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"
13 #include "becopystat.h"
15 #define DEBUG_LVL 0 //SET_LEVEL_1
16 static firm_dbg_module_t *dbg = NULL;
18 #define SLOTS_PINNED_GLOBAL 256
19 #define SLOTS_CONFLICTS 8
20 #define SLOTS_CHANGED_NODES 32
22 #define MIN(a,b) ((a<b)?a:b)
23 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
24 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
27 * Modeling additional conflicts between nodes. NOT live range interference
29 typedef struct _conflict_t {
30 const ir_node *n1, *n2;
34 * If an irn is changed, the changes first get stored in a node_stat_t,
35 * to allow undo of changes (=drop new data) in case of conflicts.
37 typedef struct _node_stat_t {
44 * Represents a node in the optimization queue.
46 typedef struct _qnode_t {
47 struct list_head queue; /**< chaining of unit_t->queue */
48 const unit_t *ou; /**< the opt unit this qnode belongs to */
49 int color; /**< target color */
50 set *conflicts; /**< contains conflict_t's. All internal conflicts */
51 int mis_size; /**< number of nodes in the mis. */
52 const ir_node **mis; /**< the nodes of unit_t->nodes[] beeing part of the max idependent set */
53 set *changed_nodes; /**< contains node_stat_t's. */
56 pset *pinned_global; /**< optimized nodes should not be altered any more */
58 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
59 const conflict_t *xx = x;
60 const conflict_t *yy = y;
61 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
65 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
66 * The next maximum independent set build, will regard it.
68 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
70 DBG((dbg, LEVEL_4, "\t %n -- %n\n", n1, n2));
72 if ((int)n1 < (int)n2) {
79 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
83 * Checks if two nodes are in a conflict.
85 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
87 /* search for live range interference */
88 if (n1!=n2 && values_interfere(n1, n2))
90 /* search for recoloring conflicts */
91 if ((int)n1 < (int)n2) {
98 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
101 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
102 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
106 * Finds a node status entry of a node if existent. Otherwise return NULL
108 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, const ir_node *irn) {
111 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
115 * Finds a node status entry of a node if existent. Otherwise it will return
116 * an initialized new entry for this node.
118 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, const ir_node *irn) {
121 find.new_color = NO_COLOR;
122 find.pinned_local = 0;
123 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
127 * Returns the virtual color of a node if set before, else returns the real color.
129 static INLINE int qnode_get_new_color(const qnode_t *qn, const ir_node *irn) {
130 node_stat_t *found = qnode_find_node(qn, irn);
132 return found->new_color;
134 return get_irn_color(irn);
138 * Sets the virtual color of a node.
140 static INLINE void qnode_set_new_color(const qnode_t *qn, const ir_node *irn, int color) {
141 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
142 found->new_color = color;
146 * Checks if a node is local pinned. A node is local pinned, iff it belongs
147 * to the same optimization unit and has been optimized before the current
150 static INLINE int qnode_is_pinned_local(const qnode_t *qn, const ir_node *irn) {
151 node_stat_t *found = qnode_find_node(qn, irn);
153 return found->pinned_local;
159 * Local-pins a node, so optimizations of further nodes of the same opt unit
160 * can handle situations in which a color change would undo prior optimizations.
162 static INLINE void qnode_pin_local(const qnode_t *qn, const ir_node *irn) {
163 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
164 found->pinned_local = 1;
168 * Possible return values of qnode_color_irn()
170 #define CHANGE_SAVE NULL
171 #define CHANGE_IMPOSSIBLE (ir_node *)1
172 #define is_conflicting_node(n) (((int)n) > 1)
175 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
176 * other nodes are changed too, as required to preserve correctness. Function is
177 * aware of local and global pinning. Recursive.
178 * @param irn The node to set the color for
179 * @param col The color to set
180 * @param trigger The irn that caused the wish to change the color of the irn
181 * @return CHANGE_SAVE iff setting the color is possible, with all transiteve effects.
182 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
183 * Else the first conflicting ir_node encountered is returned.
185 * ASSUMPTION: Assumes that a life range of a single value can't be spilt into
186 * several smaller intervals where other values can live in between.
187 * This should be true in SSA.
