2 * This file is part of libFirm.
3 * Copyright (C) 2012 University of Karlsruhe.
8 * @brief First simple copy minimization heuristics.
12 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
13 * examined. A qnode has a 'target color', nodes out of the opt unit and
14 * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges'
15 * A 'max indep set' is determined from these. We try to color this mis using a
16 * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges'
17 * and the qnode is reinserted in the queue. The first qnode colored without
18 * conflicts is the best one.
24 #include "raw_bitset.h"
27 #include "becopyopt_t.h"
28 #include "becopystat.h"
29 #include "beintlive_t.h"
33 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
35 /** Defines an invalid register index. */
38 #define SEARCH_FREE_COLORS
40 #define SLOTS_PINNED_GLOBAL 64
41 #define SLOTS_CONFLICTS 8
42 #define SLOTS_CHANGED_NODES 32
44 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
45 #define HASH_CONFLICT(c) (hash_irn(c.n1) ^ hash_irn(c.n2))
48 * Modeling additional conflicts between nodes. NOT live range interference
50 typedef struct conflict_t {
51 const ir_node *n1, *n2;
55 * If an irn is changed, the changes first get stored in a node_stat_t,
56 * to allow undo of changes (=drop new data) in case of conflicts.
58 typedef struct node_stat_t {
65 * Represents a node in the optimization queue.
67 typedef struct qnode_t {
68 struct list_head queue; /**< chaining of unit_t->queue */
69 const unit_t *ou; /**< the opt unit this node belongs to */
70 int color; /**< target color */
71 set *conflicts; /**< contains conflict_t's. All internal conflicts */
72 int mis_costs; /**< costs of nodes/copies in the mis. */
73 int mis_size; /**< size of the array below */
74 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
75 set *changed_nodes; /**< contains node_stat_t's. */
78 static pset *pinned_global; /**< optimized nodes should not be altered any more */
80 static inline int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
83 return be_ifg_connected(env->ifg, a, b);
85 be_lv_t *lv = be_get_irg_liveness(env->irg);
86 return be_values_interfere(lv, a, b);
90 static int set_cmp_conflict_t(const void *x, const void *y, size_t size)
92 const conflict_t *xx = (const conflict_t*)x;
93 const conflict_t *yy = (const conflict_t*)y;
96 return xx->n1 != yy->n1 || xx->n2 != yy->n2;
100 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
101 * The next maximum independent set build, will regard it.
103 static inline void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2)
106 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
108 if (get_irn_idx(n1) < get_irn_idx(n2)) {
115 (void)set_insert(conflict_t, qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
119 * Checks if two nodes are in a conflict.
121 static inline int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2)
124 /* search for live range interference */
125 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
127 /* search for recoloring conflicts */
128 if (get_irn_idx(n1) < get_irn_idx(n2)) {
135 return set_find(conflict_t, qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)) != 0;
138 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size)
141 return ((const node_stat_t*)x)->irn != ((const node_stat_t*)y)->irn;
145 * Finds a node status entry of a node if existent. Otherwise return NULL
147 static inline const node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn)
151 return set_find(node_stat_t, qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
155 * Finds a node status entry of a node if existent. Otherwise it will return
156 * an initialized new entry for this node.
158 static inline node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn)
162 find.new_color = NO_COLOR;
163 find.pinned_local = 0;
164 return set_insert(node_stat_t, qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
168 * Returns the virtual color of a node if set before, else returns the real color.
170 static inline int qnode_get_new_color(const qnode_t *qn, ir_node *irn)
172 const node_stat_t *found = qnode_find_node(qn, irn);
174 return found->new_color;
176 return get_irn_col(irn);
180 * Sets the virtual color of a node.
182 static inline void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color)
184 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
185 found->new_color = color;
186 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
190 * Checks if a node is local pinned. A node is local pinned, iff it belongs
191 * to the same optimization unit and has been optimized before the current
194 static inline int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn)
196 const node_stat_t *found = qnode_find_node(qn, irn);
198 return found->pinned_local;
204 * Local-pins a node, so optimizations of further nodes of the same opt unit
205 * can handle situations in which a color change would undo prior optimizations.
207 static inline void qnode_pin_local(const qnode_t *qn, ir_node *irn)
209 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
210 found->pinned_local = 1;
211 if (found->new_color == NO_COLOR)
212 found->new_color = get_irn_col(irn);
217 * Possible return values of qnode_color_irn()
219 #define CHANGE_SAVE NULL
220 #define CHANGE_IMPOSSIBLE (ir_node *)1
223 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
224 * other nodes are changed too, as required to preserve correctness. Function is
225 * aware of local and global pinning. Recursive.
227 * If irn == trigger the color @p col must be used. (the first recoloring)
228 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
230 * @param irn The node to set the color for
231 * @param col The color to set
232 * @param trigger The irn that caused the wish to change the color of the irn
233 * External callers must call with trigger = irn
235 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
236 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
237 * Else the first conflicting ir_node encountered is returned.
