4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
23 #include "irprintf_t.h"
28 #include "iredges_t.h"
34 #include "bespillbelady.h"
36 #include "besched_t.h"
40 #include "bechordal_t.h"
49 #define DBG_WORKSET 128
50 #define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
51 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
53 typedef struct _workset_t workset_t;
55 typedef struct _belady_env_t {
57 const arch_env_t *arch;
58 const arch_register_class_t *cls;
59 int n_regs; /** number of regs in this reg-class */
61 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
62 be_uses_t *uses; /**< env for the next-use magic */
63 ir_node *instr; /**< current instruction */
64 unsigned instr_nr; /**< current instruction number (relative to block start) */
65 pset *used; /**< holds the values used (so far) in the current BB */
66 pset *copies; /**< holds all copies placed due to phi-spilling */
68 spill_env_t *senv; /**< see bespill.h */
69 pset *reloads; /**< all reload nodes placed */
74 int len; /**< current length */
75 loc_t vals[1]; /**< inlined array of the values/distances in this working set */
78 void workset_print(const workset_t *w)
82 for(i = 0; i < w->len; ++i) {
83 ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time);
88 * Alloc a new workset on obstack @p ob with maximum size @p max
90 static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
92 size_t size = sizeof(*res) + (bel->n_regs-1)*sizeof(res->vals[0]);
93 res = obstack_alloc(ob, size);
100 * Alloc a new instance on obstack and make it equal to @param ws
102 static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
104 size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
105 res = obstack_alloc(ob, size);
106 memcpy(res, ws, size);
111 * Do NOT alloc anything. Make @param tgt equal to @param src.
112 * returns @param tgt for convinience
114 static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
115 size_t size = sizeof(*src) + (src->bel->n_regs-1)*sizeof(src->vals[0]);
116 memcpy(tgt, src, size);
121 * Overwrites the current content array of @param ws with the
122 * @param count locations given at memory @param locs.
123 * Set the length of @param ws to count.
125 #define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0]));
129 * Inserts the value @p val into the workset, iff it is not
130 * already contained. The workset must not be full.
132 static INLINE void workset_insert(workset_t *ws, ir_node *val) {
134 /* check for current regclass */
135 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
136 DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val));
140 /* check if val is already contained */
141 for(i=0; i<ws->len; ++i)
142 if (ws->vals[i].irn == val)
146 assert(ws->len < ws->bel->n_regs && "Workset already full!");
147 ws->vals[ws->len++].irn = val;
151 * Inserts all values in array @p vals of length @p cnt
152 * into the workset. There must be enough space for the
155 static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
158 for(o=0; o<cnt; ++o) {
159 ir_node *val = vals[o];
160 DBG((dbg, DBG_TRACE, "Bulk insert %+F\n", val));
161 /* check for current regclass */
162 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
163 DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
167 /* check if val is already contained */
168 for(i=0; i<ws->len; ++i)
169 if (ws->vals[i].irn == val) {
170 DBG((dbg, DBG_TRACE, "Already contained\n"));
175 assert(ws->len < ws->bel->n_regs && "Workset does not have enough room!");
176 ws->vals[ws->len++].irn = val;
177 DBG((dbg, DBG_TRACE, "Inserted\n"));
180 /*epsilon statement :)*/;
185 * Removes all entries from this workset
187 #define workset_clear(ws) (ws)->len = 0;
190 * Removes the value @p val from the workset if present.
192 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
194 for(i=0; i<ws->len; ++i)
195 if (ws->vals[i].irn == val) {
196 ws->vals[i] = ws->vals[--ws->len];
201 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
203 for(i=0; i<ws->len; ++i)
204 if (ws->vals[i].irn == val)
210 * Iterates over all values in the working set.
211 * @p ws The workset to iterate
212 * @p v A variable to put the current value in
213 * @p i An integer for internal use
215 #define workset_foreach(ws, v, i) for(i=0; \
216 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
219 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
220 #define workset_set_length(ws, length) (ws)->len = length
221 #define workset_get_length(ws) ((ws)->len)
222 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
223 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
225 typedef struct _block_info_t {
226 workset_t *ws_start, *ws_end;
230 static INLINE void *new_block_info(struct obstack *ob) {
231 block_info_t *res = obstack_alloc(ob, sizeof(*res));
232 res->ws_start = NULL;
238 #define get_block_info(blk) ((block_info_t *)get_irn_link(blk))
239 #define set_block_info(blk, info) set_irn_link(blk, info)
241 static int is_mem_phi(const ir_node *irn, void *data) {
243 ir_node *blk = get_nodes_block(irn);
245 DBG((dbg, DBG_SPILL, "Is %+F a mem-phi?\n", irn));
246 sws = get_block_info(blk)->ws_start;
247 DBG((dbg, DBG_SPILL, " %d\n", !workset_contains(sws, irn)));
248 return !workset_contains(sws, irn);
252 * @return The distance to the next use
253 * Or 0 if irn is an ignore node
256 static INLINE unsigned get_distance(belady_env_t *bel, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
258 arch_irn_flags_t fl = arch_irn_get_flags(bel->arch, def);
259 unsigned dist = be_get_next_use(bel->uses, from, from_step, def, skip_from_uses);
261 if(!USES_IS_INIFINITE(dist) && (fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0)
268 * Performs the actions necessary to grant the request that:
269 * - new_vals can be held in registers
270 * - as few as possible other values are disposed
271 * - the worst values get disposed
273 * @p is_usage indicates that the values in new_vals are used (not defined)
274 * In this case reloads must be performed
276 static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
278 int i, len, max_allowed, demand, iter;
279 workset_t *ws = bel->ws;
280 ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert));
283 * 1. Identify the number of needed slots and the values to reload
286 workset_foreach(new_vals, val, iter) {
287 /* mark value as used */
289 pset_insert_ptr(bel->used, val);
291 if (!workset_contains(ws, val)) {
292 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
293 to_insert[demand++] = val;
295 be_add_reload(bel->senv, val, bel->instr);
297 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
300 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
303 * 2. Make room for at least 'demand' slots
305 len = workset_get_length(ws);
306 max_allowed = bel->n_regs - demand;
308 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
310 /* Only make more free room if we do not have enough */
311 if (len > max_allowed) {
312 /* get current next-use distance */
313 for (i=0; i<ws->len; ++i)
314 workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
316 /* sort entries by increasing nextuse-distance*/
319 /* Logic for not needed live-ins: If a value is disposed
320 before its first usage, remove it from start workset */
321 for (i=max_allowed; i<ws->len; ++i) {
322 ir_node *irn = ws->vals[i].irn;
324 if (!pset_find_ptr(bel->used, irn)) {
325 ir_node *curr_bb = get_nodes_block(bel->instr);
326 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
327 workset_remove(ws_start, irn);
329 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
331 DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn));
335 /* kill the last 'demand' entries in the array */
336 workset_set_length(ws, max_allowed);
340 * 3. Insert the new values into the workset
342 workset_bulk_insert(bel->ws, demand, to_insert);
345 static void belady(ir_node *blk, void *env);
348 * Collects all values live-in at block @p blk and all phi results in this block.
349 * Then it adds the best values (at most n_regs) to the blocks start_workset.
350 * The phis among the remaining values get spilled: Introduce psudo-copies of
351 * their args to break interference and make it possible to spill them to the
354 static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
355 belady_env_t *bel = env;
356 ir_node *irn, *first;
358 int i, count, ws_count;
359 loc_t loc, *starters;
360 ir_graph *irg = get_irn_irg(blk);
362 block_info_t *res = get_block_info(blk);
364 /* Have we seen this block before? */
368 /* Create the block info for this block. */
369 res = new_block_info(&bel->ob);
370 set_block_info(blk, res);
373 /* Get all values living at the block start sorted by next use*/
376 DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
377 first = sched_first(blk);
379 sched_foreach(blk, irn) {
380 if (is_Phi(irn) && arch_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) {
382 loc.time = get_distance(bel, first, 0, irn, 0);
383 obstack_grow(&ob, &loc, sizeof(loc));
384 DBG((dbg, DBG_START, " %+F:\n", irn));
390 live_foreach(blk, li) {
391 if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) {
392 loc.irn = (ir_node *)li->irn;
393 loc.time = get_distance(bel, first, 0, li->irn, 0);
394 obstack_grow(&ob, &loc, sizeof(loc));
395 DBG((dbg, DBG_START, " %+F:\n", li->irn));
400 starters = obstack_finish(&ob);
401 qsort(starters, count, sizeof(starters[0]), loc_compare);
404 /* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */
405 if (get_Block_n_cfgpreds(blk) == 1 && blk != get_irg_start_block(get_irn_irg(blk))) {
406 ir_node *pred_blk = get_Block_cfgpred_block(blk, 0);
407 block_info_t *pred_info = get_block_info(pred_blk);
409 /* if pred block has not been processed yet, do it now */
411 belady(pred_blk, bel);
412 pred_info = get_block_info(pred_blk);
415 /* now we have an end_set of pred */
416 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
417 res->ws_start = workset_clone(&bel->ob, pred_info->ws_end);
421 /* Else we want the start_set to be the values used 'the closest' */
423 /* Copy the best ones from starters to start workset */
424 ws_count = MIN(count, bel->n_regs);
425 res->ws_start = new_workset(&bel->ob, bel);
426 workset_bulk_fill(res->ws_start, ws_count, starters);
430 /* The phis of this block which are not in the start set have to be spilled later.
431 * Therefore we add temporary copies in the pred_blocks so the spills can spill
432 * into the same spill slot.
433 * After spilling these copies get deleted. */
434 for (i=workset_get_length(res->ws_start); i<count; ++i) {
437 irn = starters[i].irn;
438 if (!is_Phi(irn) || get_nodes_block(irn) != blk)
441 DBG((dbg, DBG_START, "For %+F:\n", irn));
443 for (max=get_irn_arity(irn), o=0; o<max; ++o) {
444 ir_node *arg = get_irn_n(irn, o);
445 ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
446 ir_node *cpy = be_new_Copy(bel->cls, irg, pred_block, arg);
447 pset_insert_ptr(bel->copies, cpy);
448 DBG((dbg, DBG_START, " place a %+F of %+F in %+F\n", cpy, arg, pred_block));
449 sched_add_before(pred_block, cpy);
450 set_irn_n(irn, o, cpy);
454 obstack_free(&ob, NULL);
460 * For the given block @p blk, decide for each values
461 * whether it is used from a register or is reloaded
464 static void belady(ir_node *blk, void *env) {
465 belady_env_t *bel = env;
469 block_info_t *blk_info;
471 /* Don't do a block twice */
472 if (get_block_info(blk))
475 /* get the starting workset for this block */
476 blk_info = compute_block_start_info(blk, bel);
478 DBG((dbg, DBG_DECIDE, "\n"));
479 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", blk));
481 workset_copy(bel->ws, blk_info->ws_start);
482 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk));
483 workset_foreach(bel->ws, irn, iter)
484 DBG((dbg, DBG_WSETS, " %+F\n", irn));
486 /* process the block from start to end */
487 DBG((dbg, DBG_WSETS, "Processing...\n"));
488 bel->used = pset_new_ptr(32);
490 new_vals = new_workset(&bel->ob, bel);
491 sched_foreach(blk, irn) {
492 assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!");
495 /* projs are handled with the tuple value.
496 * Phis are no real instr (see insert_starters())
497 * instr_nr does not increase */
498 if (is_Proj(irn) || is_Phi(irn)) {
499 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
502 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
504 /* set instruction in the workset */
507 /* allocate all values _used_ by this instruction */
508 workset_clear(new_vals);
509 workset_bulk_insert(new_vals, get_irn_arity(irn)+1, get_irn_in(irn));
510 displace(bel, new_vals, 1);
512 /* allocate all values _defined_ by this instruction */
513 workset_clear(new_vals);
514 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
516 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
517 workset_insert(new_vals, proj);
519 workset_insert(new_vals, irn);
521 displace(bel, new_vals, 0);
527 /* Remember end-workset for this block */
528 blk_info->ws_end = workset_clone(&bel->ob, bel->ws);
529 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk));
530 workset_foreach(blk_info->ws_end, irn, iter)
531 DBG((dbg, DBG_WSETS, " %+F\n", irn));
535 * 'decide' is block-local and makes assumptions
536 * about the set of live-ins. Thus we must adapt the
537 * live-outs to the live-ins at each block-border.
