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.
41 #include "raw_bitset.h"
44 #include "becopyopt_t.h"
45 #include "becopystat.h"
46 #include "beintlive_t.h"
49 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
51 #define SEARCH_FREE_COLORS
53 #define SLOTS_PINNED_GLOBAL 64
54 #define SLOTS_CONFLICTS 8
55 #define SLOTS_CHANGED_NODES 32
57 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
58 #define HASH_CONFLICT(c) (hash_irn(c.n1) ^ hash_irn(c.n2))
61 * Modeling additional conflicts between nodes. NOT live range interference
63 typedef struct _conflict_t {
64 const ir_node *n1, *n2;
68 * If an irn is changed, the changes first get stored in a node_stat_t,
69 * to allow undo of changes (=drop new data) in case of conflicts.
71 typedef struct _node_stat_t {
78 * Represents a node in the optimization queue.
80 typedef struct _qnode_t {
81 struct list_head queue; /**< chaining of unit_t->queue */
82 const unit_t *ou; /**< the opt unit this node belongs to */
83 int color; /**< target color */
84 set *conflicts; /**< contains conflict_t's. All internal conflicts */
85 int mis_costs; /**< costs of nodes/copies in the mis. */
86 int mis_size; /**< size of the array below */
87 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
88 set *changed_nodes; /**< contains node_stat_t's. */
91 static pset *pinned_global; /**< optimized nodes should not be altered any more */
93 static INLINE int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
96 return be_ifg_connected(env->ifg, a, b);
98 return values_interfere(env->birg, a, b);
101 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
102 const conflict_t *xx = x;
103 const conflict_t *yy = y;
106 return xx->n1 != yy->n1 || xx->n2 != yy->n2;
110 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
111 * The next maximum independent set build, will regard it.
113 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
115 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
117 if (get_irn_idx(n1) < get_irn_idx(n2)) {
124 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
128 * Checks if two nodes are in a conflict.
130 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
132 /* search for live range interference */
133 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
135 /* search for recoloring conflicts */
136 if (get_irn_idx(n1) < get_irn_idx(n2)) {
143 return set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)) != 0;
146 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
148 return ((const node_stat_t*)x)->irn != ((const node_stat_t*)y)->irn;
152 * Finds a node status entry of a node if existent. Otherwise return NULL
154 static INLINE const node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
157 return set_find(qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
161 * Finds a node status entry of a node if existent. Otherwise it will return
162 * an initialized new entry for this node.
164 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
167 find.new_color = NO_COLOR;
168 find.pinned_local = 0;
169 return set_insert(qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
173 * Returns the virtual color of a node if set before, else returns the real color.
175 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
176 const node_stat_t *found = qnode_find_node(qn, irn);
178 return found->new_color;
180 return get_irn_col(qn->ou->co, irn);
184 * Sets the virtual color of a node.
186 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
187 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
188 found->new_color = color;
189 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
193 * Checks if a node is local pinned. A node is local pinned, iff it belongs
194 * to the same optimization unit and has been optimized before the current
197 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
198 const node_stat_t *found = qnode_find_node(qn, irn);
200 return found->pinned_local;
206 * Local-pins a node, so optimizations of further nodes of the same opt unit
207 * can handle situations in which a color change would undo prior optimizations.
209 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
210 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
211 found->pinned_local = 1;
212 if (found->new_color == NO_COLOR)
213 found->new_color = get_irn_col(qn->ou->co, irn);
218 * Possible return values of qnode_color_irn()
220 #define CHANGE_SAVE NULL
221 #define CHANGE_IMPOSSIBLE (ir_node *)1
224 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
225 * other nodes are changed too, as required to preserve correctness. Function is
226 * aware of local and global pinning. Recursive.
228 * If irn == trigger the color @p col must be used. (the first recoloring)
229 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
231 * @param irn The node to set the color for
232 * @param col The color to set
233 * @param trigger The irn that caused the wish to change the color of the irn
234 * External callers must call with trigger = irn
236 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
237 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
238 * Else the first conflicting ir_node encountered is returned.
