bitset_clear(free_cols, irn_col);
263 /* Exclude all colors used by adjacent nodes */
264 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
265 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
267 free_col = bitset_next_set(free_cols, 0);
269 if (free_col != -1) {
270 qnode_set_new_color(qn, irn, free_col);
274 #endif /* SEARCH_FREE_COLORS */
276 /* If target color is not allocatable changing color is impossible */
277 if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
278 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
279 return CHANGE_IMPOSSIBLE;
283 * If we arrive here changing color may be possible, but there may be conflicts.
284 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
286 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
287 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
288 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
289 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
290 if (sub_res != CHANGE_SAVE) {
291 be_ifg_neighbours_break(ifg, iter);
298 * If we arrive here, all conflicts were resolved.
299 * So it is save to change this irn
301 qnode_set_new_color(qn, irn, col);
307 * Tries to set the colors for all members of this queue node;
308 * to the target color qn->color
309 * @returns 1 iff all members colors could be set
312 static int qnode_try_color(const qnode_t *qn) {
314 for (i=0; i<qn->mis_size; ++i) {
315 ir_node *test_node, *confl_node;
317 test_node = qn->mis[i];
318 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
319 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
321 if (confl_node == CHANGE_SAVE) {
322 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
323 qnode_pin_local(qn, test_node);
324 } else if (confl_node == CHANGE_IMPOSSIBLE) {
325 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
326 qnode_add_conflict(qn, test_node, test_node);
328 if (qnode_is_pinned_local(qn, confl_node)) {
329 /* changing test_node would change back a node of current ou */
330 if (confl_node == qn->ou->nodes[0]) {
331 /* Adding a conflict edge between testnode and conflnode
332 * would introduce a root -- arg interference.
333 * So remove the arg of the qn */
334 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
335 qnode_add_conflict(qn, test_node, test_node);
337 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
338 qnode_add_conflict(qn, confl_node, test_node);
341 if (pset_find_ptr(pinned_global, confl_node)) {
342 /* changing test_node would change back a node of a prior ou */
343 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
344 qnode_add_conflict(qn, test_node, test_node);
348 if (confl_node != CHANGE_SAVE)
355 * Determines a maximum weighted independent set with respect to
356 * the interference and conflict edges of all nodes in a qnode.
358 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
359 ir_node **safe, **unsafe;
360 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
361 bitset_t *curr, *best;
362 int max, next, pos, curr_weight, best_weight = 0;
364 /* assign the nodes into two groups.
365 * safe: node has no interference, hence it is in every max stable set.
366 * unsafe: node has an interference
368 safe = alloca((ou->node_count-1) * sizeof(*safe));
371 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
372 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
374 for(i=1; i<ou->node_count; ++i) {
376 for(o=1; o<ou->node_count; ++o) {
377 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
379 unsafe_costs[unsafe_count] = ou->costs[i];
380 unsafe[unsafe_count] = ou->nodes[i];
388 safe_costs += ou->costs[i];
389 safe[safe_count++] = ou->nodes[i];
395 /* now compute the best set out of the unsafe nodes*/
396 best = bitset_alloca(unsafe_count);
398 if (unsafe_count > MIS_HEUR_TRIGGER) {
399 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
400 for (i=0; i<unsafe_count; ++i) {
402 /* check if it is a stable set */
403 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
404 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
405 bitset_clear(best, i); /* clear the bit and try next one */
409 /* compute the weight */
410 bitset_foreach(best, pos)
411 best_weight += unsafe_costs[pos];
413 /* Exact Algorithm: Brute force */
414 curr = bitset_alloca(unsafe_count);
415 bitset_set_all(curr);
416 while ((max = bitset_popcnt(curr)) != 0) {
417 /* check if curr is a stable set */
418 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
419 for (o=bitset_next_set(curr, i); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to ou_max_ind_set_costs(): NOT (curr, i+1) */
420 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
423 /* if we arrive here, we have a stable set */
424 /* compute the weigth of the stable set*/
426 bitset_foreach(curr, pos)
427 curr_weight += unsafe_costs[pos];
430 if (curr_weight > best_weight) {
431 best_weight = curr_weight;
432 bitset_copy(best, curr);
440 /* transfer the best set into the qn */
441 qn->mis_size = 1+safe_count+bitset_popcnt(best);
442 qn->mis_costs = safe_costs+best_weight;
443 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
445 for (i=0; i<safe_count; ++i)
446 qn->mis[next++] = safe[i];
447 bitset_foreach(best, pos)
448 qn->mis[next++] = unsafe[pos];
452 * Creates a new qnode
454 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
455 qnode_t *qn = xmalloc(sizeof(*qn));
458 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
459 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
460 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
465 * Frees space used by a queue node
467 static INLINE void free_qnode(qnode_t *qn) {
468 del_set(qn->conflicts);
469 del_set(qn->changed_nodes);
475 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
476 * ordered by field 'size' (the size of the mis) in decreasing order.
478 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
479 struct list_head *lh;
481 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
482 /* root node is not in qnode */
487 qnode_max_ind_set(qn, ou);
488 /* do the insertion */
489 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
491 while (lh->next != &ou->queue) {
492 qnode_t *curr = list_entry_queue(lh->next);
493 if (curr->mis_costs <= qn->mis_costs)
497 list_add(&qn->queue, lh);
501 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
502 * costs of copy instructions placed during SSA-destruction and lowering.
503 * Works only for opt units with exactly 1 root node, which is the
504 * case for approximately 80% of all phi classes and 100% of register constrained
505 * nodes. (All other phi classes are reduced to this case.)
507 static void ou_optimize(unit_t *ou) {
509 qnode_t *curr = NULL, *tmp;
510 const arch_env_t *aenv = ou->co->aenv;
511 const arch_register_class_t *cls = ou->co->cls;
512 bitset_t *pos_regs = bitset_alloca(cls->n_regs);
513 bitset_t *ign_regs = bitset_alloca(cls->n_regs);
515 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
516 for (i=0; i<ou->node_count; ++i)
517 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
520 INIT_LIST_HEAD(&ou->queue);
522 arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
524 /* exclude ingore colors */
525 arch_put_non_ignore_regs(aenv, cls, ign_regs);
526 bitset_and(pos_regs, ign_regs);
528 assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?");
530 /* create new qnode */
531 bitset_foreach(pos_regs, i)
532 ou_insert_qnode(ou, new_qnode(ou, i));
535 while (!list_empty(&ou->queue)) {
536 /* get head of queue */
537 curr = list_entry_queue(ou->queue.next);
538 list_del(&curr->queue);
539 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
542 if (qnode_try_color(curr))
545 /* no success, so re-insert */
546 del_set(curr->changed_nodes);
547 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
548 ou_insert_qnode(ou, curr);
551 /* apply the best found qnode */
552 if (curr->mis_size >= 2) {
554 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
555 DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d << %d << %d\n", curr->color, ou->min_nodes_costs, ou->all_nodes_costs - curr->mis_costs, ou->all_nodes_costs));
556 /* globally pin root and all args which have the same color */
557 pset_insert_ptr(pinned_global, ou->nodes[0]);
558 for (i=1; i<ou->node_count; ++i) {
559 ir_node *irn = ou->nodes[i];
560 int nc = qnode_get_new_color(curr, irn);
561 if (nc != NO_COLOR && nc == root_col)
562 pset_insert_ptr(pinned_global, irn);
565 /* set color of all changed nodes */
566 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
567 /* NO_COLOR is possible, if we had an undo */
568 if (ns->new_color != NO_COLOR) {
569 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
570 set_irn_col(ou->co, ns->irn, ns->new_color);
575 /* free best qnode (curr) and queue */
577 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
581 int co_solve_heuristic(copy_opt_t *co) {
583 dbg = firm_dbg_register("ir.be.copyoptheur");
585 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
586 list_for_each_entry(unit_t, curr, &co->units, units)
587 if (curr->node_count > 1)
590 del_pset(pinned_global);