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'. A 'max indep set' is determined form these. We try to
10 * color this mis using a color-exchanging mechanism. Occuring conflicts are
11 * modeled with 'conflict edges' and the qnode is reinserted in the queue. The
12 * first qnode colored without conflicts is the best one.
26 #include "becopyopt.h"
27 #include "becopystat.h"
29 #define DEBUG_LVL 0 //SET_LEVEL_1
30 static firm_dbg_module_t *dbg = NULL;
32 #define SLOTS_PINNED_GLOBAL 256
33 #define SLOTS_CONFLICTS 8
34 #define SLOTS_CHANGED_NODES 32
36 #define MIN(a,b) ((a<b)?(a):(b))
37 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
38 #define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
41 * Modeling additional conflicts between nodes. NOT live range interference
43 typedef struct _conflict_t {
44 const ir_node *n1, *n2;
48 * If an irn is changed, the changes first get stored in a node_stat_t,
49 * to allow undo of changes (=drop new data) in case of conflicts.
51 typedef struct _node_stat_t {
58 * Represents a node in the optimization queue.
60 typedef struct _qnode_t {
61 struct list_head queue; /**< chaining of unit_t->queue */
62 const unit_t *ou; /**< the opt unit this qnode belongs to */
63 int color; /**< target color */
64 set *conflicts; /**< contains conflict_t's. All internal conflicts */
65 int mis_size; /**< number of nodes in the mis. */
66 ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */
67 set *changed_nodes; /**< contains node_stat_t's. */
70 pset *pinned_global; /**< optimized nodes should not be altered any more */
72 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
73 const conflict_t *xx = x;
74 const conflict_t *yy = y;
75 return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2);
79 * If a local pinned conflict occurs, a new edge in the conflict graph is added.
80 * The next maximum independent set build, will regard it.
82 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
84 DBG((dbg, LEVEL_4, "\t %n -- %n\n", n1, n2));
86 if ((int)n1 < (int)n2) {
93 set_insert(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
97 * Checks if two nodes are in a conflict.
99 static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
101 /* search for live range interference */
102 if (n1!=n2 && nodes_interfere(qn->ou->co->chordal_env, n1, n2))
104 /* search for recoloring conflicts */
105 if ((int)n1 < (int)n2) {
112 return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c));
115 static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) {
116 return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn;
120 * Finds a node status entry of a node if existent. Otherwise return NULL
122 static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) {
125 return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
129 * Finds a node status entry of a node if existent. Otherwise it will return
130 * an initialized new entry for this node.
132 static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) {
135 find.new_color = NO_COLOR;
136 find.pinned_local = 0;
137 return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn));
141 * Returns the virtual color of a node if set before, else returns the real color.
143 static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
144 node_stat_t *found = qnode_find_node(qn, irn);
146 return found->new_color;
148 return get_irn_col(qn->ou->co, irn);
152 * Sets the virtual color of a node.
154 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
155 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
156 found->new_color = color;
160 * Checks if a node is local pinned. A node is local pinned, iff it belongs
161 * to the same optimization unit and has been optimized before the current
164 static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
165 node_stat_t *found = qnode_find_node(qn, irn);
167 return found->pinned_local;
173 * Local-pins a node, so optimizations of further nodes of the same opt unit
174 * can handle situations in which a color change would undo prior optimizations.
176 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
177 node_stat_t *found = qnode_find_or_insert_node(qn, irn);
178 found->pinned_local = 1;
182 * Possible return values of qnode_color_irn()
184 #define CHANGE_SAVE NULL
185 #define CHANGE_IMPOSSIBLE (ir_node *)1
186 #define is_conflicting_node(n) (((int)n) > 1)
189 * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
190 * other nodes are changed too, as required to preserve correctness. Function is
191 * aware of local and global pinning. Recursive.
192 * @param irn The node to set the color for
193 * @param col The color to set
194 * @param trigger The irn that caused the wish to change the color of the irn
195 * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
196 * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
197 * Else the first conflicting ir_node encountered is returned.
199 * ASSUMPTION: Assumes that a life range of a single value can't be split into
200 * several smaller intervals where other values can live in between.
201 * This should be true in SSA.