241 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 ir_node *sub_res, *curr;
247 be_ifg_t *ifg = chordal_env->ifg;
248 void *iter = be_ifg_neighbours_iter_alloca(ifg);
251 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
253 /* If the target color is already set do nothing */
254 if (irn_col == col) {
255 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
259 /* If the irn is pinned, changing color is impossible */
260 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
261 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", 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);
271 const arch_register_req_t *req;
275 /* Get all possible colors */
276 bitset_copy(free_cols, co->cenv->ignore_colors);
277 bitset_flip_all(free_cols);
279 /* Exclude colors not assignable to the irn */
280 req = arch_get_register_req(irn, -1);
281 if (arch_register_req_is(req, limited)) {
282 bitset_t *limited = bitset_alloca(cls->n_regs);
283 rbitset_copy_to_bitset(req->limited, limited);
284 bitset_and(free_cols, limited);
287 /* Exclude the color of the irn, because it must _change_ its color */
288 bitset_clear(free_cols, irn_col);
290 /* Exclude all colors used by adjacent nodes */
291 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
292 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
294 free_col = bitset_next_set(free_cols, 0);
296 if (free_col != -1) {
297 qnode_set_new_color(qn, irn, free_col);
301 #endif /* SEARCH_FREE_COLORS */
303 /* If target color is not allocatable changing color is impossible */
304 if (!arch_reg_is_allocatable(co->aenv, irn, -1, arch_register_for_index(cls, col))) {
305 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
306 return CHANGE_IMPOSSIBLE;
310 * If we arrive here changing color may be possible, but there may be conflicts.
311 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
313 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
314 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
315 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
316 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
317 if (sub_res != CHANGE_SAVE) {
318 be_ifg_neighbours_break(ifg, iter);
325 * If we arrive here, all conflicts were resolved.
326 * So it is save to change this irn
328 qnode_set_new_color(qn, irn, col);
334 * Tries to set the colors for all members of this queue node;
335 * to the target color qn->color
336 * @returns 1 iff all members colors could be set
339 static int qnode_try_color(const qnode_t *qn) {
341 for (i=0; i<qn->mis_size; ++i) {
342 ir_node *test_node, *confl_node;
344 test_node = qn->mis[i];
345 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
346 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
348 if (confl_node == CHANGE_SAVE) {
349 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
350 qnode_pin_local(qn, test_node);
351 } else if (confl_node == CHANGE_IMPOSSIBLE) {
352 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
353 qnode_add_conflict(qn, test_node, test_node);
356 if (qnode_is_pinned_local(qn, confl_node)) {
357 /* changing test_node would change back a node of current ou */
358 if (confl_node == qn->ou->nodes[0]) {
359 /* Adding a conflict edge between testnode and conflnode
360 * would introduce a root -- arg interference.
361 * So remove the arg of the qn */
362 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
363 qnode_add_conflict(qn, test_node, test_node);
365 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
366 qnode_add_conflict(qn, confl_node, test_node);
369 if (pset_find_ptr(pinned_global, confl_node)) {
370 /* changing test_node would change back a node of a prior ou */
371 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
372 qnode_add_conflict(qn, test_node, test_node);
381 * Determines a maximum weighted independent set with respect to
382 * the interference and conflict edges of all nodes in a qnode.
384 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
385 ir_node **safe, **unsafe;
386 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
387 bitset_t *curr, *best;
389 int max, next, curr_weight, best_weight = 0;
391 /* assign the nodes into two groups.
392 * safe: node has no interference, hence it is in every max stable set.
