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 Spillslot coalescer.
23 * @author Matthias Braun
37 #include "unionfind.h"
43 #include "bespillslots.h"
44 #include "bechordal_t.h"
45 #include "bestatevent.h"
47 #include "beintlive_t.h"
51 #define DBG_COALESCING 1
52 #define DBG_INTERFERENCES 2
54 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
56 typedef struct spill_t {
58 const ir_mode *mode; /**< mode of the spilled value */
59 int alignment; /**< alignment for the spilled value */
60 int spillslot; /**< index into spillslot_unionfind structure */
63 typedef struct affinity_edge_t {
74 affinity_edge_t **affinity_edges;
76 set_frame_entity_func set_frame_entity;
77 bool at_begin; /**< frame entities should be allocate at
78 the beginning of the stackframe */
81 /** Compare 2 affinity edges (used in quicksort) */
82 static int cmp_affinity(const void *d1, const void *d2)
84 const affinity_edge_t * const *e1 = (const affinity_edge_t**)d1;
85 const affinity_edge_t * const *e2 = (const affinity_edge_t**)d2;
87 /* sort in descending order */
88 return (*e1)->affinity < (*e2)->affinity ? 1 : -1;
91 static int cmp_spill(const void* d1, const void* d2, size_t size)
93 const spill_t* s1 = (const spill_t*)d1;
94 const spill_t* s2 = (const spill_t*)d2;
97 return s1->spill != s2->spill;
100 static spill_t *get_spill(be_fec_env_t *env, ir_node *node)
103 int hash = hash_irn(node);
106 res = (spill_t*)set_find(env->spills, &spill, sizeof(spill), hash);
112 static inline ir_node *get_memory_edge(const ir_node *node)
116 arity = get_irn_arity(node);
117 for (i = arity - 1; i >= 0; --i) {
118 ir_node *arg = get_irn_n(node, i);
119 if (get_irn_mode(arg) == mode_M)
126 static spill_t *collect_spill(be_fec_env_t *env, ir_node *node,
127 const ir_mode *mode, int align)
130 int hash = hash_irn(node);
132 /* insert into set of spills if not already there */
134 res = (spill_t*)set_find(env->spills, &spill, sizeof(spill), hash);
137 spill.spillslot = set_count(env->spills);
139 spill.alignment = align;
140 res = (spill_t*)set_insert(env->spills, &spill, sizeof(spill), hash);
141 DB((dbg, DBG_COALESCING, "Slot %d: %+F\n", spill.spillslot, node));
143 assert(res->mode == mode);
144 assert(res->alignment == align);
150 static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
151 const ir_mode *mode, int align)
155 int hash = hash_irn(node);
156 const ir_exec_freq *exec_freq = be_get_irg_exec_freq(env->irg);
158 assert(is_Phi(node));
161 res = (spill_t*)set_find(env->spills, &spill, sizeof(spill), hash);
163 assert(res->mode == mode);
164 assert(res->alignment == align);
168 spill.spillslot = set_count(env->spills);
170 spill.alignment = align;
171 DB((dbg, DBG_COALESCING, "Slot %d: %+F\n", spill.spillslot, node));
172 res = (spill_t*)set_insert(env->spills, &spill, sizeof(spill), hash);
174 /* collect attached spills and mem-phis */
175 arity = get_irn_arity(node);
176 for (i = 0; i < arity; ++i) {
177 affinity_edge_t *affinty_edge;
178 ir_node *arg = get_irn_n(node, i);
182 arg_spill = collect_memphi(env, arg, mode, align);
184 arg_spill = collect_spill(env, arg, mode, align);
187 /* add an affinity edge */
188 affinty_edge = OALLOC(&env->obst, affinity_edge_t);
189 affinty_edge->affinity = get_block_execfreq(exec_freq, get_nodes_block(arg));
190 affinty_edge->slot1 = res->spillslot;
191 affinty_edge->slot2 = arg_spill->spillslot;
192 ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
198 void be_node_needs_frame_entity(be_fec_env_t *env, ir_node *node,
199 const ir_mode *mode, int align)
201 ir_node *spillnode = get_memory_edge(node);
203 assert(spillnode != NULL);
205 /* walk upwards and collect all phis and spills on this way */
206 if (is_Phi(spillnode)) {
207 collect_memphi(env, spillnode, mode, align);
209 collect_spill(env, spillnode, mode, align);
212 ARR_APP1(ir_node *, env->reloads, node);
217 static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
218 int* spillslot_unionfind, int s1, int s2)
224 /* merge spillslots and interferences */
225 res = uf_union(spillslot_unionfind, s1, s2);
226 /* we assume that we always merge s2 to s1 so swap s1, s2 if necessary */
233 bitset_or(interferences[s1], interferences[s2]);
235 /* update other interferences */
236 spillcount = set_count(env->spills);
237 for (i = 0; i < spillcount; ++i) {
238 bitset_t *intfs = interferences[i];
239 if (bitset_is_set(intfs, s2))
240 bitset_set(intfs, s1);
246 static int my_values_interfere2(ir_graph *irg, const ir_node *a,
249 be_lv_t *lv = be_get_irg_liveness(irg);
251 int a2b = _value_dominates(a, b);
252 int b2a = _value_dominates(b, a);
254 /* If there is no dominance relation, they do not interfere. */
255 if ((a2b | b2a) > 0) {
256 const ir_edge_t *edge;
260 * Adjust a and b so, that a dominates b if
261 * a dominates b or vice versa.