189 static const ir_node *qnode_color_irn(const qnode_t *qn, const ir_node *irn, int col, const ir_node *trigger) {
191 struct obstack confl_ob;
192 ir_node **confl, *cn;
195 DBG((dbg, LEVEL_3, "\t %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
196 obstack_init(&confl_ob);
197 irn_col = qnode_get_new_color(qn, irn);
201 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
205 if (!qn->ou->co->isa->is_reg_allocatable(irn, arch_register_for_index(qn->ou->co->cls, col)))
208 /* get all nodes which would conflict with this change */
214 irn_bl = get_nodes_block(irn);
216 /* first check for a conflicting node which is 'living in' the irns block */
219 pset *live_ins = get_live_in(irn_bl);
220 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
221 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
222 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
223 obstack_ptr_grow(&confl_ob, n);
224 pset_break(live_ins);
229 /* setup the queue of blocks. */
231 obstack_ptr_grow(&q, irn_bl);
235 /* process the queue. The code below checks for every block dominated
236 * by the irns one, and in which the irn is live, if there are
237 * conflicting nodes */
239 ir_node *curr_bl, *sub_bl;
242 curr_bl = ((ir_node **)obstack_base(&q))[out++];
244 /* Add to the result all nodes in the block, which have
245 * the target color and interfere with the irn */
246 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
247 ir_node *n = get_irn_out(curr_bl, i);
248 if (is_allocatable_irn(n) && n != trigger && qnode_get_new_color(qn, n) == col && values_interfere(irn, n)) {
249 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
250 obstack_ptr_grow(&confl_ob, n);
254 /* If irn lives out check i-dominated blocks where the irn lives in */
256 if (is_live_out(curr_bl, irn)) {
257 dominates_for_each(curr_bl, sub_bl)
258 if (is_live_in(sub_bl, irn)) {
259 obstack_ptr_grow(&q, sub_bl);
264 obstack_free(&q, NULL);
265 obstack_ptr_grow(&confl_ob, NULL);
266 confl = (ir_node **) obstack_finish(&confl_ob);
269 /* process all nodes which would conflict with this change */
270 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
271 const ir_node *sub_res;
273 /* try to color the conflicting node cn with the color of the irn itself */
274 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
275 if (sub_res != CHANGE_SAVE) {
280 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
283 DBG((dbg, LEVEL_3, "\t %n save\n", irn));
284 obstack_free(&confl_ob, NULL);
285 qnode_set_new_color(qn, irn, col);
289 DBG((dbg, LEVEL_3, "\t %n impossible\n", irn));
290 obstack_free(&confl_ob, NULL);
291 return CHANGE_IMPOSSIBLE;
294 DBG((dbg, LEVEL_3, "\t %n conflicting\n", irn));
295 obstack_free(&confl_ob, NULL);
300 * Tries to set the colors for all members of this queue node;
301 * to the target color qn->color
302 * @returns 1 iff all members colors could be set
305 static int qnode_try_color(const qnode_t *qn) {
307 for (i=0; i<qn->mis_size; ++i) {
308 const ir_node *test_node, *confl_node;
310 test_node = qn->mis[i];
311 DBG((dbg, LEVEL_3, "\t Testing %n\n", test_node));
312 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
314 if (confl_node == CHANGE_SAVE) {
315 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
316 qnode_pin_local(qn, test_node);
317 } else if (confl_node == CHANGE_IMPOSSIBLE) {
318 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
319 qnode_add_conflict(qn, test_node, test_node);
321 if (qnode_is_pinned_local(qn, confl_node)) {
322 /* changing test_node would change back a node of current ou */
323 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
324 qnode_add_conflict(qn, confl_node, test_node);
326 if (pset_find_ptr(pinned_global, confl_node)) {
327 /* changing test_node would change back a node of a prior ou */
328 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
329 qnode_add_conflict(qn, test_node, test_node);
333 if (confl_node != CHANGE_SAVE)
340 * Determines a maximum independent set with respect to the interference and
341 * conflict edges of all nodes in a qnode.
