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"
56 #include "bestatevent.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;
71 ir_node *can_spill_after;
73 ir_node *rematted_node;
74 int remat_cost_delta; /** costs needed for rematerialization,
75 compared to placing a reload */
78 typedef struct spill_t spill_t;
81 ir_node *after; /**< spill has to be placed after this node (or earlier) */
85 typedef struct spill_info_t spill_info_t;
87 ir_node *to_spill; /**< the value that should get spilled */
88 reloader_t *reloaders; /**< list of places where the value should get
90 spill_t *spills; /**< list of latest places where spill must be
92 double spill_costs; /**< costs needed for spilling the value */
93 const arch_register_class_t *reload_cls; /** the register class in which the
94 reload should be placed */
95 bool spilled_phi; /* true when the whole Phi has been spilled and
96 will be replaced with a PhiM. false if only the
97 value of the Phi gets spilled */
101 const arch_env_t *arch_env;
104 int spill_cost; /**< the cost of a single spill node */
105 int reload_cost; /**< the cost of a reload node */
106 set *spills; /**< all spill_info_t's, which must be
108 spill_info_t **mem_phis; /**< set of all spilled phis. */
110 unsigned spill_count;
111 unsigned reload_count;
112 unsigned remat_count;
113 unsigned spilled_phi_count;
117 * Compare two spill infos.
119 static int cmp_spillinfo(const void *x, const void *y, size_t size)
121 const spill_info_t *xx = (const spill_info_t*)x;
122 const spill_info_t *yy = (const spill_info_t*)y;
125 return xx->to_spill != yy->to_spill;
129 * Returns spill info for a specific value (the value that is to be spilled)
131 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
133 spill_info_t info, *res;
134 int hash = hash_irn(value);
136 info.to_spill = value;
137 res = set_find(spill_info_t, env->spills, &info, sizeof(info), hash);
140 info.reloaders = NULL;
142 info.spill_costs = -1;
143 info.reload_cls = NULL;
144 info.spilled_phi = false;
145 res = set_insert(spill_info_t, env->spills, &info, sizeof(info), hash);
151 spill_env_t *be_new_spill_env(ir_graph *irg)
153 const arch_env_t *arch_env = be_get_irg_arch_env(irg);
155 spill_env_t *env = XMALLOC(spill_env_t);
156 env->spills = new_set(cmp_spillinfo, 1024);
158 env->arch_env = arch_env;
159 env->mem_phis = NEW_ARR_F(spill_info_t*, 0);
160 env->spill_cost = arch_env->spill_cost;
161 env->reload_cost = arch_env->reload_cost;
162 obstack_init(&env->obst);
164 env->spill_count = 0;
165 env->reload_count = 0;
166 env->remat_count = 0;
167 env->spilled_phi_count = 0;
172 void be_delete_spill_env(spill_env_t *env)
174 del_set(env->spills);
175 DEL_ARR_F(env->mem_phis);
176 obstack_free(&env->obst, NULL);
182 * | _ \| | __ _ ___ ___ | _ \ ___| | ___ __ _ __| |___
183 * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
184 * | __/| | (_| | (_| __/ | _ < __/ | (_) | (_| | (_| \__ \
185 * |_| |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
189 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
191 spill_info_t *spill_info = get_spillinfo(env, to_spill);
196 assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
197 DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
199 /* Just for safety make sure that we do not insert the spill in front of a phi */
200 assert(!is_Phi(sched_next(after)));
202 /* spills that are dominated by others are not needed */
204 s = spill_info->spills;
205 for ( ; s != NULL; s = s->next) {
206 /* no need to add this spill if it is dominated by another */
207 if (value_dominates(s->after, after)) {
208 DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
211 /* remove spills that we dominate */
212 if (value_dominates(after, s->after)) {
213 DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
215 last->next = s->next;
217 spill_info->spills = s->next;
224 spill = OALLOC(&env->obst, spill_t);
225 spill->after = after;
226 spill->next = spill_info->spills;
229 spill_info->spills = spill;
232 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
233 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
239 assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
241 info = get_spillinfo(env, to_spill);
243 if (is_Phi(to_spill)) {
246 /* create spillinfos for the phi arguments */
247 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
248 ir_node *arg = get_irn_n(to_spill, i);
249 get_spillinfo(env, arg);
253 assert(!is_Proj(before) && !be_is_Keep(before));
255 /* put reload into list */
256 rel = OALLOC(&env->obst, reloader_t);
257 rel->next = info->reloaders;
258 rel->reloader = before;
259 rel->rematted_node = NULL;
260 rel->can_spill_after = can_spill_after;
