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 First simple copy minimization heuristics.
23 * @author Daniel Grund
27 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
28 * examined. A qnode has a 'target color', nodes out of the opt unit and
29 * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges'
30 * A 'max indep set' is determined from these. We try to color this mis using a
31 * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges'
32 * and the qnode is reinserted in the queue. The first qnode colored without
33 * conflicts is the best one.
39 #include "raw_bitset.h"
42 #include "becopyopt_t.h"
43 #include "becopystat.h"
44 #include "beintlive_t.h"
48 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
50 #define SEARCH_FREE_COLORS
52 #define SLOTS_PINNED_GLOBAL 64
53 #define SLOTS_CONFLICTS 8
54 #define SLOTS_CHANGED_NODES 32
56 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
57 #define HASH_CONFLICT(c) (hash_irn(c.n1) ^ hash_irn(c.n2))
60 * Modeling additional conflicts between nodes. NOT live range interference
62 typedef struct _conflict_t {
63 const ir_node *n1, *n2;
67 * If an irn is changed, the changes first get stored in a node_stat_t,
68 * to allow undo of changes (=drop new data) in case of conflicts.
70 typedef struct _node_stat_t {
77 * Represents a node in the optimization queue.
79 typedef struct _qnode_t {
80 struct list_head queue; /**< chaining of unit_t->queue */
81 const unit_t *ou; /**< the opt unit this node belongs to */
82 int color; /**< target color */
83 set *conflicts; /**< contains conflict_t's. All internal conflicts */
84 int mis_costs; /**< costs of nodes/copies in the mis. */
85 int mis_size; /**< size of the array below */
86 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
87 set *changed_nodes; /**< contains node_stat_t's. */
90 static pset *pinned_global; /**< optimized nodes should not be altered any more */
92 static inline int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
95 return be_ifg_connected(env->ifg, a, b);
97 return be_values_interfere(env->birg->lv, a, b);
100 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
101 const conflict_t *xx = x;
102 const conflict_t *yy = y;
105 return xx->n1 != yy->n1 || xx->n2 != yy->n2;
109 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
110 * The next maximum independent set build, will regard it.
112 static inline void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
114 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
116 if (get_irn_idx(n1) < get_irn_idx(n2)) {
123 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
127 * Checks if two nodes are in a conflict.
129 static inline int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
131 /* search for live range interference */
132 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
134 /* search for recoloring conflicts */
135 if (get_irn_idx(n1) < get_irn_idx(n2)) {
142 return set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)) != 0;
145 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
147 return ((const node_stat_t*)x)->irn != ((const node_stat_t*)y)->irn;
151 * Finds a node status entry of a node if existent. Otherwise return NULL
153 static inline const node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
156 return set_find(qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
160 * Finds a node status entry of a node if existent. Otherwise it will return
161 * an initialized new entry for this node.
163 static inline node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
166 find.new_color = NO_COLOR;
167 find.pinned_local = 0;
168 return set_insert(qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
172 * Returns the virtual color of a node if set before, else returns the real color.
174 static inline int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
175 const node_stat_t *found = qnode_find_node(qn, irn);
177 return found->new_color;
179 return get_irn_col(irn);
183 * Sets the virtual color of a node.
185 static inline void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
186 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
187 found->new_color = color;
188 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
192 * Checks if a node is local pinned. A node is local pinned, iff it belongs
193 * to the same optimization unit and has been optimized before the current
196 static inline int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
197 const node_stat_t *found = qnode_find_node(qn, irn);
199 return found->pinned_local;
205 * Local-pins a node, so optimizations of further nodes of the same opt unit
206 * can handle situations in which a color change would undo prior optimizations.
208 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) {
241 copy_opt_t *co = qn->ou->co;
242 const be_chordal_env_t *chordal_env = co->cenv;
243 const arch_register_class_t *cls = co->cls;
244 int irn_col = qnode_get_new_color(qn, irn);
245 ir_node *sub_res, *curr;
246 be_ifg_t *ifg = chordal_env->ifg;
247 void *iter = be_ifg_neighbours_iter_alloca(ifg);
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 #ifdef SEARCH_FREE_COLORS
265 /* If we resolve conflicts (recursive calls) we can use any unused color.