343 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
344 int all_size, curr_size, i, o;
346 const ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
348 /* all contains all nodes not removed in this qn */
350 for (i=0; i<ou->node_count; ++i)
351 if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i]))
352 all[all_size++] = ou->nodes[i];
354 /* which[i] says which element to take out of all[] and put into curr[i] */
355 which = alloca(all_size*sizeof(*which));
356 for (curr_size=0; curr_size<all_size; ++curr_size)
357 which[curr_size] = curr_size;
359 /* stores the currently examined set */
360 curr = alloca(all_size*sizeof(*curr));
362 while (1) { /* this loop will terminate because at least a single node'll be a max indep. set */
363 /* build current set */
364 for (i=0; i<curr_size; ++i)
365 curr[i] = all[which[all_size-curr_size+i]];
367 /* check current set */
368 for (i=0; i<curr_size; ++i)
369 for (o=i+1; o<curr_size; ++o)
370 if (qnode_are_conflicting(qn, curr[i], curr[o]))
373 /* We had no conflict. This is the max indep. set */
374 qn->mis_size = curr_size;
375 for (i=0; i<curr_size; ++i)
376 qn->mis[i] = curr[i];
380 /* We had a conflict. Generate next set */
381 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
383 for (i=0; i<curr_size; ++i)
384 which[all_size-curr_size+i] = i;
391 } while (!(which[pos] = (which[pos]+1) % all_size));
393 for (i=pos+1; i<all_size; ++i)
394 which[i] = MIN(which[i-1]+1, all_size-1);
397 for (i=all_size-curr_size; i<all_size-1; ++i)
398 if (which[i]>=which[i+1]) {
408 * Creates a new qnode
410 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
411 qnode_t *qn = malloc(sizeof(*qn));
414 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
415 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
416 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
421 * Frees space used by a queue node
423 static INLINE void free_qnode(qnode_t *qn) {
424 del_set(qn->conflicts);
425 del_set(qn->changed_nodes);
431 * Inserts a qnode in the sorted queue of the outimization unit. Queue is
432 * ordered by field 'size' (the size of the mis) in decreasing order.
434 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
435 struct list_head *lh;
437 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
438 /* root node is not in qnode */
443 qnode_max_ind_set(qn, ou);
445 /* set ou->mis_size for lower bound compution */
446 if (ou->mis_size < qn->mis_size)
447 ou->mis_size = qn->mis_size;
449 /* do the insertion */
450 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with size %d\n", qn->color, qn->mis_size));
452 while (lh->next != &ou->queue) {
453 qnode_t *curr = list_entry_queue(lh->next);
454 if (curr->mis_size <= qn->mis_size)
458 list_add(&qn->queue, lh);
462 * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
463 * number of copy instructions placed during SSA-destruction and lowering.
464 * Works only for opt units with exactly 1 root node, which is the
465 * case for approximately 80% of all phi classes and all register constrained
466 * nodes. (All other phi classes are reduced to this case.)
468 static void ou_optimize(unit_t *ou) {
471 bitset_t *pos_regs = bitset_alloca(ou->co->cls->n_regs);
473 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
474 for (i=0; i<ou->node_count; ++i)
475 DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
478 INIT_LIST_HEAD(&ou->queue);
479 ou->co->isa->get_allocatable_regs(ou->nodes[0], ou->co->cls, pos_regs);
480 bitset_foreach(pos_regs, i)
481 ou_insert_qnode(ou, new_qnode(ou, i));
484 while (!list_empty(&ou->queue)) {
485 /* get head of queue */
486 curr = list_entry_queue(ou->queue.next);
487 list_del(&curr->queue);
488 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with size %d\n", curr->color, curr->mis_size));
491 if (qnode_try_color(curr))
493 /* no success, so re-insert */
494 del_set(curr->changed_nodes);
495 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
496 ou_insert_qnode(ou, curr);
499 /* apply the best found qnode */
500 if (curr->mis_size >= 2) {
501 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));
502 /* globally pin root and eventually others */
503 pset_insert_ptr(pinned_global, ou->nodes[0]);
504 for (i=1; i<ou->node_count; ++i) {
505 const ir_node *irn = ou->nodes[i];
506 int nc = qnode_get_new_color(curr, irn);
507 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
508 pset_insert_ptr(pinned_global, irn);
511 /* set color of all changed nodes */
513 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
514 /* NO_COLOR is possible, if we had an undo */
515 if (ns->new_color != NO_COLOR) {
516 DBG((dbg, LEVEL_2, "\t color(%n) := %d\n", ns->irn, ns->new_color));
517 set_irn_color(ns->irn, ns->new_color);
522 /* free best qnode (curr) and queue */
524 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
528 void co_heur_opt(copy_opt_t *co) {
530 dbg = firm_dbg_register("ir.be.copyoptheur");
531 firm_dbg_set_mask(dbg, DEBUG_LVL);
532 if (!strcmp(co->name, DEBUG_IRG))
533 firm_dbg_set_mask(dbg, -1);
535 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
536 list_for_each_entry(unit_t, curr, &co->units, units)
537 if (curr->node_count > 1)
540 del_pset(pinned_global);