203 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
205 struct obstack confl_ob;
206 ir_node **confl, *cn;
208 const arch_env_t *arch_env = qn->ou->co->env;
209 const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env;
211 DBG((dbg, LEVEL_3, "\t %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
212 obstack_init(&confl_ob);
213 irn_col = qnode_get_new_color(qn, irn);
217 if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
221 if (!arch_reg_is_allocatable(arch_env,
223 arch_pos_make_out(0),
224 arch_register_for_index(qn->ou->co->cls, col)))
227 /* get all nodes which would conflict with this change */
233 irn_bl = get_nodes_block(irn);
235 /* first check for a conflicting node which is 'living in' the irns block */
238 pset *live_ins = get_live_in(irn_bl);
239 for (n = pset_first(live_ins); n; n = pset_next(live_ins))
240 if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), qn->ou->co->cls)
241 && n != trigger && qnode_get_new_color(qn, n) == col
242 && nodes_interfere(chordal_env, irn, n)) {
244 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
245 obstack_ptr_grow(&confl_ob, n);
246 pset_break(live_ins);
251 /* setup the queue of blocks. */
253 obstack_ptr_grow(&q, irn_bl);
257 /* process the queue. The code below checks for every block dominated
258 * by the irns one, and in which the irn is live, if there are
259 * conflicting nodes */
261 ir_node *curr_bl, *sub_bl;
264 curr_bl = ((ir_node **)obstack_base(&q))[out++];
266 /* Add to the result all nodes in the block, which have
267 * the target color and interfere with the irn */
268 for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
269 ir_node *n = get_irn_out(curr_bl, i);
270 if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), qn->ou->co->cls)
271 && n != trigger && qnode_get_new_color(qn, n) == col
272 && nodes_interfere(chordal_env, irn, n)) {
274 DBG((dbg, LEVEL_4, "\t %n\ttroubles\n", n));
275 obstack_ptr_grow(&confl_ob, n);
279 /* If irn lives out check i-dominated blocks where the irn lives in */
281 if (is_live_out(curr_bl, irn)) {
282 dominates_for_each(curr_bl, sub_bl)
283 if (is_live_in(sub_bl, irn)) {
284 obstack_ptr_grow(&q, sub_bl);
289 obstack_free(&q, NULL);
290 obstack_ptr_grow(&confl_ob, NULL);
291 confl = (ir_node **) obstack_finish(&confl_ob);
294 /* process all nodes which would conflict with this change */
295 for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
298 /* try to color the conflicting node cn with the color of the irn itself */
299 sub_res = qnode_color_irn(qn, cn, irn_col, irn);
300 if (sub_res != CHANGE_SAVE) {
305 /* if we arrive here all sub changes can be applied, so it's save to change this irn */
308 DBG((dbg, LEVEL_3, "\t %n save\n", irn));
309 obstack_free(&confl_ob, NULL);
310 qnode_set_new_color(qn, irn, col);
314 DBG((dbg, LEVEL_3, "\t %n impossible\n", irn));
315 obstack_free(&confl_ob, NULL);
316 return CHANGE_IMPOSSIBLE;
319 DBG((dbg, LEVEL_3, "\t %n conflicting\n", irn));
320 obstack_free(&confl_ob, NULL);
325 * Tries to set the colors for all members of this queue node;
326 * to the target color qn->color
327 * @returns 1 iff all members colors could be set
330 static int qnode_try_color(const qnode_t *qn) {
332 for (i=0; i<qn->mis_size; ++i) {
333 ir_node *test_node, *confl_node;
335 test_node = qn->mis[i];
336 DBG((dbg, LEVEL_3, "\t Testing %n\n", test_node));
337 confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
339 if (confl_node == CHANGE_SAVE) {
340 DBG((dbg, LEVEL_3, "\t Save --> pin local\n"));
341 qnode_pin_local(qn, test_node);
342 } else if (confl_node == CHANGE_IMPOSSIBLE) {
343 DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n"));
344 qnode_add_conflict(qn, test_node, test_node);
346 if (qnode_is_pinned_local(qn, confl_node)) {
347 /* changing test_node would change back a node of current ou */
348 DBG((dbg, LEVEL_3, "\t Conflicting local --> add conflict\n"));
349 qnode_add_conflict(qn, confl_node, test_node);
351 if (pset_find_ptr(pinned_global, confl_node)) {
352 /* changing test_node would change back a node of a prior ou */
353 DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n"));
354 qnode_add_conflict(qn, test_node, test_node);
358 if (confl_node != CHANGE_SAVE)
365 * Determines a maximum independent set with respect to the interference and
366 * conflict edges of all nodes in a qnode.