261 rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
263 info->reloaders = rel;
264 assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
265 info->reload_cls = reload_cls;
267 DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
268 to_spill, before, allow_remat ? "" : " not"));
271 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
272 const arch_register_class_t *reload_cls, int allow_remat)
274 be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
278 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
280 ir_node *last = sched_last(block);
282 /* we might have keeps behind the jump... */
283 while (be_is_Keep(last)) {
284 last = sched_prev(last);
285 assert(!sched_is_end(last));
288 assert(is_cfop(last));
290 /* add the reload before the (cond-)jump */
295 * determine final spill position: it should be after all phis, keep nodes
296 * and behind nodes marked as prolog
298 static ir_node *determine_spill_point(ir_node *node)
300 node = skip_Proj(node);
302 ir_node *next = sched_next(node);
303 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
311 * Returns the point at which you can insert a node that should be executed
312 * before block @p block when coming from pred @p pos.
314 static ir_node *get_block_insertion_point(ir_node *block, int pos)
318 /* simply add the reload to the beginning of the block if we only have 1
319 * predecessor. We don't need to check for phis as there can't be any in a
320 * block with only 1 pred. */
321 if (get_Block_n_cfgpreds(block) == 1) {
322 assert(!is_Phi(sched_first(block)));
323 return sched_first(block);
326 /* We have to reload the value in pred-block */
327 predblock = get_Block_cfgpred_block(block, pos);
328 return be_get_end_of_block_insertion_point(predblock);
331 void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
332 const ir_node *block,
333 const arch_register_class_t *reload_cls,
336 ir_node *before = be_get_end_of_block_insertion_point(block);
337 be_add_reload(env, to_spill, before, reload_cls, allow_remat);
340 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
341 int pos, const arch_register_class_t *reload_cls,
344 ir_node *before = get_block_insertion_point(block, pos);
345 be_add_reload(env, to_spill, before, reload_cls, allow_remat);
348 void be_spill_phi(spill_env_t *env, ir_node *node)
354 assert(is_Phi(node));
356 info = get_spillinfo(env, node);
357 info->spilled_phi = true;
358 ARR_APP1(spill_info_t*, env->mem_phis, info);
360 /* create spills for the phi arguments */
361 block = get_nodes_block(node);
362 for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
363 ir_node *arg = get_irn_n(node, i);
366 /* some backends have virtual noreg/unknown nodes that are not scheduled
367 * and simply always available. */
368 if (!sched_is_scheduled(arg)) {
369 ir_node *pred_block = get_Block_cfgpred_block(block, i);
370 insert = be_get_end_of_block_insertion_point(pred_block);
371 insert = sched_prev(insert);
373 insert = determine_spill_point(arg);
376 be_add_spill(env, arg, insert);
382 * / ___|_ __ ___ __ _| |_ ___ / ___| _ __ (_) | |___
383 * | | | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
384 * | |___| | | __/ (_| | || __/ ___) | |_) | | | \__ \
385 * \____|_| \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
389 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
394 * @param senv the spill environment
395 * @param irn the node that should be spilled
396 * @param ctx_irn an user of the spilled node
398 * @return a be_Spill node
400 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
402 ir_node *to_spill = spillinfo->to_spill;
403 const ir_node *insn = skip_Proj_const(to_spill);
406 /* determine_spill_costs must have been run before */
407 assert(spillinfo->spill_costs >= 0);
409 /* some backends have virtual noreg/unknown nodes that are not scheduled
410 * and simply always available. */
411 if (!sched_is_scheduled(insn)) {
412 /* override spillinfos or create a new one */
413 ir_graph *irg = get_irn_irg(to_spill);
414 spillinfo->spills->spill = get_irg_no_mem(irg);
415 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
419 DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
420 spill = spillinfo->spills;
421 for ( ; spill != NULL; spill = spill->next) {
422 ir_node *after = spill->after;
423 after = determine_spill_point(after);
425 spill->spill = arch_env_new_spill(env->arch_env, to_spill, after);
426 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
429 DBG((dbg, LEVEL_1, "\n"));
432 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
435 * If the first usage of a Phi result would be out of memory
436 * there is no sense in allocating a register for it.