240 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger)
242 copy_opt_t *co = qn->ou->co;
243 const be_chordal_env_t *chordal_env = co->cenv;
244 const arch_register_class_t *cls = co->cls;
245 int irn_col = qnode_get_new_color(qn, irn);
246 be_ifg_t *ifg = chordal_env->ifg;
247 neighbours_iter_t iter;
248 const arch_register_req_t *req;
250 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
252 /* If the target color is already set do nothing */
253 if (irn_col == col) {
254 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
258 /* If the irn is pinned, changing color is impossible */
259 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
260 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
264 req = arch_get_irn_register_req(irn);
265 #ifdef SEARCH_FREE_COLORS
266 /* If we resolve conflicts (recursive calls) we can use any unused color.
267 * In case of the first call @p col must be used.
269 if (irn != trigger) {
270 bitset_t *free_cols = bitset_alloca(cls->n_regs);
273 /* Get all possible colors */
274 bitset_copy(free_cols, co->cenv->allocatable_regs);
276 /* Exclude colors not assignable to the irn */
277 if (arch_register_req_is(req, limited)) {
278 bitset_t *limited = bitset_alloca(cls->n_regs);
279 rbitset_copy_to_bitset(req->limited, limited);
280 bitset_and(free_cols, limited);
283 /* Exclude the color of the irn, because it must _change_ its color */
284 bitset_clear(free_cols, irn_col);
286 /* Exclude all colors used by adjacent nodes */
287 be_ifg_foreach_neighbour(ifg, &iter, irn, curr)
288 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
290 free_col = bitset_next_set(free_cols, 0);
292 if (free_col != -1) {
293 qnode_set_new_color(qn, irn, free_col);
297 #endif /* SEARCH_FREE_COLORS */
299 /* If target color is not allocatable changing color is impossible */
300 if (!arch_reg_is_allocatable(req, arch_register_for_index(cls, col))) {
301 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
302 return CHANGE_IMPOSSIBLE;
306 * If we arrive here changing color may be possible, but there may be conflicts.
307 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
309 be_ifg_foreach_neighbour(ifg, &iter, irn, curr) {
310 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
311 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
312 ir_node *const sub_res = qnode_color_irn(qn, curr, irn_col, irn);
313 if (sub_res != CHANGE_SAVE) {
314 be_ifg_neighbours_break(&iter);
321 * If we arrive here, all conflicts were resolved.
322 * So it is save to change this irn
324 qnode_set_new_color(qn, irn, col);
330 * Tries to set the colors for all members of this queue node;
331 * to the target color qn->color
332 * @returns 1 iff all members colors could be set
335 static int qnode_try_color(const qnode_t *qn)
338 for (i=0; i<qn->mis_size; ++i) {
339 ir_node *test_node, *confl_node;
341 test_node = qn->mis[i];
342 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
343 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
345 if (confl_node == CHANGE_SAVE) {
346 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
347 qnode_pin_local(qn, test_node);
348 } else if (confl_node == CHANGE_IMPOSSIBLE) {
349 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
350 qnode_add_conflict(qn, test_node, test_node);
353 if (qnode_is_pinned_local(qn, confl_node)) {
354 /* changing test_node would change back a node of current ou */
355 if (confl_node == qn->ou->nodes[0]) {
356 /* Adding a conflict edge between testnode and conflnode
357 * would introduce a root -- arg interference.
358 * So remove the arg of the qn */
359 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
360 qnode_add_conflict(qn, test_node, test_node);
362 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
363 qnode_add_conflict(qn, confl_node, test_node);
366 if (pset_find_ptr(pinned_global, confl_node)) {
367 /* changing test_node would change back a node of a prior ou */
368 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
369 qnode_add_conflict(qn, test_node, test_node);
378 * Determines a maximum weighted independent set with respect to
379 * the interference and conflict edges of all nodes in a qnode.
381 static inline void qnode_max_ind_set(qnode_t *qn, const unit_t *ou)
383 ir_node **safe, **unsafe;
384 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
385 bitset_t *curr, *best;
386 int next, curr_weight, best_weight = 0;
388 /* assign the nodes into two groups.
389 * safe: node has no interference, hence it is in every max stable set.
