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
39 #include "unionfind.h"
44 #include "bespillslots.h"
45 #include "bechordal_t.h"
46 #include "bejavacoal.h"
47 #include "bestatevent.h"
48 #include "bespilloptions.h"
50 #include "beintlive_t.h"
54 #define DBG_COALESCING 1
55 #define DBG_INTERFERENCES 2
57 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
59 typedef struct _spill_t {
61 const ir_mode *mode; /**< mode of the spilled value */
62 int alignment; /**< alignment for the spilled value */
63 int spillslot; /**< index into spillslot_unionfind structure */
66 typedef struct _affinity_edge_t {
71 struct _be_fec_env_t {
73 const arch_env_t *arch_env;
77 affinity_edge_t **affinity_edges;
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 = d1;
85 const affinity_edge_t * const *e2 = 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 = d1;
94 const spill_t* s2 = 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 = 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 = set_find(env->spills, &spill, sizeof(spill), hash);
137 spill.spillslot = set_count(env->spills);
139 spill.alignment = align;
140 res = set_insert(env->spills, &spill, sizeof(spill), hash);
142 assert(res->mode == mode);
143 assert(res->alignment == align);
149 static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
150 const ir_mode *mode, int align)
154 int hash = hash_irn(node);
155 const ir_exec_freq *exec_freq = be_get_birg_exec_freq(env->birg);
157 assert(is_Phi(node));
160 res = set_find(env->spills, &spill, sizeof(spill), hash);
162 assert(res->mode == mode);
163 assert(res->alignment == align);
167 spill.spillslot = set_count(env->spills);
169 spill.alignment = align;
170 res = set_insert(env->spills, &spill, sizeof(spill), hash);
172 /* collect attached spills and mem-phis */
173 arity = get_irn_arity(node);
174 for(i = 0; i < arity; ++i) {
175 affinity_edge_t *affinty_edge;
176 ir_node *arg = get_irn_n(node, i);
180 arg_spill = collect_memphi(env, arg, mode, align);
182 arg_spill = collect_spill(env, arg, mode, align);
185 /* add an affinity edge */
186 affinty_edge = obstack_alloc(&env->obst, sizeof(affinty_edge[0]));
187 affinty_edge->affinity = get_block_execfreq(exec_freq, get_nodes_block(arg));
188 affinty_edge->slot1 = res->spillslot;
189 affinty_edge->slot2 = arg_spill->spillslot;
190 ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
196 void be_node_needs_frame_entity(be_fec_env_t *env, ir_node *node,
197 const ir_mode *mode, int align)
199 ir_node *spillnode = get_memory_edge(node);
202 assert(spillnode != NULL);
204 /* walk upwards and collect all phis and spills on this way */
205 if (is_Phi(spillnode)) {
206 spill = collect_memphi(env, spillnode, mode, align);
208 spill = collect_spill(env, spillnode, mode, align);
211 ARR_APP1(ir_node *, env->reloads, node);
216 static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
217 int* spillslot_unionfind, int s1, int s2)
223 /* merge spillslots and interferences */
224 res = uf_union(spillslot_unionfind, s1, s2);
225 /* we assume that we always merge s2 to s1 so swap s1, s2 if necessary */
232 bitset_or(interferences[s1], interferences[s2]);
234 /* update other interferences */
235 spillcount = set_count(env->spills);
236 for(i = 0; i < spillcount; ++i) {
237 bitset_t *intfs = interferences[i];
238 if(bitset_is_set(intfs, s2))
239 bitset_set(intfs, s1);
245 static int my_values_interfere2(be_irg_t *birg, const ir_node *a,
248 be_lv_t *lv = be_get_birg_liveness(birg);
250 int a2b = _value_dominates(a, b);
251 int b2a = _value_dominates(b, a);
253 /* If there is no dominance relation, they do not interfere. */
254 if((a2b | b2a) > 0) {
255 const ir_edge_t *edge;
259 * Adjust a and b so, that a dominates b if
260 * a dominates b or vice versa.
263 const ir_node *t = a;
268 bb = get_nodes_block(b);
271 * If a is live end in b's block it is
272 * live at b's definition (a dominates b)
274 if(be_is_live_end(lv, bb, a))
278 * Look at all usages of a.
