4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
7 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
8 * examined. A qnode has a 'target color', nodes out of the opt unit and
9 * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges'
10 * A 'max indep set' is determined form these. We try to color this mis using a
11 * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges'
12 * and the qnode is reinserted in the queue. The first qnode colored without
13 * conflicts is the best one.
28 #include "becopyopt_t.h"
29 #include "becopystat.h"
30 #include "benodesets.h"
32 #include "raw_bitset.h"
34 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
36 #define SEARCH_FREE_COLORS
38 #define SLOTS_PINNED_GLOBAL 64
39 #define SLOTS_CONFLICTS 8
40 #define SLOTS_CHANGED_NODES 32
42 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
43 #define HASH_CONFLICT(c) (nodeset_hash(c.n1) ^ nodeset_hash(c.n2))
46 * Modeling additional conflicts between nodes. NOT live range interference
48 typedef struct _conflict_t {
49 const ir_node *n1, *n2;
53 * If an irn is changed, the changes first get stored in a node_stat_t,
54 * to allow undo of changes (=drop new data) in case of conflicts.
56 typedef struct _node_stat_t {
63 * Represents a node in the optimization queue.
65 typedef struct _qnode_t {
66 struct list_head queue; /**< chaining of unit_t->queue */
67 const unit_t *ou; /**< the opt unit this qnode belongs to */
68 int color; /**< target color */
69 set *conflicts; /**< contains conflict_t's. All internal conflicts */
70 int mis_costs; /**< costs of nodes/copies in the mis. */
71 int mis_size; /**< size of the array below */
72 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
73 set *changed_nodes; /**< contains node_stat_t's. */
76 static pset *pinned_global; /**< optimized nodes should not be altered any more */
78 static INLINE int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
81 return be_ifg_connected(env->ifg, a, b);
83 return values_interfere(env->birg->lv, a, b);
86 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
87 const conflict_t *xx = x;
88 const conflict_t *yy = y;
89 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
93 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
94 * The next maximum independent set build, will regard it.
96 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
98 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
100 if ((int)n1 < (int)n2) {
107 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
111 * Checks if two nodes are in a conflict.
113 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
115 /* search for live range interference */
116 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
118 /* search for recoloring conflicts */
119 if ((int)n1 < (int)n2) {
126 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
129 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
130 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
134 * Finds a node status entry of a node if existent. Otherwise return NULL
136 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
139 return set_find(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
143 * Finds a node status entry of a node if existent. Otherwise it will return
144 * an initialized new entry for this node.
146 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
149 find.new_color = NO_COLOR;
150 find.pinned_local = 0;
151 return set_insert(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn));
155 * Returns the virtual color of a node if set before, else returns the real color.
157 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
158 node_stat_t *found = qnode_find_node(qn, irn);
160 return found->new_color;
162 return get_irn_col(qn->ou->co, irn);
166 * Sets the virtual color of a node.
168 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
169 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
170 found->new_color = color;
171 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
175 * Checks if a node is local pinned. A node is local pinned, iff it belongs
176 * to the same optimization unit and has been optimized before the current
179 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
180 node_stat_t *found = qnode_find_node(qn, irn);
182 return found->pinned_local;
188 * Local-pins a node, so optimizations of further nodes of the same opt unit
189 * can handle situations in which a color change would undo prior optimizations.
191 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
192 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
193 found->pinned_local = 1;
194 if (found->new_color == NO_COLOR)
195 found->new_color = get_irn_col(qn->ou->co, irn);
200 * Possible return values of qnode_color_irn()
202 #define CHANGE_SAVE NULL
203 #define CHANGE_IMPOSSIBLE (ir_node *)1
204 #define is_conflicting_node(n) (((int)n) > 1)
207 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
208 * other nodes are changed too, as required to preserve correctness. Function is
209 * aware of local and global pinning. Recursive.
211 * If irn == trigger the color @p col must be used. (the first recoloring)
212 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
214 * @param irn The node to set the color for
215 * @param col The color to set
216 * @param trigger The irn that caused the wish to change the color of the irn
217 * External callers must call with trigger = irn
219 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
220 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
221 * Else the first conflicting ir_node encountered is returned.
224 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
225 copy_opt_t *co = qn->ou->co;
226 const be_chordal_env_t *chordal_env = co->cenv;
227 const arch_register_class_t *cls = co->cls;
228 const arch_env_t *arch_env = co->aenv;
229 int irn_col = qnode_get_new_color(qn, irn);
230 ir_node *sub_res, *curr;
231 be_ifg_t *ifg = chordal_env->ifg;
232 void *iter = be_ifg_neighbours_iter_alloca(ifg);
235 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
237 /* If the target color is already set do nothing */
238 if (irn_col == col) {
239 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
243 /* If the irn is pinned, changing color is impossible */
244 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
245 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
249 #ifdef SEARCH_FREE_COLORS
250 /* If we resolve conflicts (recursive calls) we can use any unused color.
251 * In case of the first call @p col must be used.
253 if (irn != trigger) {
254 bitset_t *free_cols = bitset_alloca(cls->n_regs);
255 const arch_register_req_t *req;
259 /* Get all possible colors */
260 bitset_copy(free_cols, co->cenv->ignore_colors);
261 bitset_flip_all(free_cols);
263 /* Exclude colors not assignable to the irn */
264 req = arch_get_register_req(arch_env, irn, -1);
265 if (arch_register_req_is(req, limited)) {
266 bitset_t *limited = bitset_alloca(cls->n_regs);
267 rbitset_copy_to_bitset(req->limited, limited);
268 bitset_and(free_cols, limited);
271 /* Exclude the color of the irn, because it must _change_ its color */
272 bitset_clear(free_cols, irn_col);
274 /* Exclude all colors used by adjacent nodes */
275 be_ifg_foreach_neighbour(ifg, iter, irn, curr)
276 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
278 free_col = bitset_next_set(free_cols, 0);
280 if (free_col != -1) {
281 qnode_set_new_color(qn, irn, free_col);
285 #endif /* SEARCH_FREE_COLORS */
287 /* If target color is not allocatable changing color is impossible */
288 if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
289 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
290 return CHANGE_IMPOSSIBLE;
294 * If we arrive here changing color may be possible, but there may be conflicts.
295 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
297 be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
298 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
299 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
300 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
301 if (sub_res != CHANGE_SAVE) {
302 be_ifg_neighbours_break(ifg, iter);
309 * If we arrive here, all conflicts were resolved.
310 * So it is save to change this irn
312 qnode_set_new_color(qn, irn, col);
318 * Tries to set the colors for all members of this queue node;
319 * to the target color qn->color
320 * @returns 1 iff all members colors could be set
323 static int qnode_try_color(const qnode_t *qn) {
325 for (i=0; i<qn->mis_size; ++i) {
326 ir_node *test_node, *confl_node;
328 test_node = qn->mis[i];
329 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
330 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
332 if (confl_node == CHANGE_SAVE) {
333 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
334 qnode_pin_local(qn, test_node);
335 } else if (confl_node == CHANGE_IMPOSSIBLE) {
336 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
337 qnode_add_conflict(qn, test_node, test_node);
339 if (qnode_is_pinned_local(qn, confl_node)) {
340 /* changing test_node would change back a node of current ou */
341 if (confl_node == qn->ou->nodes[0]) {
342 /* Adding a conflict edge between testnode and conflnode
343 * would introduce a root -- arg interference.
344 * So remove the arg of the qn */
345 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
346 qnode_add_conflict(qn, test_node, test_node);
348 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
349 qnode_add_conflict(qn, confl_node, test_node);
352 if (pset_find_ptr(pinned_global, confl_node)) {
353 /* changing test_node would change back a node of a prior ou */
354 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
355 qnode_add_conflict(qn, test_node, test_node);
359 if (confl_node != CHANGE_SAVE)
366 * Determines a maximum weighted independent set with respect to
367 * the interference and conflict edges of all nodes in a qnode.
369 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
370 ir_node **safe, **unsafe;
371 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
372 bitset_t *curr, *best;
373 int max, next, pos, curr_weight, best_weight = 0;
375 /* assign the nodes into two groups.
376 * safe: node has no interference, hence it is in every max stable set.