390 * unsafe: node has an interference
392 safe = ALLOCAN(ir_node*, ou->node_count - 1);
395 unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
396 unsafe_costs = ALLOCAN(int, ou->node_count - 1);
398 for (i=1; i<ou->node_count; ++i) {
400 for (o=1; o<ou->node_count; ++o) {
401 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
403 unsafe_costs[unsafe_count] = ou->costs[i];
404 unsafe[unsafe_count] = ou->nodes[i];
412 safe_costs += ou->costs[i];
413 safe[safe_count++] = ou->nodes[i];
419 /* now compute the best set out of the unsafe nodes*/
420 best = bitset_alloca(unsafe_count);
422 if (unsafe_count > MIS_HEUR_TRIGGER) {
423 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
424 for (i=0; i<unsafe_count; ++i) {
426 /* check if it is a stable set */
427 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
428 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
429 bitset_clear(best, i); /* clear the bit and try next one */
433 /* compute the weight */
434 bitset_foreach(best, pos)
435 best_weight += unsafe_costs[pos];
437 /* Exact Algorithm: Brute force */
438 curr = bitset_alloca(unsafe_count);
439 bitset_set_all(curr);
440 while (!bitset_is_empty(curr)) {
441 /* check if curr is a stable set */
442 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
443 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) */
444 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
447 /* if we arrive here, we have a stable set */
448 /* compute the weight of the stable set*/
450 bitset_foreach(curr, pos)
451 curr_weight += unsafe_costs[pos];
454 if (curr_weight > best_weight) {
455 best_weight = curr_weight;
456 bitset_copy(best, curr);
464 /* transfer the best set into the qn */
465 qn->mis_size = 1+safe_count+bitset_popcount(best);
466 qn->mis_costs = safe_costs+best_weight;
467 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
469 for (i=0; i<safe_count; ++i)
470 qn->mis[next++] = safe[i];
471 bitset_foreach(best, pos)
472 qn->mis[next++] = unsafe[pos];
476 * Creates a new qnode
478 static inline qnode_t *new_qnode(const unit_t *ou, int color)
480 qnode_t *qn = XMALLOC(qnode_t);
483 qn->mis = XMALLOCN(ir_node*, ou->node_count);
484 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
485 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
490 * Frees space used by a queue node
492 static inline void free_qnode(qnode_t *qn)
494 del_set(qn->conflicts);
495 del_set(qn->changed_nodes);
501 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
502 * ordered by field 'size' (the size of the mis) in decreasing order.
504 static inline void ou_insert_qnode(unit_t *ou, qnode_t *qn)
506 struct list_head *lh;
508 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
509 /* root node is not in qnode */
514 qnode_max_ind_set(qn, ou);
515 /* do the insertion */
516 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
518 while (lh->next != &ou->queue) {
519 qnode_t *curr = list_entry_queue(lh->next);
520 if (curr->mis_costs <= qn->mis_costs)
524 list_add(&qn->queue, lh);
528 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
529 * costs of copy instructions placed during SSA-destruction and lowering.
530 * Works only for opt units with exactly 1 root node, which is the
531 * case for approximately 80% of all phi classes and 100% of register constrained
532 * nodes. (All other phi classes are reduced to this case.)
534 static void ou_optimize(unit_t *ou)
536 qnode_t *curr = NULL;
537 const arch_register_req_t *req;
538 bitset_t const* allocatable_regs;
543 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
544 for (i=0; i<ou->node_count; ++i)
545 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
548 INIT_LIST_HEAD(&ou->queue);
550 req = arch_get_irn_register_req(ou->nodes[0]);
551 allocatable_regs = ou->co->cenv->allocatable_regs;
552 n_regs = req->cls->n_regs;
553 if (arch_register_req_is(req, limited)) {
554 unsigned const* limited = req->limited;
556 for (idx = 0; idx != n_regs; ++idx) {
557 if (!bitset_is_set(allocatable_regs, idx))
559 if (!rbitset_is_set(limited, idx))
562 ou_insert_qnode(ou, new_qnode(ou, idx));
565 for (idx = 0; idx != n_regs; ++idx) {
566 if (!bitset_is_set(allocatable_regs, idx))
569 ou_insert_qnode(ou, new_qnode(ou, idx));
575 assert(!list_empty(&ou->queue));
576 /* get head of queue */
577 curr = list_entry_queue(ou->queue.next);
578 list_del(&curr->queue);
579 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
582 if (qnode_try_color(curr))
585 /* no success, so re-insert */
586 del_set(curr->changed_nodes);
587 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
588 ou_insert_qnode(ou, curr);
591 /* apply the best found qnode */
592 if (curr->mis_size >= 2) {
593 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
594 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));
595 /* globally pin root and all args which have the same color */
596 pset_insert_ptr(pinned_global, ou->nodes[0]);
597 for (i=1; i<ou->node_count; ++i) {
598 ir_node *irn = ou->nodes[i];
599 int nc = qnode_get_new_color(curr, irn);
600 if (nc != NO_COLOR && nc == root_col)
601 pset_insert_ptr(pinned_global, irn);
604 /* set color of all changed nodes */
605 foreach_set(curr->changed_nodes, node_stat_t, ns) {
606 /* NO_COLOR is possible, if we had an undo */
607 if (ns->new_color != NO_COLOR) {
608 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
609 set_irn_col(ou->co->cls, ns->irn, ns->new_color);
614 /* free best qnode (curr) and queue */
616 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
621 * Solves the problem using a heuristic approach
622 * Uses the OU data structure
624 int co_solve_heuristic(copy_opt_t *co)
628 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
629 list_for_each_entry(unit_t, curr, &co->units, units)
630 if (curr->node_count > 1)
633 del_pset(pinned_global);
637 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur)
638 void be_init_copyheur(void)
640 static co_algo_info copyheur = {
641 co_solve_heuristic, 0
644 be_register_copyopt("heur1", ©heur);
645 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");