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 DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
50 static firm_dbg_module_t *dbg = NULL;
52 #define MIN(a,b) (((a)<(b))?(a):(b))
54 typedef struct _workset_t workset_t;
56 typedef struct _belady_env_t {
58 const arch_env_t *arch;
59 const arch_register_class_t *cls;
60 int n_regs; /** number of regs in this reg-class */
62 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
63 be_uses_t *uses; /**< env for the next-use magic */
64 ir_node *instr; /**< current instruction */
65 unsigned instr_nr; /**< current instruction number (relative to block start) */
66 pset *used; /**< holds the values used (so far) in the current BB */
67 pset *copies; /**< holds all copies placed due to phi-spilling */
69 spill_env_t *senv; /* see bespill.h */
70 pset *reloads; /**< all reload nodes placed */
75 int len; /**< current length */
76 loc_t vals[1]; /**< inlined array of the values/distances in this working set */
79 void workset_print(const workset_t *w)
83 for(i = 0; i < w->len; ++i) {
84 ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time);
89 * Alloc a new workset on obstack @p ob with maximum size @p max
91 static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
93 size_t size = sizeof(*res) + (bel->n_regs-1)*sizeof(res->vals[0]);
94 res = obstack_alloc(ob, size);
101 * Alloc a new instance on obstack and make it equal to @param ws
103 static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
105 size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
106 res = obstack_alloc(ob, size);
107 memcpy(res, ws, size);
112 * Do NOT alloc anything. Make @param tgt equal to @param src.
113 * returns @param tgt for convinience
115 static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
116 size_t size = sizeof(*src) + (src->bel->n_regs-1)*sizeof(src->vals[0]);
117 memcpy(tgt, src, size);
122 * Overwrites the current content array of @param ws with the
123 * @param count locations given at memory @param locs.
124 * Set the length of @param ws to count.
126 #define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0]));
130 * Inserts the value @p val into the workset, iff it is not
131 * already contained. The workset must not be full.
133 static INLINE void workset_insert(workset_t *ws, ir_node *val) {
135 /* check for current regclass */
136 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
137 DBG((dbg, DBG_DECIDE, "Dropped %+F\n", val));
141 /* check if val is already contained */
142 for(i=0; i<ws->len; ++i)
143 if (ws->vals[i].irn == val)
147 assert(ws->len < ws->bel->n_regs && "Workset already full!");
148 ws->vals[ws->len++].irn = val;
152 * Inserts all values in array @p vals of length @p cnt
153 * into the workset. There must be enough space for the
156 static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
159 for(o=0; o<cnt; ++o) {
160 ir_node *val = vals[o];
161 DBG((dbg, DBG_TRACE, "Bulk insert %+F\n", val));
162 /* check for current regclass */
163 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
164 DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
168 /* check if val is already contained */
169 for(i=0; i<ws->len; ++i)
170 if (ws->vals[i].irn == val) {
171 DBG((dbg, DBG_TRACE, "Already contained\n"));
176 assert(ws->len < ws->bel->n_regs && "Workset does not have enough room!");
177 ws->vals[ws->len++].irn = val;
178 DBG((dbg, DBG_TRACE, "Inserted\n"));
181 /*epsilon statement :)*/;
186 * Removes all entries from this workset
188 #define workset_clear(ws) (ws)->len = 0;
191 * Removes the value @p val from the workset if present.
193 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
195 for(i=0; i<ws->len; ++i)
196 if (ws->vals[i].irn == val) {
197 ws->vals[i] = ws->vals[--ws->len];
202 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
204 for(i=0; i<ws->len; ++i)
205 if (ws->vals[i].irn == val)
211 * Iterates over all values in the working set.
212 * @p ws The workset to iterate
213 * @p v A variable to put the current value in
214 * @p i An integer for internal use
216 #define workset_foreach(ws, v, i) for(i=0; \
217 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
220 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
221 #define workset_set_length(ws, length) (ws)->len = length
222 #define workset_get_length(ws) ((ws)->len)
223 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
224 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
226 typedef struct _block_info_t {
227 workset_t *ws_start, *ws_end;
231 static INLINE void *new_block_info(struct obstack *ob) {
232 block_info_t *res = obstack_alloc(ob, sizeof(*res));
233 res->ws_start = NULL;
239 #define get_block_info(blk) ((block_info_t *)get_irn_link(blk))
240 #define set_block_info(blk, info) set_irn_link(blk, info)
242 static int is_mem_phi(const ir_node *irn, void *data) {
244 ir_node *blk = get_nodes_block(irn);
246 DBG((dbg, DBG_SPILL, "Is %+F a mem-phi?\n", irn));
247 sws = get_block_info(blk)->ws_start;
248 DBG((dbg, DBG_SPILL, " %d\n", !workset_contains(sws, irn)));
249 return !workset_contains(sws, irn);
253 * @return The distance to the next use
254 * Or 0 if irn is an ignore node
257 static INLINE unsigned get_distance(belady_env_t *bel, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
259 arch_irn_flags_t fl = arch_irn_get_flags(bel->arch, def);
260 if((fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0)
263 return be_get_next_use(bel->uses, from, from_step, def, skip_from_uses);
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));
298 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 /* Only make more free room if we do not have enough */
308 if (len > max_allowed) {
309 /* get current next-use distance */
310 for (i=0; i<ws->len; ++i)
311 workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
313 /* sort entries by increasing nextuse-distance*/
316 /* Logic for not needed live-ins: If a value is disposed
317 before its first usage, remove it from start workset */
318 for (i=max_allowed; i<ws->len; ++i) {
319 ir_node *irn = ws->vals[i].irn;
320 if (!pset_find_ptr(bel->used, irn)) {
321 ir_node *curr_bb = get_nodes_block(bel->instr);
322 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
323 workset_remove(ws_start, irn);
325 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
327 DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn));
330 /* kill the last 'demand' entries in the array */
331 workset_set_length(ws, max_allowed);
335 * 3. Insert the new values into the workset
337 workset_bulk_insert(bel->ws, demand, to_insert);
340 static void belady(ir_node *blk, void *env);
343 * Collects all values live-in at block @p blk and all phi results in this block.
