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 Beladys spillalgorithm version 2.
23 * @author Sebastian Hack, Matthias Braun, Daniel Grund
27 * The main differences to the original Belady are:
28 * - The workset is empty at the start of a block
29 * There is no attempt to fill it with variables which
30 * are not used in the block.
31 * - There is a global pass which tries to use the remaining
32 * capacity of the blocks to let global variables live through
41 #include "irnodeset.h"
43 #include "irprintf_t.h"
49 #include "iredges_t.h"
50 #include "irphase_t.h"
62 #include "bechordal_t.h"
64 #include "beloopana.h"
67 #include "bespillutil.h"
70 #include "lc_opts_enum.h"
79 #define DBG_WORKSET 128
80 #define DBG_GLOBAL 256
82 #define ALREADY_SPILLED_FACTOR 2
85 #define LIVE_END (DEAD-1)
86 #define REMAT_DIST (DEAD-2)
88 static int already_spilled_factor = 2;
89 static int remat_live_range_ext = 1;
90 static int global_pass_enabled = 1;
92 static const lc_opt_table_entry_t options[] = {
93 LC_OPT_ENT_INT ("asf", "already spilled factor", &already_spilled_factor),
94 LC_OPT_ENT_BOOL ("remat", "rematerializable ops get infinite long live ranges", &remat_live_range_ext),
95 LC_OPT_ENT_BOOL ("global", "enable/disable the global pass", &global_pass_enabled),
99 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
102 * An association between a node and a point in time.
104 typedef struct _loc_t {
105 ir_node *irn; /**< A node. */
106 unsigned time; /**< A use time.
107 In the global pass this is used
108 as the version number and not as a time.
109 Only to save space...
113 typedef struct _workset_t {
114 int len; /**< current length */
115 loc_t vals[0]; /**< inlined array of the values/distances in this working set */
118 typedef struct _belady_env_t {
122 const arch_env_t *arch;
123 const arch_register_class_t *cls;
127 ir_node **blocks; /**< Array of all blocks. */
128 int n_blocks; /**< Number of blocks in the graph. */
129 int n_regs; /**< number of regs in this reg-class */
130 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
131 ir_node *instr; /**< current instruction */
132 int instr_nr; /**< current instruction number (relative to block start) */
134 spill_env_t *senv; /**< see bespill.h */
135 bitset_t *spilled; /**< bitset to keep all the irns which have already been spilled. */
136 ir_nodeset_t *extra_spilled; /** All nodes for which a special spill location has been computed. */
140 static int loc_compare(const void *a, const void *b)
144 return (p->time > q->time) - (p->time < q->time);
147 static inline void workset_print(const workset_t *w)
151 for(i = 0; i < w->len; ++i) {
152 ir_fprintf(stderr, "%+F %d\n", w->vals[i].irn, w->vals[i].time);
157 * Alloc a new workset on obstack @p ob with maximum size @p max
159 static inline workset_t *new_workset(belady_env_t *env, struct obstack *ob)
161 return OALLOCFZ(ob, workset_t, vals, env->n_regs);
165 * Alloc a new instance on obstack and make it equal to @param ws
167 static inline workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws)
169 workset_t *res = OALLOCF(ob, workset_t, vals, env->n_regs);
170 memcpy(res, ws, sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]));
175 * Do NOT alloc anything. Make @param tgt equal to @param src.
176 * returns @param tgt for convenience
178 static inline workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src)
180 size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]);
181 memcpy(tgt, src, size);
186 * Overwrites the current content array of @param ws with the
187 * @param count locations given at memory @param locs.
188 * Set the length of @param ws to count.
190 static inline void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
192 workset->len = count;
193 memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
197 * Inserts the value @p val into the workset, iff it is not
198 * already contained. The workset must not be full.
200 static inline void workset_insert(belady_env_t *env, workset_t *ws, ir_node *val)
203 /* check for current regclass */
204 if (!arch_irn_consider_in_reg_alloc(env->cls, val)) {
205 // DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
209 /* check if val is already contained */
210 for(i=0; i<ws->len; ++i)
211 if (ws->vals[i].irn == val)
215 assert(ws->len < env->n_regs && "Workset already full!");
216 ws->vals[ws->len++].irn = val;
220 * Removes all entries from this workset
222 static inline void workset_clear(workset_t *ws)
228 * Removes the value @p val from the workset if present.
230 static inline void workset_remove(workset_t *ws, ir_node *val)
233 for(i=0; i<ws->len; ++i) {
234 if (ws->vals[i].irn == val) {
235 ws->vals[i] = ws->vals[--ws->len];
241 static inline int workset_get_index(const workset_t *ws, const ir_node *val)
244 for(i=0; i<ws->len; ++i) {
245 if (ws->vals[i].irn == val)
253 * Iterates over all values in the working set.