279 * If there's one usage of a in the block of b, then
280 * we check, if this use is dominated by b, if that's true
281 * a and b interfere. Note that b must strictly dominate the user,
282 * since if b is the last user of in the block, b and a do not
284 * Uses of a not in b's block can be disobeyed, because the
285 * check for a being live at the end of b's block is already
288 foreach_out_edge(a, edge) {
289 const ir_node *user = get_edge_src_irn(edge);
291 const ir_edge_t *edge2;
292 foreach_out_edge(user, edge2) {
293 const ir_node *user2 = get_edge_src_irn(edge2);
294 assert(!is_Sync(user2));
295 if(get_nodes_block(user2) == bb && !is_Phi(user2) &&
296 _value_strictly_dominates(b, user2))
300 if(get_nodes_block(user) == bb && !is_Phi(user) &&
301 _value_strictly_dominates(b, user))
311 * same as values_interfere but with special handling for Syncs
313 static int my_values_interfere(be_irg_t *birg, ir_node *a, ir_node *b)
316 int i, arity = get_irn_arity(a);
317 for(i = 0; i < arity; ++i) {
318 ir_node *in = get_irn_n(a, i);
319 if(my_values_interfere(birg, in, b))
323 } else if(is_Sync(b)) {
324 int i, arity = get_irn_arity(b);
325 for(i = 0; i < arity; ++i) {
326 ir_node *in = get_irn_n(b, i);
327 /* a is not a sync, so no need for my_values_interfere */
328 if(my_values_interfere2(birg, a, in))
334 return my_values_interfere2(birg, a, b);
338 * A greedy coalescing algorithm for spillslots:
339 * 1. Sort the list of affinity edges
340 * 2. Try to merge slots with affinity edges (most expensive slots first)
341 * 3. Try to merge everything else that is possible
343 static void do_greedy_coalescing(be_fec_env_t *env)
349 int affinity_edge_count;
350 bitset_t **interferences;
351 int* spillslot_unionfind;
353 spillcount = set_count(env->spills);
357 DBG((dbg, DBG_COALESCING, "Coalescing %d spillslots\n", spillcount));
359 interferences = alloca(spillcount * sizeof(interferences[0]));
360 spillslot_unionfind = alloca(spillcount * sizeof(spillslot_unionfind[0]));
361 spilllist = alloca(spillcount * sizeof(spilllist[0]));
363 uf_init(spillslot_unionfind, 0, spillcount);
366 memset(spilllist, 0, spillcount * sizeof(spilllist[0]));
369 for(spill = set_first(env->spills), i = 0; spill != NULL;
370 spill = set_next(env->spills), ++i) {
371 assert(spill->spillslot < spillcount);
372 spilllist[spill->spillslot] = spill;
375 for(i = 0; i < spillcount; ++i) {
376 interferences[i] = bitset_alloca(spillcount);
379 /* construct interferences */
380 for (i = 0; i < spillcount; ++i) {
381 ir_node *spill1 = spilllist[i]->spill;
383 if (is_NoMem(spill1))
386 for(i2 = i+1; i2 < spillcount; ++i2) {
387 ir_node *spill2 = spilllist[i2]->spill;
389 if (is_NoMem(spill2))
392 if (my_values_interfere(env->birg, spill1, spill2)) {
393 DBG((dbg, DBG_INTERFERENCES,
394 "Slot %d and %d interfere\n", i, i2));
396 bitset_set(interferences[i], i2);
397 bitset_set(interferences[i2], i);
402 /* sort affinity edges */
403 affinity_edge_count = ARR_LEN(env->affinity_edges);
404 qsort(env->affinity_edges, affinity_edge_count,
405 sizeof(env->affinity_edges[0]), cmp_affinity);
407 /*dump_interference_graph(env, interferences, "before"); */
409 /* try to merge affine nodes */
410 for(i = 0; i < affinity_edge_count; ++i) {
411 const affinity_edge_t *edge = env->affinity_edges[i];
412 int s1 = uf_find(spillslot_unionfind, edge->slot1);
413 int s2 = uf_find(spillslot_unionfind, edge->slot2);
415 /* test if values interfere */
416 if (bitset_is_set(interferences[s1], s2)) {
417 assert(bitset_is_set(interferences[s2], s1));
421 DBG((dbg, DBG_COALESCING,
422 "Merging %d and %d because of affinity edge\n", s1, s2));
424 merge_interferences(env, interferences, spillslot_unionfind, s1, s2);
427 /* try to merge as much remaining spillslots as possible */
428 for(i = 0; i < spillcount; ++i) {
429 int s1 = uf_find(spillslot_unionfind, i);
433 for(i2 = i+1; i2 < spillcount; ++i2) {
434 int s2 = uf_find(spillslot_unionfind, i2);
438 /* test if values interfere
439 * we have to test n1-n2 and n2-n1, because only 1 side gets updated
440 * when node merging occurs
442 if(bitset_is_set(interferences[s1], s2)) {
443 assert(bitset_is_set(interferences[s2], s1));
447 DBG((dbg, DBG_COALESCING,
448 "Merging %d and %d because it is possible\n", s1, s2));