344 * Then it adds the best values (at most n_regs) to the blocks start_workset.
345 * The phis among the remaining values get spilled: Introduce psudo-copies of
346 * their args to break interference and make it possible to spill them to the
349 static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
350 belady_env_t *bel = env;
351 ir_node *irn, *first;
353 int i, count, ws_count;
354 loc_t loc, *starters;
355 ir_graph *irg = get_irn_irg(blk);
357 block_info_t *res = get_block_info(blk);
359 /* Have we seen this block before? */
363 /* Create the block info for this block. */
364 res = new_block_info(&bel->ob);
365 set_block_info(blk, res);
368 /* Get all values living at the block start sorted by next use*/
371 DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
372 first = sched_first(blk);
374 sched_foreach(blk, irn)
375 if (is_Phi(irn) && arch_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) {
377 loc.time = get_distance(bel, first, 0, irn, 0);
378 obstack_grow(&ob, &loc, sizeof(loc));
379 DBG((dbg, DBG_START, " %+F:\n", irn));
384 live_foreach(blk, li)
385 if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) {
386 loc.irn = (ir_node *)li->irn;
387 loc.time = get_distance(bel, first, 0, li->irn, 0);
388 obstack_grow(&ob, &loc, sizeof(loc));
389 DBG((dbg, DBG_START, " %+F:\n", irn));
393 starters = obstack_finish(&ob);
394 qsort(starters, count, sizeof(starters[0]), loc_compare);
397 /* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */
398 if (get_Block_n_cfgpreds(blk) == 1 && blk != get_irg_start_block(get_irn_irg(blk))) {
399 ir_node *pred_blk = get_Block_cfgpred_block(blk, 0);
400 block_info_t *pred_info = get_block_info(pred_blk);
402 /* if pred block has not been processed yet, do it now */
404 belady(pred_blk, bel);
405 pred_info = get_block_info(pred_blk);
408 /* now we have an end_set of pred */
409 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
410 res->ws_start = workset_clone(&bel->ob, pred_info->ws_end);
414 /* Else we want the start_set to be the values used 'the closest' */
416 /* Copy the best ones from starters to start workset */
417 ws_count = MIN(count, bel->n_regs);
418 res->ws_start = new_workset(&bel->ob, bel);
419 workset_bulk_fill(res->ws_start, ws_count, starters);
423 /* The phis of this block which are not in the start set have to be spilled later.
424 * Therefore we add temporary copies in the pred_blocks so the spills can spill
425 * into the same spill slot.
426 * After spilling these copies get deleted. */
427 for (i=workset_get_length(res->ws_start); i<count; ++i) {
430 irn = starters[i].irn;
431 if (!is_Phi(irn) || get_nodes_block(irn) != blk)
434 DBG((dbg, DBG_START, "For %+F:\n", irn));
436 for (max=get_irn_arity(irn), o=0; o<max; ++o) {
437 ir_node *arg = get_irn_n(irn, o);
438 ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
439 ir_node *cpy = be_new_Copy(bel->cls, irg, pred_block, arg);
440 pset_insert_ptr(bel->copies, cpy);
441 DBG((dbg, DBG_START, " place a %+F of %+F in %+F\n", cpy, arg, pred_block));
442 sched_add_before(pred_block, cpy);
443 set_irn_n(irn, o, cpy);
447 obstack_free(&ob, NULL);
453 * For the given block @p blk, decide for each values
454 * whether it is used from a register or is reloaded
457 static void belady(ir_node *blk, void *env) {
458 belady_env_t *bel = env;
462 block_info_t *blk_info;
464 /* Don't do a block twice */
465 if (get_block_info(blk))
468 /* get the starting workset for this block */
469 blk_info = compute_block_start_info(blk, bel);
471 DBG((dbg, DBG_DECIDE, "\n"));
472 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", blk));
474 workset_copy(bel->ws, blk_info->ws_start);
475 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk));
476 workset_foreach(bel->ws, irn, iter)
477 DBG((dbg, DBG_WSETS, " %+F\n", irn));
479 /* process the block from start to end */
480 DBG((dbg, DBG_WSETS, "Processing...\n"));
481 bel->used = pset_new_ptr(32);
483 new_vals = new_workset(&bel->ob, bel);
484 sched_foreach(blk, irn) {
485 assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!");
488 /* projs are handled with the tuple value.
489 * Phis are no real instr (see insert_starters())
490 * instr_nr does not increase */
491 if (is_Proj(irn) || is_Phi(irn)) {
492 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
495 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
497 /* set instruction in the workset */
500 /* allocate all values _used_ by this instruction */
501 workset_clear(new_vals);
502 workset_bulk_insert(new_vals, get_irn_arity(irn)+1, get_irn_in(irn));
503 displace(bel, new_vals, 1);
505 /* allocate all values _defined_ by this instruction */
506 workset_clear(new_vals);
507 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
509 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
510 workset_insert(new_vals, proj);
512 workset_insert(new_vals, irn);
514 displace(bel, new_vals, 0);
520 /* Remember end-workset for this block */
521 blk_info->ws_end = workset_clone(&bel->ob, bel->ws);
522 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk));
523 workset_foreach(blk_info->ws_end, irn, iter)
524 DBG((dbg, DBG_WSETS, " %+F\n", irn));
528 * 'decide' is block-local and makes assumptions
529 * about the set of live-ins. Thus we must adapt the
530 * live-outs to the live-ins at each block-border.
532 static void fix_block_borders(ir_node *blk, void *env) {
534 belady_env_t *bel = env;
535 int i, max, iter, iter2;
537 DBG((dbg, DBG_FIX, "\n"));
538 DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
540 wsb = get_block_info(blk)->ws_start;
542 /* process all pred blocks */
543 for (i=0, max=get_irn_arity(blk); i<max; ++i) {
544 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
545 workset_t *wsp = get_block_info(pred)->ws_end;
547 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
549 workset_foreach(wsb, irnb, iter) {
550 /* if irnb is a phi of the current block we reload
551 * the corresponding argument, else irnb itself */
552 if(is_Phi(irnb) && blk == get_nodes_block(irnb))
553 irnb = get_irn_n(irnb, i);
555 /* Unknowns are available everywhere */
556 if(get_irn_opcode(irnb) == iro_Unknown)
559 /* check if irnb is in a register at end of pred */
560 workset_foreach(wsp, irnp, iter2)
564 /* irnb is in memory at the end of pred, so we have to reload it */
565 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
566 be_add_reload_on_edge(bel->senv, irnb, blk, i);
569 /*epsilon statement :)*/;
575 * Removes all used reloads from bel->reloads.
576 * The remaining nodes in bel->reloads will be removed from the graph.
578 static void rescue_used_reloads(ir_node *irn, void *env) {
579 pset *rlds = (pset *)env;
580 if (pset_find_ptr(rlds, irn))
581 pset_remove_ptr(rlds, irn);
585 * Removes all copies introduced for phi-spills
587 static void remove_copies(belady_env_t *bel) {
590 for (irn = pset_first(bel->copies); irn; irn = pset_next(bel->copies)) {
593 assert(be_is_Copy(irn));
594 assert(get_irn_n_edges(irn) == 1 && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
596 user = get_irn_edge(get_irn_irg(irn), irn, 0)->src;
598 src = get_irn_n(irn, be_pos_Copy_orig);
599 set_irn_n(user, 0, src);
604 * Finds all unused reloads and remove them from the schedule
605 * Also removes spills if they are not used anymore after removing reloads
607 static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) {
610 irg_walk_graph(irg, rescue_used_reloads, NULL, bel->reloads);
611 for(irn = pset_first(bel->reloads); irn; irn = pset_next(bel->reloads)) {
613 DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn)));
615 spill = get_irn_n(irn, be_pos_Reload_mem);
618 set_irn_n(irn, 0, new_Bad());
621 /* if spill not used anymore, remove it too
622 * test of regclass is necessary since spill may be a phi-M */
623 if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) {
624 set_irn_n(spill, 0, new_Bad());
630 void be_spill_belady(const be_chordal_env_t *chordal_env) {
633 dbg = firm_dbg_register("ir.be.spillbelady");
635 /* init belady env */
636 obstack_init(&bel.ob);
637 bel.arch = chordal_env->birg->main_env->arch_env;
638 bel.cls = chordal_env->cls;
639 bel.n_regs = arch_register_class_n_regs(bel.cls);
640 bel.ws = new_workset(&bel.ob, &bel);
641 bel.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls);
642 bel.senv = be_new_spill_env(dbg, chordal_env, is_mem_phi, NULL);
643 bel.reloads = pset_new_ptr_default();
644 bel.copies = pset_new_ptr_default();
646 DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
649 be_clear_links(chordal_env->irg);
650 irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel);
651 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
652 be_insert_spills_reloads(bel.senv, bel.reloads);
653 remove_unused_reloads(chordal_env->irg, &bel);
657 del_pset(bel.reloads);
658 be_delete_spill_env(bel.senv);
659 be_end_uses(bel.uses);
660 obstack_free(&bel.ob, NULL);