254 * @p ws The workset to iterate
255 * @p v A variable to put the current value in
256 * @p i An integer for internal use
258 #define workset_foreach(ws, v, i) for(i=0; \
259 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
262 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
263 #define workset_get_time(ws, i) (ws)->vals[i].time
264 #define workset_set_length(ws, length) (ws)->len = length
265 #define workset_get_length(ws) ((ws)->len)
266 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
267 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
268 #define workset_contains(ws, n) (workset_get_index(ws, n) >= 0)
270 typedef struct _bring_in_t bring_in_t;
272 typedef struct _block_info_t {
277 workset_t *ws_end; /**< The end set after the local belady pass. */
278 double exec_freq; /**< The execution frequency of this block. */
280 double reload_cost; /**< Cost of a reload in this block. */
281 ir_node *first_non_in; /**< First node in block which is not a phi. */
282 ir_node *last_ins; /**< The instruction before which end of
283 block reloads will be inserted. */
285 int pressure; /**< The amount of registers which remain free
286 in this block. This capacity can be used to let
287 global variables, transported into other blocks,
288 live through this block. */
290 int front_pressure; /**< The pressure right before the first
291 real (non-phi) node. At the beginning
292 of the global pass, this is 0. */
293 struct list_head br_head; /**< List head for all bring_in variables. */
294 int free_at_jump; /**< registers free at jump. */
298 static inline void *new_block_info(belady_env_t *bel, int id)
300 ir_node *bl = bel->blocks[id];
301 block_info_t *res = OALLOCZ(&bel->ob, block_info_t);
302 res->first_non_in = NULL;
303 res->last_ins = NULL;
307 res->exec_freq = get_block_execfreq(bel->ef, bl);
308 res->reload_cost = bel->arch->reload_cost * res->exec_freq;
309 res->free_at_jump = bel->n_regs;
310 INIT_LIST_HEAD(&res->br_head);
311 set_irn_link(bl, res);
315 #define get_block_info(block) ((block_info_t *)get_irn_link(block))
316 #define set_block_info(block, info) set_irn_link(block, info)
318 static inline ir_node *block_info_get_last_ins(block_info_t *bi)
321 bi->last_ins = be_get_end_of_block_insertion_point(bi->bl);
326 typedef struct _next_use_t {
327 unsigned is_first_use : 1; /**< Indicate that this use is the first
328 in the block. Needed to identify
329 transport in values for the global
331 sched_timestep_t step; /**< The time step of the use. */
333 struct _next_use_t *next; /**< The next use int this block
337 static void *next_use_init(ir_phase *phase, const ir_node *irn, void *old)
345 static void build_next_uses(block_info_t *bi)
349 sched_renumber(bi->bl);
351 phase_init(&bi->next_uses, "next uses", bi->bel->irg, PHASE_DEFAULT_GROWTH, next_use_init, NULL);
352 sched_foreach_reverse(bi->bl, irn) {
358 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
359 ir_node *op = get_irn_n(irn, i);
360 next_use_t *curr = phase_get_irn_data(&bi->next_uses, op);
361 next_use_t *use = phase_alloc(&bi->next_uses, sizeof(use[0]));
363 use->is_first_use = 1;
364 use->step = sched_get_time_step(irn);
369 curr->is_first_use = 0;
370 assert(curr->step >= use->step);
373 phase_set_irn_data(&bi->next_uses, op, use);
378 #define get_current_use(bi, irn) phase_get_irn_data(&(bi)->next_uses, (irn))
380 static inline void advance_current_use(block_info_t *bi, const ir_node *irn)
382 next_use_t *use = get_current_use(bi, irn);
385 phase_set_irn_data(&bi->next_uses, irn, use->next);
388 static __attribute__((unused)) int block_freq_gt(const void *a, const void *b)
390 const ir_node * const *p = a;
391 const ir_node * const *q = b;
392 block_info_t *pi = get_block_info(*p);
393 block_info_t *qi = get_block_info(*q);
394 double diff = qi->exec_freq - pi->exec_freq;
395 return (diff > 0) - (diff < 0);
398 static int block_freq_dfs_gt(const void *a, const void *b)
400 const ir_node * const *p = a;
401 const ir_node * const *q = b;
402 block_info_t *pi = get_block_info(*p);
403 block_info_t *qi = get_block_info(*q);
406 if ((pi->exec_freq > 1.0 && qi->exec_freq > 1.0)
407 || (pi->exec_freq <= 1.0 && qi->exec_freq <= 1.0)) {
409 const dfs_t *dfs = pi->bel->dfs;
410 int pp = dfs_get_post_num(dfs, pi->bl);
411 int pq = dfs_get_post_num(dfs, qi->bl);
415 diff = qi->exec_freq - pi->exec_freq;
416 return (diff > 0) - (diff < 0);
421 | __ ) _ __(_)_ __ __ _ |_ _|_ __
422 | _ \| '__| | '_ \ / _` | | || '_ \
423 | |_) | | | | | | | (_| | | || | | |
424 |____/|_| |_|_| |_|\__, | |___|_| |_|
427 Data structures to represent bring in variables.
431 ir_node *irn; /**< The node to bring in. */
432 block_info_t *bi; /**< The block to which bring in should happen. */
433 int pressure_so_far; /**< The maximal pressure till the first use of irn in bl. */
434 ir_node *first_use; /**< The first user of irn in bl. */
435 sched_timestep_t use_step; /**< Schedule step of the first use. */
437 int is_remat : 1; /**< Is rematerializable. */
438 int sect_pressure; /**< Offset to maximum pressure in block. */
439 struct list_head list;
440 struct list_head sect_list;
441 bring_in_t *sect_head;
444 static inline bring_in_t *new_bring_in(block_info_t *bi, ir_node *irn, const next_use_t *use)
446 bring_in_t *br = OALLOC(&bi->bel->ob, bring_in_t);
449 br->first_use = use->irn;
450 br->use_step = use->step;
451 br->is_remat = be_is_rematerializable(bi->bel->senv, irn, use->irn);
452 br->pressure_so_far = bi->pressure;
453 br->sect_pressure = bi->front_pressure;
456 INIT_LIST_HEAD(&br->list);
457 INIT_LIST_HEAD(&br->sect_list);
458 list_add_tail(&br->list, &bi->br_head);
462 static int bring_in_cmp(const void *a, const void *b)
464 const bring_in_t *p = *(const bring_in_t * const *) a;
465 const bring_in_t *q = *(const bring_in_t * const *) b;
468 /* if one of both is a remat node, it will be done after the other. */
469 if (p->is_remat != q->is_remat)
470 return p->is_remat - q->is_remat;
472 /* in the same block, the one further in the front has to be processed first!