264 const ir_node *t = a;
269 bb = get_nodes_block(b);
272 * If a is live end in b's block it is
273 * live at b's definition (a dominates b)
275 if (be_is_live_end(lv, bb, a))
279 * Look at all usages of a.
280 * If there's one usage of a in the block of b, then
281 * we check, if this use is dominated by b, if that's true
282 * a and b interfere. Note that b must strictly dominate the user,
283 * since if b is the last user of in the block, b and a do not
285 * Uses of a not in b's block can be disobeyed, because the
286 * check for a being live at the end of b's block is already
289 foreach_out_edge(a, edge) {
290 const ir_node *user = get_edge_src_irn(edge);
292 const ir_edge_t *edge2;
293 foreach_out_edge(user, edge2) {
294 const ir_node *user2 = get_edge_src_irn(edge2);
295 assert(!is_Sync(user2));
296 if (get_nodes_block(user2) == bb && !is_Phi(user2) &&
297 _value_strictly_dominates(b, user2))
301 if (get_nodes_block(user) == bb && !is_Phi(user) &&
302 _value_strictly_dominates(b, user))
312 * same as values_interfere but with special handling for Syncs
314 static int my_values_interfere(ir_graph *irg, ir_node *a, ir_node *b)
317 int i, arity = get_irn_arity(a);
318 for (i = 0; i < arity; ++i) {
319 ir_node *in = get_irn_n(a, i);
320 if (my_values_interfere(irg, in, b))
324 } else if (is_Sync(b)) {
325 int i, arity = get_irn_arity(b);
326 for (i = 0; i < arity; ++i) {
327 ir_node *in = get_irn_n(b, i);
328 /* a is not a sync, so no need for my_values_interfere */
329 if (my_values_interfere2(irg, a, in))
335 return my_values_interfere2(irg, a, b);
339 * A greedy coalescing algorithm for spillslots:
340 * 1. Sort the list of affinity edges
341 * 2. Try to merge slots with affinity edges (most expensive slots first)
342 * 3. Try to merge everything else that is possible
344 static void do_greedy_coalescing(be_fec_env_t *env)
350 int affinity_edge_count;
351 bitset_t **interferences;
352 int* spillslot_unionfind;
355 spillcount = set_count(env->spills);
361 DB((dbg, DBG_COALESCING, "Coalescing %d spillslots\n", spillcount));
363 interferences = OALLOCN(&data, bitset_t*, spillcount);
364 spillslot_unionfind = OALLOCN(&data, int, spillcount);
365 spilllist = OALLOCN(&data, spill_t*, spillcount);
367 uf_init(spillslot_unionfind, spillcount);
370 memset(spilllist, 0, spillcount * sizeof(spilllist[0]));
374 foreach_set(env->spills, spill_t*, spill) {
375 assert(spill->spillslot < spillcount);
376 spilllist[spill->spillslot] = spill;
380 for (i = 0; i < spillcount; ++i) {
381 interferences[i] = bitset_obstack_alloc(&data, spillcount);
384 /* construct interferences */
385 for (i = 0; i < spillcount; ++i) {
386 ir_node *spill1 = spilllist[i]->spill;
388 if (is_NoMem(spill1))
391 for (i2 = i+1; i2 < spillcount; ++i2) {
392 ir_node *spill2 = spilllist[i2]->spill;
394 if (is_NoMem(spill2))
397 if (my_values_interfere(env->irg, spill1, spill2)) {
398 DB((dbg, DBG_INTERFERENCES,
399 "Slot %d and %d interfere\n", i, i2));
401 bitset_set(interferences[i], i2);
402 bitset_set(interferences[i2], i);
407 /* sort affinity edges */
408 affinity_edge_count = ARR_LEN(env->affinity_edges);
409 qsort(env->affinity_edges, affinity_edge_count,
410 sizeof(env->affinity_edges[0]), cmp_affinity);
412 /*dump_interference_graph(env, interferences, "before"); */
414 /* try to merge affine nodes */
415 for (i = 0; i < affinity_edge_count; ++i) {
416 const affinity_edge_t *edge = env->affinity_edges[i];
417 int s1 = uf_find(spillslot_unionfind, edge->slot1);
418 int s2 = uf_find(spillslot_unionfind, edge->slot2);
420 /* test if values interfere */
421 if (bitset_is_set(interferences[s1], s2)) {
422 assert(bitset_is_set(interferences[s2], s1));
426 DB((dbg, DBG_COALESCING,
427 "Merging %d and %d because of affinity edge\n", s1, s2));
429 merge_interferences(env, interferences, spillslot_unionfind, s1, s2);
432 /* try to merge as much remaining spillslots as possible */