368 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
369 int all_size, curr_size, i, o;
371 ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
373 /* all contains all nodes not removed in this qn */
375 for (i=0; i<ou->node_count; ++i)
376 if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i]))
377 all[all_size++] = ou->nodes[i];
379 /* which[i] says which element to take out of all[] and put into curr[i] */
380 which = alloca(all_size*sizeof(*which));
381 for (curr_size=0; curr_size<all_size; ++curr_size)
382 which[curr_size] = curr_size;
384 /* stores the currently examined set */
385 curr = alloca(all_size*sizeof(*curr));
387 while (1) { /* this loop will terminate because at least a single node will be a max indep. set */
388 /* build current set */
389 for (i=0; i<curr_size; ++i)
390 curr[i] = all[which[all_size-curr_size+i]];
392 /* check current set */
393 for (i=0; i<curr_size; ++i)
394 for (o=i+1; o<curr_size; ++o)
395 if (qnode_are_conflicting(qn, curr[i], curr[o]))
398 /* We had no conflict. This is the max indep. set */
399 qn->mis_size = curr_size;
400 for (i=0; i<curr_size; ++i)
401 qn->mis[i] = curr[i];
405 /* We had a conflict. Generate next set */
406 if (which[all_size-curr_size+1] == all_size-curr_size+1) {
408 for (i=0; i<curr_size; ++i)
409 which[all_size-curr_size+i] = i;
416 } while (!(which[pos] = (which[pos]+1) % all_size));
418 for (i=pos+1; i<all_size; ++i)
419 which[i] = MIN(which[i-1]+1, all_size-1);
422 for (i=all_size-curr_size; i<all_size-1; ++i)
423 if (which[i]>=which[i+1]) {
433 * Creates a new qnode
435 static INLINE qnode_t *new_qnode(const unit_t *ou, int color) {
436 qnode_t *qn = xmalloc(sizeof(*qn));
439 qn->mis = malloc(ou->node_count * sizeof(*qn->mis));
440 qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS);
441 qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
446 * Frees space used by a queue node
448 static INLINE void free_qnode(qnode_t *qn) {
449 del_set(qn->conflicts);
450 del_set(qn->changed_nodes);
456 * Inserts a qnode in the sorted queue of the optimization unit. Queue is
457 * ordered by field 'size' (the size of the mis) in decreasing order.
459 static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
460 struct list_head *lh;
462 if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) {
463 /* root node is not in qnode */
468 qnode_max_ind_set(qn, ou);
470 /* set ou->mis_size for lower bound compution */
471 if (ou->mis_size < qn->mis_size)
472 ou->mis_size = qn->mis_size;
474 /* do the insertion */
475 DBG((dbg, LEVEL_4, "\t Insert qnode color %d with size %d\n", qn->color, qn->mis_size));
477 while (lh->next != &ou->queue) {
478 qnode_t *curr = list_entry_queue(lh->next);
479 if (curr->mis_size <= qn->mis_size)
483 list_add(&qn->queue, lh);
487 * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
488 * number of copy instructions placed during SSA-destruction and lowering.
489 * Works only for opt units with exactly 1 root node, which is the
490 * case for approximately 80% of all phi classes and all register constrained
491 * nodes. (All other phi classes are reduced to this case.)
493 static void ou_optimize(unit_t *ou) {
496 bitset_t *pos_regs = bitset_alloca(ou->co->cls->n_regs);
498 DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
499 for (i=0; i<ou->node_count; ++i)
500 DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
503 INIT_LIST_HEAD(&ou->queue);
504 arch_get_allocatable_regs(ou->co->env, ou->nodes[0], arch_pos_make_out(0), ou->co->cls, pos_regs);
505 bitset_foreach(pos_regs, i)
506 ou_insert_qnode(ou, new_qnode(ou, i));
509 while (!list_empty(&ou->queue)) {
510 /* get head of queue */
511 curr = list_entry_queue(ou->queue.next);
512 list_del(&curr->queue);
513 DBG((dbg, LEVEL_2, "\t Examine qnode color %d with size %d\n", curr->color, curr->mis_size));
516 if (qnode_try_color(curr))
518 /* no success, so re-insert */
519 del_set(curr->changed_nodes);
520 curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
521 ou_insert_qnode(ou, curr);
524 /* apply the best found qnode */
525 if (curr->mis_size >= 2) {
528 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));
529 /* globally pin root and eventually others */
530 pset_insert_ptr(pinned_global, ou->nodes[0]);
531 for (i=1; i<ou->node_count; ++i) {
532 ir_node *irn = ou->nodes[i];
533 int nc = qnode_get_new_color(curr, irn);
534 if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
535 pset_insert_ptr(pinned_global, irn);
538 /* set color of all changed nodes */
539 for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
540 /* NO_COLOR is possible, if we had an undo */
541 if (ns->new_color != NO_COLOR) {
542 DBG((dbg, LEVEL_2, "\t color(%n) := %d\n", ns->irn, ns->new_color));
543 set_irn_col(ou->co, ns->irn, ns->new_color);
548 /* free best qnode (curr) and queue */
550 list_for_each_entry_safe(qnode_t, curr, tmp, &ou->queue, queue)
554 void co_heur_opt(copy_opt_t *co) {
556 dbg = firm_dbg_register("ir.be.copyoptheur");
557 firm_dbg_set_mask(dbg, DEBUG_LVL);
558 if (!strcmp(co->name, DEBUG_IRG))
559 firm_dbg_set_mask(dbg, -1);
561 pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
562 list_for_each_entry(unit_t, curr, &co->units, units)
563 if (curr->node_count > 1)
566 del_pset(pinned_global);