473 * Otherwise the front_pressure 'trick' is not exact. */
475 return p->use_step - q->use_step;
477 fp = p->bi->exec_freq;
478 fq = q->bi->exec_freq;
480 /* if both have the same frequency, inspect the frequency of the definition */
482 double fdp = get_block_info(get_nodes_block(p->irn))->exec_freq;
483 double fdq = get_block_info(get_nodes_block(q->irn))->exec_freq;
485 /* if the defs of both have the same freq, we go for reverse dfs post order. */
487 const dfs_t *dfs = p->bi->bel->dfs;
488 int pp = dfs_get_post_num(dfs, p->bi->bl);
489 int pq = dfs_get_post_num(dfs, q->bi->bl);
493 return (fdq > fdp) - (fdq < fdp);
496 return (fq > fp) - (fq < fp);
499 static inline unsigned get_curr_distance(block_info_t *bi, const ir_node *irn, int is_usage)
501 belady_env_t *env = bi->bel;
502 sched_timestep_t curr_step = sched_get_time_step(env->instr);
503 next_use_t *use = get_current_use(bi, irn);
504 int flags = arch_irn_get_flags(irn);
506 assert(!arch_irn_is_ignore(irn));
508 /* We have to keep non-spillable nodes in the working set */
509 if(flags & arch_irn_flags_dont_spill)
512 if (!is_usage && use && use->step == curr_step)
516 unsigned res = use->step - curr_step;
518 assert(use->step >= curr_step);
521 if (remat_live_range_ext && be_is_rematerializable(env->senv, irn, use->irn))
523 else if (bitset_contains_irn(env->spilled, irn))
524 res *= already_spilled_factor;
530 return be_is_live_end(env->lv, bi->bl, irn) ? LIVE_END : DEAD;
533 static inline int is_local_phi(const ir_node *bl, const ir_node *irn)
535 return is_Phi(irn) && get_nodes_block(irn) == bl;
539 * Check, if the value is something that is transported into a block.
540 * That is, the value is defined elsewhere or defined by a Phi in the block.
541 * @param env The belady environment.
542 * @param bl The block in question.
543 * @param irn The node in question.
544 * @return 1, if node is something transported into @p bl, 0 if not.
545 * @note The function will only give correct answers in the case
546 * where @p irn is unused in the block @p bl which is always
547 * the case in our usage scenario.
549 static inline int is_transport_in(const ir_node *bl, const ir_node *irn)
551 return get_nodes_block(irn) != bl || is_Phi(irn);
555 * Performs the actions necessary to grant the request that:
556 * - new_vals can be held in registers
557 * - as few as possible other values are disposed
558 * - the worst values get disposed
560 * @p is_usage indicates that the values in new_vals are used (not defined)
561 * In this case reloads must be performed
563 static void displace(block_info_t *bi, workset_t *new_vals, int is_usage)
565 belady_env_t *env = bi->bel;
566 workset_t *ws = env->ws;
567 ir_node **to_insert = ALLOCAN(ir_node*, env->n_regs);
569 int i, len, max_allowed, demand, iter;
573 1. Identify the number of needed slots and the values to reload
576 workset_foreach(new_vals, val, iter) {
577 /* mark value as used */
579 if (! workset_contains(ws, val)) {
580 DBG((dbg, DBG_DECIDE, "\t\tinsert %+F\n", val));
581 to_insert[demand++] = val;
583 next_use_t *use = get_current_use(bi, val);
586 * if we use a value which is transported in this block, i.e. a
587 * phi defined here or a live in, for the first time, we check
588 * if there is room for that guy to survive from the block's
589 * entrance to here or not.
592 assert(sched_get_time_step(env->instr) == (int) use->step);
593 if (is_transport_in(bi->bl, val) && use->is_first_use) {
594 bring_in_t *bri = new_bring_in(bi, val, use);
595 bri->first_use = env->instr;
597 /* reset the section pressure, since a new section starts. */
598 bi->front_pressure = 0;
600 DBG((dbg, DBG_DECIDE, "\t\tbring in node %+F, pressure %d:\n", val, bi->pressure));
601 DBG((dbg, DBG_DECIDE, "\t\tno reload. must be considered at block start\n"));
605 bitset_add_irn(env->spilled, val);
606 DBG((dbg, DBG_SPILL, "\t\tReload %+F before %+F\n", val, env->instr));
607 be_add_reload(env->senv, val, env->instr, env->cls, 1);
611 assert(is_usage && "Defined value already in workset?!?");
612 DBG((dbg, DBG_DECIDE, "\t\tskip %+F\n", val));
615 DBG((dbg, DBG_DECIDE, "\t\tdemand = %d\n", demand));
618 2. Make room for at least 'demand' slots
620 len = workset_get_length(ws);
621 max_allowed = env->n_regs - demand;
623 /* Only make more free room if we do not have enough */
624 if (len > max_allowed) {
625 DBG((dbg, DBG_DECIDE, "\t\tdisposing %d values\n", len - max_allowed));
627 /* get current next-use distance */
628 for (i = 0; i < ws->len; ++i) {
629 ir_node *val = workset_get_val(ws, i);
630 unsigned dist = get_curr_distance(bi, val, is_usage);
631 workset_set_time(ws, i, dist);
634 /* sort entries by increasing nextuse-distance*/
637 /* kill the last 'demand' entries in the array */
638 workset_set_length(ws, max_allowed);
642 3. Insert the new values into the workset
643 Also, we update the pressure in the block info.
644 That is important for the global pass to decide
645 how many values can live through the block.
