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 */
67 spill_env_t *senv; /**< see bespill.h */
72 int len; /**< current length */
73 loc_t vals[1]; /**< inlined array of the values/distances in this working set */
76 void workset_print(const workset_t *w)
80 for(i = 0; i < w->len; ++i) {
81 ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time);
86 * Alloc a new workset on obstack @p ob with maximum size @p max
88 static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
90 size_t size = sizeof(*res) + (bel->n_regs-1)*sizeof(res->vals[0]);
91 res = obstack_alloc(ob, size);
98 * Alloc a new instance on obstack and make it equal to @param ws
100 static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
102 size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
103 res = obstack_alloc(ob, size);
104 memcpy(res, ws, size);
109 * Do NOT alloc anything. Make @param tgt equal to @param src.
110 * returns @param tgt for convinience
112 static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
113 size_t size = sizeof(*src) + (src->bel->n_regs-1)*sizeof(src->vals[0]);
114 memcpy(tgt, src, size);
119 * Overwrites the current content array of @param ws with the
120 * @param count locations given at memory @param locs.
121 * Set the length of @param ws to count.
123 #define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0]));
127 * Inserts the value @p val into the workset, iff it is not
128 * already contained. The workset must not be full.
130 static INLINE void workset_insert(workset_t *ws, ir_node *val) {
132 /* check for current regclass */
133 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
134 DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val));
138 /* check if val is already contained */
139 for(i=0; i<ws->len; ++i)
140 if (ws->vals[i].irn == val)
144 assert(ws->len < ws->bel->n_regs && "Workset already full!");
145 ws->vals[ws->len++].irn = val;
149 * Inserts all values in array @p vals of length @p cnt
150 * into the workset. There must be enough space for the
153 static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
156 for(o=0; o<cnt; ++o) {
157 ir_node *val = vals[o];
158 DBG((dbg, DBG_TRACE, "Bulk insert %+F\n", val));
159 /* check for current regclass */
160 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
161 DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
165 /* check if val is already contained */
166 for(i=0; i<ws->len; ++i)
167 if (ws->vals[i].irn == val) {
168 DBG((dbg, DBG_TRACE, "Already contained\n"));
173 assert(ws->len < ws->bel->n_regs && "Workset does not have enough room!");
174 ws->vals[ws->len++].irn = val;
175 DBG((dbg, DBG_TRACE, "Inserted\n"));
178 /*epsilon statement :)*/;
183 * Removes all entries from this workset
185 #define workset_clear(ws) (ws)->len = 0;
188 * Removes the value @p val from the workset if present.
190 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
192 for(i=0; i<ws->len; ++i)
193 if (ws->vals[i].irn == val) {
194 ws->vals[i] = ws->vals[--ws->len];
199 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
201 for(i=0; i<ws->len; ++i)
202 if (ws->vals[i].irn == val)
208 * Iterates over all values in the working set.
209 * @p ws The workset to iterate
210 * @p v A variable to put the current value in
211 * @p i An integer for internal use
213 #define workset_foreach(ws, v, i) for(i=0; \
214 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
217 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
218 #define workset_set_length(ws, length) (ws)->len = length
219 #define workset_get_length(ws) ((ws)->len)
220 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
221 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
223 typedef struct _block_info_t {
224 workset_t *ws_start, *ws_end;
228 static INLINE void *new_block_info(struct obstack *ob) {
229 block_info_t *res = obstack_alloc(ob, sizeof(*res));
230 res->ws_start = NULL;
236 #define get_block_info(blk) ((block_info_t *)get_irn_link(blk))
237 #define set_block_info(blk, info) set_irn_link(blk, info)
240 * @return The distance to the next use
241 * Or 0 if irn is an ignore node
243 static INLINE unsigned get_distance(belady_env_t *bel, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
245 arch_irn_flags_t fl = arch_irn_get_flags(bel->arch, def);
246 unsigned dist = be_get_next_use(bel->uses, from, from_step, def, skip_from_uses);
248 if(!USES_IS_INIFINITE(dist) && (fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0)
255 * Performs the actions necessary to grant the request that:
256 * - new_vals can be held in registers
257 * - as few as possible other values are disposed
258 * - the worst values get disposed
