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 /** Defines an invalid register index. */
53 #define SEARCH_FREE_COLORS
55 #define SLOTS_PINNED_GLOBAL 64
56 #define SLOTS_CONFLICTS 8
57 #define SLOTS_CHANGED_NODES 32
59 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
60 #define HASH_CONFLICT(c) (hash_irn(c.n1) ^ hash_irn(c.n2))
63 * Modeling additional conflicts between nodes. NOT live range interference
65 typedef struct conflict_t {
66 const ir_node *n1, *n2;
70 * If an irn is changed, the changes first get stored in a node_stat_t,
71 * to allow undo of changes (=drop new data) in case of conflicts.
73 typedef struct node_stat_t {
80 * Represents a node in the optimization queue.
82 typedef struct qnode_t {
83 struct list_head queue; /**< chaining of unit_t->queue */
84 const unit_t *ou; /**< the opt unit this node belongs to */
85 int color; /**< target color */
86 set *conflicts; /**< contains conflict_t's. All internal conflicts */
87 int mis_costs; /**< costs of nodes/copies in the mis. */
88 int mis_size; /**< size of the array below */
89 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
90 set *changed_nodes; /**< contains node_stat_t's. */
93 static pset *pinned_global; /**< optimized nodes should not be altered any more */
95 static inline int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
98 return be_ifg_connected(env->ifg, a, b);
100 be_lv_t *lv = be_get_irg_liveness(env->irg);
101 return be_values_interfere(lv, a, b);
105 static int set_cmp_conflict_t(const void *x, const void *y, size_t size)
107 const conflict_t *xx = (const conflict_t*)x;
108 const conflict_t *yy = (const conflict_t*)y;
111 return xx->n1 != yy->n1 || xx->n2 != yy->n2;
115 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
116 * The next maximum independent set build, will regard it.
118 static inline void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2)
121 DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2));
123 if (get_irn_idx(n1) < get_irn_idx(n2)) {
130 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
134 * Checks if two nodes are in a conflict.
136 static inline int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2)
139 /* search for live range interference */
140 if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
142 /* search for recoloring conflicts */
143 if (get_irn_idx(n1) < get_irn_idx(n2)) {
150 return set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)) != 0;
153 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size)
156 return ((const node_stat_t*)x)->irn != ((const node_stat_t*)y)->irn;
160 * Finds a node status entry of a node if existent. Otherwise return NULL
162 static inline const node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn)
166 return (const node_stat_t*)set_find(qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
170 * Finds a node status entry of a node if existent. Otherwise it will return
171 * an initialized new entry for this node.
173 static inline node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn)
177 find.new_color = NO_COLOR;
178 find.pinned_local = 0;
179 return (node_stat_t*)set_insert(qn->changed_nodes, &find, sizeof(find), hash_irn(irn));
183 * Returns the virtual color of a node if set before, else returns the real color.
185 static inline int qnode_get_new_color(const qnode_t *qn, ir_node *irn)
187 const node_stat_t *found = qnode_find_node(qn, irn);
189 return found->new_color;
191 return get_irn_col(irn);
195 * Sets the virtual color of a node.
197 static inline void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color)
199 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
200 found->new_color = color;
201 DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
205 * Checks if a node is local pinned. A node is local pinned, iff it belongs
206 * to the same optimization unit and has been optimized before the current
209 static inline int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn)
211 const node_stat_t *found = qnode_find_node(qn, irn);
213 return found->pinned_local;
219 * Local-pins a node, so optimizations of further nodes of the same opt unit
220 * can handle situations in which a color change would undo prior optimizations.
222 static inline void qnode_pin_local(const qnode_t *qn, ir_node *irn)
224 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
225 found->pinned_local = 1;
226 if (found->new_color == NO_COLOR)
227 found->new_color = get_irn_col(irn);
232 * Possible return values of qnode_color_irn()
234 #define CHANGE_SAVE NULL
235 #define CHANGE_IMPOSSIBLE (ir_node *)1
238 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
239 * other nodes are changed too, as required to preserve correctness. Function is
240 * aware of local and global pinning. Recursive.
242 * If irn == trigger the color @p col must be used. (the first recoloring)
243 * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
245 * @param irn The node to set the color for
246 * @param col The color to set
247 * @param trigger The irn that caused the wish to change the color of the irn
248 * External callers must call with trigger = irn
250 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
251 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
252 * Else the first conflicting ir_node encountered is returned.
255 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger)
257 copy_opt_t *co = qn->ou->co;
258 const be_chordal_env_t *chordal_env = co->cenv;
259 const arch_register_class_t *cls = co->cls;
260 int irn_col = qnode_get_new_color(qn, irn);
261 ir_node *sub_res, *curr;
262 be_ifg_t *ifg = chordal_env->ifg;
263 neighbours_iter_t iter;
266 DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
268 /* If the target color is already set do nothing */
269 if (irn_col == col) {
270 DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
274 /* If the irn is pinned, changing color is impossible */
275 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
276 DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
280 #ifdef SEARCH_FREE_COLORS
281 /* If we resolve conflicts (recursive calls) we can use any unused color.