647 for (i = 0; i < demand; ++i)
648 workset_insert(env, env->ws, to_insert[i]);
650 /* TODO: simplify this expression? */
651 bi->pressure = MAX(bi->pressure, workset_get_length(env->ws));
652 bi->front_pressure = MAX(bi->front_pressure, workset_get_length(env->ws));
656 * For the given block @p block, decide for each values
657 * whether it is used from a register or is reloaded
660 static void belady(belady_env_t *env, int id)
662 block_info_t *block_info = new_block_info(env, id);
663 const ir_node *block = block_info->bl;
669 DBG((dbg, DBG_WSETS, "Belady on %+F\n", block_info->bl));
670 new_vals = new_workset(env, &env->ob);
671 workset_clear(env->ws);
673 /* build the next use information for this block. */
674 build_next_uses(block_info);
677 block_info->first_non_in = NULL;
679 /* process the block from start to end */
680 sched_foreach(block, irn) {
682 assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
684 /* Projs are handled with the tuple value.
685 * Phis are no real instr (see insert_starters())
686 * instr_nr does not increase */
687 if (is_Proj(irn) || is_Phi(irn))
689 DBG((dbg, DBG_DECIDE, "\t%+F\n", irn));
691 if (!block_info->first_non_in)
692 block_info->first_non_in = irn;
694 /* set instruction in the workset */
697 /* allocate all values _used_ by this instruction */
698 workset_clear(new_vals);
699 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
700 workset_insert(env, new_vals, get_irn_n(irn, i));
702 DBG((dbg, DBG_DECIDE, "\t* uses\n"));
703 displace(block_info, new_vals, 1);
706 * set all used variables to the next use in their next_use_t list
707 * Also, kill all dead variables from the workset. They are only
708 * augmenting the pressure. Note, that a variable is dead
709 * if it has no further use in this block and is *not* live end
711 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
712 ir_node *op = get_irn_n(irn, i);
713 next_use_t *use = get_current_use(block_info, op);
716 if (!use->next && !be_is_live_end(env->lv, block, op))
717 workset_remove(env->ws, op);
719 advance_current_use(block_info, op);
722 /* allocate all values _defined_ by this instruction */
723 workset_clear(new_vals);
724 if (get_irn_mode(irn) == mode_T) { /* special handling for Tuples and Projs */
725 const ir_edge_t *edge;
727 foreach_out_edge(irn, edge) {
728 ir_node *proj = get_edge_src_irn(edge);
729 workset_insert(env, new_vals, proj);
732 workset_insert(env, new_vals, irn);
734 DBG((dbg, DBG_DECIDE, "\t* defs\n"));
735 displace(block_info, new_vals, 0);
737 if (is_op_forking(get_irn_op(env->instr))) {
738 for (i = get_irn_arity(env->instr) - 1; i >= 0; --i) {
739 ir_node *op = get_irn_n(env->instr, i);
740 block_info->free_at_jump -= arch_irn_consider_in_reg_alloc(env->cls, op);
747 phase_free(&block_info->next_uses);
749 /* Remember end-workset for this block */
750 block_info->ws_end = workset_clone(env, &env->ob, env->ws);
751 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
752 workset_foreach(block_info->ws_end, irn, iter)
753 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(block_info->ws_end, iter)));
754 DBG((dbg, DBG_WSETS, "Max pressure in block: %d\n", block_info->pressure));
756 /* now, initialize the front pressure to 0. */
757 block_info->front_pressure = 0;
762 |_ _| |__ ___ __ _| | ___ | |__ __ _| | | _ \ __ _ _ __| |_
763 | | | '_ \ / _ \ / _` | |/ _ \| '_ \ / _` | | | |_) / _` | '__| __|
764 | | | | | | __/ | (_| | | (_) | |_) | (_| | | | __/ (_| | | | |_
765 |_| |_| |_|\___| \__, |_|\___/|_.__/ \__,_|_| |_| \__,_|_| \__|
770 #define workset_set_version(ws, i, t) ((ws)->vals[(i)].time = (t))
771 #define workset_get_version(ws, i) ((ws)->vals[(i)].time)
773 #define ver_oldest (0)
774 #define ver_youngest ((unsigned) -1)
775 #define ver_make_newer(v) ((v) + 1)
776 #define ver_is_older(v, w) ((v) < (w))
777 #define ver_is_younger(v, w) ((v) > (w))
785 typedef struct _block_state_t {
786 struct _block_state_t *next;
787 struct _block_state_t *next_intern;
790 workset_t *end_state;
793 typedef struct _irn_action_t {
794 struct _irn_action_t *next;
800 typedef struct _global_end_state_t {
808 unsigned *bs_tops_vers;
809 block_state_t **bs_tops;
810 block_state_t *bs_top;
811 irn_action_t *ia_top;
812 } global_end_state_t;
816 block_state_t *bs_top;
817 irn_action_t *ia_top;
820 static inline block_state_t *get_block_state(global_end_state_t *ges, const block_info_t *bi)
823 assert(!ver_is_younger(ges->bs_tops_vers[id], ges->version));
824 return ver_is_older(ges->bs_tops_vers[id], ges->version) ? NULL : ges->bs_tops[bi->id];
827 static inline const workset_t *get_end_state(global_end_state_t *ges, block_info_t *bi)
829 block_state_t *bs = get_block_state(ges, bi);
830 return bs ? bs->end_state : bi->ws_end;
833 static block_state_t *new_block_state(global_end_state_t *ges, block_info_t *bi)
835 block_state_t *bs = get_block_state(ges, bi);
836 block_state_t *nw = OALLOC(&ges->obst, block_state_t);
838 nw->next_intern = bs;
839 nw->next = ges->bs_top;
843 nw->pressure = bs->pressure;
844 nw->end_state = workset_clone(ges->env, &ges->obst, bs->end_state);
847 nw->pressure = bi->pressure;
848 nw->end_state = workset_clone(ges->env, &ges->obst, bi->ws_end);
852 ges->bs_tops[bi->id] = nw;
853 ges->bs_tops_vers[bi->id] = ges->version;
857 static irn_action_t *new_irn_action(global_end_state_t *ges, ir_node *irn, const ir_node *bl)
859 irn_action_t *ia = OALLOC(&ges->obst, irn_action_t);
863 ia->act = irn_act_none;
864 ia->next = ges->ia_top;
869 static inline rollback_info_t trans_begin(global_end_state_t *ges)
872 rb.obst_level = obstack_base(&ges->obst);
873 rb.bs_top = ges->bs_top;
874 rb.ia_top = ges->ia_top;
878 static inline void trans_rollback(global_end_state_t *ges, rollback_info_t *rb)
882 /* unwind all the stacks indiced with the block number */
883 for (bs = ges->bs_top; bs != rb->bs_top; bs = bs->next) {
884 unsigned id = bs->bi->id;
885 ges->bs_tops[id] = bs->next_intern;
888 ges->ia_top = rb->ia_top;
889 ges->bs_top = rb->bs_top;
890 obstack_free(&ges->obst, rb->obst_level);
894 static double can_bring_in(global_end_state_t *ges, ir_node *bl, ir_node *irn, double limit, int level);
896 static double can_make_available_at_end(global_end_state_t *ges, ir_node *bl, ir_node *irn, double limit, int level)
898 block_info_t *bi = get_block_info(bl);
899 const workset_t *end = get_end_state(ges, bi);
903 DBG((dbg, DBG_GLOBAL, "\t%2Dcan make avail %+F at end of %+F\n", level, irn, bl));
906 * to make the value available at end,
907 * we have several cases here.