377 * unsafe: node has an interference
379 safe = alloca((ou->node_count-1) * sizeof(*safe));
382 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
383 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
385 for(i=1; i<ou->node_count; ++i) {
387 for(o=1; o<ou->node_count; ++o) {
388 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
390 unsafe_costs[unsafe_count] = ou->costs[i];
391 unsafe[unsafe_count] = ou->nodes[i];
399 safe_costs += ou->costs[i];
400 safe[safe_count++] = ou->nodes[i];
406 /* now compute the best set out of the unsafe nodes*/
407 best = bitset_alloca(unsafe_count);
409 if (unsafe_count > MIS_HEUR_TRIGGER) {
410 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
411 for (i=0; i<unsafe_count; ++i) {
413 /* check if it is a stable set */
414 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
415 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
416 bitset_clear(best, i); /* clear the bit and try next one */
420 /* compute the weight */
421 bitset_foreach(best, pos)
422 best_weight += unsafe_costs[pos];
424 /* Exact Algorithm: Brute force */
425 curr = bitset_alloca(unsafe_count);
426 bitset_set_all(curr);
427 while ((max = bitset_popcnt(curr)) != 0) {
428 /* check if curr is a stable set */
429 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
430 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) */
431 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
434 /* if we arrive here, we have a stable set */
435 /* compute the weigth of the stable set*/
437 bitset_foreach(curr, pos)
438 curr_weight += unsafe_costs[pos];
441 if (curr_weight > best_weight) {
442 best_weight = curr_weight;
443 bitset_copy(best, curr);
451 /* transfer the best set into the qn */
452 qn->mis_size = 1+safe_count+bitset_popcnt(best);
453 qn->mis_costs = safe_costs+best_weight;
454 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
456 for (i=0; i<safe_count; ++i)
457 qn->mis[next++] = safe[i];
458 bitset_foreach(best, pos)
459 qn->mis[next++] = unsafe[pos];
463 * Creates a new qnode
465 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
466 qnode_t *qn = xmalloc(sizeof(*qn));
469 qn->mis = xmalloc(ou->node_count * sizeof(*qn->mis));
470 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
471 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
476 * Frees space used by a queue node
478 static INLINE void free_qnode(qnode_t *qn) {
479 del_set(qn->conflicts);
480 del_set(qn->changed_nodes);
486 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
487 * ordered by field 'size' (the size of the mis) in decreasing order.
489 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
490 struct list_head *lh;
492 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
493 /* root node is not in qnode */
498 qnode_max_ind_set(qn, ou);
499 /* do the insertion */
500 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
502 while (lh->next != &ou->queue) {
503 qnode_t *curr = list_entry_queue(lh->next);
504 if (curr->mis_costs <= qn->mis_costs)
508 list_add(&qn->queue, lh);
512 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
513 * costs of copy instructions placed during SSA-destruction and lowering.
514 * Works only for opt units with exactly 1 root node, which is the
515 * case for approximately 80% of all phi classes and 100% of register constrained
516 * nodes. (All other phi classes are reduced to this case.)
518 static void ou_optimize(unit_t *ou) {
520 qnode_t *curr = NULL, *tmp;
521 const arch_env_t *aenv = ou->co->aenv;
522 const arch_register_class_t *cls = ou->co->cls;
523 bitset_t *pos_regs = bitset_alloca(cls->n_regs);
525 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
526 for (i=0; i<ou->node_count; ++i)
527 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
530 INIT_LIST_HEAD(&ou->queue);
532 arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
534 /* exclude ingore colors */
535 bitset_andnot(pos_regs, ou->co->cenv->ignore_colors);
537 assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?");
539 /* create new qnode */
540 bitset_foreach(pos_regs, i)
541 ou_insert_qnode(ou, new_qnode(ou, i));
544 while (!list_empty(&ou->queue)) {
545 /* get head of queue */
546 curr = list_entry_queue(ou->queue.next);
547 list_del(&curr->queue);
548 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
551 if (qnode_try_color(curr))
554 /* no success, so re-insert */
555 del_set(curr->changed_nodes);
556 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
557 ou_insert_qnode(ou, curr);
560 /* apply the best found qnode */
561 if (curr->mis_size >= 2) {
563 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
564 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));
565 /* globally pin root and all args which have the same color */
566 pset_insert_ptr(pinned_global, ou->nodes[0]);
567 for (i=1; i<ou->node_count; ++i) {
568 ir_node *irn = ou->nodes[i];
569 int nc = qnode_get_new_color(curr, irn);
570 if (nc != NO_COLOR && nc == root_col)
571 pset_insert_ptr(pinned_global, irn);
574 /* set color of all changed nodes */
575 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
576 /* NO_COLOR is possible, if we had an undo */
577 if (ns->new_color != NO_COLOR) {
578 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
579 set_irn_col(ou->co, ns->irn, ns->new_color);
584 /* free best qnode (curr) and queue */
586 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
590 int co_solve_heuristic(copy_opt_t *co) {
592 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");
596 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
597 list_for_each_entry(unit_t, curr, &co->units, units)
598 if (curr->node_count > 1)
601 del_pset(pinned_global);