450 if(merge_interferences(env, interferences, spillslot_unionfind, s1, s2) != 0) {
451 /* we can break the loop here, because s2 is the new supernode
452 * now and we'll test s2 again later anyway */
458 /* assign spillslots to spills */
459 for(i = 0; i < spillcount; ++i) {
460 spill_t *spill = spilllist[i];
462 spill->spillslot = uf_find(spillslot_unionfind, i);
465 /*dump_interference_graph(env, interferences, "after");*/
470 typedef struct _spill_slot_t {
476 typedef struct _memperm_entry_t {
481 struct _memperm_entry_t *next;
484 typedef struct _memperm_t {
487 memperm_entry_t *entries;
490 static int cmp_memperm(const void* d1, const void* d2, size_t size)
492 const memperm_t* e1 = d1;
493 const memperm_t* e2 = d2;
496 return e1->block != e2->block;
499 static memperm_t *get_memperm(be_fec_env_t *env, ir_node *block)
501 memperm_t entry, *res;
505 hash = hash_irn(block);
507 res = set_find(env->memperms, &entry, sizeof(entry), hash);
510 entry.entrycount = 0;
511 entry.entries = NULL;
512 res = set_insert(env->memperms, &entry, sizeof(entry), hash);
518 static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
520 ir_graph *irg = be_get_birg_irg(env->birg);
521 ir_type *frame = get_irg_frame_type(irg);
522 ir_entity *res = frame_alloc_area(frame, slot->size, slot->align, 0);
524 /* adjust size of the entity type... */
525 ir_type *enttype = get_entity_type(res);
526 set_type_size_bytes(enttype, slot->size);
534 * Enlarges a spillslot (if necessary) so that it can carry a value of size
535 * @p othersize and alignment @p otheralign.
537 static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
539 if(othersize > slot->size) {
540 slot->size = othersize;
542 if(otheralign > slot->align) {
543 if(otheralign % slot->align != 0)
544 slot->align *= otheralign;
546 slot->align = otheralign;
547 } else if(slot->align % otheralign != 0) {
548 slot->align *= otheralign;
553 static void assign_spill_entity(const arch_env_t *arch_env, ir_node *node,
561 arity = get_irn_arity(node);
562 for(i = 0; i < arity; ++i) {
563 ir_node *in = get_irn_n(node, i);
566 assign_spill_entity(arch_env, in, entity);
571 /* beware: we might have Stores with Memory Proj's, ia32 fisttp for instance */
572 node = skip_Proj(node);
573 assert(arch_get_frame_entity(arch_env, node) == NULL);
574 arch_set_frame_entity(arch_env, node, entity);
578 * Create stack entities for the spillslots and assign them to the spill and
581 static void assign_spillslots(be_fec_env_t *env)
583 const arch_env_t *arch_env = env->arch_env;
587 spill_slot_t* spillslots;
589 spillcount = set_count(env->spills);
590 spillslots = alloca(spillcount * sizeof(spillslots[0]));
592 memset(spillslots, 0, spillcount * sizeof(spillslots[0]));
594 /* construct spillslots */
595 for(spill = set_first(env->spills); spill != NULL;
596 spill = set_next(env->spills)) {
598 int slotid = spill->spillslot;
599 const ir_mode *mode = spill->mode;
600 spill_slot_t *slot = & (spillslots[slotid]);
601 int size = get_mode_size_bytes(mode);
602 int align = spill->alignment;
604 if(slot->align == 0 && slot->size == 0) {
608 enlarge_spillslot(slot, align, size);
612 for(spill = set_first(env->spills); spill != NULL;
613 spill = set_next(env->spills)) {
615 ir_node *node = spill->spill;
616 int slotid = spill->spillslot;
619 slot = &spillslots[slotid];
620 if(slot->entity == NULL) {
621 create_stack_entity(env, slot);
626 ir_node *block = get_nodes_block(node);
628 /* should be a PhiM */
629 assert(is_Phi(node));
631 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
632 ir_node *arg = get_irn_n(node, i);
633 ir_node *predblock = get_Block_cfgpred_block(block, i);
637 argspill = get_spill(env, arg);
638 assert(argspill != NULL);
640 argslotid = argspill->spillslot;
641 if(slotid != argslotid) {
643 memperm_entry_t *entry;
644 spill_slot_t *argslot = &spillslots[argslotid];
645 if(argslot->entity == NULL) {
646 create_stack_entity(env, argslot);
649 memperm = get_memperm(env, predblock);
651 entry = obstack_alloc(&env->obst, sizeof(entry[0]));
654 entry->in = argslot->entity;
655 entry->out = slot->entity;
656 entry->next = memperm->entries;
657 memperm->entrycount++;
658 memperm->entries = entry;
662 assign_spill_entity(arch_env, node, slot->entity);