393 * unsafe: node has an interference
395 safe = alloca((ou->node_count-1) * sizeof(*safe));
398 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
399 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
401 for(i=1; i<ou->node_count; ++i) {
403 for(o=1; o<ou->node_count; ++o) {
404 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
406 unsafe_costs[unsafe_count] = ou->costs[i];
407 unsafe[unsafe_count] = ou->nodes[i];
415 safe_costs += ou->costs[i];
416 safe[safe_count++] = ou->nodes[i];
422 /* now compute the best set out of the unsafe nodes*/
423 best = bitset_alloca(unsafe_count);
425 if (unsafe_count > MIS_HEUR_TRIGGER) {
426 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
427 for (i=0; i<unsafe_count; ++i) {
429 /* check if it is a stable set */
430 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
431 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
432 bitset_clear(best, i); /* clear the bit and try next one */
436 /* compute the weight */
437 bitset_foreach(best, pos)
438 best_weight += unsafe_costs[pos];
440 /* Exact Algorithm: Brute force */
441 curr = bitset_alloca(unsafe_count);
442 bitset_set_all(curr);
443 while ((max = bitset_popcnt(curr)) != 0) {
444 /* check if curr is a stable set */
445 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
446 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) */
447 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
450 /* if we arrive here, we have a stable set */
451 /* compute the weigth of the stable set*/
453 bitset_foreach(curr, pos)
454 curr_weight += unsafe_costs[pos];
457 if (curr_weight > best_weight) {
458 best_weight = curr_weight;
459 bitset_copy(best, curr);
467 /* transfer the best set into the qn */
468 qn->mis_size = 1+safe_count+bitset_popcnt(best);
469 qn->mis_costs = safe_costs+best_weight;
470 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
472 for (i=0; i<safe_count; ++i)
473 qn->mis[next++] = safe[i];
474 bitset_foreach(best, pos)
475 qn->mis[next++] = unsafe[pos];
479 * Creates a new qnode
481 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
482 qnode_t *qn = XMALLOC(qnode_t);
485 qn->mis = XMALLOCN(ir_node*, ou->node_count);
486 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
487 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
492 * Frees space used by a queue node
494 static INLINE void free_qnode(qnode_t *qn) {
495 del_set(qn->conflicts);
496 del_set(qn->changed_nodes);
502 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
503 * ordered by field 'size' (the size of the mis) in decreasing order.
505 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, *tmp;
537 const arch_env_t *aenv = ou->co->aenv;
538 const arch_register_class_t *cls = ou->co->cls;
540 bitset_t *pos_regs = bitset_alloca(cls->n_regs);
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 arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
551 /* exclude ignore colors */
552 bitset_andnot(pos_regs, ou->co->cenv->ignore_colors);
554 assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?");
556 /* create new qnode */
557 bitset_foreach(pos_regs, idx)
558 ou_insert_qnode(ou, new_qnode(ou, idx));
562 assert(!list_empty(&ou->queue));
563 /* get head of queue */
564 curr = list_entry_queue(ou->queue.next);
565 list_del(&curr->queue);
566 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
569 if (qnode_try_color(curr))
572 /* no success, so re-insert */
573 del_set(curr->changed_nodes);
574 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
575 ou_insert_qnode(ou, curr);
578 /* apply the best found qnode */
579 if (curr->mis_size >= 2) {
581 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
582 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));
583 /* globally pin root and all args which have the same color */
584 pset_insert_ptr(pinned_global, ou->nodes[0]);
585 for (i=1; i<ou->node_count; ++i) {
586 ir_node *irn = ou->nodes[i];
587 int nc = qnode_get_new_color(curr, irn);
588 if (nc != NO_COLOR && nc == root_col)
589 pset_insert_ptr(pinned_global, irn);
592 /* set color of all changed nodes */
593 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
594 /* NO_COLOR is possible, if we had an undo */
595 if (ns->new_color != NO_COLOR) {
596 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
597 set_irn_col(ou->co, ns->irn, ns->new_color);
602 /* free best qnode (curr) and queue */
604 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
608 int co_solve_heuristic(copy_opt_t *co) {
610 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");
614 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
615 list_for_each_entry(unit_t, curr, &co->units, units)
616 if (curr->node_count > 1)
619 del_pset(pinned_global);