909 * - we already visited that block.
910 * - If the value is in the final end set, return 0.
911 * somebody else already allocated it there.
912 * - If not and the final end set is already full,
913 * we cannot make the value available at the end
914 * of this block. return INFINITY.
915 * - Else (value not in final end set and there is room):
916 * 1) The value is in a register at the end of the local Belady pass.
917 * Allocate a slot in the final end set and return 0.
918 * 2) The value is not in the Belady end set:
919 * If the block's capacity is < k then check what it costs
920 * to transport the value from upper blocks to this block.
921 * Compare that against the reload cost in this block. If
922 * cheaper, do the other thing. If not, reload it here.
925 /* if the end set contains it already, it is in a reg and it costs nothing
926 * to load it to one. */
927 index = workset_get_index(end, irn);
929 unsigned ver = workset_get_version(end, index);
930 DBG((dbg, DBG_GLOBAL, "\t%2Dnode is in the end set and is %s fixed\n",
931 level, ver_is_older(ver, ges->version) ? "already" : "not yet"));
934 * if the version is older, the value is already fixed
935 * and cannot be removed from the end set.
937 * If not, we will create a new block state for that block since
938 * we modify it by giving the end state a new version.
940 if (ver_is_younger(ver, ges->version)) {
941 block_state_t *bs = new_block_state(ges, bi);
942 workset_set_version(bs->end_state, index, ges->version);
950 * Now we have two options:
951 * 1) Reload the value at the end of the block.
952 * Therefore, perhaps, we have to erase another one from the workset.
953 * This may only be done if it has not been fixed.
954 * Since fixed means that a previous pass has decided that that value
955 * *has* to stay in the end set.
956 * 2) we can try, if the capacity of the block allows it, to let
957 * the value live through the block and make it available at
960 * First, we test the local (reload in this block) alternative
961 * and compare against the other alternative.
962 * Of course, we chose the cheaper one.
966 int n_regs = bi->free_at_jump;
967 int len = workset_get_length(end);
974 * look if there is room in the end array
975 * for the variable. Note that this does not
976 * mean that the var can live through the block.
977 * There is just room at the *end*
980 DBG((dbg, DBG_GLOBAL, "\t%2Dthe end set has %d free slots\n", level, n_regs - len));
983 for (i = 0; i < len; ++i) {
984 unsigned ver = workset_get_version(end, i);
985 if (ver_is_younger(ver, ges->version))
990 DBG((dbg, DBG_GLOBAL, "\t%2D%+F (slot %d) can be erased from the end set\n",
991 level, end->vals[i].irn, i));
997 * finally there is some room. we can at least reload the value.
998 * but we will try to let or live through anyhow.
1001 irn_action_t *vs = new_irn_action(ges, irn, bi->bl);
1002 block_state_t *bs = new_block_state(ges, bi);
1003 workset_t *end = bs->end_state;
1004 ir_node *ins_before = block_info_get_last_ins(bi);
1005 double reload_here = be_get_reload_costs(bi->bel->senv, irn, ins_before);
1006 int pressure_ok = bs->pressure < n_regs;
1008 if (reload_here < bi->reload_cost)
1012 * No matter what we do, the value will be in the end set
1013 * if the block from now on (of course only regarding the
1014 * current state). Enter it and set the new length
1017 end->vals[slot].irn = irn;
1018 workset_set_version(end, slot, ges->version);
1019 workset_set_length(end, MAX(workset_get_length(end), slot + 1));
1021 vs->act = irn_act_reload;
1024 DBG((dbg, DBG_GLOBAL, "\t%2Dthere is a free slot. capacity=%d, reload here=%f, pressure %s\n",
1025 level, n_regs - bs->pressure, reload_here, pressure_ok ? "ok" : "insufficient"));
1028 /* look if we can bring the value in. */
1029 if (pressure_ok && reload_here > 0.0) {
1030 rollback_info_t rb = trans_begin(ges);
1031 double new_limit = MIN(reload_here, limit);
1033 vs->act = irn_act_live_through;
1035 res = can_bring_in(ges, bl, irn, new_limit, level + 1);
1038 * if bring in is too expensive re-adjust the pressure
1039 * and roll back the state
1041 if (res >= reload_here) {
1043 vs->act = irn_act_reload;
1044 trans_rollback(ges, &rb);
1050 DBG((dbg, DBG_GLOBAL, "\t%2D%s\n", level,
1051 vs->act == irn_act_reload ? "reloading" : "bringing in"));
1056 DBG((dbg, DBG_GLOBAL, "\t%2D-> %f\n", level, res));
1060 static double can_bring_in(global_end_state_t *ges, ir_node *bl, ir_node *irn, double limit, int level)
1062 belady_env_t *env = ges->env;
1063 double glob_costs = HUGE_VAL;
1065 DBG((dbg, DBG_GLOBAL, "\t%2Dcan bring in (max %f) for %+F at block %+F\n", level, limit, irn, bl));
1067 if (is_transport_in(bl, irn)) {
1068 int i, n = get_irn_arity(bl);
1069 rollback_info_t rb = trans_begin(ges);
1072 for (i = 0; i < n; ++i) {
1073 ir_node *pr = get_Block_cfgpred_block(bl, i);
1074 ir_node *op = is_local_phi(bl, irn) ? get_irn_n(irn, i) : irn;
1078 * there might by Unknowns as operands of Phis in that case
1079 * we set the costs to zero, since they won't get spilled.
1081 if (arch_irn_consider_in_reg_alloc(env->cls, op))
1082 c = can_make_available_at_end(ges, pr, op, limit - glob_costs, level + 1);
1088 if (glob_costs >= limit) {
1089 glob_costs = HUGE_VAL;
1090 trans_rollback(ges, &rb);
1097 DBG((dbg, DBG_GLOBAL, "\t%2D-> %f\n", level, glob_costs));
1101 static void materialize_and_commit_end_state(global_end_state_t *ges)
1103 belady_env_t *env = ges->env;
1107 DBG((dbg, DBG_GLOBAL, "\tmaterializing\n"));
1110 * Perform all the variable actions.
1112 for (ia = ges->ia_top; ia != NULL; ia = ia->next) {
1114 case irn_act_live_through:
1116 block_info_t *bi = get_block_info(ia->bl);
1119 if (is_local_phi(ia->bl, ia->irn)) {
1120 bitset_add_irn(ges->succ_phis, ia->irn);
1121 DBG((dbg, DBG_GLOBAL, "\t\tlive through phi kept alive: %+F\n", ia->irn));
1124 list_for_each_entry_reverse(bring_in_t, iter, &bi->br_head, list)
1125 ++iter->sect_pressure;
1126 ++bi->front_pressure;
1129 case irn_act_reload:
1130 be_add_reload_at_end(env->senv, ia->irn, ia->bl, env->cls, 1);
1131 DBG((dbg, DBG_GLOBAL, "\t\tadding reload of %+F at end of %+F\n", ia->irn, ia->bl));
1134 DBG((dbg, DBG_GLOBAL, "\t\t%+F is in the end set of %+F\n", ia->irn, ia->bl));
1139 * Commit the block end states
1141 for (bs = ges->bs_top; bs != NULL; bs = bs->next) {
1142 block_info_t *bi = bs->bi;
1144 if (!bitset_is_set(ges->committed, bi->id)) {
1145 DBG((dbg, DBG_GLOBAL, "\t\tcommiting workset of %+F with version %x\n", bi->bl, ges->version));
1146 // bes->bs->end_state->vals[idx].version = ges->version;
1147 workset_copy(env, bi->ws_end, bs->end_state);
1148 DBG((dbg, DBG_GLOBAL, "\t\told pressure: %d, new pressure: %d, end length: %d\n",
1149 bi->pressure, bs->pressure, workset_get_length(bs->end_state)));
1150 bi->pressure = bs->pressure;
1151 bitset_set(ges->committed, bi->id);
1155 /* clear the committed bitset. the next call is expecting it. */
1156 bitset_clear_all(ges->committed);
1159 static ir_node *better_spilled_here(const bring_in_t *br)
1161 const block_info_t *bi = br->bi;
1162 double spill_ef = get_block_info(get_nodes_block(br->irn))->exec_freq;
1165 * If the bring in node is a phi in the bring in block,
1166 * we look at all definitions and sum up their execution frequencies,
1167 * since spills will be placed there.
1168 * (except for the case where an operand is also a phi which is spilled :-( )
1169 * If that cost is higher than spilling the phi in that block, we opt for
1170 * bringing the phi into the block and spill it there.
1172 if (is_local_phi(bi->bl, br->irn)) {
1173 ir_node *bl = bi->bl;
1177 for (i = get_Block_n_cfgpreds(bl) - 1; i >= 0; --i)
1178 spill_ef += get_block_info(get_Block_cfgpred_block(bl, i))->exec_freq;
1181 return bi->exec_freq < spill_ef ? sched_prev(bi->first_non_in) : NULL;
1184 static int get_max_pressure_so_far(const block_info_t *bi, const bring_in_t *br)
1186 const struct list_head *l;
1189 assert(br->bi == bi);
1190 for (l = &br->list; l != &bi->br_head; l = l->prev) {
1191 br = list_entry(l, bring_in_t, list);
1192 res = MAX(res, br->sect_pressure);
1195 /* finally consider the front pressure distance and add the reference line (the global block pressure) */
1196 return MAX(res, bi->front_pressure);
1199 #define block_last_bring_in(bi) list_entry((bi)->br_head.prev, bring_in_t, list)
1202 static int get_block_max_pressure(const block_info_t *bi)
1204 int max = get_max_pressure_so_far(bi, block_last_bring_in(bi));
1205 return MAX(bi->pressure, max);
1210 * Try to bring a variable into a block.
1211 * @param ges The state of all end sets.
1212 * @param block The block.
1213 * @param irn The variable.