433 for (i = 0; i < spillcount; ++i) {
434 int s1 = uf_find(spillslot_unionfind, i);
438 for (i2 = i+1; i2 < spillcount; ++i2) {
439 int s2 = uf_find(spillslot_unionfind, i2);
443 /* test if values interfere
444 * we have to test n1-n2 and n2-n1, because only 1 side gets updated
445 * when node merging occurs
447 if (bitset_is_set(interferences[s1], s2)) {
448 assert(bitset_is_set(interferences[s2], s1));
452 DB((dbg, DBG_COALESCING,
453 "Merging %d and %d because it is possible\n", s1, s2));
455 if (merge_interferences(env, interferences, spillslot_unionfind, s1, s2) != 0) {
456 /* we can break the loop here, because s2 is the new supernode
457 * now and we'll test s2 again later anyway */
463 /* assign spillslots to spills */
464 for (i = 0; i < spillcount; ++i) {
465 spill_t *spill = spilllist[i];
467 spill->spillslot = uf_find(spillslot_unionfind, i);
470 /*dump_interference_graph(env, interferences, "after");*/
471 obstack_free(&data, 0);
476 typedef struct spill_slot_t {
482 typedef struct memperm_entry_t {
487 struct memperm_entry_t *next;
490 typedef struct memperm_t {
493 memperm_entry_t *entries;
496 static int cmp_memperm(const void* d1, const void* d2, size_t size)
498 const memperm_t* e1 = (const memperm_t*)d1;
499 const memperm_t* e2 = (const memperm_t*)d2;
502 return e1->block != e2->block;
505 static memperm_t *get_memperm(be_fec_env_t *env, ir_node *block)
507 memperm_t entry, *res;
511 hash = hash_irn(block);
513 res = (memperm_t*)set_find(env->memperms, &entry, sizeof(entry), hash);
516 entry.entrycount = 0;
517 entry.entries = NULL;
518 res = (memperm_t*)set_insert(env->memperms, &entry, sizeof(entry), hash);
524 static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
526 ir_graph *irg = env->irg;
527 ir_type *frame = get_irg_frame_type(irg);
528 ir_entity *res = frame_alloc_area(frame, slot->size, slot->align,
531 /* adjust size of the entity type... */
532 ir_type *enttype = get_entity_type(res);
533 set_type_size_bytes(enttype, slot->size);
541 * Enlarges a spillslot (if necessary) so that it can carry a value of size
542 * @p othersize and alignment @p otheralign.
544 static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
546 if (othersize > slot->size) {
547 slot->size = othersize;
549 if (otheralign > slot->align) {
550 if (otheralign % slot->align != 0)
551 slot->align *= otheralign;
553 slot->align = otheralign;
554 } else if (slot->align % otheralign != 0) {
555 slot->align *= otheralign;
559 static void assign_spill_entity(be_fec_env_t *env,
560 ir_node *node, ir_entity *entity)
567 arity = get_irn_arity(node);
568 for (i = 0; i < arity; ++i) {
569 ir_node *in = get_irn_n(node, i);
572 assign_spill_entity(env, in, entity);
577 /* beware: we might have Stores with Memory Proj's, ia32 fisttp for
579 node = skip_Proj(node);
580 assert(arch_get_frame_entity(node) == NULL);
581 env->set_frame_entity(node, entity);
585 * Create stack entities for the spillslots and assign them to the spill and
588 static void assign_spillslots(be_fec_env_t *env)
590 int spillcount = set_count(env->spills);
591 spill_slot_t *spillslots = ALLOCANZ(spill_slot_t, spillcount);
595 /* construct spillslots */
596 foreach_set(env->spills, spill_t*, spill) {
597 int slotid = spill->spillslot;
598 const ir_mode *mode = spill->mode;
599 spill_slot_t *slot = & (spillslots[slotid]);
600 int size = get_mode_size_bytes(mode);
601 int align = spill->alignment;
603 if (slot->align == 0 && slot->size == 0) {
607 enlarge_spillslot(slot, align, size);
611 foreach_set(env->spills, spill_t*, spill) {
612 ir_node *node = spill->spill;
613 int slotid = spill->spillslot;
616 slot = &spillslots[slotid];
617 if (slot->entity == NULL) {
618 create_stack_entity(env, slot);
623 ir_node *block = get_nodes_block(node);
625 /* should be a PhiM */
626 assert(get_irn_mode(node) == mode_M);
628 for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
629 ir_node *arg = get_irn_n(node, i);
630 ir_node *predblock = get_Block_cfgpred_block(block, i);
634 argspill = get_spill(env, arg);
635 assert(argspill != NULL);
637 argslotid = argspill->spillslot;
638 if (slotid != argslotid) {
640 memperm_entry_t *entry;
641 spill_slot_t *argslot = &spillslots[argslotid];
642 if (argslot->entity == NULL) {
643 create_stack_entity(env, argslot);
646 memperm = get_memperm(env, predblock);
648 entry = OALLOC(&env->obst, memperm_entry_t);
651 entry->in = argslot->entity;
652 entry->out = slot->entity;
653 entry->next = memperm->entries;
654 memperm->entrycount++;
655 memperm->entries = entry;
659 assign_spill_entity(env, node, slot->entity);
663 for (i = 0; i < ARR_LEN(env->reloads); ++i) {
664 ir_node *reload = env->reloads[i];
665 ir_node *spillnode = get_memory_edge(reload);
666 spill_t *spill = get_spill(env, spillnode);
667 const spill_slot_t *slot = & spillslots[spill->spillslot];
669 assert(slot->entity != NULL);
671 env->set_frame_entity(reload, slot->entity);
676 * Returns the last node in a block which is no control flow changing node
678 static ir_node *get_end_of_block_insertion_point(ir_node* block)
680 ir_node* ins = sched_last(block);
681 while (is_Proj(ins) && get_irn_mode(ins) == mode_X) {
682 ins = sched_prev(ins);
688 ir_node *prev = sched_prev(ins);
698 static void create_memperms(be_fec_env_t *env)
702 foreach_set(env->memperms, memperm_t*, memperm) {
703 ir_node **nodes = ALLOCAN(ir_node*, memperm->entrycount);
704 memperm_entry_t *entry;
706 ir_node *mempermnode;
709 assert(memperm->entrycount > 0);
711 for (entry = memperm->entries, i = 0; entry != NULL; entry = entry->next, ++i) {
712 ir_node* arg = get_irn_n(entry->node, entry->pos);
716 mempermnode = be_new_MemPerm(memperm->block, memperm->entrycount,
719 /* insert node into schedule */
720 blockend = get_end_of_block_insertion_point(memperm->block);
721 sched_add_before(blockend, mempermnode);
722 stat_ev_dbl("mem_perm", memperm->entrycount);
725 for (entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
727 ir_node* arg = get_irn_n(entry->node, entry->pos);
729 be_set_MemPerm_in_entity(mempermnode, i, entry->in);
730 be_set_MemPerm_out_entity(mempermnode, i, entry->out);
731 proj = new_r_Proj(mempermnode, get_irn_mode(arg), i);
733 set_irn_n(entry->node, entry->pos, proj);
738 static int count_spillslots(const be_fec_env_t *env)
740 const spill_t *spill;
741 int spillcount = set_count(env->spills);
742 bitset_t *counted = bitset_alloca(spillcount);
746 foreach_set(env->spills, spill_t*, spill) {
747 int spillslot = spill->spillslot;
748 if (!bitset_is_set(counted, spillslot)) {
750 bitset_set(counted, spillslot);
757 be_fec_env_t *be_new_frame_entity_coalescer(ir_graph *irg)
759 be_fec_env_t *env = XMALLOCZ(be_fec_env_t);
761 be_liveness_assure_chk(be_assure_liveness(irg));
763 obstack_init(&env->obst);
765 env->spills = new_set(cmp_spill, 10);
766 env->reloads = NEW_ARR_F(ir_node*, 0);
767 env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
768 env->memperms = new_set(cmp_memperm, 10);
773 void be_free_frame_entity_coalescer(be_fec_env_t *env)
775 del_set(env->memperms);
776 DEL_ARR_F(env->reloads);
777 DEL_ARR_F(env->affinity_edges);
778 del_set(env->spills);
779 obstack_free(&env->obst, NULL);
784 void be_assign_entities(be_fec_env_t *env,
785 set_frame_entity_func set_frame_entity,
786 bool alloc_entities_at_begin)
788 env->set_frame_entity = set_frame_entity;
789 env->at_begin = alloc_entities_at_begin;
791 stat_ev_dbl("spillslots", set_count(env->spills));
793 if (be_coalesce_spill_slots) {
794 do_greedy_coalescing(env);
797 stat_ev_dbl("spillslots_after_coalescing", count_spillslots(env));
799 assign_spillslots(env);
801 create_memperms(env);
804 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots)
805 void be_init_spillslots(void)
807 FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");