282 * In case of the first call @p col must be used.
284 if (irn != trigger) {
285 bitset_t *free_cols = bitset_alloca(cls->n_regs);
286 const arch_register_req_t *req;
290 /* Get all possible colors */
291 bitset_copy(free_cols, co->cenv->allocatable_regs);
293 /* Exclude colors not assignable to the irn */
294 req = arch_get_register_req_out(irn);
295 if (arch_register_req_is(req, limited)) {
296 bitset_t *limited = bitset_alloca(cls->n_regs);
297 rbitset_copy_to_bitset(req->limited, limited);
298 bitset_and(free_cols, limited);
301 /* Exclude the color of the irn, because it must _change_ its color */
302 bitset_clear(free_cols, irn_col);
304 /* Exclude all colors used by adjacent nodes */
305 be_ifg_foreach_neighbour(ifg, &iter, irn, curr)
306 bitset_clear(free_cols, qnode_get_new_color(qn, curr));
308 free_col = bitset_next_set(free_cols, 0);
310 if (free_col != -1) {
311 qnode_set_new_color(qn, irn, free_col);
315 #endif /* SEARCH_FREE_COLORS */
317 /* If target color is not allocatable changing color is impossible */
318 if (!arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, col))) {
319 DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
320 return CHANGE_IMPOSSIBLE;
324 * If we arrive here changing color may be possible, but there may be conflicts.
325 * Try to color all conflicting nodes 'curr' with the color of the irn itself.
327 be_ifg_foreach_neighbour(ifg, &iter, irn, curr) {
328 DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
329 if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
330 sub_res = qnode_color_irn(qn, curr, irn_col, irn);
331 if (sub_res != CHANGE_SAVE) {
332 be_ifg_neighbours_break(&iter);
339 * If we arrive here, all conflicts were resolved.
340 * So it is save to change this irn
342 qnode_set_new_color(qn, irn, col);
348 * Tries to set the colors for all members of this queue node;
349 * to the target color qn->color
350 * @returns 1 iff all members colors could be set
353 static int qnode_try_color(const qnode_t *qn)
356 for (i=0; i<qn->mis_size; ++i) {
357 ir_node *test_node, *confl_node;
359 test_node = qn->mis[i];
360 DBG((dbg, LEVEL_3, "\t Testing %+F\n", test_node));
361 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
363 if (confl_node == CHANGE_SAVE) {
364 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
365 qnode_pin_local(qn, test_node);
366 } else if (confl_node == CHANGE_IMPOSSIBLE) {
367 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
368 qnode_add_conflict(qn, test_node, test_node);
371 if (qnode_is_pinned_local(qn, confl_node)) {
372 /* changing test_node would change back a node of current ou */
373 if (confl_node == qn->ou->nodes[0]) {
374 /* Adding a conflict edge between testnode and conflnode
375 * would introduce a root -- arg interference.
376 * So remove the arg of the qn */
377 DBG((dbg, LEVEL_3, "\t Conflicting local with phi --> remove from qnode\n"));
378 qnode_add_conflict(qn, test_node, test_node);
380 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
381 qnode_add_conflict(qn, confl_node, test_node);
384 if (pset_find_ptr(pinned_global, confl_node)) {
385 /* changing test_node would change back a node of a prior ou */
386 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
387 qnode_add_conflict(qn, test_node, test_node);
396 * Determines a maximum weighted independent set with respect to
397 * the interference and conflict edges of all nodes in a qnode.
399 static inline void qnode_max_ind_set(qnode_t *qn, const unit_t *ou)
401 ir_node **safe, **unsafe;
402 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
403 bitset_t *curr, *best;
405 int next, curr_weight, best_weight = 0;
407 /* assign the nodes into two groups.
408 * safe: node has no interference, hence it is in every max stable set.
