2 * Copyright (C) 1995-2007 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.
41 #include "raw_bitset.h"
44 #include "becopyopt_t.h"
45 #include "becopystat.h"
46 #include "benodesets.h"
47 #include "beintlive_t.h"
50 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
52 #define SEARCH_FREE_COLORS
54 #define SLOTS_PINNED_GLOBAL 64
55 #define SLOTS_CONFLICTS 8
56 #define SLOTS_CHANGED_NODES 32
58 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
59 #define HASH_CONFLICT(c) (nodeset_hash(c.n1) ^ nodeset_hash(c.n2))
62 * Modeling additional conflicts between nodes. NOT live range interference
64 typedef struct _conflict_t {
65 const ir_node *n1, *n2;
69 * If an irn is changed, the changes first get stored in a node_stat_t,
70 * to allow undo of changes (=drop new data) in case of conflicts.
72 typedef struct _node_stat_t {
79 * Represents a node in the optimization queue.
81 typedef struct _qnode_t {
82 struct list_head queue; /**< chaining of unit_t->queue */
83 const unit_t *ou; /**< the opt unit this node belongs to */
84 int color; /**< target color */
85 set *conflicts; /**< contains conflict_t's. All internal conflicts */
86 int mis_costs; /**< costs of nodes/copies in the mis. */
87 int mis_size; /**< size of the array below */
88 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
89 set *changed_nodes; /**< contains node_stat_t's. */
92 static pset *pinned_global; /**< optimized nodes should not be altered any more */
94 static INLINE int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
97 return be_ifg_connected(env->ifg, a, b);
99 return values_interfere(env->birg, a, b);
102 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
103 const conflict_t *xx = x;
104 const conflict_t *yy = y;
107 return xx->n1 != yy->n1 || xx->n2 != yy->n2;
111 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
112 * The next maximum independent set build, will regard it.
114 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
116 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
118 if (get_irn_node_nr(n1) < get_irn_node_nr(n2)) {
125 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
129 * Checks if two nodes are in a conflict.
131 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
133 /* search for live range interference */
134 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
136 /* search for recoloring conflicts */
137 if (get_irn_node_nr(n1) < get_irn_node_nr(n2)) {
144 return set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)) != 0;
147 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
149 return ((const node_stat_t*)x)->irn != ((const node_stat_t*)y)->irn;
153 * Finds a node status entry of a node if existent. Otherwise return NULL
155 static INLINE const node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
158 return set_find(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
162 * Finds a node status entry of a node if existent. Otherwise it will return
163 * an initialized new entry for this node.
165 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
168 find.new_color = NO_COLOR;
169 find.pinned_local = 0;
170 return set_insert(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
174 * Returns the virtual color of a node if set before, else returns the real color.
176 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
177 const node_stat_t *found = qnode_find_node(qn, irn);
179 return found->new_color;
181 return get_irn_col(qn->ou->co, irn);
185 * Sets the virtual color of a node.
187 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
188 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
189 found->new_color = color;
190 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
194 * Checks if a node is local pinned. A node is local pinned, iff it belongs
195 * to the same optimization unit and has been optimized before the current
198 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
199 const node_stat_t *found = qnode_find_node(qn, irn);
201 return found->pinned_local;
207 * Local-pins a node, so optimizations of further nodes of the same opt unit
208 * can handle situations in which a color change would undo prior optimizations.
210 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
211 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
212 found->pinned_local = 1;
213 if (found->new_color == NO_COLOR)
214 found->new_color = get_irn_col(qn->ou->co, irn);
219 * Possible return values of qnode_color_irn()
221 #define CHANGE_SAVE NULL
222 #define CHANGE_IMPOSSIBLE (ir_node *)1
225 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
226 * other nodes are changed too, as required to preserve correctness. Function is
227 * aware of local and global pinning. Recursive.
229 * If irn == trigger the color @p col must be used. (the first recoloring)
230 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
232 * @param irn The node to set the color for
233 * @param col The color to set
234 * @param trigger The irn that caused the wish to change the color of the irn
235 * External callers must call with trigger = irn
237 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
238 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
239 * Else the first conflicting ir_node encountered is returned.