666 for(i = 0; i < ARR_LEN(env->reloads); ++i) {
667 ir_node *reload = env->reloads[i];
668 ir_node *spillnode = get_memory_edge(reload);
669 spill_t *spill = get_spill(env, spillnode);
670 const spill_slot_t *slot = & spillslots[spill->spillslot];
672 assert(slot->entity != NULL);
674 arch_set_frame_entity(arch_env, reload, slot->entity);
679 * Returns the last node in a block which is no control flow changing node
681 static ir_node *get_end_of_block_insertion_point(ir_node* block)
683 ir_node* ins = sched_last(block);
684 while(is_Proj(ins) && get_irn_mode(ins) == mode_X) {
685 ins = sched_prev(ins);
691 ir_node *prev = sched_prev(ins);
701 static void create_memperms(be_fec_env_t *env)
703 const arch_env_t *arch_env = env->arch_env;
704 ir_graph *irg = be_get_birg_irg(env->birg);
707 for(memperm = set_first(env->memperms); memperm != NULL; memperm = set_next(env->memperms)) {
709 memperm_entry_t *entry;
711 ir_node** nodes = alloca(memperm->entrycount * sizeof(nodes[0]));
712 ir_node* mempermnode;
714 assert(memperm->entrycount > 0);
716 for(entry = memperm->entries, i = 0; entry != NULL; entry = entry->next, ++i) {
717 ir_node* arg = get_irn_n(entry->node, entry->pos);
721 mempermnode = be_new_MemPerm(arch_env, irg, memperm->block,
722 memperm->entrycount, nodes);
724 /* insert node into schedule */
725 blockend = get_end_of_block_insertion_point(memperm->block);
726 sched_add_before(blockend, mempermnode);
727 stat_ev_dbl("mem_perm", memperm->entrycount);
730 for(entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
732 ir_node* arg = get_irn_n(entry->node, entry->pos);
734 be_set_MemPerm_in_entity(mempermnode, i, entry->in);
735 be_set_MemPerm_out_entity(mempermnode, i, entry->out);
736 set_irg_current_block(irg, memperm->block);
737 proj = new_Proj(mempermnode, get_irn_mode(arg), i);
739 set_irn_n(entry->node, entry->pos, proj);
744 static int count_spillslots(const be_fec_env_t *env)
746 const spill_t *spill;
747 int spillcount = set_count(env->spills);
748 bitset_t *counted = bitset_alloca(spillcount);
752 for(spill = set_first(env->spills); spill != NULL;
753 spill = set_next(env->spills)) {
754 int spillslot = spill->spillslot;
755 if(!bitset_is_set(counted, spillslot)) {
757 bitset_set(counted, spillslot);
764 be_fec_env_t *be_new_frame_entity_coalescer(be_irg_t *birg)
766 const arch_env_t *arch_env = birg->main_env->arch_env;
767 be_fec_env_t *env = xmalloc(sizeof(env[0]));
769 be_liveness_assure_chk(be_assure_liveness(birg));
771 obstack_init(&env->obst);
772 env->arch_env = arch_env;
774 env->spills = new_set(cmp_spill, 10);
775 env->reloads = NEW_ARR_F(ir_node*, 0);
776 env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
777 env->memperms = new_set(cmp_memperm, 10);
782 void be_free_frame_entity_coalescer(be_fec_env_t *env)
784 del_set(env->memperms);
785 DEL_ARR_F(env->reloads);
786 DEL_ARR_F(env->affinity_edges);
787 del_set(env->spills);
788 obstack_free(&env->obst, NULL);
793 void be_assign_entities(be_fec_env_t *env)
795 stat_ev_dbl("spillslots", set_count(env->spills));
797 if(be_coalesce_spill_slots) {
798 do_greedy_coalescing(env);
801 stat_ev_dbl("spillslots_after_coalescing", count_spillslots(env));
803 assign_spillslots(env);
805 create_memperms(env);
809 * This walker function searches for reloads and collects all the spills
810 * and memphis attached to them.
812 static void collect_spills_walker(ir_node *node, void *data)
814 be_fec_env_t *env = data;
815 const arch_env_t *arch_env = env->arch_env;
817 const arch_register_class_t *cls;
820 /* classify returns classification of the irn the proj is attached to */
824 if (!arch_irn_class_is(arch_env, node, reload))
827 mode = get_irn_mode(node);
828 cls = arch_get_irn_reg_class(arch_env, node, -1);
829 align = arch_env_get_reg_class_alignment(arch_env, cls);
831 be_node_needs_frame_entity(env, node, mode, align);
834 void be_coalesce_spillslots(be_irg_t *birg)
836 be_fec_env_t *env = be_new_frame_entity_coalescer(birg);
838 /* collect reloads */
839 irg_walk_graph(birg->irg, NULL, collect_spills_walker, env);
841 be_assign_entities(env);
843 be_free_frame_entity_coalescer(env);
846 void be_init_spillslots(void)
848 FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");
851 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots);