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.
23 * @author Daniel Grund, Matthias Braun
32 #include "irprintf_t.h"
38 #include "iredges_t.h"
41 #include "irnodeset.h"
48 #include "besched_t.h"
52 #include "bechordal_t.h"
53 #include "bespilloptions.h"
54 #include "beloopana.h"
66 #define DBG_WORKSET 128
67 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
69 /* factor to weight the different costs of reloading/rematerializing a node
70 (see bespill.h be_get_reload_costs_no_weight) */
71 #define RELOAD_COST_FACTOR 10
74 value_not_reloaded, /* the value has not been reloaded */
75 value_partially_reloaded, /* the value has been reloaded on some paths */
76 value_reloaded /* the value has been reloaded on all paths */
80 * An association between a node and a point in time.
82 typedef struct loc_t {
84 unsigned time; /**< A use time (see beuses.h). */
85 reloaded_state_t reloaded; /**< the value is a reloaded value */
88 typedef struct _workset_t {
89 int len; /**< current length */
90 loc_t vals[0]; /**< inlined array of the values/distances in this working set */
93 static struct obstack obst;
94 static const arch_env_t *arch_env;
95 static const arch_register_class_t *cls;
96 static const be_lv_t *lv;
97 static be_loopana_t *loop_ana;
99 static workset_t *ws; /**< the main workset used while
100 processing a block. */
101 static be_uses_t *uses; /**< env for the next-use magic */
102 static ir_node *instr; /**< current instruction */
103 static unsigned instr_nr; /**< current instruction number
104 (relative to block start) */
105 static ir_nodeset_t used;
106 static spill_env_t *senv; /**< see bespill.h */
107 static pdeq *worklist;
109 static int loc_compare(const void *a, const void *b)
113 return p->time - q->time;
116 void workset_print(const workset_t *w)
120 for(i = 0; i < w->len; ++i) {
121 ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time);
126 * Alloc a new workset on obstack @p ob with maximum size @p max
128 static workset_t *new_workset(void)
131 size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
133 res = obstack_alloc(&obst, size);
134 memset(res, 0, size);
139 * Alloc a new instance on obstack and make it equal to @param workset
141 static workset_t *workset_clone(workset_t *workset)
144 size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]);
145 res = obstack_alloc(&obst, size);
146 memcpy(res, workset, size);
151 * Copy workset @param src to @param tgt
153 static void workset_copy(workset_t *dest, const workset_t *src)
155 size_t size = sizeof(*src) + n_regs * sizeof(src->vals[0]);
156 memcpy(dest, src, size);
160 * Overwrites the current content array of @param ws with the
161 * @param count locations given at memory @param locs.
162 * Set the length of @param ws to count.
164 static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs)
166 workset->len = count;
167 memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
171 * Inserts the value @p val into the workset, iff it is not
172 * already contained. The workset must not be full.
174 static void workset_insert(workset_t *workset, ir_node *val, int reloaded)
178 /* check for current regclass */
179 assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val));
181 /* check if val is already contained */
182 for (i = 0; i < workset->len; ++i) {
183 loc = &workset->vals[i];
184 if (loc->node == val) {
186 loc->reloaded = reloaded;
193 assert(workset->len < n_regs && "Workset already full!");
194 loc = &workset->vals[workset->len];
196 loc->reloaded = reloaded;
197 loc->time = 6666; /* undefined yet */
202 * Removes all entries from this workset
204 static void workset_clear(workset_t *workset)
210 * Removes the value @p val from the workset if present.
212 static INLINE void workset_remove(workset_t *workset, ir_node *val)
215 for(i = 0; i < workset->len; ++i) {
216 if (workset->vals[i].node == val) {
217 workset->vals[i] = workset->vals[--workset->len];
223 static INLINE int workset_contains(const workset_t *ws, const ir_node *val)
227 for(i=0; i<ws->len; ++i) {
228 if (ws->vals[i].node == val)
236 * Iterates over all values in the working set.
237 * @p ws The workset to iterate
238 * @p v A variable to put the current value in
239 * @p i An integer for internal use
241 #define workset_foreach(ws, v, i) for(i=0; \
242 v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; \
245 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
246 #define workset_get_time(ws, i) (ws)->vals[i].time
247 #define workset_set_length(ws, length) (ws)->len = length
248 #define workset_get_length(ws) ((ws)->len)
249 #define workset_get_val(ws, i) ((ws)->vals[i].node)
250 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
252 typedef struct _block_info_t
254 workset_t *start_workset;
255 workset_t *end_workset;
259 static void *new_block_info(void)
261 block_info_t *res = obstack_alloc(&obst, sizeof(res[0]));
262 memset(res, 0, sizeof(res[0]));
267 #define get_block_info(block) ((block_info_t *)get_irn_link(block))
268 #define set_block_info(block, info) set_irn_link(block, info)
271 * @return The distance to the next use or 0 if irn has dont_spill flag set
273 static INLINE unsigned get_distance(ir_node *from, unsigned from_step,
274 const ir_node *def, int skip_from_uses)
277 int flags = arch_irn_get_flags(arch_env, def);
281 assert(! (flags & arch_irn_flags_ignore));
283 use = be_get_next_use(uses, from, from_step, def, skip_from_uses);
284 if(USES_IS_INFINITE(use.time))
285 return USES_INFINITY;
287 /* We have to keep nonspillable nodes in the workingset */
288 if(flags & arch_irn_flags_dont_spill)
291 costs = be_get_reload_costs_no_weight(senv, def, use.before);
292 assert(costs * RELOAD_COST_FACTOR < 1000);
293 time = use.time + 1000 - (costs * RELOAD_COST_FACTOR);
299 * Performs the actions necessary to grant the request that:
300 * - new_vals can be held in registers
301 * - as few as possible other values are disposed
302 * - the worst values get disposed
304 * @p is_usage indicates that the values in new_vals are used (not defined)
305 * In this case reloads must be performed
307 static void displace(workset_t *new_vals, int is_usage)
309 ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
317 /* 1. Identify the number of needed slots and the values to reload */
319 workset_foreach(new_vals, val, iter) {
320 /* mark value as used */
322 ir_nodeset_insert(&used, val);
324 if (! workset_contains(ws, val)) {
325 DB((dbg, DBG_DECIDE, " insert %+F\n", val));
327 DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
328 be_add_reload(senv, val, instr, cls, 1);
331 DB((dbg, DBG_DECIDE, " %+F already in workset\n", val));
333 /* remove the value from the current workset so it is not accidently
335 workset_remove(ws, val);
337 to_insert[demand++] = val;
340 /* 2. Make room for at least 'demand' slots */
341 len = workset_get_length(ws);
342 spills_needed = len + demand - n_regs;
343 assert(spills_needed <= len);
345 /* Only make more free room if we do not have enough */
346 if (spills_needed > 0) {
347 ir_node *curr_bb = get_nodes_block(instr);
348 workset_t *ws_start = get_block_info(curr_bb)->start_workset;
350 DB((dbg, DBG_DECIDE, " disposing %d values\n", spills_needed));
352 /* calculate current next-use distance for live values */
353 for (i = 0; i < len; ++i) {
354 ir_node *val = workset_get_val(ws, i);
355 unsigned dist = get_distance(instr, instr_nr, val, !is_usage);
356 workset_set_time(ws, i, dist);
359 /* sort entries by increasing nextuse-distance*/
362 /* Logic for not needed live-ins: If a value is disposed
363 * before its first usage, remove it from start workset
364 * We don't do this for phis though */
365 for (i = len - spills_needed; i < len; ++i) {
366 ir_node *val = ws->vals[i].node;
368 DB((dbg, DBG_DECIDE, " disposing node %+F (%u)\n", val,
369 workset_get_time(ws, i)));
371 if(!USES_IS_INFINITE(ws->vals[i].time)
372 && !ws->vals[i].reloaded) {
373 //be_add_spill(senv, val, instr);
376 if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) {
377 workset_remove(ws_start, val);
378 DB((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", val));
382 /* kill the last 'demand' entries in the array */
383 workset_set_length(ws, len - spills_needed);
386 /* 3. Insert the new values into the workset */
387 for (i = 0; i < demand; ++i) {
388 ir_node *val = to_insert[i];
390 workset_insert(ws, val, 1);
394 /** Decides whether a specific node should be in the start workset or not
396 * @param env belady environment
398 * @param node the node to test
399 * @param loop the loop of the node
401 static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node,
404 be_next_use_t next_use;
407 loc.time = USES_INFINITY;
409 //loc.reloaded = rand() % 2; /* provoke a bug... */
412 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
413 loc.time = USES_INFINITY;
417 /* We have to keep nonspillable nodes in the workingset */
418 if(arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
420 DB((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time));
424 next_use = be_get_next_use(uses, first, 0, node, 0);
425 if(USES_IS_INFINITE(next_use.time)) {
426 // the nodes marked as live in shouldn't be dead, so it must be a phi
427 assert(is_Phi(node));
428 loc.time = USES_INFINITY;
429 DB((dbg, DBG_START, " %+F not taken (dead)\n", node));
431 be_spill_phi(senv, node);
436 loc.time = next_use.time;
438 if(next_use.outermost_loop >= get_loop_depth(loop)) {
439 DB((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
441 loc.time = USES_PENDING;
442 DB((dbg, DBG_START, " %+F delayed (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
448 * Computes the start-workset for a block with multiple predecessors. We assume
449 * that at least 1 of the predeccesors is a back-edge which means we're at the
450 * beginning of a loop. We try to reload as much values as possible now so they
451 * don't get reloaded inside the loop.
453 static void compute_live_ins(const ir_node *block)
455 ir_loop *loop = get_irn_loop(block);
461 int i, len, ws_count;
462 int free_slots, free_pressure_slots;
465 //int n_pred_worksets;
466 //workset_t **pred_worksets;
468 /* Collect all values living at start of block */
469 starters = NEW_ARR_F(loc_t, 0);
470 delayed = NEW_ARR_F(loc_t, 0);
472 DB((dbg, DBG_START, "Living at start of %+F:\n", block));
473 first = sched_first(block);
475 /* check all Phis first */
476 sched_foreach(block, node) {
480 loc = to_take_or_not_to_take(first, node, loop);
482 if (! USES_IS_INFINITE(loc.time)) {
483 if (USES_IS_PENDING(loc.time))
484 ARR_APP1(loc_t, delayed, loc);
486 ARR_APP1(loc_t, starters, loc);
490 /* check all Live-Ins */
491 be_lv_foreach(lv, block, be_lv_state_in, i) {
492 ir_node *node = be_lv_get_irn(lv, block, i);
494 loc = to_take_or_not_to_take(first, node, loop);
496 if (! USES_IS_INFINITE(loc.time)) {
497 if (USES_IS_PENDING(loc.time))
498 ARR_APP1(loc_t, delayed, loc);
500 ARR_APP1(loc_t, starters, loc);
504 pressure = be_get_loop_pressure(loop_ana, cls, loop);
505 assert(ARR_LEN(delayed) <= (signed)pressure);
506 free_slots = n_regs - ARR_LEN(starters);
507 free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed));
508 free_slots = MIN(free_slots, free_pressure_slots);
510 /* so far we only put nodes into the starters list that are used inside
511 * the loop. If register pressure in the loop is low then we can take some
512 * values and let them live through the loop */
514 qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare);
516 for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
517 DB((dbg, DBG_START, " delayed %+F taken\n", delayed[i].node));
518 ARR_APP1(loc_t, starters, delayed[i]);
519 delayed[i].node = NULL;
523 /* spill phis (the actual phis not just their values) that are in this block
524 * but not in the start workset */
525 for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
526 ir_node *node = delayed[i].node;
527 if(node == NULL || !is_Phi(node) || get_nodes_block(node) != block)
530 DB((dbg, DBG_START, " spilling delayed phi %+F\n", node));
531 be_spill_phi(senv, node);
535 /* Sort start values by first use */
536 qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
538 /* Copy the best ones from starters to start workset */
539 ws_count = MIN(ARR_LEN(starters), n_regs);
541 workset_bulk_fill(ws, ws_count, starters);
543 /* spill phis (the actual phis not just their values) that are in this block
544 * but not in the start workset */
545 len = ARR_LEN(starters);
546 for (i = ws_count; i < len; ++i) {
547 ir_node *node = starters[i].node;
548 if (! is_Phi(node) || get_nodes_block(node) != block)
551 DB((dbg, DBG_START, " spilling phi %+F\n", node));
552 be_spill_phi(senv, node);
558 /* determine reloaded status of the values: If there's 1 pred block (which
559 * is no backedge) where the value is reloaded then we must set it to
560 * reloaded here. We place spills in all pred where the value was not yet
561 * reloaded to be sure we have a spill on each path */
563 arity = get_irn_arity(block);
564 pred_worksets = alloca(sizeof(pred_worksets[0]) * arity);
565 for(i = 0; i < arity; ++i) {
566 ir_node *pred_block = get_Block_cfgpred_block(block, i);
567 block_info_t *pred_info = get_block_info(pred_block);
568 if(pred_info == NULL)
571 pred_worksets[n_pred_worksets] = pred_info->end_workset;
575 for(i = 0; i < ws_count; ++i) {
576 loc_t *loc = &ws->vals[i];
577 ir_node *value = loc->node;
581 /* phis from this block aren't reloaded */
582 if(get_nodes_block(value) == block) {
583 assert(is_Phi(value));
584 loc->reloaded = value_not_reloaded;
588 /* was the value reloaded on any of the other inputs */
590 arity = get_Block_n_cfgpreds(block);
591 for(n = 0; n < n_pred_worksets; ++n) {
592 workset_t *pred_workset = pred_worksets[n];
593 int p_len = workset_get_length(pred_workset);
596 for(p = 0; p < p_len; ++p) {
597 loc_t *l = &pred_workset->vals[p];
598 if(l->node == value) {
615 * For the given block @p block, decide for each values
616 * whether it is used from a register or is reloaded
619 static void belady(ir_node *block)
624 block_info_t *block_info;
626 int has_backedges = 0;
628 const ir_edge_t *edge;
630 /* no need to process a block twice */
631 if(get_block_info(block) != NULL) {
635 /* check if all predecessor blocks are processed yet (though for backedges
636 * we have to make an exception as we can't process them first) */
637 arity = get_Block_n_cfgpreds(block);
638 for(i = 0; i < arity; ++i) {
639 ir_node *pred_block = get_Block_cfgpred_block(block, i);
640 block_info_t *pred_info = get_block_info(pred_block);
642 if(pred_info == NULL) {
643 /* process predecessor first (it will be in the queue already) */
644 if(!is_backedge(block, i)) {
650 (void) has_backedges;
653 } else if(arity == 1) {
654 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
655 block_info_t *pred_info = get_block_info(pred_block);
657 assert(pred_info != NULL);
658 workset_copy(ws, pred_info->end_workset);
660 /* we need 2 heuristics here, for the case when all predecessor blocks
661 * are known and when some are backedges (and therefore can't be known
663 compute_live_ins(block);
666 DB((dbg, DBG_DECIDE, "\n"));
667 DB((dbg, DBG_DECIDE, "Decide for %+F\n", block));
669 block_info = new_block_info();
670 set_block_info(block, block_info);
672 DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
673 workset_foreach(ws, irn, iter) {
674 DB((dbg, DBG_WSETS, " %+F (%u)\n", irn,
675 workset_get_time(ws, iter)));
678 block_info->start_workset = workset_clone(ws);
680 /* process the block from start to end */
681 DB((dbg, DBG_WSETS, "Processing...\n"));
682 ir_nodeset_init(&used);
684 /* TODO: this leaks (into the obstack)... */
685 new_vals = new_workset();
687 sched_foreach(block, irn) {
689 assert(workset_get_length(ws) <= n_regs);
691 /* Phis are no real instr (see insert_starters()) */
695 DB((dbg, DBG_DECIDE, " ...%+F\n", irn));
697 /* set instruction in the workset */
700 /* allocate all values _used_ by this instruction */
701 workset_clear(new_vals);
702 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
703 ir_node *in = get_irn_n(irn, i);
704 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in))
707 /* (note that reloaded_value is irrelevant here) */
708 workset_insert(new_vals, in, 0);
710 displace(new_vals, 1);
712 /* allocate all values _defined_ by this instruction */
713 workset_clear(new_vals);
714 if (get_irn_mode(irn) == mode_T) {
715 const ir_edge_t *edge;
717 foreach_out_edge(irn, edge) {
718 ir_node *proj = get_edge_src_irn(edge);
719 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
721 workset_insert(new_vals, proj, 0);
724 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
726 workset_insert(new_vals, irn, 0);
728 displace(new_vals, 0);
732 ir_nodeset_destroy(&used);
734 /* Remember end-workset for this block */
735 block_info->end_workset = workset_clone(ws);
736 DB((dbg, DBG_WSETS, "End workset for %+F:\n", block));
737 workset_foreach(ws, irn, iter)
738 DB((dbg, DBG_WSETS, " %+F (%u)\n", irn,
739 workset_get_time(ws, iter)));
741 /* add successor blocks into worklist */
742 foreach_block_succ(block, edge) {
743 ir_node *succ = get_edge_src_irn(edge);
744 pdeq_putr(worklist, succ);
749 * 'decide' is block-local and makes assumptions
750 * about the set of live-ins. Thus we must adapt the
751 * live-outs to the live-ins at each block-border.
753 static void fix_block_borders(ir_node *block, void *data)
755 workset_t *start_workset;
761 DB((dbg, DBG_FIX, "\n"));
762 DB((dbg, DBG_FIX, "Fixing %+F\n", block));
764 start_workset = get_block_info(block)->start_workset;
766 /* process all pred blocks */
767 arity = get_irn_arity(block);
768 for (i = 0; i < arity; ++i) {
769 ir_node *pred = get_Block_cfgpred_block(block, i);
770 workset_t *pred_end_workset = get_block_info(pred)->end_workset;
773 DB((dbg, DBG_FIX, " Pred %+F\n", pred));
775 /* spill all values not used anymore */
776 workset_foreach(pred_end_workset, node, iter) {
780 workset_foreach(start_workset, n2, iter2) {
785 /* note that we do not look at phi inputs, becuase the values
786 * will be either live-end and need no spill or
787 * they have other users in which must be somewhere else in the
792 if(!found && be_is_live_out(lv, pred, node)
793 && !pred_end_workset->vals[iter].reloaded) {
794 ir_node *insert_point
795 = be_get_end_of_block_insertion_point(pred);
796 DB((dbg, DBG_SPILL, "Spill %+F before %+F\n", node,
798 be_add_spill(senv, node, insert_point);
803 /* reload missing values in predecessors */
804 workset_foreach(start_workset, node, iter) {
805 /* if node is a phi of the current block we reload
806 * the corresponding argument, else node itself */
807 if(is_Phi(node) && block == get_nodes_block(node)) {
808 node = get_irn_n(node, i);
810 /* we might have unknowns as argument for the phi */
811 if(!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
815 /* check if node is in a register at end of pred */
816 if(workset_contains(pred_end_workset, node))
819 /* node is not in memory at the end of pred -> reload it */
820 DB((dbg, DBG_FIX, " reload %+F\n", node));
821 DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
822 be_add_reload_on_edge(senv, node, block, i, cls, 1);
827 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls)
829 ir_graph *irg = be_get_birg_irg(birg);
831 be_liveness_assure_sets(be_assure_liveness(birg));
833 /* construct control flow loop tree */
834 if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
835 construct_cf_backedges(irg);
840 /* init belady env */
842 arch_env = birg->main_env->arch_env;
844 lv = be_get_birg_liveness(birg);
845 n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
847 uses = be_begin_uses(irg, lv);
848 loop_ana = be_new_loop_pressure(birg);
849 senv = be_new_spill_env(birg);
850 worklist = new_pdeq();
852 pdeq_putr(worklist, get_irg_start_block(irg));
854 while(!pdeq_empty(worklist)) {
855 ir_node *block = pdeq_getl(worklist);
858 /* end block might not be reachable in endless loops */
859 belady(get_irg_end_block(irg));
863 /* belady was block-local, fix the global flow by adding reloads on the
865 irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
867 /* Insert spill/reload nodes into the graph and fix usages */
868 be_insert_spills_reloads(senv);
871 be_delete_spill_env(senv);
873 be_free_loop_pressure(loop_ana);
874 obstack_free(&obst, NULL);
877 void be_init_spillbelady(void)
879 static be_spiller_t belady_spiller = {
883 be_register_spiller("belady", &belady_spiller);
884 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
887 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);