2 * Copyright (C) 1995-2007 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"
45 #include "bespillslots.h"
46 #include "bechordal_t.h"
47 #include "bejavacoal.h"
48 #include "benodesets.h"
49 #include "bestatevent.h"
50 #include "bespilloptions.h"
56 #define DBG_COALESCING 1
57 #define DBG_INTERFERENCES 2
60 static firm_dbg_module_t *dbg = NULL;
63 typedef struct _spill_t {
65 /** mode of the spilled value */
67 /** alignment for the spilled value */
69 /** index into spillslot_unionfind unionfind structure */
73 typedef struct _affinity_edge_t {
78 struct _be_fec_env_t {
80 const arch_env_t *arch_env;
84 affinity_edge_t **affinity_edges;
88 /** Compare 2 affinity edges (used in quicksort) */
89 static int cmp_affinity(const void *d1, const void *d2)
91 const affinity_edge_t * const *e1 = d1;
92 const affinity_edge_t * const *e2 = d2;
94 // sort in descending order
95 return (*e1)->affinity < (*e2)->affinity ? 1 : -1;
98 static int cmp_spill(const void* d1, const void* d2, size_t size)
100 const spill_t* s1 = d1;
101 const spill_t* s2 = d2;
102 return s1->spill != s2->spill;
105 static spill_t *get_spill(be_fec_env_t *env, ir_node *node)
108 int hash = nodeset_hash(node);
111 res = set_find(env->spills, &spill, sizeof(spill), hash);
117 * ____ _ _ _ ____ _ _ _
118 * / ___|___ | | | ___ ___| |_ / ___| _ __ (_) | |___
119 * | | / _ \| | |/ _ \/ __| __| \___ \| '_ \| | | / __|
120 * | |__| (_) | | | __/ (__| |_ ___) | |_) | | | \__ \
121 * \____\___/|_|_|\___|\___|\__| |____/| .__/|_|_|_|___/
125 static INLINE ir_node *get_memory_edge(const ir_node *node)
129 arity = get_irn_arity(node);
130 for(i = arity - 1; i >= 0; --i) {
131 ir_node *arg = get_irn_n(node, i);
132 if(get_irn_mode(arg) == mode_M)
139 static spill_t *collect_spill(be_fec_env_t *env, ir_node *node,
140 const ir_mode *mode, int align)
143 int hash = nodeset_hash(node);
145 /* insert into set of spills if not already there */
147 res = set_find(env->spills, &spill, sizeof(spill), hash);
150 spill.spillslot = set_count(env->spills);
152 spill.alignment = align;
153 res = set_insert(env->spills, &spill, sizeof(spill), hash);
155 assert(res->mode == mode);
156 assert(res->alignment == align);
162 static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
163 const ir_mode *mode, int align)
167 int hash = nodeset_hash(node);
168 const ir_exec_freq *exec_freq = be_get_birg_exec_freq(env->birg);
170 assert(is_Phi(node));
173 res = set_find(env->spills, &spill, sizeof(spill), hash);
175 assert(res->mode == mode);
176 assert(res->alignment == align);
180 spill.spillslot = set_count(env->spills);
182 spill.alignment = align;
183 res = set_insert(env->spills, &spill, sizeof(spill), hash);
185 // collect attached spills and mem-phis
186 arity = get_irn_arity(node);
187 for(i = 0; i < arity; ++i) {
188 affinity_edge_t *affinty_edge;
189 ir_node *arg = get_irn_n(node, i);
193 arg_spill = collect_memphi(env, arg, mode, align);
195 arg_spill = collect_spill(env, arg, mode, align);
198 // add an affinity edge
199 affinty_edge = obstack_alloc(&env->obst, sizeof(affinty_edge[0]));
200 affinty_edge->affinity = get_block_execfreq(exec_freq, get_nodes_block(arg));
201 affinty_edge->slot1 = res->spillslot;
202 affinty_edge->slot2 = arg_spill->spillslot;
203 ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
209 void be_node_needs_frame_entity(be_fec_env_t *env, ir_node *node,
210 const ir_mode *mode, int align)
212 ir_node *spillnode = get_memory_edge(node);
215 assert(spillnode != NULL);
217 if (is_Phi(spillnode)) {
218 spill = collect_memphi(env, spillnode, mode, align);
220 spill = collect_spill(env, spillnode, mode, align);
223 ARR_APP1(ir_node *, env->reloads, node);
228 * / ___|___ __ _| | ___ ___ ___ ___ / ___|| | ___ | |_ ___
229 * | | / _ \ / _` | |/ _ \/ __|/ __/ _ \ \___ \| |/ _ \| __/ __|
230 * | |__| (_) | (_| | | __/\__ \ (_| __/ ___) | | (_) | |_\__ \
231 * \____\___/ \__,_|_|\___||___/\___\___| |____/|_|\___/ \__|___/
234 static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
235 int* spillslot_unionfind, int s1, int s2)
241 // merge spillslots and interferences
242 res = uf_union(spillslot_unionfind, s1, s2);
243 // we assume that we always merge s2 to s1 so swap s1, s2 if necessary
250 bitset_or(interferences[s1], interferences[s2]);
252 // update other interferences
253 spillcount = set_count(env->spills);
254 for(i = 0; i < spillcount; ++i) {
255 bitset_t *intfs = interferences[i];
256 if(bitset_is_set(intfs, s2))
257 bitset_set(intfs, s1);
264 * A greedy coalescing algorithm for spillslots:
265 * 1. Sort the list of affinity edges
266 * 2. Try to merge slots with affinity edges (most expensive slots first)
267 * 3. Try to merge everything else that is possible
269 static void do_greedy_coalescing(be_fec_env_t *env)
275 int affinity_edge_count;
276 bitset_t **interferences;
277 int* spillslot_unionfind;
278 const be_lv_t *lv = be_get_birg_liveness(env->birg);
280 spillcount = set_count(env->spills);
284 DBG((dbg, DBG_COALESCING, "Coalescing %d spillslots\n", spillcount));
286 interferences = alloca(spillcount * sizeof(interferences[0]));
287 spillslot_unionfind = alloca(spillcount * sizeof(spillslot_unionfind[0]));
288 spilllist = alloca(spillcount * sizeof(spilllist[0]));
290 uf_init(spillslot_unionfind, 0, spillcount);
293 memset(spilllist, 0, spillcount * sizeof(spilllist[0]));
296 for(spill = set_first(env->spills), i = 0; spill != NULL; spill = set_next(env->spills), ++i) {
297 assert(spill->spillslot < spillcount);
298 spilllist[spill->spillslot] = spill;
301 for(i = 0; i < spillcount; ++i) {
302 interferences[i] = bitset_alloca(spillcount);
305 /* construct interferences */
306 for (i = 0; i < spillcount; ++i) {
307 ir_node *spill1 = spilllist[i]->spill;
309 if (is_NoMem(spill1))
312 for(i2 = i+1; i2 < spillcount; ++i2) {
313 ir_node *spill2 = spilllist[i2]->spill;
315 if (is_NoMem(spill2))
318 if (values_interfere(lv, spill1, spill2)) {
319 DBG((dbg, DBG_INTERFERENCES, "Slot %d and %d interfere\n", i, i2));
320 bitset_set(interferences[i], i2);
321 bitset_set(interferences[i2], i);
326 /* sort affinity edges */
327 affinity_edge_count = ARR_LEN(env->affinity_edges);
328 qsort(env->affinity_edges, affinity_edge_count, sizeof(env->affinity_edges[0]), cmp_affinity);
330 //dump_interference_graph(env, interferences, "before");
332 /* try to merge affine nodes */
333 for(i = 0; i < affinity_edge_count; ++i) {
334 const affinity_edge_t *edge = env->affinity_edges[i];
335 int s1 = uf_find(spillslot_unionfind, edge->slot1);
336 int s2 = uf_find(spillslot_unionfind, edge->slot2);
338 /* test if values interfere */
339 if (bitset_is_set(interferences[s1], s2)) {
340 assert(bitset_is_set(interferences[s2], s1));
344 DBG((dbg, DBG_COALESCING, "Merging %d and %d because of affinity edge\n", s1, s2));
346 merge_interferences(env, interferences, spillslot_unionfind, s1, s2);
349 // try to merge as much remaining spillslots as possible
350 for(i = 0; i < spillcount; ++i) {
351 int s1 = uf_find(spillslot_unionfind, i);
355 for(i2 = i+1; i2 < spillcount; ++i2) {
356 int s2 = uf_find(spillslot_unionfind, i2);
360 /* test if values interfere
361 * we have to test n1-n2 and n2-n1, because only 1 side gets updated
362 * when node merging occurs
364 if(bitset_is_set(interferences[s1], s2)) {
365 assert(bitset_is_set(interferences[s2], s1));
369 DBG((dbg, DBG_COALESCING, "Merging %d and %d because it is possible\n", s1, s2));
371 if(merge_interferences(env, interferences, spillslot_unionfind, s1, s2) != 0) {
372 // we can break the loop here, because s2 is the new supernode now
373 // and we'll test s2 again later anyway
379 // assign spillslots to spills
380 for(i = 0; i < spillcount; ++i) {
381 spill_t *spill = spilllist[i];
383 spill->spillslot = uf_find(spillslot_unionfind, i);
386 //dump_interference_graph(env, interferences, "after");
391 * / \ ___ ___(_) __ _ _ __ | ____|_ __ | |_(_) |_(_) ___ ___
392 * / _ \ / __/ __| |/ _` | '_ \ | _| | '_ \| __| | __| |/ _ \/ __|
393 * / ___ \\__ \__ \ | (_| | | | | | |___| | | | |_| | |_| | __/\__ \
394 * /_/ \_\___/___/_|\__, |_| |_| |_____|_| |_|\__|_|\__|_|\___||___/
398 typedef struct _spill_slot_t {
404 typedef struct _memperm_entry_t {
409 struct _memperm_entry_t *next;
412 typedef struct _memperm_t {
415 memperm_entry_t *entries;
418 static int cmp_memperm(const void* d1, const void* d2, size_t size)
420 const memperm_t* e1 = d1;
421 const memperm_t* e2 = d2;
422 return e1->block != e2->block;
425 static memperm_t *get_memperm(be_fec_env_t *env, ir_node *block)
427 memperm_t entry, *res;
431 hash = nodeset_hash(block);
433 res = set_find(env->memperms, &entry, sizeof(entry), hash);
436 entry.entrycount = 0;
437 entry.entries = NULL;
438 res = set_insert(env->memperms, &entry, sizeof(entry), hash);
444 static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
446 ir_graph *irg = be_get_birg_irg(env->birg);
447 ir_type *frame = get_irg_frame_type(irg);
448 ir_entity *res = frame_alloc_area(frame, slot->size, slot->align, 0);
450 /* adjust size of the entity type... */
451 ir_type *enttype = get_entity_type(res);
452 set_type_size_bytes(enttype, slot->size);
460 * Enlarges a spillslot (if necessary) so that it can carry a value of size
461 * @p othersize and alignment @p otheralign.
463 static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
465 if(othersize > slot->size) {
466 slot->size = othersize;
468 if(otheralign > slot->align) {
469 if(otheralign % slot->align != 0)
470 slot->align *= otheralign;
472 slot->align = otheralign;
473 } else if(slot->align % otheralign != 0) {
474 slot->align *= otheralign;
479 * Create stack entities for the spillslots and assign them to the spill and
482 static void assign_spillslots(be_fec_env_t *env)
484 const arch_env_t *arch_env = env->arch_env;
488 spill_slot_t* spillslots;
490 spillcount = set_count(env->spills);
491 spillslots = alloca(spillcount * sizeof(spillslots[0]));
493 memset(spillslots, 0, spillcount * sizeof(spillslots[0]));
495 // construct spillslots
496 for(spill = set_first(env->spills); spill != NULL; spill = set_next(env->spills)) {
497 int slotid = spill->spillslot;
498 const ir_mode *mode = spill->mode;
499 spill_slot_t *slot = & (spillslots[slotid]);
500 int size = get_mode_size_bytes(mode);
501 int align = spill->alignment;
503 if(slot->align == 0 && slot->size == 0) {
507 enlarge_spillslot(slot, align, size);
511 for(spill = set_first(env->spills); spill != NULL; spill = set_next(env->spills)) {
513 ir_node *node = spill->spill;
514 int slotid = spill->spillslot;
516 slot = &spillslots[slotid];
517 if(slot->entity == NULL) {
518 create_stack_entity(env, slot);
523 ir_node *block = get_nodes_block(node);
526 assert(is_Phi(node));
528 for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
529 ir_node *arg = get_irn_n(node, i);
530 ir_node *predblock = get_Block_cfgpred_block(block, i);
534 argspill = get_spill(env, arg);
535 assert(argspill != NULL);
537 argslotid = argspill->spillslot;
538 if(slotid != argslotid) {
540 memperm_entry_t *entry;
541 spill_slot_t *argslot = &spillslots[argslotid];
542 if(argslot->entity == NULL) {
543 create_stack_entity(env, argslot);
546 memperm = get_memperm(env, predblock);
548 entry = obstack_alloc(&env->obst, sizeof(entry[0]));
551 entry->in = argslot->entity;
552 entry->out = slot->entity;
553 entry->next = memperm->entries;
554 memperm->entrycount++;
555 memperm->entries = entry;
560 arch_set_frame_entity(arch_env, node, slot->entity);
564 for(i = 0; i < ARR_LEN(env->reloads); ++i) {
565 ir_node* reload = env->reloads[i];
566 ir_node* spillnode = get_memory_edge(reload);
567 spill_t *spill = get_spill(env, spillnode);
568 const spill_slot_t *slot = & spillslots[spill->spillslot];
570 assert(slot->entity != NULL);
572 arch_set_frame_entity(arch_env, reload, slot->entity);
577 * Returns the last node in a block which is no control flow changing node
579 static ir_node *get_end_of_block_insertion_point(ir_node* block)
581 ir_node* ins = sched_last(block);
582 while(is_Proj(ins) && get_irn_mode(ins) == mode_X) {
583 ins = sched_prev(ins);
589 ir_node *prev = sched_prev(ins);
599 static void create_memperms(be_fec_env_t *env)
601 const arch_env_t *arch_env = env->arch_env;
602 ir_graph *irg = be_get_birg_irg(env->birg);
605 for(memperm = set_first(env->memperms); memperm != NULL; memperm = set_next(env->memperms)) {
607 memperm_entry_t *entry;
609 ir_node** nodes = alloca(memperm->entrycount * sizeof(nodes[0]));
610 ir_node* mempermnode;
612 assert(memperm->entrycount > 0);
614 for(entry = memperm->entries, i = 0; entry != NULL; entry = entry->next, ++i) {
615 ir_node* arg = get_irn_n(entry->node, entry->pos);
619 mempermnode = be_new_MemPerm(arch_env, irg, memperm->block,
620 memperm->entrycount, nodes);
622 // insert node into schedule
623 blockend = get_end_of_block_insertion_point(memperm->block);
624 sched_add_before(blockend, mempermnode);
625 be_stat_ev("mem_perm", memperm->entrycount);
628 for(entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
630 ir_node* arg = get_irn_n(entry->node, entry->pos);
632 be_set_MemPerm_in_entity(mempermnode, i, entry->in);
633 be_set_MemPerm_out_entity(mempermnode, i, entry->out);
634 set_irg_current_block(irg, memperm->block);
635 proj = new_Proj(mempermnode, get_irn_mode(arg), i);
636 sched_add_before(blockend, proj);
638 set_irn_n(entry->node, entry->pos, proj);
643 static int count_spillslots(const be_fec_env_t *env)
645 const spill_t *spill;
646 int spillcount = set_count(env->spills);
647 bitset_t *counted = bitset_alloca(spillcount);
651 for(spill = set_first(env->spills); spill != NULL;
652 spill = set_next(env->spills)) {
653 int spillslot = spill->spillslot;
654 if(!bitset_is_set(counted, spillslot)) {
656 bitset_set(counted, spillslot);
663 be_fec_env_t *be_new_frame_entity_coalescer(be_irg_t *birg)
665 const arch_env_t *arch_env = birg->main_env->arch_env;
666 be_fec_env_t *env = xmalloc(sizeof(env[0]));
668 be_assure_liveness(birg);
670 obstack_init(&env->obst);
671 env->arch_env = arch_env;
673 env->spills = new_set(cmp_spill, 10);
674 env->reloads = NEW_ARR_F(ir_node*, 0);
675 env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
676 env->memperms = new_set(cmp_memperm, 10);
681 void be_free_frame_entity_coalescer(be_fec_env_t *env)
683 del_set(env->memperms);
684 DEL_ARR_F(env->reloads);
685 DEL_ARR_F(env->affinity_edges);
686 del_set(env->spills);
687 obstack_free(&env->obst, NULL);
692 void be_assign_entities(be_fec_env_t *env)
694 if(be_stat_ev_is_active()) {
695 int count = set_count(env->spills);
696 be_stat_ev("spillslots", count);
699 if(be_coalesce_spill_slots) {
700 do_greedy_coalescing(env);
703 if(be_stat_ev_is_active()) {
704 int count = count_spillslots(env);
705 be_stat_ev("spillslots_after_coalescing", count);
708 assign_spillslots(env);
710 create_memperms(env);
714 * This walker function searches for reloads and collects all the spills
715 * and memphis attached to them.
717 static void collect_spills_walker(ir_node *node, void *data)
719 be_fec_env_t *env = data;
720 const arch_env_t *arch_env = env->arch_env;
722 const arch_register_class_t *cls;
725 /* classify returns classification of the irn the proj is attached to */
729 if (!arch_irn_class_is(arch_env, node, reload))
732 mode = get_irn_mode(node);
733 cls = arch_get_irn_reg_class(arch_env, node, -1);
734 align = arch_isa_get_reg_class_alignment(arch_env_get_isa(arch_env), cls);
736 be_node_needs_frame_entity(env, node, mode, align);
739 void be_coalesce_spillslots(be_irg_t *birg)
741 be_fec_env_t *env = be_new_frame_entity_coalescer(birg);
743 /* collect reloads */
744 irg_walk_graph(birg->irg, NULL, collect_spills_walker, env);
746 be_assign_entities(env);
748 be_free_frame_entity_coalescer(env);
751 void be_init_spillslots(void)
753 FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");
756 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots);