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 implementation of the spill/reload placement abstraction layer
23 * @author Daniel Grund, Sebastian Hack, Matthias Braun
34 #include "iredges_t.h"
35 #include "irbackedge_t.h"
45 #include "irnodeset.h"
52 #include "bespillutil.h"
55 #include "bechordal_t.h"
57 #include "bessaconstr.h"
60 #include "beintlive_t.h"
64 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
66 #define REMAT_COST_INFINITE 1000
68 typedef struct reloader_t reloader_t;
72 ir_node *rematted_node;
73 int remat_cost_delta; /** costs needed for rematerialization,
74 compared to placing a reload */
77 typedef struct spill_t spill_t;
80 ir_node *after; /**< spill has to be placed after this node (or earlier) */
84 typedef struct spill_info_t spill_info_t;
86 ir_node *to_spill; /**< the value that should get spilled */
87 reloader_t *reloaders; /**< list of places where the value should get
89 spill_t *spills; /**< list of latest places where spill must be
91 double spill_costs; /**< costs needed for spilling the value */
92 const arch_register_class_t *reload_cls; /** the register class in which the
93 reload should be placed */
94 bool spilled_phi; /* true when the whole Phi has been spilled and
95 will be replaced with a PhiM. false if only the
96 value of the Phi gets spilled */
100 const arch_env_t *arch_env;
103 int spill_cost; /**< the cost of a single spill node */
104 int reload_cost; /**< the cost of a reload node */
105 set *spills; /**< all spill_info_t's, which must be
107 spill_info_t **mem_phis; /**< set of all spilled phis. */
109 unsigned spill_count;
110 unsigned reload_count;
111 unsigned remat_count;
112 unsigned spilled_phi_count;
116 * Compare two spill infos.
118 static int cmp_spillinfo(const void *x, const void *y, size_t size)
120 const spill_info_t *xx = (const spill_info_t*)x;
121 const spill_info_t *yy = (const spill_info_t*)y;
124 return xx->to_spill != yy->to_spill;
128 * Returns spill info for a specific value (the value that is to be spilled)
130 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
132 spill_info_t info, *res;
133 int hash = hash_irn(value);
135 info.to_spill = value;
136 res = set_find(spill_info_t, env->spills, &info, sizeof(info), hash);
139 info.reloaders = NULL;
141 info.spill_costs = -1;
142 info.reload_cls = NULL;
143 info.spilled_phi = false;
144 res = set_insert(spill_info_t, env->spills, &info, sizeof(info), hash);
150 spill_env_t *be_new_spill_env(ir_graph *irg)
152 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
154 spill_env_t *env = XMALLOC(spill_env_t);
155 env->spills = new_set(cmp_spillinfo, 1024);
157 env->arch_env = arch_env;
158 env->mem_phis = NEW_ARR_F(spill_info_t*, 0);
159 env->spill_cost = arch_env->spill_cost;
160 env->reload_cost = arch_env->reload_cost;
161 obstack_init(&env->obst);
163 env->spill_count = 0;
164 env->reload_count = 0;
165 env->remat_count = 0;
166 env->spilled_phi_count = 0;
171 void be_delete_spill_env(spill_env_t *env)
173 del_set(env->spills);
174 DEL_ARR_F(env->mem_phis);
175 obstack_free(&env->obst, NULL);
181 * | _ \| | __ _ ___ ___ | _ \ ___| | ___ __ _ __| |___
182 * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
183 * | __/| | (_| | (_| __/ | _ < __/ | (_) | (_| | (_| \__ \
184 * |_| |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
188 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
190 spill_info_t *spill_info = get_spillinfo(env, to_spill);
195 assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
196 DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
198 /* Just for safety make sure that we do not insert the spill in front of a phi */
199 assert(!is_Phi(sched_next(after)));
201 /* spills that are dominated by others are not needed */
203 s = spill_info->spills;
204 for ( ; s != NULL; s = s->next) {
205 /* no need to add this spill if it is dominated by another */
206 if (value_dominates(s->after, after)) {
207 DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
210 /* remove spills that we dominate */
211 if (value_dominates(after, s->after)) {
212 DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
214 last->next = s->next;
216 spill_info->spills = s->next;
223 spill = OALLOC(&env->obst, spill_t);
224 spill->after = after;
225 spill->next = spill_info->spills;
228 spill_info->spills = spill;
231 void be_add_reload(spill_env_t *env, ir_node *to_spill, ir_node *before, const arch_register_class_t *reload_cls, int allow_remat)
236 assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
238 info = get_spillinfo(env, to_spill);
240 if (is_Phi(to_spill)) {
243 /* create spillinfos for the phi arguments */
244 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
245 ir_node *arg = get_irn_n(to_spill, i);
246 get_spillinfo(env, arg);
250 assert(!be_is_Keep(before));
252 /* put reload into list */
253 rel = OALLOC(&env->obst, reloader_t);
254 rel->next = info->reloaders;
255 rel->reloader = before;
256 rel->rematted_node = NULL;
257 rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
259 info->reloaders = rel;
260 assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
261 info->reload_cls = reload_cls;
263 DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
264 to_spill, before, allow_remat ? "" : " not"));
267 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
269 ir_node *last = sched_last(block);
271 /* we might have keeps behind the jump... */
272 while (be_is_Keep(last)) {
273 last = sched_prev(last);
274 assert(!sched_is_end(last));
277 assert(is_cfop(last));
279 /* add the reload before the (cond-)jump */
284 * determine final spill position: it should be after all phis, keep nodes
285 * and behind nodes marked as prolog
287 static ir_node *determine_spill_point(ir_node *node)
289 node = skip_Proj(node);
291 ir_node *next = sched_next(node);
292 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
300 * Returns the point at which you can insert a node that should be executed
301 * before block @p block when coming from pred @p pos.
303 static ir_node *get_block_insertion_point(ir_node *block, int pos)
307 /* simply add the reload to the beginning of the block if we only have 1
308 * predecessor. We don't need to check for phis as there can't be any in a
309 * block with only 1 pred. */
310 if (get_Block_n_cfgpreds(block) == 1) {
311 assert(!is_Phi(sched_first(block)));
312 return sched_first(block);
315 /* We have to reload the value in pred-block */
316 predblock = get_Block_cfgpred_block(block, pos);
317 return be_get_end_of_block_insertion_point(predblock);
320 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
321 int pos, const arch_register_class_t *reload_cls,
324 ir_node *before = get_block_insertion_point(block, pos);
325 be_add_reload(env, to_spill, before, reload_cls, allow_remat);
328 void be_spill_phi(spill_env_t *env, ir_node *node)
334 assert(is_Phi(node));
336 info = get_spillinfo(env, node);
337 info->spilled_phi = true;
338 ARR_APP1(spill_info_t*, env->mem_phis, info);
340 /* create spills for the phi arguments */
341 block = get_nodes_block(node);
342 for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
343 ir_node *arg = get_irn_n(node, i);
346 /* some backends have virtual noreg/unknown nodes that are not scheduled
347 * and simply always available. */
348 if (!sched_is_scheduled(arg)) {
349 ir_node *pred_block = get_Block_cfgpred_block(block, i);
350 insert = be_get_end_of_block_insertion_point(pred_block);
351 insert = sched_prev(insert);
353 insert = determine_spill_point(arg);
356 be_add_spill(env, arg, insert);
362 * / ___|_ __ ___ __ _| |_ ___ / ___| _ __ (_) | |___
363 * | | | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
364 * | |___| | | __/ (_| | || __/ ___) | |_) | | | \__ \
365 * \____|_| \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
369 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
374 * @param senv the spill environment
375 * @param irn the node that should be spilled
376 * @param ctx_irn an user of the spilled node
378 * @return a be_Spill node
380 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
382 ir_node *to_spill = spillinfo->to_spill;
383 const ir_node *insn = skip_Proj_const(to_spill);
386 /* determine_spill_costs must have been run before */
387 assert(spillinfo->spill_costs >= 0);
389 /* some backends have virtual noreg/unknown nodes that are not scheduled
390 * and simply always available. */
391 if (!sched_is_scheduled(insn)) {
392 /* override spillinfos or create a new one */
393 ir_graph *irg = get_irn_irg(to_spill);
394 spillinfo->spills->spill = get_irg_no_mem(irg);
395 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
399 DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
400 spill = spillinfo->spills;
401 for ( ; spill != NULL; spill = spill->next) {
402 ir_node *after = spill->after;
403 after = determine_spill_point(after);
405 spill->spill = arch_env_new_spill(env->arch_env, to_spill, after);
406 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
409 DBG((dbg, LEVEL_1, "\n"));
412 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
415 * If the first usage of a Phi result would be out of memory
416 * there is no sense in allocating a register for it.
417 * Thus we spill it and all its operands to the same spill slot.
418 * Therefore the phi/dataB becomes a phi/Memory
420 * @param senv the spill environment
421 * @param phi the Phi node that should be spilled
422 * @param ctx_irn an user of the spilled node
424 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
426 ir_graph *irg = env->irg;
427 ir_node *phi = spillinfo->to_spill;
428 ir_node *block = get_nodes_block(phi);
432 assert(!get_opt_cse());
433 DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
435 /* build a new PhiM */
436 int const arity = get_Phi_n_preds(phi);
437 ir_node **const ins = ALLOCAN(ir_node*, arity);
438 ir_node *const unknown = new_r_Unknown(irg, mode_M);
439 for (i = 0; i < arity; ++i) {
443 /* override or replace spills list... */
444 spill = OALLOC(&env->obst, spill_t);
445 spill->after = determine_spill_point(phi);
446 spill->spill = be_new_Phi(block, arity, ins, mode_M, arch_no_register_req);
448 sched_add_after(block, spill->spill);
450 spillinfo->spills = spill;
451 env->spilled_phi_count++;
453 for (i = 0; i < arity; ++i) {
454 ir_node *arg = get_irn_n(phi, i);
455 spill_info_t *arg_info = get_spillinfo(env, arg);
457 determine_spill_costs(env, arg_info);
458 spill_node(env, arg_info);
460 set_irn_n(spill->spill, i, arg_info->spills->spill);
462 DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
469 * @param senv the spill environment
470 * @param to_spill the node that should be spilled
472 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
474 /* node is already spilled */
475 if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
478 if (spillinfo->spilled_phi) {
479 spill_phi(env, spillinfo);
481 spill_irn(env, spillinfo);
488 * | _ \ ___ _ __ ___ __ _| |_ ___ _ __(_) __ _| (_)_______
489 * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_ / _ \
490 * | _ < __/ | | | | | (_| | || __/ | | | (_| | | |/ / __/
491 * |_| \_\___|_| |_| |_|\__,_|\__\___|_| |_|\__,_|_|_/___\___|
496 * Tests whether value @p arg is available before node @p reloader
497 * @returns 1 if value is available, 0 otherwise
499 static int is_value_available(spill_env_t *env, const ir_node *arg,
500 const ir_node *reloader)
502 if (is_Unknown(arg) || is_NoMem(arg))
505 if (arch_irn_is(skip_Proj_const(arg), spill))
508 if (arg == get_irg_frame(env->irg))
513 if (get_irn_mode(arg) == mode_T)
517 * Ignore registers are always available
519 if (arch_irn_is_ignore(arg))
526 * Check if a node is rematerializable. This tests for the following conditions:
528 * - The node itself is rematerializable
529 * - All arguments of the node are available or also rematerialisable
530 * - The costs for the rematerialisation operation is less or equal a limit
532 * Returns the costs needed for rematerialisation or something
533 * >= REMAT_COST_INFINITE if remat is not possible.
535 static int check_remat_conditions_costs(spill_env_t *env,
536 const ir_node *spilled, const ir_node *reloader, int parentcosts)
541 const ir_node *insn = skip_Proj_const(spilled);
543 assert(!be_is_Spill(insn));
544 if (!arch_irn_is(insn, rematerializable))
545 return REMAT_COST_INFINITE;
547 if (be_is_Reload(insn)) {
550 costs += arch_get_op_estimated_cost(insn);
552 if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
553 return REMAT_COST_INFINITE;
555 /* never rematerialize a node which modifies the flags.
556 * (would be better to test whether the flags are actually live at point
559 if (arch_irn_is(insn, modify_flags)) {
560 return REMAT_COST_INFINITE;
564 for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
565 ir_node *arg = get_irn_n(insn, i);
567 if (is_value_available(env, arg, reloader))
570 /* we have to rematerialize the argument as well */
573 /* we only support rematerializing 1 argument at the moment,
574 * as multiple arguments could increase register pressure */
575 return REMAT_COST_INFINITE;
578 costs += check_remat_conditions_costs(env, arg, reloader,
579 parentcosts + costs);
580 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
581 return REMAT_COST_INFINITE;
588 * Re-materialize a node.
590 * @param env the spill environment
591 * @param spilled the node that was spilled
592 * @param reloader a irn that requires a reload
594 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
600 ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
601 for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
602 ir_node *arg = get_irn_n(spilled, i);
604 if (is_value_available(env, arg, reloader)) {
607 ins[i] = do_remat(env, arg, reloader);
608 /* don't count the argument rematerialization as an extra remat */
613 /* create a copy of the node */
614 ir_node *const bl = get_nodes_block(reloader);
615 res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
616 get_irn_op(spilled), get_irn_mode(spilled),
617 get_irn_arity(spilled), ins);
618 copy_node_attr(env->irg, spilled, res);
619 arch_env_mark_remat(env->arch_env, res);
621 DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
623 if (! is_Proj(res)) {
624 /* insert in schedule */
626 sched_add_before(reloader, res);
633 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
635 ir_node *block = get_nodes_block(before);
636 double freq = get_block_execfreq(block);
639 return env->spill_cost * freq;
642 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
643 const ir_node *before)
646 /* is the node rematerializable? */
647 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
648 if (costs < (unsigned) env->reload_cost)
652 return env->reload_cost;
655 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
657 ir_node *block = get_nodes_block(before);
658 double freq = get_block_execfreq(block);
661 /* is the node rematerializable? */
662 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
663 if (costs < env->reload_cost)
667 return env->reload_cost * freq;
670 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
671 const ir_node *before)
673 return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
676 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
677 ir_node *block, int pos)
679 ir_node *before = get_block_insertion_point(block, pos);
680 return be_get_reload_costs(env, to_spill, before);
683 ir_node *be_new_spill(ir_node *value, ir_node *after)
685 ir_graph *irg = get_irn_irg(value);
686 ir_node *frame = get_irg_frame(irg);
687 const arch_register_class_t *cls = arch_get_irn_reg_class(value);
688 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
689 ir_node *block = get_block(after);
691 = be_new_Spill(cls, cls_frame, block, frame, value);
693 sched_add_after(after, spill);
697 ir_node *be_new_reload(ir_node *value, ir_node *spill, ir_node *before)
699 ir_graph *irg = get_irn_irg(value);
700 ir_node *frame = get_irg_frame(irg);
701 ir_node *block = get_block(before);
702 const arch_register_class_t *cls = arch_get_irn_reg_class(value);
703 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
704 ir_mode *mode = get_irn_mode(value);
707 assert(be_is_Spill(spill) || is_Phi(spill));
708 assert(get_irn_mode(spill) == mode_M);
710 reload = be_new_Reload(cls, cls_frame, block, frame, spill, mode);
711 sched_add_before(before, reload);
718 * |_ _|_ __ ___ ___ _ __| |_ | _ \ ___| | ___ __ _ __| |___
719 * | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
720 * | || | | \__ \ __/ | | |_ | _ < __/ | (_) | (_| | (_| \__ \
721 * |___|_| |_|___/\___|_| \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
726 * analyzes how to best spill a node and determine costs for that
728 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
730 ir_node *to_spill = spillinfo->to_spill;
731 const ir_node *insn = skip_Proj_const(to_spill);
732 ir_node *spill_block;
734 double spill_execfreq;
736 /* already calculated? */
737 if (spillinfo->spill_costs >= 0)
740 assert(!arch_irn_is(insn, dont_spill));
741 assert(!be_is_Reload(insn));
743 /* some backends have virtual noreg/unknown nodes that are not scheduled
744 * and simply always available.
745 * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
746 * predecessor (of a PhiM) but this test might match other things too...
748 if (!sched_is_scheduled(insn)) {
749 ir_graph *irg = get_irn_irg(to_spill);
750 /* override spillinfos or create a new one */
751 spill_t *spill = OALLOC(&env->obst, spill_t);
754 spill->spill = get_irg_no_mem(irg);
756 spillinfo->spills = spill;
757 spillinfo->spill_costs = 0;
759 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
763 spill_block = get_nodes_block(insn);
764 spill_execfreq = get_block_execfreq(spill_block);
766 if (spillinfo->spilled_phi) {
767 /* TODO calculate correct costs...
768 * (though we can't remat this node anyway so no big problem) */
769 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
773 if (spillinfo->spills != NULL) {
775 double spills_execfreq;
777 /* calculate sum of execution frequencies of individual spills */
779 s = spillinfo->spills;
780 for ( ; s != NULL; s = s->next) {
781 ir_node *spill_block = get_block(s->after);
782 double freq = get_block_execfreq(spill_block);
784 spills_execfreq += freq;
787 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
788 spills_execfreq * env->spill_cost,
789 spill_execfreq * env->spill_cost));
791 /* multi-/latespill is advantageous -> return*/
792 if (spills_execfreq < spill_execfreq) {
793 DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
794 spillinfo->spill_costs = spills_execfreq * env->spill_cost;
799 /* override spillinfos or create a new one */
800 spill = OALLOC(&env->obst, spill_t);
801 spill->after = determine_spill_point(to_spill);
805 spillinfo->spills = spill;
806 spillinfo->spill_costs = spill_execfreq * env->spill_cost;
807 DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
810 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
812 const spill_info_t *si = get_spillinfo(env, irn);
813 ir_node *start_block = get_irg_start_block(get_irn_irg(irn));
814 int n_blocks = get_Block_dom_max_subtree_pre_num(start_block);
815 bitset_t *reloads = bitset_alloca(n_blocks);
822 /* Fill the bitset with the dominance pre-order numbers
823 * of the blocks the reloads are located in. */
824 for (r = si->reloaders; r != NULL; r = r->next) {
825 ir_node *bl = get_nodes_block(r->reloader);
826 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
829 /* Now, cancel out all the blocks that are dominated by each spill.
830 * If the bitset is not empty after that, we have reloads that are
831 * not dominated by any spill. */
832 for (s = si->spills; s != NULL; s = s->next) {
833 ir_node *bl = get_nodes_block(s->after);
834 int start = get_Block_dom_tree_pre_num(bl);
835 int end = get_Block_dom_max_subtree_pre_num(bl);
837 bitset_clear_range(reloads, start, end);
840 if (!bitset_is_empty(reloads))
841 be_add_spill(env, si->to_spill, si->to_spill);
844 void be_insert_spills_reloads(spill_env_t *env)
846 size_t n_mem_phis = ARR_LEN(env->mem_phis);
849 be_timer_push(T_RA_SPILL_APPLY);
851 /* create all phi-ms first, this is needed so, that phis, hanging on
852 spilled phis work correctly */
853 for (i = 0; i < n_mem_phis; ++i) {
854 spill_info_t *info = env->mem_phis[i];
855 spill_node(env, info);
858 /* process each spilled node */
859 foreach_set(env->spills, spill_info_t, si) {
860 ir_node *to_spill = si->to_spill;
861 ir_node **copies = NEW_ARR_F(ir_node*, 0);
862 double all_remat_costs = 0; /** costs when we would remat all nodes */
863 bool force_remat = false;
866 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
868 determine_spill_costs(env, si);
870 /* determine possibility of rematerialisations */
872 /* calculate cost savings for each indivial value when it would
873 be rematted instead of reloaded */
874 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
877 int remat_cost_delta;
879 ir_node *reloader = rld->reloader;
881 if (rld->rematted_node != NULL) {
882 DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
883 rld->rematted_node, reloader));
886 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
887 DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
889 all_remat_costs = REMAT_COST_INFINITE;
893 remat_cost = check_remat_conditions_costs(env, to_spill,
895 if (remat_cost >= REMAT_COST_INFINITE) {
896 DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
898 rld->remat_cost_delta = REMAT_COST_INFINITE;
899 all_remat_costs = REMAT_COST_INFINITE;
903 remat_cost_delta = remat_cost - env->reload_cost;
904 rld->remat_cost_delta = remat_cost_delta;
905 block = is_Block(reloader) ? reloader : get_nodes_block(reloader);
906 freq = get_block_execfreq(block);
907 all_remat_costs += remat_cost_delta * freq;
908 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
909 "%d (rel %f)\n", reloader, remat_cost_delta,
910 remat_cost_delta * freq));
912 if (all_remat_costs < REMAT_COST_INFINITE) {
913 /* we don't need the costs for the spill if we can remat
915 all_remat_costs -= si->spill_costs;
917 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
918 env->spill_cost, si->spill_costs));
921 if (all_remat_costs < 0) {
922 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
928 /* go through all reloads for this spill */
929 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
930 ir_node *copy; /* a reload is a "copy" of the original value */
932 if (rld->rematted_node != NULL) {
933 copy = rld->rematted_node;
934 sched_add_before(rld->reloader, copy);
935 } else if (be_do_remats &&
936 (force_remat || rld->remat_cost_delta < 0)) {
937 copy = do_remat(env, to_spill, rld->reloader);
939 /* make sure we have a spill */
942 /* create a reload, use the first spill for now SSA
943 * reconstruction for memory comes below */
944 assert(si->spills != NULL);
945 copy = arch_env_new_reload(env->arch_env, si->to_spill,
946 si->spills->spill, rld->reloader);
950 DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
951 copy, to_spill, rld->reloader));
952 ARR_APP1(ir_node*, copies, copy);
955 /* if we had any reloads or remats, then we need to reconstruct the
956 * SSA form for the spilled value */
957 if (ARR_LEN(copies) > 0) {
958 be_ssa_construction_env_t senv;
959 /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
961 be_ssa_construction_init(&senv, env->irg);
962 be_ssa_construction_add_copy(&senv, to_spill);
963 be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
964 be_ssa_construction_fix_users(&senv, to_spill);
966 be_ssa_construction_destroy(&senv);
968 /* need to reconstruct SSA form if we had multiple spills */
969 if (si->spills != NULL && si->spills->next != NULL) {
973 be_ssa_construction_env_t senv;
975 be_ssa_construction_init(&senv, env->irg);
977 for ( ; spill != NULL; spill = spill->next) {
978 /* maybe we rematerialized the value and need no spill */
979 if (spill->spill == NULL)
981 be_ssa_construction_add_copy(&senv, spill->spill);
984 if (spill_count > 1) {
985 /* all reloads are attached to the first spill, fix them now */
986 be_ssa_construction_fix_users(&senv, si->spills->spill);
989 be_ssa_construction_destroy(&senv);
993 si->reloaders = NULL;
996 stat_ev_dbl("spill_spills", env->spill_count);
997 stat_ev_dbl("spill_reloads", env->reload_count);
998 stat_ev_dbl("spill_remats", env->remat_count);
999 stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1001 /* Matze: In theory be_ssa_construction should take care of the liveness...
1002 * try to disable this again in the future */
1003 be_invalidate_live_sets(env->irg);
1005 be_remove_dead_nodes_from_schedule(env->irg);
1007 be_timer_pop(T_RA_SPILL_APPLY);
1010 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill)
1011 void be_init_spill(void)
1013 FIRM_DBG_REGISTER(dbg, "firm.be.spill");