539 static void fix_block_borders(ir_node *blk, void *env) {
541 belady_env_t *bel = env;
542 int i, max, iter, iter2;
544 DBG((dbg, DBG_FIX, "\n"));
545 DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
547 wsb = get_block_info(blk)->ws_start;
549 /* process all pred blocks */
550 for (i=0, max=get_irn_arity(blk); i<max; ++i) {
551 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
552 workset_t *wsp = get_block_info(pred)->ws_end;
554 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
556 workset_foreach(wsb, irnb, iter) {
557 /* if irnb is a phi of the current block we reload
558 * the corresponding argument, else irnb itself */
559 if(is_Phi(irnb) && blk == get_nodes_block(irnb))
560 irnb = get_irn_n(irnb, i);
562 /* Unknowns are available everywhere */
563 if(get_irn_opcode(irnb) == iro_Unknown)
566 /* check if irnb is in a register at end of pred */
567 workset_foreach(wsp, irnp, iter2) {
572 /* irnb is in memory at the end of pred, so we have to reload it */
573 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
574 be_add_reload_on_edge(bel->senv, irnb, blk, i);
577 /*epsilon statement :)*/;
583 * Removes all used reloads from bel->reloads.
584 * The remaining nodes in bel->reloads will be removed from the graph.
586 static void rescue_used_reloads(ir_node *irn, void *env) {
587 pset *rlds = (pset *)env;
588 if (pset_find_ptr(rlds, irn))
589 pset_remove_ptr(rlds, irn);
593 * Removes all copies introduced for phi-spills
595 static void remove_copies(belady_env_t *bel) {
598 foreach_pset(bel->copies, irn) {
601 assert(be_is_Copy(irn));
602 assert(get_irn_n_edges(irn) == 1 && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
604 user = get_irn_edge(get_irn_irg(irn), irn, 0)->src;
606 src = be_get_Copy_op(irn);
607 set_irn_n(user, 0, src);
612 * Finds all unused reloads and remove them from the schedule
613 * Also removes spills if they are not used anymore after removing reloads
615 static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) {
618 irg_walk_graph(irg, rescue_used_reloads, NULL, bel->reloads);
619 foreach_pset(bel->reloads, irn) {
621 DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn)));
623 if (be_is_Reload(irn))
624 spill = get_irn_n(irn, be_pos_Reload_mem);
627 set_irn_n(irn, 0, new_Bad());
630 if (be_is_Reload(irn)) {
631 /* if spill not used anymore, remove it too
632 * test of regclass is necessary since spill may be a phi-M */
633 if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) {
634 set_irn_n(spill, 0, new_Bad());
641 void be_spill_belady(const be_chordal_env_t *chordal_env) {
642 be_spill_belady_spill_env(chordal_env, NULL);
645 void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
648 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
650 /* init belady env */
651 obstack_init(&bel.ob);
652 bel.arch = chordal_env->birg->main_env->arch_env;
653 bel.cls = chordal_env->cls;
654 bel.n_regs = arch_register_class_n_regs(bel.cls);
655 bel.ws = new_workset(&bel.ob, &bel);
656 bel.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls);
657 if(spill_env == NULL) {
658 bel.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL);
660 bel.senv = spill_env;
662 DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);)
663 bel.reloads = pset_new_ptr_default();
664 bel.copies = pset_new_ptr_default();
666 DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
669 be_clear_links(chordal_env->irg);
670 irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel);
671 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
672 be_insert_spills_reloads(bel.senv, bel.reloads);
673 remove_unused_reloads(chordal_env->irg, &bel);
677 del_pset(bel.reloads);
678 if(spill_env == NULL)
679 be_delete_spill_env(bel.senv);
680 be_end_uses(bel.uses);
681 obstack_free(&bel.ob, NULL);