242 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
243 copy_opt_t *co = qn->ou->co;
244 const be_chordal_env_t *chordal_env = co->cenv;
245 const arch_register_class_t *cls = co->cls;
246 const arch_env_t *arch_env = co->aenv;
247 int irn_col = qnode_get_new_color(qn, irn);
248 ir_node *sub_res, *curr;
249 be_ifg_t *ifg = chordal_env->ifg;
250 void *iter = be_ifg_neighbours_iter_alloca(ifg);
253 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
255 /* If the target color is already set do nothing */
256 if (irn_col == col) {
257 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
261 /* If the irn is pinned, changing color is impossible */
262 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
263 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
267 #ifdef SEARCH_FREE_COLORS
268 /* If we resolve conflicts (recursive calls) we can use any unused color.
269 * In case of the first call @p col must be used.
271 if (irn != trigger) {
272 bitset_t *free_cols = bitset_alloca(cls->n_regs);
273 const arch_register_req_t *req;
277 /* Get all possible colors */
278 bitset_copy(free_cols, co->cenv->ignore_colors);
279 bitset_flip_all(free_cols);
281 /* Exclude colors not assignable to the irn */
282 req = arch_get_register_req(arch_env, irn, -1);
283 if (arch_register_req_is(req, limited)) {
284 bitset_t *limited = bitset_alloca(cls->n_regs);
285 rbitset_copy_to_bitset(req->limited, limited);
286 bitset_and(free_cols, limited);
289 /* Exclude the color of the irn, because it must _change_ its color */
290 bitset_clear(free_cols, irn_col);
292 /* Exclude all colors used by adjacent nodes */
293 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
294 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
296 free_col = bitset_next_set(free_cols, 0);
298 if (free_col != -1) {
299 qnode_set_new_color(qn, irn, free_col);
303 #endif /* SEARCH_FREE_COLORS */
305 /* If target color is not allocatable changing color is impossible */
306 if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
307 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
308 return CHANGE_IMPOSSIBLE;
312 * If we arrive here changing color may be possible, but there may be conflicts.
313 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
315 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
316 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
317 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
318 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
319 if (sub_res != CHANGE_SAVE) {
320 be_ifg_neighbours_break(ifg, iter);
327 * If we arrive here, all conflicts were resolved.
328 * So it is save to change this irn
330 qnode_set_new_color(qn, irn, col);
336 * Tries to set the colors for all members of this queue node;
337 * to the target color qn->color
338 * @returns 1 iff all members colors could be set
341 static int qnode_try_color(const qnode_t *qn) {
343 for (i=0; i<qn->mis_size; ++i) {
344 ir_node *test_node, *confl_node;
346 test_node = qn->mis[i];
347 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
348 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
350 if (confl_node == CHANGE_SAVE) {
351 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
352 qnode_pin_local(qn, test_node);
353 } else if (confl_node == CHANGE_IMPOSSIBLE) {
354 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
355 qnode_add_conflict(qn, test_node, test_node);
358 if (qnode_is_pinned_local(qn, confl_node)) {
359 /* changing test_node would change back a node of current ou */
360 if (confl_node == qn->ou->nodes[0]) {
361 /* Adding a conflict edge between testnode and conflnode
362 * would introduce a root -- arg interference.
363 * So remove the arg of the qn */
364 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
365 qnode_add_conflict(qn, test_node, test_node);
367 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
368 qnode_add_conflict(qn, confl_node, test_node);
371 if (pset_find_ptr(pinned_global, confl_node)) {
372 /* changing test_node would change back a node of a prior ou */
373 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
374 qnode_add_conflict(qn, test_node, test_node);
383 * Determines a maximum weighted independent set with respect to
384 * the interference and conflict edges of all nodes in a qnode.
386 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
387 ir_node **safe, **unsafe;
388 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
389 bitset_t *curr, *best;
391 int max, next, curr_weight, best_weight = 0;
393 /* assign the nodes into two groups.
394 * safe: node has no interference, hence it is in every max stable set.