1215 static void optimize_variable(global_end_state_t *ges, bring_in_t *br)
1217 block_info_t *bi = br->bi;
1218 ir_node *irn = br->irn;
1219 ir_node *bl = bi->bl;
1220 belady_env_t *env = ges->env;
1221 void *reset_level = obstack_base(&ges->obst);
1222 int k = env->n_regs;
1223 int pressure_upto_use = get_max_pressure_so_far(bi, br);
1224 int front_pressure = bi->front_pressure;
1225 ir_node *better_spill_loc = NULL;
1227 assert(front_pressure <= k);
1228 assert(pressure_upto_use <= k);
1230 DBG((dbg, DBG_GLOBAL, "fixing %+F at %+F (%f), front pr: %d, pr to use: %d, first use: %x\n",
1231 irn, bl, bi->exec_freq, front_pressure, pressure_upto_use, br->first_use));
1233 // assert(!is_local_phi(bl, irn) || !bitset_contains_irn(ges->succ_phis, irn));
1236 * if we cannot bring the value to the use, let's see if it would be worthwhile
1237 * to bring the value to the beginning of the block to have a better spill
1240 * better _spilled_here will return a node where the value can be spilled after
1241 * or NULL if this block does not provide a better spill location.
1244 if (pressure_upto_use >= k && front_pressure < k && !bitset_contains_irn(env->spilled, irn))
1245 better_spill_loc = better_spilled_here(br);
1249 * If either we can bring the value to the use or we should try
1250 * to bring it here to do the spill here, let's try to bring it in.
1252 if (better_spill_loc || pressure_upto_use < k) {
1254 double bring_in_costs, local_costs;
1255 rollback_info_t trans;
1258 /* process all variables which shall be in a reg at
1259 * the beginning of the block in the order of the next use. */
1260 local_costs = be_get_reload_costs(env->senv, irn, br->first_use);
1262 /* reset the lists */
1266 /* if the variable will live into this block, we must adapt the pressure.
1267 * The new pressure is the MAX of:
1268 * 1) the total block pressure
1269 * 2) the pressure so far + the front pressure increase + 1
1271 * If the second is larger than the first,
1272 * we have to increment the total block pressure and hence
1273 * save the old pressure to restore it in case of failing to
1274 * bring the variable into the block in a register.
1276 trans = trans_begin(ges);
1277 bs = new_block_state(ges, bi);
1278 pressure_inc = MAX(bs->pressure, better_spill_loc ? front_pressure : pressure_upto_use + 1);
1279 bs->pressure = pressure_inc;
1282 assert(bi->pressure <= k);
1283 DBG((dbg, DBG_GLOBAL, "\ttrans in var %+F, version %x\n", irn, ges->version));
1284 bring_in_costs = can_bring_in(ges, bl, irn, local_costs, 1);
1285 DBG((dbg, DBG_GLOBAL, "\tbring in: %f, local: %f\n", bring_in_costs, local_costs));
1288 * Following cases can now occur:
1289 * 1) There is room and costs ok
1290 * 2) Cannot bring to use but can spill at begin and costs are ok
1291 * 3) neither of both worked.
1293 * following actions can be taken:
1295 * b) mark phi as succeeded if node was phi
1296 * c) insert reload at use location
1297 * d) give a spill location hint
1299 * this is the case/action matrix
1307 /* the costs were acceptable... */
1308 if (bring_in_costs < local_costs) {
1313 * case 1 and first part of case 2:
1314 * commit all the changes done. this manifests the bring-in action.
1315 * if the transport-in was a phi (that is actually used in block)
1316 * mark it in the succ_phis set to *not* phi spill it.
1318 materialize_and_commit_end_state(ges);
1319 if (is_local_phi(bl, irn))
1320 bitset_add_irn(ges->succ_phis, irn);
1322 DBG((dbg, DBG_GLOBAL, "\t-> bring it in.", pressure_inc));
1324 /* second half of case 2 */
1325 if (pressure_upto_use >= k) {
1326 DBG((dbg, DBG_GLOBAL, "\t-> use blocked. local reload: %+F, try spill at: %+F\n",
1327 br->first_use, better_spill_loc));
1328 be_add_reload(env->senv, irn, br->first_use, env->cls, 1);
1329 be_add_spill(env->senv, irn, better_spill_loc);
1330 ir_nodeset_insert(env->extra_spilled, irn);
1334 * go from the last bring in use to the first and add all the variables
1335 * which additionally live through the block to their pressure.
1336 * at the point were the actually treated use is, we have to increase
1337 * the pressure by one more as the brought in value starts to count.
1338 * Finally, adjust the front pressure as well.
1341 list_for_each_entry_reverse(bring_in_t, iter, &bi->br_head, list) {
1343 pressure_inc += pressure_upto_use < k;
1344 iter->sect_pressure += pressure_inc;
1345 check = MAX(check, iter->sect_pressure);
1346 DBG((dbg, DBG_GLOBAL, "\tinc section pressure of %+F by %d to %d\n", iter->first_use, pressure_inc, iter->sect_pressure));
1348 bi->front_pressure += pressure_inc;
1349 assert(MAX(check, bi->front_pressure) <= bi->pressure);
1350 DBG((dbg, DBG_GLOBAL, "\t-> result: p: %d, fp: %d\n", bi->pressure, bi->front_pressure));
1353 /* case 3: nothing worked. insert normal reload and rollback. */
1355 DBG((dbg, DBG_GLOBAL, "\t-> bring in was too expensive. local reload: %+F\n", br->first_use));
1356 be_add_reload(env->senv, irn, br->first_use, env->cls, 1);
1357 bitset_add_irn(env->spilled, irn);
1358 trans_rollback(ges, &trans);
1362 /* there was no opportunity for optimization at all. reload and be sad ... */
1364 DBG((dbg, DBG_GLOBAL, "\t-> can\'t do anything but reload before %+F\n", br->first_use));
1365 be_add_reload(env->senv, irn, br->first_use, env->cls, 1);
1366 bitset_add_irn(env->spilled, irn);
1369 DBG((dbg, DBG_GLOBAL, "\n"));
1371 /* reset the obstack and create a new version. */
1372 obstack_free(&ges->obst, reset_level);
1373 ges->version = ver_make_newer(ges->version);
1376 static bring_in_t **determine_global_order(belady_env_t *env)
1382 for (i = env->n_blocks - 1; i >= 0; --i) {
1383 block_info_t *bi = get_block_info(env->blocks[i]);
1384 list_for_each_entry(bring_in_t, elm, &bi->br_head, list) {
1385 obstack_ptr_grow(&env->ob, elm);
1390 obstack_ptr_grow(&env->ob, NULL);
1391 res = obstack_finish(&env->ob);
1392 qsort(res, n, sizeof(res[0]), bring_in_cmp);
1399 static void global_assign(belady_env_t *env)
1401 ir_nodeset_iterator_t iter;
1402 global_end_state_t ges;
1408 * sort the blocks according to execution frequency.
1409 * That's not necessary for belady() but for the global pass later on.
1411 qsort(env->blocks, env->n_blocks, sizeof(env->blocks[0]), block_freq_dfs_gt);
1413 memset(&ges, 0, sizeof(ges));
1414 obstack_init(&ges.obst);
1416 ges.version = ver_make_newer(ver_oldest);
1417 ges.succ_phis = bitset_irg_obstack_alloc(&ges.obst, env->irg);
1418 ges.committed = bitset_obstack_alloc(&ges.obst, env->n_blocks);
1419 ges.bs_tops = OALLOCN(&ges.obst, block_state_t*, env->n_blocks);
1420 ges.bs_tops_vers = OALLOCN(&ges.obst, unsigned, env->n_blocks);
1422 /* invalidate all state stack pointer versions */
1423 for (i = 0; i < env->n_blocks; ++i) {
1424 block_info_t *bi = get_block_info(env->blocks[i]);
1425 ges.bs_tops_vers[i] = ver_oldest;
1427 /* Set all block end sets entries to the youngest version */
1428 for (j = workset_get_length(bi->ws_end) - 1; j >= 0; --j)
1429 workset_set_version(bi->ws_end, j, ver_youngest);
1432 /* determine order and optimize them */
1433 for (br = determine_global_order(env); *br; ++br)
1434 optimize_variable(&ges, *br);
1437 * Now we spill phis which cannot be kept since they were replaced
1438 * by reloads at the block entrances.
1440 for (i = 0; i < env->n_blocks; ++i) {
1441 ir_node *bl = env->blocks[i];
1444 sched_foreach(bl, irn) {
1448 if (arch_irn_consider_in_reg_alloc(env->cls, irn)
1449 && !bitset_contains_irn(ges.succ_phis, irn))
1450 be_spill_phi(env->senv, irn);
1454 /* check dominance for specially spilled nodes. */
1455 foreach_ir_nodeset (env->extra_spilled, irn, iter)
1456 make_spill_locations_dominate_irn(env->senv, irn);
1459 static void collect_blocks(ir_node *bl, void *data)
1461 belady_env_t *env = data;
1463 obstack_ptr_grow(&env->ob, bl);
1467 * Do spilling for a register class on a graph using the belady heuristic.
1468 * In the transformed graph, the register pressure never exceeds the number
1469 * of available registers.
1471 * @param birg The backend graph
1472 * @param cls The register class to spill
1474 void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls)
1476 ir_graph *irg = be_get_birg_irg(birg);
1480 /* some special classes contain only ignore regs, nothing to do then */
1481 n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1485 be_clear_links(irg);
1487 /* init belady env */
1488 obstack_init(&env.ob);
1490 env.arch = be_get_birg_arch_env(birg);
1492 env.lv = be_get_birg_liveness(birg);
1493 env.dfs = env.lv->dfs;
1494 env.n_regs = n_regs;
1495 env.ws = new_workset(&env, &env.ob);
1496 env.senv = be_new_spill_env(birg);
1497 env.ef = be_get_birg_exec_freq(birg);
1498 env.spilled = bitset_irg_obstack_alloc(&env.ob, irg);
1499 env.extra_spilled = ir_nodeset_new(64);
1502 irg_block_walk_graph(irg, NULL, collect_blocks, &env);
1503 obstack_ptr_grow(&env.ob, NULL);
1504 env.blocks = obstack_finish(&env.ob);
1506 /* renumbering in the blocks gives nicer debug output as number are smaller. */
1507 #ifdef DEBUG_libfirm
1508 for (i = 0; i < env.n_blocks; ++i)
1509 sched_renumber(env.blocks[i]);
1512 /* Fix high register pressure in blocks with belady algorithm */
1513 for (i = 0; i < env.n_blocks; ++i)
1516 global_assign(&env);
1518 /* check dominance for specially spilled nodes. */
1520 ir_nodeset_iterator_t iter;
1523 foreach_ir_nodeset (env.extra_spilled, irn, iter)
1524 make_spill_locations_dominate_irn(env.senv, irn);
1527 /* Insert spill/reload nodes into the graph and fix usages */
1528 be_insert_spills_reloads(env.senv);
1531 be_delete_spill_env(env.senv);
1532 ir_nodeset_del(env.extra_spilled);
1534 obstack_free(&env.ob, NULL);
1537 void be_init_spillbelady2(void)
1539 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1540 lc_opt_entry_t *spill_grp = lc_opt_get_grp(be_grp, "spill");
1541 lc_opt_entry_t *bel2_grp = lc_opt_get_grp(spill_grp, "belady2");
1543 static be_spiller_t belady_spiller = {
1547 lc_opt_add_table(bel2_grp, options);
1548 be_register_spiller("belady2", &belady_spiller);
1549 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady2");
1552 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady2);