409 * unsafe: node has an interference
411 safe = ALLOCAN(ir_node*, ou->node_count - 1);
414 unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
415 unsafe_costs = ALLOCAN(int, ou->node_count - 1);
417 for (i=1; i<ou->node_count; ++i) {
419 for (o=1; o<ou->node_count; ++o) {
420 if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
422 unsafe_costs[unsafe_count] = ou->costs[i];
423 unsafe[unsafe_count] = ou->nodes[i];
431 safe_costs += ou->costs[i];
432 safe[safe_count++] = ou->nodes[i];
438 /* now compute the best set out of the unsafe nodes*/
439 best = bitset_alloca(unsafe_count);
441 if (unsafe_count > MIS_HEUR_TRIGGER) {
442 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
443 for (i=0; i<unsafe_count; ++i) {
445 /* check if it is a stable set */
446 for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
447 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
448 bitset_clear(best, i); /* clear the bit and try next one */
452 /* compute the weight */
453 bitset_foreach(best, pos)
454 best_weight += unsafe_costs[pos];
456 /* Exact Algorithm: Brute force */
457 curr = bitset_alloca(unsafe_count);
458 bitset_set_all(curr);
459 while (!bitset_is_empty(curr)) {
460 /* check if curr is a stable set */
461 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
462 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) */
463 if (qnode_are_conflicting(qn, unsafe[i], unsafe[o]))
466 /* if we arrive here, we have a stable set */
467 /* compute the weight of the stable set*/
469 bitset_foreach(curr, pos)
470 curr_weight += unsafe_costs[pos];
473 if (curr_weight > best_weight) {
474 best_weight = curr_weight;
475 bitset_copy(best, curr);
483 /* transfer the best set into the qn */
484 qn->mis_size = 1+safe_count+bitset_popcount(best);
485 qn->mis_costs = safe_costs+best_weight;
486 qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
488 for (i=0; i<safe_count; ++i)
489 qn->mis[next++] = safe[i];
490 bitset_foreach(best, pos)
491 qn->mis[next++] = unsafe[pos];
495 * Creates a new qnode
497 static inline qnode_t *new_qnode(const unit_t *ou, int color)
499 qnode_t *qn = XMALLOC(qnode_t);
502 qn->mis = XMALLOCN(ir_node*, ou->node_count);
503 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
504 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
509 * Frees space used by a queue node
511 static inline void free_qnode(qnode_t *qn)
513 del_set(qn->conflicts);
514 del_set(qn->changed_nodes);
520 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
521 * ordered by field 'size' (the size of the mis) in decreasing order.
523 static inline void ou_insert_qnode(unit_t *ou, qnode_t *qn)
525 struct list_head *lh;
527 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
528 /* root node is not in qnode */
533 qnode_max_ind_set(qn, ou);
534 /* do the insertion */
535 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
537 while (lh->next != &ou->queue) {
538 qnode_t *curr = list_entry_queue(lh->next);
539 if (curr->mis_costs <= qn->mis_costs)
543 list_add(&qn->queue, lh);
547 * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
548 * costs of copy instructions placed during SSA-destruction and lowering.
549 * Works only for opt units with exactly 1 root node, which is the
550 * case for approximately 80% of all phi classes and 100% of register constrained
551 * nodes. (All other phi classes are reduced to this case.)
553 static void ou_optimize(unit_t *ou)
555 qnode_t *curr = NULL;
557 const arch_register_req_t *req;
558 bitset_t const* allocatable_regs;
563 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
564 for (i=0; i<ou->node_count; ++i)
565 DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
568 INIT_LIST_HEAD(&ou->queue);
570 req = arch_get_register_req_out(ou->nodes[0]);
571 allocatable_regs = ou->co->cenv->allocatable_regs;
572 n_regs = req->cls->n_regs;
573 if (arch_register_req_is(req, limited)) {
574 unsigned const* limited = req->limited;
576 for (idx = 0; idx != n_regs; ++idx) {
577 if (!bitset_is_set(allocatable_regs, idx))
579 if (!rbitset_is_set(limited, idx))
582 ou_insert_qnode(ou, new_qnode(ou, idx));
585 for (idx = 0; idx != n_regs; ++idx) {
586 if (!bitset_is_set(allocatable_regs, idx))
589 ou_insert_qnode(ou, new_qnode(ou, idx));
595 assert(!list_empty(&ou->queue));
596 /* get head of queue */
597 curr = list_entry_queue(ou->queue.next);
598 list_del(&curr->queue);
599 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
602 if (qnode_try_color(curr))
605 /* no success, so re-insert */
606 del_set(curr->changed_nodes);
607 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
608 ou_insert_qnode(ou, curr);
611 /* apply the best found qnode */
612 if (curr->mis_size >= 2) {
614 int root_col = qnode_get_new_color(curr, ou->nodes[0]);
615 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));
616 /* globally pin root and all args which have the same color */
617 pset_insert_ptr(pinned_global, ou->nodes[0]);
618 for (i=1; i<ou->node_count; ++i) {
619 ir_node *irn = ou->nodes[i];
620 int nc = qnode_get_new_color(curr, irn);
621 if (nc != NO_COLOR && nc == root_col)
622 pset_insert_ptr(pinned_global, irn);
625 /* set color of all changed nodes */
626 for (ns = (node_stat_t*)set_first(curr->changed_nodes); ns != NULL;
627 ns = (node_stat_t*)set_next(curr->changed_nodes)) {
628 /* NO_COLOR is possible, if we had an undo */
629 if (ns->new_color != NO_COLOR) {
630 DBG((dbg, LEVEL_1, "\t color(%+F) := %d\n", ns->irn, ns->new_color));
631 set_irn_col(ou->co, ns->irn, ns->new_color);
636 /* free best qnode (curr) and queue */
638 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
643 * Solves the problem using a heuristic approach
644 * Uses the OU data structure
646 int co_solve_heuristic(copy_opt_t *co)
652 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
653 list_for_each_entry(unit_t, curr, &co->units, units)
654 if (curr->node_count > 1)
657 del_pset(pinned_global);
661 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur)
662 void be_init_copyheur(void)
664 static co_algo_info copyheur = {
665 co_solve_heuristic, 0
668 be_register_copyopt("heur1", ©heur);
669 FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur");