2 * Author: Daniel Grund, Matthias Braun
4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
23 #include "irprintf_t.h"
29 #include "iredges_t.h"
35 #include "bespillbelady.h"
37 #include "besched_t.h"
41 #include "bechordal_t.h"
50 #define DBG_WORKSET 128
51 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
54 * An association between a node and a point in time.
56 typedef struct _loc_t {
57 ir_node *irn; /**< A node. */
58 unsigned time; /**< A use time (see beuses.h). */
61 typedef struct _workset_t {
62 int len; /**< current length */
63 loc_t vals[0]; /**< inlined array of the values/distances in this working set */
66 typedef struct _belady_env_t {
68 const be_chordal_env_t *cenv;
69 const arch_env_t *arch;
70 const arch_register_class_t *cls;
71 int n_regs; /** number of regs in this reg-class */
73 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
74 be_uses_t *uses; /**< env for the next-use magic */
75 ir_node *instr; /**< current instruction */
76 unsigned instr_nr; /**< current instruction number (relative to block start) */
79 spill_env_t *senv; /**< see bespill.h */
82 static int loc_compare(const void *a, const void *b)
86 return p->time - q->time;
89 void workset_print(const workset_t *w)
93 for(i = 0; i < w->len; ++i) {
94 ir_fprintf(stderr, "%+F %d\n", w->vals[i].irn, w->vals[i].time);
99 * Alloc a new workset on obstack @p ob with maximum size @p max
101 static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) {
103 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
104 res = obstack_alloc(ob, size);
105 memset(res, 0, size);
110 * Alloc a new instance on obstack and make it equal to @param ws
112 static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) {
114 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
115 res = obstack_alloc(ob, size);
116 memcpy(res, ws, size);
121 * Do NOT alloc anything. Make @param tgt equal to @param src.
122 * returns @param tgt for convenience
124 static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src) {
125 size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]);
126 memcpy(tgt, src, size);
131 * Overwrites the current content array of @param ws with the
132 * @param count locations given at memory @param locs.
133 * Set the length of @param ws to count.
135 static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) {
136 workset->len = count;
137 memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
141 * Inserts the value @p val into the workset, iff it is not
142 * already contained. The workset must not be full.
144 static INLINE void workset_insert(belady_env_t *env, workset_t *ws, ir_node *val) {
146 /* check for current regclass */
147 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
148 DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
152 /* check if val is already contained */
153 for(i=0; i<ws->len; ++i)
154 if (ws->vals[i].irn == val)
158 assert(ws->len < env->n_regs && "Workset already full!");
159 ws->vals[ws->len++].irn = val;
163 * Removes all entries from this workset
165 static INLINE void workset_clear(workset_t *ws) {
170 * Removes the value @p val from the workset if present.
172 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
174 for(i=0; i<ws->len; ++i) {
175 if (ws->vals[i].irn == val) {
176 ws->vals[i] = ws->vals[--ws->len];
182 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
184 for(i=0; i<ws->len; ++i) {
185 if (ws->vals[i].irn == val)
193 * Iterates over all values in the working set.
194 * @p ws The workset to iterate
195 * @p v A variable to put the current value in
196 * @p i An integer for internal use
198 #define workset_foreach(ws, v, i) for(i=0; \
199 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
202 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
203 #define workset_get_time(ws, i) (ws)->vals[i].time
204 #define workset_set_length(ws, length) (ws)->len = length
205 #define workset_get_length(ws) ((ws)->len)
206 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
207 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
209 typedef struct _block_info_t {
210 workset_t *ws_start, *ws_end;
215 static INLINE void *new_block_info(struct obstack *ob) {
216 block_info_t *res = obstack_alloc(ob, sizeof(*res));
217 res->ws_start = NULL;
224 #define get_block_info(block) ((block_info_t *)get_irn_link(block))
225 #define set_block_info(block, info) set_irn_link(block, info)
228 * @return The distance to the next use or 0 if irn has dont_spill flag set
230 static INLINE unsigned get_distance(belady_env_t *env, ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
233 int flags = arch_irn_get_flags(env->arch, def);
235 assert(! (flags & arch_irn_flags_ignore));
237 use = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
238 if(USES_IS_INFINITE(use.time))
239 return USES_INFINITY;
241 /* We have to keep nonspillable nodes in the workingset */
242 if(flags & arch_irn_flags_dont_spill)
250 * Fix to remove dead nodes (especially don't spill nodes) from workset.
252 static void fix_dead_values(workset_t *ws, ir_node *irn) {
255 ir_node *block = get_nodes_block(irn);
257 DBG((dbg, DBG_DECIDE, "fixing dead values at %+F:\n", irn));
259 workset_foreach(ws, node, idx) {
260 const ir_edge_t *edge;
263 /* skip already fixed nodes */
264 if (workset_get_time(ws, idx) == INT_MAX)
267 /* check all users */
268 foreach_out_edge(node, edge) {
269 ir_node *user = get_edge_src_irn(edge);
271 if ((get_nodes_block(user) != block) || /* user is in a different block */
272 (sched_is_scheduled(user) && sched_comes_after(irn, user)) || /* user is scheduled after irn */
273 user == irn) /* irn is the user */
274 { /* => don't fix distance */
280 /* all users scheduled prior to current irn in in same block as irn -> fix */
282 workset_set_time(ws, idx, INT_MAX);
283 DBG((dbg, DBG_DECIDE, "\tfixing time for %+F to INT_MAX\n", node));
291 * Performs the actions necessary to grant the request that:
292 * - new_vals can be held in registers
293 * - as few as possible other values are disposed
294 * - the worst values get disposed
296 * @p is_usage indicates that the values in new_vals are used (not defined)
297 * In this case reloads must be performed
299 static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) {
301 int i, len, max_allowed, demand, iter;
303 workset_t *ws = env->ws;
304 ir_node **to_insert = alloca(env->n_regs * sizeof(*to_insert));
307 1. Identify the number of needed slots and the values to reload
310 workset_foreach(new_vals, val, iter) {
311 /* mark value as used */
313 pset_insert_ptr(env->used, val);
315 if (! workset_contains(ws, val)) {
316 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
317 to_insert[demand++] = val;
319 DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr));
320 be_add_reload(env->senv, val, env->instr, env->cls);
324 assert(is_usage || "Defined value already in workset?!?");
325 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
328 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
331 2. Make room for at least 'demand' slots
333 len = workset_get_length(ws);
334 max_allowed = env->n_regs - demand;
336 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
338 /* Only make more free room if we do not have enough */
339 if (len > max_allowed) {
340 /* get current next-use distance */
341 for (i = 0; i < ws->len; ++i) {
342 unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage);
343 workset_set_time(ws, i, dist);
348 FIX for don't spill nodes:
349 Problem is that get_distance always returns 0 for those nodes even if they are not
350 needed anymore (all their usages have already been visited).
351 Even if we change this behavior, get_distance doesn't distinguish between not
352 used anymore (dead) and live out of block.
353 Solution: Set distances of all nodes having all their usages in schedule prior to
354 current instruction to MAX_INT.
356 fix_dead_values(ws, env->instr);
359 /* sort entries by increasing nextuse-distance*/
363 Logic for not needed live-ins: If a value is disposed
364 before its first usage, remove it from start workset
365 We don't do this for phis though
367 for (i = max_allowed; i < ws->len; ++i) {
368 ir_node *irn = ws->vals[i].irn;
370 DBG((dbg, DBG_DECIDE, " disposing %+F (%u)\n", irn, workset_get_time(ws, i)));
375 if (! pset_find_ptr(env->used, irn)) {
376 ir_node *curr_bb = get_nodes_block(env->instr);
377 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
378 workset_remove(ws_start, irn);
380 DBG((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", irn));
384 /* kill the last 'demand' entries in the array */
385 workset_set_length(ws, max_allowed);
389 3. Insert the new values into the workset
391 for (i = 0; i < demand; ++i)
392 workset_insert(env, env->ws, to_insert[i]);
395 static void belady(ir_node *block, void *env);
397 static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first, ir_node *node, ir_node *block, ir_loop *loop) {
398 be_next_use_t next_use;
400 loc.time = USES_INFINITY;
402 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) {
403 loc.time = USES_INFINITY;
409 /* We have to keep nonspillable nodes in the workingset */
410 if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) {
412 DBG((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time));
416 next_use = be_get_next_use(env->uses, first, 0, node, 0);
417 if(USES_IS_INFINITE(next_use.time)) {
418 // the nodes marked as live in shouldn't be dead, so it must be a phi
419 loc.time = USES_INFINITY;
420 DBG((dbg, DBG_START, " %+F not taken (dead)\n", node));
421 assert(is_Phi(node));
425 loc.time = next_use.time;
427 if(next_use.outermost_loop >= get_loop_depth(loop)) {
428 DBG((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
430 // ARR_APP1(loc_t, starters, loc);
432 loc.time = USES_INFINITY;
433 DBG((dbg, DBG_START, " %+F not taken (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
439 * Computes set of live-ins for each block with multiple predecessors
440 * and notifies spill algorithm which phis need to be spilled
442 static void compute_live_ins(ir_node *block, void *data) {
443 belady_env_t *env = data;
444 block_info_t *block_info;
445 ir_node *first, *irn;
446 loc_t loc, *starters;
447 int i, len, ws_count;
448 ir_loop *loop = get_irn_loop(block);
449 be_lv_t *lv = env->cenv->birg->lv;
451 if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
454 block_info = new_block_info(&env->ob);
455 set_block_info(block, block_info);
457 /* Collect all values living at start of block */
458 starters = NEW_ARR_F(loc_t, 0);
460 /* rebuild schedule time information, because it seems to be broken */
461 // Matze: is this still true?