260 * @p is_usage indicates that the values in new_vals are used (not defined)
261 * In this case reloads must be performed
263 static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
265 int i, len, max_allowed, demand, iter;
266 workset_t *ws = bel->ws;
267 ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert));
270 * 1. Identify the number of needed slots and the values to reload
273 workset_foreach(new_vals, val, iter) {
274 /* mark value as used */
276 pset_insert_ptr(bel->used, val);
278 if (!workset_contains(ws, val)) {
279 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
280 to_insert[demand++] = val;
282 be_add_reload(bel->senv, val, bel->instr);
284 assert(is_usage || "Defined value already in workset?!?");
285 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
288 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
291 * 2. Make room for at least 'demand' slots
293 len = workset_get_length(ws);
294 max_allowed = bel->n_regs - demand;
296 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
298 /* Only make more free room if we do not have enough */
299 if (len > max_allowed) {
300 /* get current next-use distance */
301 for (i=0; i<ws->len; ++i)
302 workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
304 /* sort entries by increasing nextuse-distance*/
307 /* Logic for not needed live-ins: If a value is disposed
308 before its first usage, remove it from start workset */
309 for (i=max_allowed; i<ws->len; ++i) {
310 ir_node *irn = ws->vals[i].irn;
312 if (!pset_find_ptr(bel->used, irn)) {
313 ir_node *curr_bb = get_nodes_block(bel->instr);
314 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
315 workset_remove(ws_start, irn);
317 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
319 DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn));
323 /* kill the last 'demand' entries in the array */
324 workset_set_length(ws, max_allowed);
328 * 3. Insert the new values into the workset
330 workset_bulk_insert(bel->ws, demand, to_insert);
333 static void belady(ir_node *blk, void *env);
336 * Collects all values live-in at block @p blk and all phi results in this block.
337 * Then it adds the best values (at most n_regs) to the blocks start_workset.
338 * The phis among the remaining values get spilled: Introduce psudo-copies of
339 * their args to break interference and make it possible to spill them to the
342 static block_info_t *compute_block_start_info(ir_node *blk, void *data) {
343 belady_env_t *env = data;
344 ir_node *irn, *first;
346 int i, count, ws_count;
347 loc_t loc, *starters;
348 ir_graph *irg = get_irn_irg(blk);
350 block_info_t *res = get_block_info(blk);
352 /* Have we seen this block before? */
356 /* Create the block info for this block. */
357 res = new_block_info(&env->ob);
358 set_block_info(blk, res);
361 /* Get all values living at the block start sorted by next use*/
364 DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
365 first = sched_first(blk);
367 sched_foreach(blk, irn) {
368 if (is_Phi(irn) && arch_get_irn_reg_class(env->arch, irn, -1) == env->cls) {
370 loc.time = get_distance(env, first, 0, irn, 0);
371 obstack_grow(&ob, &loc, sizeof(loc));
372 DBG((dbg, DBG_START, " %+F:\n", irn));
378 live_foreach(blk, li) {
379 if (live_is_in(li) && arch_get_irn_reg_class(env->arch, li->irn, -1) == env->cls) {
380 loc.irn = (ir_node *)li->irn;
381 loc.time = get_distance(env, first, 0, li->irn, 0);
382 obstack_grow(&ob, &loc, sizeof(loc));
383 DBG((dbg, DBG_START, " %+F:\n", li->irn));
388 starters = obstack_finish(&ob);
389 qsort(starters, count, sizeof(starters[0]), loc_compare);
392 /* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */
393 if (get_Block_n_cfgpreds(blk) == 1 && blk != get_irg_start_block(get_irn_irg(blk))) {
394 ir_node *pred_blk = get_Block_cfgpred_block(blk, 0);
395 block_info_t *pred_info = get_block_info(pred_blk);
397 /* if pred block has not been processed yet, do it now */
399 belady(pred_blk, env);
400 pred_info = get_block_info(pred_blk);
403 /* now we have an end_set of pred */
404 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
405 res->ws_start = workset_clone(&env->ob, pred_info->ws_end);
409 /* Else we want the start_set to be the values used 'the closest' */
411 /* Copy the best ones from starters to start workset */
412 ws_count = MIN(count, env->n_regs);
413 res->ws_start = new_workset(&env->ob, env);
414 workset_bulk_fill(res->ws_start, ws_count, starters);
419 /* The phis of this block which are not in the start set have to be spilled later.
420 * Therefore we add temporary copies in the pred_blocks so the spills can spill
421 * into the same spill slot.
422 * After spilling these copies get deleted. */
423 for (i=workset_get_length(res->ws_start); i<count; ++i) {
424 irn = starters[i].irn;
425 if (!is_Phi(irn) || get_nodes_block(irn) != blk)
428 be_spill_phi(env->senv, irn);
431 obstack_free(&ob, NULL);
437 * For the given block @p blk, decide for each values
438 * whether it is used from a register or is reloaded
441 static void belady(ir_node *blk, void *env) {
442 belady_env_t *bel = env;
446 block_info_t *blk_info;
448 /* Don't do a block twice */
449 if (get_block_info(blk))
452 /* get the starting workset for this block */
453 blk_info = compute_block_start_info(blk, bel);
455 DBG((dbg, DBG_DECIDE, "\n"));
456 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", blk));
458 workset_copy(bel->ws, blk_info->ws_start);
459 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk));
460 workset_foreach(bel->ws, irn, iter)
461 DBG((dbg, DBG_WSETS, " %+F\n", irn));
463 /* process the block from start to end */
464 DBG((dbg, DBG_WSETS, "Processing...\n"));
465 bel->used = pset_new_ptr(32);
467 new_vals = new_workset(&bel->ob, bel);
468 sched_foreach(blk, irn) {
469 assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!");
471 /* projs are handled with the tuple value.
472 * Phis are no real instr (see insert_starters())
473 * instr_nr does not increase */
474 if (is_Proj(irn) || is_Phi(irn)) {
475 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
478 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
480 /* set instruction in the workset */
483 /* allocate all values _used_ by this instruction */
484 workset_clear(new_vals);
485 workset_bulk_insert(new_vals, get_irn_arity(irn)+1, get_irn_in(irn));
486 displace(bel, new_vals, 1);
488 /* allocate all values _defined_ by this instruction */
489 workset_clear(new_vals);
490 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
492 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
493 workset_insert(new_vals, proj);
495 workset_insert(new_vals, irn);
497 displace(bel, new_vals, 0);
503 /* Remember end-workset for this block */
504 blk_info->ws_end = workset_clone(&bel->ob, bel->ws);
505 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk));
506 workset_foreach(blk_info->ws_end, irn, iter)
507 DBG((dbg, DBG_WSETS, " %+F\n", irn));
511 * 'decide' is block-local and makes assumptions
512 * about the set of live-ins. Thus we must adapt the
513 * live-outs to the live-ins at each block-border.
515 static void fix_block_borders(ir_node *blk, void *env) {
517 belady_env_t *bel = env;
518 int i, max, iter, iter2;
520 DBG((dbg, DBG_FIX, "\n"));
521 DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
523 wsb = get_block_info(blk)->ws_start;
525 /* process all pred blocks */
526 for (i=0, max=get_irn_arity(blk); i<max; ++i) {
527 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
528 workset_t *wsp = get_block_info(pred)->ws_end;
530 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
532 workset_foreach(wsb, irnb, iter) {
533 /* if irnb is a phi of the current block we reload
534 * the corresponding argument, else irnb itself */
535 if(is_Phi(irnb) && blk == get_nodes_block(irnb))
536 irnb = get_irn_n(irnb, i);
538 /* Unknowns are available everywhere */
539 if(get_irn_opcode(irnb) == iro_Unknown)
542 /* check if irnb is in a register at end of pred */
543 workset_foreach(wsp, irnp, iter2) {
548 /* irnb is in memory at the end of pred, so we have to reload it */
549 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
550 be_add_reload_on_edge(bel->senv, irnb, blk, i);
553 /*epsilon statement :)*/;
558 void be_spill_belady(const be_chordal_env_t *chordal_env) {
559 be_spill_belady_spill_env(chordal_env, NULL);
562 void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
565 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
567 /* init belady env */
568 obstack_init(&bel.ob);
569 bel.arch = chordal_env->birg->main_env->arch_env;
570 bel.cls = chordal_env->cls;
571 bel.n_regs = arch_register_class_n_regs(bel.cls);
572 bel.ws = new_workset(&bel.ob, &bel);
573 bel.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls);
574 if(spill_env == NULL) {
575 bel.senv = be_new_spill_env(chordal_env);
577 bel.senv = spill_env;
579 DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);)
581 DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
584 be_clear_links(chordal_env->irg);
585 irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel);
586 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
587 be_insert_spills_reloads(bel.senv);
589 be_remove_dead_nodes_from_schedule(chordal_env->irg);
592 if(spill_env == NULL)
593 be_delete_spill_env(bel.senv);
594 be_end_uses(bel.uses);
595 obstack_free(&bel.ob, NULL);