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"
30 #include "benodesets.h"
33 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
35 #define SEARCH_FREE_COLORS
37 #define SLOTS_PINNED_GLOBAL 64
38 #define SLOTS_CONFLICTS 8
39 #define SLOTS_CHANGED_NODES 32
41 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
42 #define HASH_CONFLICT(c) (nodeset_hash(c.n1) ^ nodeset_hash(c.n2))
45 * Modeling additional conflicts between nodes. NOT live range interference
47 typedef struct _conflict_t {
48 const ir_node *n1, *n2;
52 * If an irn is changed, the changes first get stored in a node_stat_t,
53 * to allow undo of changes (=drop new data) in case of conflicts.
55 typedef struct _node_stat_t {
62 * Represents a node in the optimization queue.
64 typedef struct _qnode_t {
65 struct list_head queue; /**< chaining of unit_t->queue */
66 const unit_t *ou; /**< the opt unit this qnode belongs to */
67 int color; /**< target color */
68 set *conflicts; /**< contains conflict_t's. All internal conflicts */
69 int mis_costs; /**< costs of nodes/copies in the mis. */
70 int mis_size; /**< size of the array below */
71 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
72 set *changed_nodes; /**< contains node_stat_t's. */
75 static pset *pinned_global; /**< optimized nodes should not be altered any more */
77 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
78 const conflict_t *xx = x;
79 const conflict_t *yy = y;
80 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
84 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
85 * The next maximum independent set build, will regard it.
87 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
89 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
91 if ((int)n1 < (int)n2) {
98 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
102 * Checks if two nodes are in a conflict.
104 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
106 /* search for live range interference */
107 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
109 /* search for recoloring conflicts */
110 if ((int)n1 < (int)n2) {
117 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
120 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
121 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
125 * Finds a node status entry of a node if existent. Otherwise return NULL
127 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
130 return set_find(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
134 * Finds a node status entry of a node if existent. Otherwise it will return
135 * an initialized new entry for this node.
137 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
140 find.new_color = NO_COLOR;
141 find.pinned_local = 0;
142 return set_insert(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
146 * Returns the virtual color of a node if set before, else returns the real color.
148 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
149 node_stat_t *found = qnode_find_node(qn, irn);
151 return found->new_color;
153 return get_irn_col(qn->ou->co, irn);
157 * Sets the virtual color of a node.
159 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
160 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
161 found->new_color = color;
162 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
166 * Checks if a node is local pinned. A node is local pinned, iff it belongs
167 * to the same optimization unit and has been optimized before the current
170 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
171 node_stat_t *found = qnode_find_node(qn, irn);
173 return found->pinned_local;
179 * Local-pins a node, so optimizations of further nodes of the same opt unit
180 * can handle situations in which a color change would undo prior optimizations.
182 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
183 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
184 found->pinned_local = 1;
185 if (found->new_color == NO_COLOR)
186 found->new_color = get_irn_col(qn->ou->co, irn);
191 * Possible return values of qnode_color_irn()
193 #define CHANGE_SAVE NULL
194 #define CHANGE_IMPOSSIBLE (ir_node *)1
195 #define is_conflicting_node(n) (((int)n) > 1)
198 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
199 * other nodes are changed too, as required to preserve correctness. Function is
200 * aware of local and global pinning. Recursive.
202 * If irn == trigger the color @p col must be used. (the first recoloring)
203 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
205 * @param irn The node to set the color for
206 * @param col The color to set
207 * @param trigger The irn that caused the wish to change the color of the irn
208 * External callers must call with trigger = irn
210 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
211 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
212 * Else the first conflicting ir_node encountered is returned.
215 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
216 copy_opt_t *co = qn->ou->co;
217 const be_chordal_env_t *chordal_env = co->cenv;
218 const arch_register_class_t *cls = co->cls;
219 const arch_env_t *arch_env = co->aenv;
220 int irn_col = qnode_get_new_color(qn, irn);
221 ir_node *sub_res, *curr;
222 be_ifg_t *ifg = chordal_env->ifg;
223 void *iter = be_ifg_neighbours_iter_alloca(ifg);
226 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
228 /* If the target color is already set do nothing */
229 if (irn_col == col) {
230 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
234 /* If the irn is pinned, changing color is impossible */
235 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
236 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
240 #ifdef SEARCH_FREE_COLORS
241 /* If we resolve conflicts (recursive calls) we can use any unused color.
242 * In case of the first call @p col must be used.
244 if (irn != trigger) {
245 bitset_t *free_cols = bitset_alloca(cls->n_regs);
246 arch_register_req_t req;
250 /* Get all possible colors */
251 bitset_copy(free_cols, co->cenv->ignore_colors);
252 bitset_flip_all(free_cols);
254 /* Exclude colors not assignable to the irn */
255 arch_get_register_req(arch_env, &req, irn, -1);
256 if (arch_register_req_is(&req, limited)) {
257 bitset_t *limited = bitset_alloca(cls->n_regs);
258 req.limited(req.limited_env, limited);
259 bitset_and(free_cols, limited);
262 /* Exclude the color of the irn, because it must _change_ its color */
263 bitset_clear(free_cols, irn_col);
265 /* Exclude all colors used by adjacent nodes */
266 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
267 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
269 free_col = bitset_next_set(free_cols, 0);
271 if (free_col != -1) {
272 qnode_set_new_color(qn, irn, free_col);
276 #endif /* SEARCH_FREE_COLORS */
278 /* If target color is not allocatable changing color is impossible */
279 if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
280 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
281 return CHANGE_IMPOSSIBLE;
285 * If we arrive here changing color may be possible, but there may be conflicts.
286 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
288 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
289 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
290 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
291 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
292 if (sub_res != CHANGE_SAVE) {
293 be_ifg_neighbours_break(ifg, iter);
300 * If we arrive here, all conflicts were resolved.
301 * So it is save to change this irn
303 qnode_set_new_color(qn, irn, col);
309 * Tries to set the colors for all members of this queue node;
310 * to the target color qn->color
311 * @returns 1 iff all members colors could be set
314 static int qnode_try_color(const qnode_t *qn) {
316 for (i=0; i<qn->mis_size; ++i) {
317 ir_node *test_node, *confl_node;
319 test_node = qn->mis[i];
320 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
321 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
323 if (confl_node == CHANGE_SAVE) {
324 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
325 qnode_pin_local(qn, test_node);
326 } else if (confl_node == CHANGE_IMPOSSIBLE) {
327 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
328 qnode_add_conflict(qn, test_node, test_node);
330 if (qnode_is_pinned_local(qn, confl_node)) {
331 /* changing test_node would change back a node of current ou */
332 if (confl_node == qn->ou->nodes[0]) {
333 /* Adding a conflict edge between testnode and conflnode
334 * would introduce a root -- arg interference.
335 * So remove the arg of the qn */
336 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
337 qnode_add_conflict(qn, test_node, test_node);
339 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
340 qnode_add_conflict(qn, confl_node, test_node);
343 if (pset_find_ptr(pinned_global, confl_node)) {
344 /* changing test_node would change back a node of a prior ou */
345 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
346 qnode_add_conflict(qn, test_node, test_node);
350 if (confl_node != CHANGE_SAVE)
357 * Determines a maximum weighted independent set with respect to
358 * the interference and conflict edges of all nodes in a qnode.
360 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
361 ir_node **safe, **unsafe;
362 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
363 bitset_t *curr, *best;
364 int max, next, pos, curr_weight, best_weight = 0;
366 /* assign the nodes into two groups.
367 * safe: node has no interference, hence it is in every max stable set.
368 * unsafe: node has an interference
370 safe = alloca((ou->node_count-1) * sizeof(*safe));
373 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
374 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
376 for(i=1; i<ou->node_count; ++i) {
378 for(o=1; o<ou->node_count; ++o) {
379 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
381 unsafe_costs[unsafe_count] = ou->costs[i];
382 unsafe[unsafe_count] = ou->nodes[i];
390 safe_costs += ou->costs[i];
391 safe[safe_count++] = ou->nodes[i];
397 /* now compute the best set out of the unsafe nodes*/
398 best = bitset_alloca(unsafe_count);
400 if (unsafe_count > MIS_HEUR_TRIGGER) {
401 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
402 for (i=0; i<unsafe_count; ++i) {
404 /* check if it is a stable set */
405 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
406 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
407 bitset_clear(best, i); /* clear the bit and try next one */
411 /* compute the weight */
412 bitset_foreach(best, pos)
413 best_weight += unsafe_costs[pos];
415 /* Exact Algorithm: Brute force */
416 curr = bitset_alloca(unsafe_count);
417 bitset_set_all(curr);
418 while ((max = bitset_popcnt(curr)) != 0) {
419 /* check if curr is a stable set */
420 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
421 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) */
422 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
425 /* if we arrive here, we have a stable set */
426 /* compute the weigth of the stable set*/
428 bitset_foreach(curr, pos)
429 curr_weight += unsafe_costs[pos];
432 if (curr_weight > best_weight) {
433 best_weight = curr_weight;
434 bitset_copy(best, curr);
442 /* transfer the best set into the qn */
443 qn->mis_size = 1+safe_count+bitset_popcnt(best);
444 qn->mis_costs = safe_costs+best_weight;
445 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
447 for (i=0; i<safe_count; ++i)
448 qn->mis[next++] = safe[i];
449 bitset_foreach(best, pos)
450 qn->mis[next++] = unsafe[pos];
454 * Creates a new qnode
456 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
457 qnode_t *qn = xmalloc(sizeof(*qn));
460 qn->mis = xmalloc(ou->node_count * sizeof(*qn->mis));
461 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
462 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
467 * Frees space used by a queue node
469 static INLINE void free_qnode(qnode_t *qn) {
470 del_set(qn->conflicts);
471 del_set(qn->changed_nodes);
477 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
478 * ordered by field 'size' (the size of the mis) in decreasing order.
480 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
481 struct list_head *lh;
483 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
484 /* root node is not in qnode */
489 qnode_max_ind_set(qn, ou);
490 /* do the insertion */
491 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
493 while (lh->next != &ou->queue) {
494 qnode_t *curr = list_entry_queue(lh->next);
495 if (curr->mis_costs <= qn->mis_costs)
499 list_add(&qn->queue, lh);
503 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
504 * costs of copy instructions placed during SSA-destruction and lowering.
505 * Works only for opt units with exactly 1 root node, which is the
506 * case for approximately 80% of all phi classes and 100% of register constrained
507 * nodes. (All other phi classes are reduced to this case.)
509 static void ou_optimize(unit_t *ou) {
511 qnode_t *curr = NULL, *tmp;
512 const arch_env_t *aenv = ou->co->aenv;
513 const arch_register_class_t *cls = ou->co->cls;
514 bitset_t *pos_regs = bitset_alloca(cls->n_regs);
516 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
517 for (i=0; i<ou->node_count; ++i)
518 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
521 INIT_LIST_HEAD(&ou->queue);
523 arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
525 /* exclude ingore colors */
526 bitset_andnot(pos_regs, ou->co->cenv->ignore_colors);
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 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");
587 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
588 list_for_each_entry(unit_t, curr, &co->units, units)
589 if (curr->node_count > 1)
592 del_pset(pinned_global);