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 */
73 int len; /**< current length */
74 loc_t vals[1]; /**< inlined array of the values/distances in this working set */
77 void workset_print(const workset_t *w)
81 for(i = 0; i < w->len; ++i) {
82 ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time);
87 * Alloc a new workset on obstack @p ob with maximum size @p max
89 static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
91 size_t size = sizeof(*res) + (bel->n_regs-1)*sizeof(res->vals[0]);
92 res = obstack_alloc(ob, size);
99 * Alloc a new instance on obstack and make it equal to @param ws
101 static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
103 size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
104 res = obstack_alloc(ob, size);
105 memcpy(res, ws, size);
110 * Do NOT alloc anything. Make @param tgt equal to @param src.
111 * returns @param tgt for convinience
113 static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
114 size_t size = sizeof(*src) + (src->bel->n_regs-1)*sizeof(src->vals[0]);
115 memcpy(tgt, src, size);
120 * Overwrites the current content array of @param ws with the
121 * @param count locations given at memory @param locs.
122 * Set the length of @param ws to count.
124 #define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0]));
128 * Inserts the value @p val into the workset, iff it is not
129 * already contained. The workset must not be full.
131 static INLINE void workset_insert(workset_t *ws, ir_node *val) {
133 /* check for current regclass */
134 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
135 DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val));
139 /* check if val is already contained */
140 for(i=0; i<ws->len; ++i)
141 if (ws->vals[i].irn == val)
145 assert(ws->len < ws->bel->n_regs && "Workset already full!");
146 ws->vals[ws->len++].irn = val;
150 * Inserts all values in array @p vals of length @p cnt
151 * into the workset. There must be enough space for the
154 static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
157 for(o=0; o<cnt; ++o) {
158 ir_node *val = vals[o];
159 DBG((dbg, DBG_TRACE, "Bulk insert %+F\n", val));
160 /* check for current regclass */
161 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
162 DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
166 /* check if val is already contained */
167 for(i=0; i<ws->len; ++i)
168 if (ws->vals[i].irn == val) {
169 DBG((dbg, DBG_TRACE, "Already contained\n"));
174 assert(ws->len < ws->bel->n_regs && "Workset does not have enough room!");
175 ws->vals[ws->len++].irn = val;
176 DBG((dbg, DBG_TRACE, "Inserted\n"));
179 /*epsilon statement :)*/;
184 * Removes all entries from this workset
186 #define workset_clear(ws) (ws)->len = 0;
189 * Removes the value @p val from the workset if present.
191 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
193 for(i=0; i<ws->len; ++i)
194 if (ws->vals[i].irn == val) {
195 ws->vals[i] = ws->vals[--ws->len];
200 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
202 for(i=0; i<ws->len; ++i)
203 if (ws->vals[i].irn == val)
209 * Iterates over all values in the working set.
210 * @p ws The workset to iterate
211 * @p v A variable to put the current value in
212 * @p i An integer for internal use
214 #define workset_foreach(ws, v, i) for(i=0; \
215 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
218 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
219 #define workset_set_length(ws, length) (ws)->len = length
220 #define workset_get_length(ws) ((ws)->len)
221 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
222 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
224 typedef struct _block_info_t {
225 workset_t *ws_start, *ws_end;
229 static INLINE void *new_block_info(struct obstack *ob) {
230 block_info_t *res = obstack_alloc(ob, sizeof(*res));
231 res->ws_start = NULL;
237 #define get_block_info(blk) ((block_info_t *)get_irn_link(blk))
238 #define set_block_info(blk, info) set_irn_link(blk, info)
240 static int is_mem_phi(const ir_node *irn, void *data) {
242 ir_node *blk = get_nodes_block(irn);
244 DBG((dbg, DBG_SPILL, "Is %+F a mem-phi?\n", irn));
245 sws = get_block_info(blk)->ws_start;
246 DBG((dbg, DBG_SPILL, " %d\n", !workset_contains(sws, irn)));
247 return !workset_contains(sws, irn);
251 * @return The distance to the next use
252 * Or 0 if irn is an ignore node
255 static INLINE unsigned get_distance(belady_env_t *bel, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
257 arch_irn_flags_t fl = arch_irn_get_flags(bel->arch, def);
258 unsigned dist = be_get_next_use(bel->uses, from, from_step, def, skip_from_uses);
260 if(!USES_IS_INIFINITE(dist) && (fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0)
267 * Performs the actions necessary to grant the request that:
268 * - new_vals can be held in registers
269 * - as few as possible other values are disposed
270 * - the worst values get disposed
272 * @p is_usage indicates that the values in new_vals are used (not defined)
273 * In this case reloads must be performed
275 static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
277 int i, len, max_allowed, demand, iter;
278 workset_t *ws = bel->ws;
279 ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert));
282 * 1. Identify the number of needed slots and the values to reload
285 workset_foreach(new_vals, val, iter) {
286 /* mark value as used */
288 pset_insert_ptr(bel->used, val);
290 if (!workset_contains(ws, val)) {
291 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
292 to_insert[demand++] = val;
294 be_add_reload(bel->senv, val, bel->instr);
296 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
299 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
302 * 2. Make room for at least 'demand' slots
304 len = workset_get_length(ws);
305 max_allowed = bel->n_regs - demand;
307 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
309 /* Only make more free room if we do not have enough */
310 if (len > max_allowed) {
311 /* get current next-use distance */
312 for (i=0; i<ws->len; ++i)
313 workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
315 /* sort entries by increasing nextuse-distance*/
318 /* Logic for not needed live-ins: If a value is disposed
319 before its first usage, remove it from start workset */
320 for (i=max_allowed; i<ws->len; ++i) {
321 ir_node *irn = ws->vals[i].irn;
323 if (!pset_find_ptr(bel->used, irn)) {
324 ir_node *curr_bb = get_nodes_block(bel->instr);
325 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
326 workset_remove(ws_start, irn);
328 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
330 DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn));
334 /* kill the last 'demand' entries in the array */
335 workset_set_length(ws, max_allowed);
339 * 3. Insert the new values into the workset
341 workset_bulk_insert(bel->ws, demand, to_insert);
344 static void belady(ir_node *blk, void *env);
347 * Collects all values live-in at block @p blk and all phi results in this block.
348 * Then it adds the best values (at most n_regs) to the blocks start_workset.
349 * The phis among the remaining values get spilled: Introduce psudo-copies of
350 * their args to break interference and make it possible to spill them to the
353 static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
354 belady_env_t *bel = env;
355 ir_node *irn, *first;
357 int i, count, ws_count;
358 loc_t loc, *starters;
359 ir_graph *irg = get_irn_irg(blk);
361 block_info_t *res = get_block_info(blk);
363 /* Have we seen this block before? */
367 /* Create the block info for this block. */
368 res = new_block_info(&bel->ob);
369 set_block_info(blk, res);
372 /* Get all values living at the block start sorted by next use*/
375 DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
376 first = sched_first(blk);
378 sched_foreach(blk, irn) {
379 if (is_Phi(irn) && arch_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) {
381 loc.time = get_distance(bel, first, 0, irn, 0);
382 obstack_grow(&ob, &loc, sizeof(loc));
383 DBG((dbg, DBG_START, " %+F:\n", irn));
389 live_foreach(blk, li) {
390 if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) {
391 loc.irn = (ir_node *)li->irn;
392 loc.time = get_distance(bel, first, 0, li->irn, 0);
393 obstack_grow(&ob, &loc, sizeof(loc));
394 DBG((dbg, DBG_START, " %+F:\n", li->irn));
399 starters = obstack_finish(&ob);
400 qsort(starters, count, sizeof(starters[0]), loc_compare);
403 /* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */
404 if (get_Block_n_cfgpreds(blk) == 1 && blk != get_irg_start_block(get_irn_irg(blk))) {
405 ir_node *pred_blk = get_Block_cfgpred_block(blk, 0);
406 block_info_t *pred_info = get_block_info(pred_blk);
408 /* if pred block has not been processed yet, do it now */
410 belady(pred_blk, bel);
411 pred_info = get_block_info(pred_blk);
414 /* now we have an end_set of pred */
415 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
416 res->ws_start = workset_clone(&bel->ob, pred_info->ws_end);
420 /* Else we want the start_set to be the values used 'the closest' */
422 /* Copy the best ones from starters to start workset */
423 ws_count = MIN(count, bel->n_regs);
424 res->ws_start = new_workset(&bel->ob, bel);
425 workset_bulk_fill(res->ws_start, ws_count, starters);
429 /* The phis of this block which are not in the start set have to be spilled later.
430 * Therefore we add temporary copies in the pred_blocks so the spills can spill
431 * into the same spill slot.
432 * After spilling these copies get deleted. */
433 for (i=workset_get_length(res->ws_start); i<count; ++i) {
436 irn = starters[i].irn;
437 if (!is_Phi(irn) || get_nodes_block(irn) != blk)
440 DBG((dbg, DBG_START, "For %+F:\n", irn));
442 for (max=get_irn_arity(irn), o=0; o<max; ++o) {
443 ir_node *arg = get_irn_n(irn, o);
444 ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
445 ir_node *cpy = be_new_Copy(bel->cls, irg, pred_block, arg);
446 pset_insert_ptr(bel->copies, cpy);
447 DBG((dbg, DBG_START, " place a %+F of %+F in %+F\n", cpy, arg, pred_block));
448 /* TODO: Place copies before jumps! */
449 sched_add_before(sched_last(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 copies introduced for phi-spills
585 static void remove_copies(belady_env_t *bel) {
588 foreach_pset(bel->copies, irn) {
590 const ir_edge_t *edge, *ne;
592 assert(be_is_Copy(irn));
594 src = be_get_Copy_op(irn);
595 foreach_out_edge_safe(irn, edge, ne) {
596 ir_node *user = get_edge_src_irn(edge);
597 int user_pos = get_edge_src_pos(edge);
599 set_irn_n(user, user_pos, src);
604 void be_spill_belady(const be_chordal_env_t *chordal_env) {
605 be_spill_belady_spill_env(chordal_env, NULL);
608 void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
611 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
613 /* init belady env */
614 obstack_init(&bel.ob);
615 bel.arch = chordal_env->birg->main_env->arch_env;
616 bel.cls = chordal_env->cls;
617 bel.n_regs = arch_register_class_n_regs(bel.cls);
618 bel.ws = new_workset(&bel.ob, &bel);
619 bel.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls);
620 if(spill_env == NULL) {
621 bel.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL);
623 bel.senv = spill_env;
624 be_set_is_spilled_phi(bel.senv, is_mem_phi, NULL);
626 DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);)
627 bel.copies = pset_new_ptr_default();
629 DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
632 be_clear_links(chordal_env->irg);
633 irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel);
634 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
635 be_insert_spills_reloads(bel.senv);
638 be_remove_dead_nodes_from_schedule(chordal_env->irg);
641 if(spill_env == NULL)
642 be_delete_spill_env(bel.senv);
643 be_end_uses(bel.uses);
644 obstack_free(&bel.ob, NULL);