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
21 * Author: Daniel Grund
23 * Copyright: (c) Universitaet Karlsruhe
24 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
26 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
27 * examined. A qnode has a 'target color', nodes out of the opt unit and
28 * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges'
29 * A 'max indep set' is determined form these. We try to color this mis using a
30 * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges'
31 * and the qnode is reinserted in the queue. The first qnode colored without
32 * conflicts is the best one.
40 #include "becopyopt_t.h"
41 #include "becopystat.h"
42 #include "benodesets.h"
44 #include "raw_bitset.h"
47 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
49 #define SEARCH_FREE_COLORS
51 #define SLOTS_PINNED_GLOBAL 64
52 #define SLOTS_CONFLICTS 8
53 #define SLOTS_CHANGED_NODES 32
55 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
56 #define HASH_CONFLICT(c) (nodeset_hash(c.n1) ^ nodeset_hash(c.n2))
59 * Modeling additional conflicts between nodes. NOT live range interference
61 typedef struct _conflict_t {
62 const ir_node *n1, *n2;
66 * If an irn is changed, the changes first get stored in a node_stat_t,
67 * to allow undo of changes (=drop new data) in case of conflicts.
69 typedef struct _node_stat_t {
76 * Represents a node in the optimization queue.
78 typedef struct _qnode_t {
79 struct list_head queue; /**< chaining of unit_t->queue */
80 const unit_t *ou; /**< the opt unit this qnode belongs to */
81 int color; /**< target color */
82 set *conflicts; /**< contains conflict_t's. All internal conflicts */
83 int mis_costs; /**< costs of nodes/copies in the mis. */
84 int mis_size; /**< size of the array below */
85 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
86 set *changed_nodes; /**< contains node_stat_t's. */
89 static pset *pinned_global; /**< optimized nodes should not be altered any more */
91 static INLINE int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
94 return be_ifg_connected(env->ifg, a, b);
96 return values_interfere(env->birg->lv, a, b);
99 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
100 const conflict_t *xx = x;
101 const conflict_t *yy = y;
102 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
106 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
107 * The next maximum independent set build, will regard it.
109 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
111 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
113 if ((int)n1 < (int)n2) {
120 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
124 * Checks if two nodes are in a conflict.
126 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
128 /* search for live range interference */
129 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
131 /* search for recoloring conflicts */
132 if ((int)n1 < (int)n2) {
139 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
142 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
143 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
147 * Finds a node status entry of a node if existent. Otherwise return NULL
149 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
152 return set_find(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
156 * Finds a node status entry of a node if existent. Otherwise it will return
157 * an initialized new entry for this node.
159 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(qn->changed_nodes, &find, sizeof(find), nodeset_hash(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) {
171 node_stat_t *found = qnode_find_node(qn, irn);
173 return found->new_color;
175 return get_irn_col(qn->ou->co, irn);
179 * Sets the virtual color of a node.
181 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
182 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
183 found->new_color = color;
184 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
188 * Checks if a node is local pinned. A node is local pinned, iff it belongs
189 * to the same optimization unit and has been optimized before the current
192 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
193 node_stat_t *found = qnode_find_node(qn, irn);
195 return found->pinned_local;
201 * Local-pins a node, so optimizations of further nodes of the same opt unit
202 * can handle situations in which a color change would undo prior optimizations.
204 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
205 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
206 found->pinned_local = 1;
207 if (found->new_color == NO_COLOR)
208 found->new_color = get_irn_col(qn->ou->co, irn);
213 * Possible return values of qnode_color_irn()
215 #define CHANGE_SAVE NULL
216 #define CHANGE_IMPOSSIBLE (ir_node *)1
217 #define is_conflicting_node(n) (((int)n) > 1)
220 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
221 * other nodes are changed too, as required to preserve correctness. Function is
222 * aware of local and global pinning. Recursive.
224 * If irn == trigger the color @p col must be used. (the first recoloring)
225 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
227 * @param irn The node to set the color for
228 * @param col The color to set
229 * @param trigger The irn that caused the wish to change the color of the irn
230 * External callers must call with trigger = irn
232 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
233 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
234 * Else the first conflicting ir_node encountered is returned.
237 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
238 copy_opt_t *co = qn->ou->co;
239 const be_chordal_env_t *chordal_env = co->cenv;
240 const arch_register_class_t *cls = co->cls;
241 const arch_env_t *arch_env = co->aenv;
242 int irn_col = qnode_get_new_color(qn, irn);
243 ir_node *sub_res, *curr;
244 be_ifg_t *ifg = chordal_env->ifg;
245 void *iter = be_ifg_neighbours_iter_alloca(ifg);
248 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
250 /* If the target color is already set do nothing */
251 if (irn_col == col) {
252 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
256 /* If the irn is pinned, changing color is impossible */
257 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
258 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
262 #ifdef SEARCH_FREE_COLORS
263 /* If we resolve conflicts (recursive calls) we can use any unused color.
264 * In case of the first call @p col must be used.
266 if (irn != trigger) {
267 bitset_t *free_cols = bitset_alloca(cls->n_regs);
268 const arch_register_req_t *req;
272 /* Get all possible colors */
273 bitset_copy(free_cols, co->cenv->ignore_colors);
274 bitset_flip_all(free_cols);
276 /* Exclude colors not assignable to the irn */
277 req = arch_get_register_req(arch_env, irn, -1);
278 if (arch_register_req_is(req, limited)) {
279 bitset_t *limited = bitset_alloca(cls->n_regs);
280 rbitset_copy_to_bitset(req->limited, limited);
281 bitset_and(free_cols, limited);
284 /* Exclude the color of the irn, because it must _change_ its color */
285 bitset_clear(free_cols, irn_col);
287 /* Exclude all colors used by adjacent nodes */
288 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
289 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
291 free_col = bitset_next_set(free_cols, 0);
293 if (free_col != -1) {
294 qnode_set_new_color(qn, irn, free_col);
298 #endif /* SEARCH_FREE_COLORS */
300 /* If target color is not allocatable changing color is impossible */
301 if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
302 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
303 return CHANGE_IMPOSSIBLE;
307 * If we arrive here changing color may be possible, but there may be conflicts.
308 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
310 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
311 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
312 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
313 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
314 if (sub_res != CHANGE_SAVE) {
315 be_ifg_neighbours_break(ifg, iter);
322 * If we arrive here, all conflicts were resolved.
323 * So it is save to change this irn
325 qnode_set_new_color(qn, irn, col);
331 * Tries to set the colors for all members of this queue node;
332 * to the target color qn->color
333 * @returns 1 iff all members colors could be set
336 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);
352 if (qnode_is_pinned_local(qn, confl_node)) {
353 /* changing test_node would change back a node of current ou */
354 if (confl_node == qn->ou->nodes[0]) {
355 /* Adding a conflict edge between testnode and conflnode
356 * would introduce a root -- arg interference.
357 * So remove the arg of the qn */
358 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
359 qnode_add_conflict(qn, test_node, test_node);
361 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
362 qnode_add_conflict(qn, confl_node, test_node);
365 if (pset_find_ptr(pinned_global, confl_node)) {
366 /* changing test_node would change back a node of a prior ou */
367 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
368 qnode_add_conflict(qn, test_node, test_node);
372 if (confl_node != CHANGE_SAVE)
379 * Determines a maximum weighted independent set with respect to
380 * the interference and conflict edges of all nodes in a qnode.
382 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 max, next, pos, 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 = alloca((ou->node_count-1) * sizeof(*safe));
395 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
396 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
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 ((max = bitset_popcnt(curr)) != 0) {
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 weigth 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_popcnt(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) {
479 qnode_t *qn = xmalloc(sizeof(*qn));
482 qn->mis = xmalloc(ou->node_count * sizeof(*qn->mis));
483 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
484 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
489 * Frees space used by a queue node
491 static INLINE void free_qnode(qnode_t *qn) {
492 del_set(qn->conflicts);
493 del_set(qn->changed_nodes);
499 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
500 * ordered by field 'size' (the size of the mis) in decreasing order.
502 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
503 struct list_head *lh;
505 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
506 /* root node is not in qnode */
511 qnode_max_ind_set(qn, ou);
512 /* do the insertion */
513 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
515 while (lh->next != &ou->queue) {
516 qnode_t *curr = list_entry_queue(lh->next);
517 if (curr->mis_costs <= qn->mis_costs)
521 list_add(&qn->queue, lh);
525 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
526 * costs of copy instructions placed during SSA-destruction and lowering.
527 * Works only for opt units with exactly 1 root node, which is the
528 * case for approximately 80% of all phi classes and 100% of register constrained
529 * nodes. (All other phi classes are reduced to this case.)
531 static void ou_optimize(unit_t *ou) {
533 qnode_t *curr = NULL, *tmp;
534 const arch_env_t *aenv = ou->co->aenv;
535 const arch_register_class_t *cls = ou->co->cls;
536 bitset_t *pos_regs = bitset_alloca(cls->n_regs);
538 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
539 for (i=0; i<ou->node_count; ++i)
540 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
543 INIT_LIST_HEAD(&ou->queue);
545 arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
547 /* exclude ingore colors */
548 bitset_andnot(pos_regs, ou->co->cenv->ignore_colors);
550 assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?");
552 /* create new qnode */
553 bitset_foreach(pos_regs, i)
554 ou_insert_qnode(ou, new_qnode(ou, i));
557 while (!list_empty(&ou->queue)) {
558 /* get head of queue */
559 curr = list_entry_queue(ou->queue.next);
560 list_del(&curr->queue);
561 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
564 if (qnode_try_color(curr))
567 /* no success, so re-insert */
568 del_set(curr->changed_nodes);
569 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
570 ou_insert_qnode(ou, curr);
573 /* apply the best found qnode */
574 if (curr->mis_size >= 2) {
576 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
577 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));
578 /* globally pin root and all args which have the same color */
579 pset_insert_ptr(pinned_global, ou->nodes[0]);
580 for (i=1; i<ou->node_count; ++i) {
581 ir_node *irn = ou->nodes[i];
582 int nc = qnode_get_new_color(curr, irn);
583 if (nc != NO_COLOR && nc == root_col)
584 pset_insert_ptr(pinned_global, irn);
587 /* set color of all changed nodes */
588 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
589 /* NO_COLOR is possible, if we had an undo */
590 if (ns->new_color != NO_COLOR) {
591 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
592 set_irn_col(ou->co, ns->irn, ns->new_color);
597 /* free best qnode (curr) and queue */
599 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
603 int co_solve_heuristic(copy_opt_t *co) {
605 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");
609 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
610 list_for_each_entry(unit_t, curr, &co->units, units)
611 if (curr->node_count > 1)
614 del_pset(pinned_global);