266 * In case of the first call @p col must be used.
268 if (irn != trigger) {
269 bitset_t *free_cols = bitset_alloca(cls->n_regs);
270 const arch_register_req_t *req;
274 /* Get all possible colors */
275 bitset_copy(free_cols, co->cenv->ignore_colors);
276 bitset_flip_all(free_cols);
278 /* Exclude colors not assignable to the irn */
279 req = arch_get_register_req_out(irn);
280 if (arch_register_req_is(req, limited)) {
281 bitset_t *limited = bitset_alloca(cls->n_regs);
282 rbitset_copy_to_bitset(req->limited, limited);
283 bitset_and(free_cols, limited);
286 /* Exclude the color of the irn, because it must _change_ its color */
287 bitset_clear(free_cols, irn_col);
289 /* Exclude all colors used by adjacent nodes */
290 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
291 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
293 free_col = bitset_next_set(free_cols, 0);
295 if (free_col != -1) {
296 qnode_set_new_color(qn, irn, free_col);
300 #endif /* SEARCH_FREE_COLORS */
302 /* If target color is not allocatable changing color is impossible */
303 if (!arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, col))) {
304 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
305 return CHANGE_IMPOSSIBLE;
309 * If we arrive here changing color may be possible, but there may be conflicts.
310 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
312 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
313 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
314 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
315 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
316 if (sub_res != CHANGE_SAVE) {
317 be_ifg_neighbours_break(ifg, iter);
324 * If we arrive here, all conflicts were resolved.
325 * So it is save to change this irn
327 qnode_set_new_color(qn, irn, col);
333 * Tries to set the colors for all members of this queue node;
334 * to the target color qn->color
335 * @returns 1 iff all members colors could be set
338 static int qnode_try_color(const qnode_t *qn) {
340 for (i=0; i<qn->mis_size; ++i) {
341 ir_node *test_node, *confl_node;
343 test_node = qn->mis[i];
344 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
345 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
347 if (confl_node == CHANGE_SAVE) {
348 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
349 qnode_pin_local(qn, test_node);
350 } else if (confl_node == CHANGE_IMPOSSIBLE) {
351 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
352 qnode_add_conflict(qn, test_node, test_node);
355 if (qnode_is_pinned_local(qn, confl_node)) {
356 /* changing test_node would change back a node of current ou */
357 if (confl_node == qn->ou->nodes[0]) {
358 /* Adding a conflict edge between testnode and conflnode
359 * would introduce a root -- arg interference.
360 * So remove the arg of the qn */
361 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
362 qnode_add_conflict(qn, test_node, test_node);
364 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
365 qnode_add_conflict(qn, confl_node, test_node);
368 if (pset_find_ptr(pinned_global, confl_node)) {
369 /* changing test_node would change back a node of a prior ou */
370 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
371 qnode_add_conflict(qn, test_node, test_node);
380 * Determines a maximum weighted independent set with respect to
381 * the interference and conflict edges of all nodes in a qnode.
383 static inline void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
384 ir_node **safe, **unsafe;
385 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
386 bitset_t *curr, *best;
388 int next, curr_weight, best_weight = 0;
390 /* assign the nodes into two groups.
391 * safe: node has no interference, hence it is in every max stable set.
392 * unsafe: node has an interference
394 safe = ALLOCAN(ir_node*, ou->node_count - 1);
397 unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
398 unsafe_costs = ALLOCAN(int, ou->node_count - 1);
400 for(i=1; i<ou->node_count; ++i) {
402 for(o=1; o<ou->node_count; ++o) {
403 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
405 unsafe_costs[unsafe_count] = ou->costs[i];
406 unsafe[unsafe_count] = ou->nodes[i];
414 safe_costs += ou->costs[i];
415 safe[safe_count++] = ou->nodes[i];
421 /* now compute the best set out of the unsafe nodes*/
422 best = bitset_alloca(unsafe_count);
424 if (unsafe_count > MIS_HEUR_TRIGGER) {
425 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
426 for (i=0; i<unsafe_count; ++i) {
428 /* check if it is a stable set */
429 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
430 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
431 bitset_clear(best, i); /* clear the bit and try next one */
435 /* compute the weight */
436 bitset_foreach(best, pos)
437 best_weight += unsafe_costs[pos];
439 /* Exact Algorithm: Brute force */
440 curr = bitset_alloca(unsafe_count);
441 bitset_set_all(curr);
442 while (!bitset_is_empty(curr)) {
443 /* check if curr is a stable set */
444 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
445 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) */
446 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
449 /* if we arrive here, we have a stable set */
450 /* compute the weigth of the stable set*/
452 bitset_foreach(curr, pos)
453 curr_weight += unsafe_costs[pos];
456 if (curr_weight > best_weight) {
457 best_weight = curr_weight;
458 bitset_copy(best, curr);
466 /* transfer the best set into the qn */
467 qn->mis_size = 1+safe_count+bitset_popcnt(best);
468 qn->mis_costs = safe_costs+best_weight;
469 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
471 for (i=0; i<safe_count; ++i)
472 qn->mis[next++] = safe[i];
473 bitset_foreach(best, pos)
474 qn->mis[next++] = unsafe[pos];
478 * Creates a new qnode
480 static inline qnode_t *new_qnode(const unit_t *ou, int color) {
481 qnode_t *qn = XMALLOC(qnode_t);
484 qn->mis = XMALLOCN(ir_node*, ou->node_count);
485 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
486 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
491 * Frees space used by a queue node
493 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) {
505 struct list_head *lh;
507 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
508 /* root node is not in qnode */
513 qnode_max_ind_set(qn, ou);
514 /* do the insertion */
515 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
517 while (lh->next != &ou->queue) {
518 qnode_t *curr = list_entry_queue(lh->next);
519 if (curr->mis_costs <= qn->mis_costs)
523 list_add(&qn->queue, lh);
527 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
528 * costs of copy instructions placed during SSA-destruction and lowering.
529 * Works only for opt units with exactly 1 root node, which is the
530 * case for approximately 80% of all phi classes and 100% of register constrained
531 * nodes. (All other phi classes are reduced to this case.)
533 static void ou_optimize(unit_t *ou) {
534 qnode_t *curr = NULL;
536 const arch_register_req_t *req;
537 bitset_t const* ignore;
542 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
543 for (i=0; i<ou->node_count; ++i)
544 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
547 INIT_LIST_HEAD(&ou->queue);
549 req = arch_get_register_req_out(ou->nodes[0]);
550 ignore = ou->co->cenv->ignore_colors;
551 n_regs = req->cls->n_regs;
552 if (arch_register_req_is(req, limited)) {
553 rawbs_base_t const* limited = req->limited;
555 for (idx = 0; idx != n_regs; ++idx) {
556 if (bitset_is_set(ignore, idx))
558 if (!rbitset_is_set(limited, idx))
561 ou_insert_qnode(ou, new_qnode(ou, idx));
564 for (idx = 0; idx != n_regs; ++idx) {
565 if (bitset_is_set(ignore, idx))
568 ou_insert_qnode(ou, new_qnode(ou, idx));
574 assert(!list_empty(&ou->queue));
575 /* get head of queue */
576 curr = list_entry_queue(ou->queue.next);
577 list_del(&curr->queue);
578 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
581 if (qnode_try_color(curr))
584 /* no success, so re-insert */
585 del_set(curr->changed_nodes);
586 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
587 ou_insert_qnode(ou, curr);
590 /* apply the best found qnode */
591 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 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
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, 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) {
629 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
630 list_for_each_entry(unit_t, curr, &co->units, units)
631 if (curr->node_count > 1)
634 del_pset(pinned_global);
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");
648 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur);