2 * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet
3 * examined. A qnode has a 'target color', nodes out of the opt unit and
4 * a 'conflict graph'. A 'max indep set' is determined form these. We try to
5 * color this mis using a color-exchanging mechanism. Occuring conflicts are
6 * modeled with 'conflict edges' and the qnode is reinserted in the queue. The
7 * first qnode colored without conflicts is the best one.
12 #include "becopyopt.h"
14 #define DEBUG_LVL 0 //SET_LEVEL_1
15 static firm_dbg_module_t *dbg = NULL;
17 #define SLOTS_PINNED_GLOBAL 256
18 #define SLOTS_CONFLICTS 8
19 #define SLOTS_CHANGED_NODES 32
21 #define MIN(a,b) ((a<b)?a:b)
22 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
23 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
26 * Modeling additional conflicts between nodes. NOT live range interference
28 typedef struct _conflict_t {
29 const ir_node *n1, *n2;
33 * If an irn is changed, the changes first get stored in a node_stat_t,
34 * to allow undo of changes (=drop new data) in case of conflicts.
36 typedef struct _node_stat_t {
43 * Represents a node in the optimization queue.
45 typedef struct _qnode_t {
46 struct list_head queue; /**< chaining of unit_t->queue */
47 int color; /**< target color */
48 set *conflicts; /**< contains conflict_t's. All internal conflicts */
49 int mis_size; /**< number of nodes in the mis. */
50 const ir_node **mis; /**< the nodes of unit_t->nodes[] beeing part of the max idependent set */
51 set *changed_nodes; /**< contains node_stat_t's. */
54 pset *pinned_global; /**< optimized nodes should not be altered any more */
56 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
57 const conflict_t *xx = x;
58 const conflict_t *yy = y;
59 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
63 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
64 * The next maximum independent set build, will regard it.
66 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
68 DBG((dbg, LEVEL_4, "\t %n -- %n\n", n1, n2));
70 if ((int)n1 < (int)n2) {
77 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
81 * Checks if two nodes are in a conflict.
83 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
85 /* search for live range interference */
86 if (n1!=n2 && values_interfere(n1, n2))
88 /* search for recoloring conflicts */
89 if ((int)n1 < (int)n2) {
96 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
99 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
100 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
104 * Finds a node status entry of a node if existent. Otherwise return NULL
106 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, const ir_node *irn) {
109 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
113 * Finds a node status entry of a node if existent. Otherwise it will return
114 * an initialized new entry for this node.
116 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, const ir_node *irn) {
119 find.new_color = NO_COLOR;
120 find.pinned_local = 0;
121 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
125 * Returns the virtual color of a node if set before, else returns the real color.
127 static INLINE int qnode_get_new_color(const qnode_t *qn, const ir_node *irn) {
128 node_stat_t *found = qnode_find_node(qn, irn);
130 return found->new_color;
132 return get_irn_color(irn);
136 * Sets the virtual color of a node.
138 static INLINE void qnode_set_new_color(const qnode_t *qn, const ir_node *irn, int color) {
139 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
140 found->new_color = color;
144 * Checks if a node is local pinned. A node is local pinned, iff it belongs
145 * to the same optimization unit and has been optimized before the current
148 static INLINE int qnode_is_pinned_local(const qnode_t *qn, const ir_node *irn) {
149 node_stat_t *found = qnode_find_node(qn, irn);
151 return found->pinned_local;
157 * Local-pins a node, so optimizations of further nodes of the same opt unit
158 * can handle situations in which a color change would undo prior optimizations.
160 static INLINE void qnode_pin_local(const qnode_t *qn, const ir_node *irn) {
161 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
162 found->pinned_local = 1;
166 * Possible return values of qnode_color_irn()
168 #define CHANGE_SAVE NULL
169 #define CHANGE_IMPOSSIBLE (ir_node *)1
170 #define is_conflicting_node(n) (((int)n) > 1)
173 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
174 * other nodes are changed too, as required to preserve correctness. Function is
175 * aware of local and global pinning. Recursive.
176 * @param irn The node to set the color for
177 * @param col The color to set
178 * @param trigger The irn that caused the wish to change the color of the irn
179 * @return CHANGE_SAVE iff setting the color is possible, with all transiteve effects.
180 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
181 * Else the first conflicting ir_node encountered is returned.
183 * ASSUMPTION: Assumes that a life range of a single value can't be spilt into
184 * several smaller intervals where other values can live in between.
185 * This should be true in SSA.
187 static const ir_node *qnode_color_irn(const qnode_t *qn, const ir_node *irn, int col, const ir_node *trigger) {
189 struct obstack confl_ob;
190 ir_node **confl, *cn;
193 DBG((dbg, LEVEL_3, "\t %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
194 obstack_init(&confl_ob);
195 irn_col = qnode_get_new_color(qn, irn);
199 if (!is_possible_color(irn, col))
201 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
206 /* get all nodes which would conflict with this change */
212 irn_bl = get_nodes_block(irn);
214 /* first check for a conflicting node which is 'living in' the irns block */
217 pset *live_ins = get_live_in(irn_bl);
218 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
219 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
220 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
221 obstack_ptr_grow(&confl_ob, n);
222 pset_break(live_ins);
227 /* setup the queue of blocks. */
229 obstack_ptr_grow(&q, irn_bl);
233 /* process the queue. The code below checks for every block dominated
234 * by the irns one, and in which the irn is live, if there are
235 * conflicting nodes */
237 ir_node *curr_bl, *sub_bl;
240 curr_bl = ((ir_node **)obstack_base(&q))[out++];
242 /* Add to the result all nodes in the block, which have
243 * the target color and interfere with the irn */
244 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
245 ir_node *n = get_irn_out(curr_bl, i);
246 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
247 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
248 obstack_ptr_grow(&confl_ob, n);
252 /* If irn lives out check i-dominated blocks where the irn lives in */
254 if (is_live_out(curr_bl, irn)) {
255 dominates_for_each(curr_bl, sub_bl)
256 if (is_live_in(sub_bl, irn)) {
257 obstack_ptr_grow(&q, sub_bl);
262 obstack_free(&q, NULL);
263 obstack_ptr_grow(&confl_ob, NULL);
264 confl = (ir_node **) obstack_finish(&confl_ob);
267 /* process all nodes which would conflict with this change */
268 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
269 const ir_node *sub_res;
271 /* try to color the conflicting node cn with the color of the irn itself */
272 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
273 if (sub_res != CHANGE_SAVE) {
278 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
281 DBG((dbg, LEVEL_3, "\t %n save\n", irn));
282 obstack_free(&confl_ob, NULL);
283 qnode_set_new_color(qn, irn, col);
287 DBG((dbg, LEVEL_3, "\t %n impossible\n", irn));
288 obstack_free(&confl_ob, NULL);
292 DBG((dbg, LEVEL_3, "\t %n conflicting\n", irn));
293 obstack_free(&confl_ob, NULL);
298 * Tries to set the colors for all members of this queue node;
299 * to the target color qn->color
300 * @returns 1 iff all members colors could be set
303 static int qnode_try_color(const qnode_t *qn) {
305 for (i=0; i<qn->mis_size; ++i) {
306 const ir_node *test_node, *confl_node;
308 test_node = qn->mis[i];
309 DBG((dbg, LEVEL_3, "\t Testing %n\n", test_node));
310 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
312 if (confl_node == CHANGE_SAVE) {
313 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
314 qnode_pin_local(qn, test_node);
315 } else if (confl_node == CHANGE_IMPOSSIBLE) {
316 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
317 qnode_add_conflict(qn, test_node, test_node);
319 if (qnode_is_pinned_local(qn, confl_node)) {
320 /* changing test_node would change back a node of current ou */
321 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
322 qnode_add_conflict(qn, confl_node, test_node);
324 if (pset_find_ptr(pinned_global, confl_node)) {
325 /* changing test_node would change back a node of a prior ou */
326 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
327 qnode_add_conflict(qn, test_node, test_node);
331 if (confl_node != CHANGE_SAVE)
338 * Determines a maximum independent set with respect to the interference and
339 * conflict edges of all nodes in a qnode.
341 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
342 int all_size, curr_size, i, o;
344 const ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
346 /* all contains all nodes not removed in this qn */
348 for (i=0; i<ou->node_count; ++i)
349 if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i]))
350 all[all_size++] = ou->nodes[i];
352 /* which[i] says which element to take out of all[] and put into curr[i] */
353 which = alloca(all_size*sizeof(*which));
354 for (curr_size=0; curr_size<all_size; ++curr_size)
355 which[curr_size] = curr_size;
357 /* stores the currently examined set */
358 curr = alloca(all_size*sizeof(*curr));
360 while (1) { /* this loop will terminate because at least a single node'll be a max indep. set */
361 /* build current set */
362 for (i=0; i<curr_size; ++i)
363 curr[i] = all[which[all_size-curr_size+i]];
365 /* check current set */
366 for (i=0; i<curr_size; ++i)
367 for (o=i+1; o<curr_size; ++o)
368 if (qnode_are_conflicting(qn, curr[i], curr[o]))
371 /* We had no conflict. This is the max indep. set */
372 qn->mis_size = curr_size;
373 for (i=0; i<curr_size; ++i)
374 qn->mis[i] = curr[i];
378 /* We had a conflict. Generate next set */
379 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
381 for (i=0; i<curr_size; ++i)
382 which[all_size-curr_size+i] = i;
389 } while (!(which[pos] = (which[pos]+1) % all_size));
391 for (i=pos+1; i<all_size; ++i)
392 which[i] = MIN(which[i-1]+1, all_size-1);
395 for (i=all_size-curr_size; i<all_size-1; ++i)
396 if (which[i]>=which[i+1]) {
406 * Creates a new qnode
408 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
409 qnode_t *qn = malloc(sizeof(*qn));
411 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
412 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
413 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
418 * Frees space used by a queue node
420 static INLINE void free_qnode(qnode_t *qn) {
421 del_set(qn->conflicts);
422 del_set(qn->changed_nodes);
428 * Inserts a qnode in the sorted queue of the outimization unit. Queue is
429 * ordered by field 'size' (the size of the mis) in decreasing order.
431 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
432 struct list_head *lh;
434 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
435 /* root node is not in qnode */
440 qnode_max_ind_set(qn, ou);
442 /* set ou->mis_size for lower bound compution */
443 if (ou->mis_size < qn->mis_size)
444 ou->mis_size = qn->mis_size;
446 /* do the insertion */
447 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with size %d\n", qn->color, qn->mis_size));
449 while (lh->next != &ou->queue) {
450 qnode_t *curr = list_entry_queue(lh->next);
451 if (curr->mis_size <= qn->mis_size)
455 list_add(&qn->queue, lh);
459 * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
460 * number of copy instructions placed during SSA-destruction and lowering.
461 * Works only for opt units with exactly 1 root node, which is the
462 * case for approximately 80% of all phi classes and all register constrained
463 * nodes. (All other phi classes are reduced to this case.)
465 static void ou_optimize(unit_t *ou) {
469 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
470 for (i=0; i<ou->node_count; ++i)
471 DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
474 INIT_LIST_HEAD(&ou->queue);
475 for (i=0; i<MAX_COLORS; ++i)
476 if (is_possible_color(ou->nodes[0], i))
477 ou_insert_qnode(ou, new_qnode(ou, i));
480 while (!list_empty(&ou->queue)) {
481 /* get head of queue */
482 curr = list_entry_queue(ou->queue.next);
483 list_del(&curr->queue);
484 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with size %d\n", curr->color, curr->mis_size));
487 if (qnode_try_color(curr))
489 /* no success, so re-insert */
490 del_set(curr->changed_nodes);
491 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
492 ou_insert_qnode(ou, curr);
495 /* apply the best found qnode */
496 if (curr->mis_size >= 2) {
497 DBG((dbg, LEVEL_1, "\t Best color: %d Copies: %d/%d\n", curr->color, ou->interf+ou->node_count-curr->mis_size, ou->interf+ou->node_count-1));
498 /* globally pin root and eventually others */
499 pset_insert_ptr(pinned_global, ou->nodes[0]);
500 for (i=1; i<ou->node_count; ++i) {
501 const ir_node *irn = ou->nodes[i];
502 int nc = qnode_get_new_color(curr, irn);
503 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
504 pset_insert_ptr(pinned_global, irn);
507 /* set color of all changed nodes */
509 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
510 /* NO_COLOR is possible, if we had an undo */
511 if (ns->new_color != NO_COLOR) {
512 DBG((dbg, LEVEL_2, "\t color(%n) := %d\n", ns->irn, ns->new_color));
513 set_irn_color(ns->irn, ns->new_color);
518 /* free best qnode (curr) and queue */
520 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
524 void co_heur_opt(copy_opt_t *co) {
526 dbg = firm_dbg_register("ir.be.copyoptheur");
527 firm_dbg_set_mask(dbg, DEBUG_LVL);
529 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
531 list_for_each_entry(unit_t, curr, &co->units, units)
534 del_pset(pinned_global);