437 * Thus we spill it and all its operands to the same spill slot.
438 * Therefore the phi/dataB becomes a phi/Memory
440 * @param senv the spill environment
441 * @param phi the Phi node that should be spilled
442 * @param ctx_irn an user of the spilled node
444 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
446 ir_graph *irg = env->irg;
447 ir_node *phi = spillinfo->to_spill;
448 ir_node *block = get_nodes_block(phi);
456 assert(!get_opt_cse());
457 DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
459 /* build a new PhiM */
460 arity = get_irn_arity(phi);
461 ins = ALLOCAN(ir_node*, arity);
462 unknown = new_r_Unknown(irg, mode_M);
463 for (i = 0; i < arity; ++i) {
467 /* override or replace spills list... */
468 spill = OALLOC(&env->obst, spill_t);
469 spill->after = determine_spill_point(phi);
470 spill->spill = be_new_Phi(block, arity, ins, mode_M, arch_no_register_req);
472 sched_add_after(block, spill->spill);
474 spillinfo->spills = spill;
475 env->spilled_phi_count++;
477 for (i = 0; i < arity; ++i) {
478 ir_node *arg = get_irn_n(phi, i);
479 spill_info_t *arg_info = get_spillinfo(env, arg);
481 determine_spill_costs(env, arg_info);
482 spill_node(env, arg_info);
484 set_irn_n(spill->spill, i, arg_info->spills->spill);
486 DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
493 * @param senv the spill environment
494 * @param to_spill the node that should be spilled
496 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
498 /* node is already spilled */
499 if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
502 if (spillinfo->spilled_phi) {
503 spill_phi(env, spillinfo);
505 spill_irn(env, spillinfo);
512 * | _ \ ___ _ __ ___ __ _| |_ ___ _ __(_) __ _| (_)_______
513 * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_ / _ \
514 * | _ < __/ | | | | | (_| | || __/ | | | (_| | | |/ / __/
515 * |_| \_\___|_| |_| |_|\__,_|\__\___|_| |_|\__,_|_|_/___\___|
520 * Tests whether value @p arg is available before node @p reloader
521 * @returns 1 if value is available, 0 otherwise
523 static int is_value_available(spill_env_t *env, const ir_node *arg,
524 const ir_node *reloader)
526 if (is_Unknown(arg) || is_NoMem(arg))
529 if (be_is_Spill(skip_Proj_const(arg)))
532 if (arg == get_irg_frame(env->irg))
537 if (get_irn_mode(arg) == mode_T)
541 * Ignore registers are always available
543 if (arch_irn_is_ignore(arg))
550 * Check if a node is rematerializable. This tests for the following conditions:
552 * - The node itself is rematerializable
553 * - All arguments of the node are available or also rematerialisable
554 * - The costs for the rematerialisation operation is less or equal a limit
556 * Returns the costs needed for rematerialisation or something
557 * >= REMAT_COST_INFINITE if remat is not possible.
559 static int check_remat_conditions_costs(spill_env_t *env,
560 const ir_node *spilled, const ir_node *reloader, int parentcosts)
565 const ir_node *insn = skip_Proj_const(spilled);
567 assert(!be_is_Spill(insn));
568 if (!arch_irn_is(insn, rematerializable))
569 return REMAT_COST_INFINITE;
571 if (be_is_Reload(insn)) {
574 costs += arch_get_op_estimated_cost(insn);
576 if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
577 return REMAT_COST_INFINITE;
579 /* never rematerialize a node which modifies the flags.
580 * (would be better to test whether the flags are actually live at point
583 if (arch_irn_is(insn, modify_flags)) {
584 return REMAT_COST_INFINITE;
588 for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
589 ir_node *arg = get_irn_n(insn, i);
591 if (is_value_available(env, arg, reloader))
594 /* we have to rematerialize the argument as well */
597 /* we only support rematerializing 1 argument at the moment,
598 * as multiple arguments could increase register pressure */
599 return REMAT_COST_INFINITE;
602 costs += check_remat_conditions_costs(env, arg, reloader,
603 parentcosts + costs);
604 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
605 return REMAT_COST_INFINITE;
612 * Re-materialize a node.
614 * @param env the spill environment
615 * @param spilled the node that was spilled
616 * @param reloader a irn that requires a reload
618 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
625 if (is_Block(reloader)) {
628 bl = get_nodes_block(reloader);
631 ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
632 for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
633 ir_node *arg = get_irn_n(spilled, i);
635 if (is_value_available(env, arg, reloader)) {
638 ins[i] = do_remat(env, arg, reloader);
639 /* don't count the argument rematerialization as an extra remat */
644 /* create a copy of the node */
645 res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
646 get_irn_op(spilled), get_irn_mode(spilled),
647 get_irn_arity(spilled), ins);
648 copy_node_attr(env->irg, spilled, res);
649 arch_env_mark_remat(env->arch_env, res);
651 DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
653 if (! is_Proj(res)) {
654 /* insert in schedule */
656 sched_add_before(reloader, res);
663 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
665 ir_node *block = get_nodes_block(before);
666 double freq = get_block_execfreq(block);
669 return env->spill_cost * freq;
672 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
673 const ir_node *before)
676 /* is the node rematerializable? */
677 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
678 if (costs < (unsigned) env->reload_cost)
682 return env->reload_cost;
685 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
687 ir_node *block = get_nodes_block(before);
688 double freq = get_block_execfreq(block);
691 /* is the node rematerializable? */
692 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
693 if (costs < env->reload_cost)
697 return env->reload_cost * freq;
700 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
701 const ir_node *before)
703 return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
706 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
707 ir_node *block, int pos)
709 ir_node *before = get_block_insertion_point(block, pos);
710 return be_get_reload_costs(env, to_spill, before);
713 ir_node *be_new_spill(ir_node *value, ir_node *after)
715 ir_graph *irg = get_irn_irg(value);
716 ir_node *frame = get_irg_frame(irg);
717 const arch_register_class_t *cls = arch_get_irn_reg_class(value);
718 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
719 ir_node *block = get_block(after);
721 = be_new_Spill(cls, cls_frame, block, frame, value);
723 sched_add_after(after, spill);
727 ir_node *be_new_reload(ir_node *value, ir_node *spill, ir_node *before)
729 ir_graph *irg = get_irn_irg(value);
730 ir_node *frame = get_irg_frame(irg);
731 ir_node *block = get_block(before);
732 const arch_register_class_t *cls = arch_get_irn_reg_class(value);
733 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
734 ir_mode *mode = get_irn_mode(value);
737 assert(be_is_Spill(spill) || is_Phi(spill));
738 assert(get_irn_mode(spill) == mode_M);
740 reload = be_new_Reload(cls, cls_frame, block, frame, spill, mode);
741 sched_add_before(before, reload);
748 * |_ _|_ __ ___ ___ _ __| |_ | _ \ ___| | ___ __ _ __| |___
749 * | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
750 * | || | | \__ \ __/ | | |_ | _ < __/ | (_) | (_| | (_| \__ \
751 * |___|_| |_|___/\___|_| \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
756 * analyzes how to best spill a node and determine costs for that
758 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
760 ir_node *to_spill = spillinfo->to_spill;
761 const ir_node *insn = skip_Proj_const(to_spill);
762 ir_node *spill_block;
764 double spill_execfreq;
766 /* already calculated? */
767 if (spillinfo->spill_costs >= 0)
770 assert(!arch_irn_is(insn, dont_spill));
771 assert(!be_is_Reload(insn));
773 /* some backends have virtual noreg/unknown nodes that are not scheduled
774 * and simply always available.
775 * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
776 * predecessor (of a PhiM) but this test might match other things too...
778 if (!sched_is_scheduled(insn)) {
779 ir_graph *irg = get_irn_irg(to_spill);
780 /* override spillinfos or create a new one */
781 spill_t *spill = OALLOC(&env->obst, spill_t);
784 spill->spill = get_irg_no_mem(irg);
786 spillinfo->spills = spill;
787 spillinfo->spill_costs = 0;
789 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
793 spill_block = get_nodes_block(insn);
794 spill_execfreq = get_block_execfreq(spill_block);
796 if (spillinfo->spilled_phi) {
797 /* TODO calculate correct costs...
798 * (though we can't remat this node anyway so no big problem) */
799 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
803 if (spillinfo->spills != NULL) {
805 double spills_execfreq;
807 /* calculate sum of execution frequencies of individual spills */
809 s = spillinfo->spills;
810 for ( ; s != NULL; s = s->next) {
811 ir_node *spill_block = get_block(s->after);
812 double freq = get_block_execfreq(spill_block);
814 spills_execfreq += freq;
817 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
818 spills_execfreq * env->spill_cost,
819 spill_execfreq * env->spill_cost));
821 /* multi-/latespill is advantageous -> return*/
822 if (spills_execfreq < spill_execfreq) {
823 DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
824 spillinfo->spill_costs = spills_execfreq * env->spill_cost;
829 /* override spillinfos or create a new one */
830 spill = OALLOC(&env->obst, spill_t);
831 spill->after = determine_spill_point(to_spill);
835 spillinfo->spills = spill;
836 spillinfo->spill_costs = spill_execfreq * env->spill_cost;
837 DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
840 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
842 const spill_info_t *si = get_spillinfo(env, irn);
843 ir_node *start_block = get_irg_start_block(get_irn_irg(irn));
844 int n_blocks = get_Block_dom_max_subtree_pre_num(start_block);
845 bitset_t *reloads = bitset_alloca(n_blocks);
852 /* Fill the bitset with the dominance pre-order numbers
853 * of the blocks the reloads are located in. */
854 for (r = si->reloaders; r != NULL; r = r->next) {
855 ir_node *bl = get_nodes_block(r->reloader);
856 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
859 /* Now, cancel out all the blocks that are dominated by each spill.
860 * If the bitset is not empty after that, we have reloads that are
861 * not dominated by any spill. */
862 for (s = si->spills; s != NULL; s = s->next) {
863 ir_node *bl = get_nodes_block(s->after);
864 int start = get_Block_dom_tree_pre_num(bl);
865 int end = get_Block_dom_max_subtree_pre_num(bl);
867 bitset_clear_range(reloads, start, end);
870 if (!bitset_is_empty(reloads))
871 be_add_spill(env, si->to_spill, si->to_spill);
874 void be_insert_spills_reloads(spill_env_t *env)
876 size_t n_mem_phis = ARR_LEN(env->mem_phis);
879 be_timer_push(T_RA_SPILL_APPLY);
881 /* create all phi-ms first, this is needed so, that phis, hanging on
882 spilled phis work correctly */
883 for (i = 0; i < n_mem_phis; ++i) {
884 spill_info_t *info = env->mem_phis[i];
885 spill_node(env, info);
888 /* process each spilled node */
889 foreach_set(env->spills, spill_info_t, si) {
890 ir_node *to_spill = si->to_spill;
891 ir_node **copies = NEW_ARR_F(ir_node*, 0);
892 double all_remat_costs = 0; /** costs when we would remat all nodes */
893 bool force_remat = false;
896 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
898 determine_spill_costs(env, si);
900 /* determine possibility of rematerialisations */
902 /* calculate cost savings for each indivial value when it would
903 be rematted instead of reloaded */
904 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
907 int remat_cost_delta;
909 ir_node *reloader = rld->reloader;
911 if (rld->rematted_node != NULL) {
912 DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
913 rld->rematted_node, reloader));
916 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
917 DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
919 all_remat_costs = REMAT_COST_INFINITE;
923 remat_cost = check_remat_conditions_costs(env, to_spill,
925 if (remat_cost >= REMAT_COST_INFINITE) {
926 DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
928 rld->remat_cost_delta = REMAT_COST_INFINITE;
929 all_remat_costs = REMAT_COST_INFINITE;
933 remat_cost_delta = remat_cost - env->reload_cost;
934 rld->remat_cost_delta = remat_cost_delta;
935 block = is_Block(reloader) ? reloader : get_nodes_block(reloader);
936 freq = get_block_execfreq(block);
937 all_remat_costs += remat_cost_delta * freq;
938 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
939 "%d (rel %f)\n", reloader, remat_cost_delta,
940 remat_cost_delta * freq));
942 if (all_remat_costs < REMAT_COST_INFINITE) {
943 /* we don't need the costs for the spill if we can remat
945 all_remat_costs -= si->spill_costs;
947 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
948 env->spill_cost, si->spill_costs));
951 if (all_remat_costs < 0) {
952 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
958 /* go through all reloads for this spill */
959 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
960 ir_node *copy; /* a reload is a "copy" of the original value */
962 if (rld->rematted_node != NULL) {
963 copy = rld->rematted_node;
964 sched_add_before(rld->reloader, copy);
965 } else if (be_do_remats &&
966 (force_remat || rld->remat_cost_delta < 0)) {
967 copy = do_remat(env, to_spill, rld->reloader);
969 /* make sure we have a spill */
972 /* create a reload, use the first spill for now SSA
973 * reconstruction for memory comes below */
974 assert(si->spills != NULL);
975 copy = arch_env_new_reload(env->arch_env, si->to_spill,
976 si->spills->spill, rld->reloader);
980 DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
981 copy, to_spill, rld->reloader));
982 ARR_APP1(ir_node*, copies, copy);
985 /* if we had any reloads or remats, then we need to reconstruct the
986 * SSA form for the spilled value */
987 if (ARR_LEN(copies) > 0) {
988 be_ssa_construction_env_t senv;
989 /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
991 be_ssa_construction_init(&senv, env->irg);
992 be_ssa_construction_add_copy(&senv, to_spill);
993 be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
994 be_ssa_construction_fix_users(&senv, to_spill);
997 /* no need to enable this as long as we invalidate liveness
998 after this function... */
999 be_ssa_construction_update_liveness_phis(&senv);
1000 be_liveness_update(to_spill);
1001 len = ARR_LEN(copies);
1002 for (i = 0; i < len; ++i) {
1003 be_liveness_update(lv, copies[i]);
1006 be_ssa_construction_destroy(&senv);
1008 /* need to reconstruct SSA form if we had multiple spills */
1009 if (si->spills != NULL && si->spills->next != NULL) {
1011 int spill_count = 0;
1013 be_ssa_construction_env_t senv;
1015 be_ssa_construction_init(&senv, env->irg);
1017 for ( ; spill != NULL; spill = spill->next) {
1018 /* maybe we rematerialized the value and need no spill */
1019 if (spill->spill == NULL)
1021 be_ssa_construction_add_copy(&senv, spill->spill);
1024 if (spill_count > 1) {
1025 /* all reloads are attached to the first spill, fix them now */
1026 be_ssa_construction_fix_users(&senv, si->spills->spill);
1029 be_ssa_construction_destroy(&senv);
1033 si->reloaders = NULL;
1036 stat_ev_dbl("spill_spills", env->spill_count);
1037 stat_ev_dbl("spill_reloads", env->reload_count);
1038 stat_ev_dbl("spill_remats", env->remat_count);
1039 stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1041 /* Matze: In theory be_ssa_construction should take care of the liveness...
1042 * try to disable this again in the future */
1043 be_invalidate_live_sets(env->irg);
1045 be_remove_dead_nodes_from_schedule(env->irg);
1047 be_timer_pop(T_RA_SPILL_APPLY);
1050 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill)
1051 void be_init_spill(void)
1053 FIRM_DBG_REGISTER(dbg, "firm.be.spill");