395 * unsafe: node has an interference
397 safe = alloca((ou->node_count-1) * sizeof(*safe));
400 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
401 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
403 for(i=1; i<ou->node_count; ++i) {
405 for(o=1; o<ou->node_count; ++o) {
406 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
408 unsafe_costs[unsafe_count] = ou->costs[i];
409 unsafe[unsafe_count] = ou->nodes[i];
417 safe_costs += ou->costs[i];
418 safe[safe_count++] = ou->nodes[i];
424 /* now compute the best set out of the unsafe nodes*/
425 best = bitset_alloca(unsafe_count);
427 if (unsafe_count > MIS_HEUR_TRIGGER) {
428 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
429 for (i=0; i<unsafe_count; ++i) {
431 /* check if it is a stable set */
432 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
433 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
434 bitset_clear(best, i); /* clear the bit and try next one */
438 /* compute the weight */
439 bitset_foreach(best, pos)
440 best_weight += unsafe_costs[pos];
442 /* Exact Algorithm: Brute force */
443 curr = bitset_alloca(unsafe_count);
444 bitset_set_all(curr);
445 while ((max = bitset_popcnt(curr)) != 0) {
446 /* check if curr is a stable set */
447 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
448 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) */
449 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
452 /* if we arrive here, we have a stable set */
453 /* compute the weigth of the stable set*/
455 bitset_foreach(curr, pos)
456 curr_weight += unsafe_costs[pos];
459 if (curr_weight > best_weight) {
460 best_weight = curr_weight;
461 bitset_copy(best, curr);
469 /* transfer the best set into the qn */
470 qn->mis_size = 1+safe_count+bitset_popcnt(best);
471 qn->mis_costs = safe_costs+best_weight;
472 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
474 for (i=0; i<safe_count; ++i)
475 qn->mis[next++] = safe[i];
476 bitset_foreach(best, pos)
477 qn->mis[next++] = unsafe[pos];
481 * Creates a new qnode
483 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
484 qnode_t *qn = xmalloc(sizeof(*qn));
487 qn->mis = xmalloc(ou->node_count * sizeof(*qn->mis));
488 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
489 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
494 * Frees space used by a queue node
496 static INLINE void free_qnode(qnode_t *qn) {
497 del_set(qn->conflicts);
498 del_set(qn->changed_nodes);
504 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
505 * ordered by field 'size' (the size of the mis) in decreasing order.
507 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
508 struct list_head *lh;
510 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
511 /* root node is not in qnode */
516 qnode_max_ind_set(qn, ou);
517 /* do the insertion */
518 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
520 while (lh->next != &ou->queue) {
521 qnode_t *curr = list_entry_queue(lh->next);
522 if (curr->mis_costs <= qn->mis_costs)
526 list_add(&qn->queue, lh);
530 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
531 * costs of copy instructions placed during SSA-destruction and lowering.
532 * Works only for opt units with exactly 1 root node, which is the
533 * case for approximately 80% of all phi classes and 100% of register constrained
534 * nodes. (All other phi classes are reduced to this case.)
536 static void ou_optimize(unit_t *ou) {
538 qnode_t *curr = NULL, *tmp;
539 const arch_env_t *aenv = ou->co->aenv;
540 const arch_register_class_t *cls = ou->co->cls;
542 bitset_t *pos_regs = bitset_alloca(cls->n_regs);
544 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
545 for (i=0; i<ou->node_count; ++i)
546 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
549 INIT_LIST_HEAD(&ou->queue);
551 arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
553 /* exclude ingore colors */
554 bitset_andnot(pos_regs, ou->co->cenv->ignore_colors);
556 assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?");
558 /* create new qnode */
559 bitset_foreach(pos_regs, idx)
560 ou_insert_qnode(ou, new_qnode(ou, idx));
564 assert(!list_empty(&ou->queue));
565 /* get head of queue */
566 curr = list_entry_queue(ou->queue.next);
567 list_del(&curr->queue);
568 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
571 if (qnode_try_color(curr))
574 /* no success, so re-insert */
575 del_set(curr->changed_nodes);
576 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
577 ou_insert_qnode(ou, curr);
580 /* apply the best found qnode */
581 if (curr->mis_size >= 2) {
583 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
584 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));
585 /* globally pin root and all args which have the same color */
586 pset_insert_ptr(pinned_global, ou->nodes[0]);
587 for (i=1; i<ou->node_count; ++i) {
588 ir_node *irn = ou->nodes[i];
589 int nc = qnode_get_new_color(curr, irn);
590 if (nc != NO_COLOR && nc == root_col)
591 pset_insert_ptr(pinned_global, irn);
594 /* set color of all changed nodes */
595 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
596 /* NO_COLOR is possible, if we had an undo */
597 if (ns->new_color != NO_COLOR) {
598 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
599 set_irn_col(ou->co, ns->irn, ns->new_color);
604 /* free best qnode (curr) and queue */
606 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
610 int co_solve_heuristic(copy_opt_t *co) {
612 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");
616 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
617 list_for_each_entry(unit_t, curr, &co->units, units)
618 if (curr->node_count > 1)
621 del_pset(pinned_global);