462 //sched_renumber(block);
464 DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
465 first = sched_first(block);
467 sched_foreach(block, irn) {
471 loc = to_take_or_not_to_take(env, first, irn, block, loop);
473 if(!USES_IS_INFINITE(loc.time)) {
474 ARR_APP1(loc_t, starters, loc);
476 be_spill_phi(env->senv, irn);
480 be_lv_foreach(lv, block, be_lv_state_in, i) {
481 ir_node *node = be_lv_get_irn(lv, block, i);
483 loc = to_take_or_not_to_take(env, first, node, block, loop);
485 if(!USES_IS_INFINITE(loc.time)) {
486 ARR_APP1(loc_t, starters, loc);
490 // Sort start values by first use
491 qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
493 /* Copy the best ones from starters to start workset */
494 ws_count = MIN(ARR_LEN(starters), env->n_regs);
495 block_info->ws_start = new_workset(env, &env->ob);
496 workset_bulk_fill(block_info->ws_start, ws_count, starters);
498 /* The phis of this block which are not in the start set have to be spilled later. */
499 len = ARR_LEN(starters);
500 for (i = ws_count; i < len; ++i) {
501 irn = starters[i].irn;
502 if (!is_Phi(irn) || get_nodes_block(irn) != block)
505 be_spill_phi(env->senv, irn);
512 * Collects all values live-in at block @p block and all phi results in this block.
513 * Then it adds the best values (at most n_regs) to the blocks start_workset.
514 * The phis among the remaining values get spilled: Introduce psudo-copies of
515 * their args to break interference and make it possible to spill them to the
518 static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) {
520 block_info_t *res, *pred_info;
522 /* Have we seen this block before? */
523 res = get_block_info(block);
527 /* Create the block info for this block. */
528 res = new_block_info(&env->ob);
529 set_block_info(block, res);
531 /* Use endset of predecessor block as startset */
532 assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block)));
533 pred_block = get_Block_cfgpred_block(block, 0);
534 pred_info = get_block_info(pred_block);
536 /* if pred block has not been processed yet, do it now */
537 if (pred_info == NULL || pred_info->processed == 0) {
538 belady(pred_block, env);
539 pred_info = get_block_info(pred_block);
542 /* now we have an end_set of pred */
543 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
544 res->ws_start = workset_clone(env, &env->ob, pred_info->ws_end);
551 * For the given block @p block, decide for each values
552 * whether it is used from a register or is reloaded
555 static void belady(ir_node *block, void *data) {
556 belady_env_t *env = data;
560 block_info_t *block_info;
562 /* make sure we have blockinfo (with startset) */
563 block_info = get_block_info(block);
564 if (block_info == NULL)
565 block_info = compute_block_start_info(env, block);
567 /* Don't do a block twice */
568 if(block_info->processed)
571 /* get the starting workset for this block */
572 DBG((dbg, DBG_DECIDE, "\n"));
573 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
575 workset_copy(env, env->ws, block_info->ws_start);
576 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
577 workset_foreach(env->ws, irn, iter)
578 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(env->ws, iter)));
580 /* process the block from start to end */
581 DBG((dbg, DBG_WSETS, "Processing...\n"));
582 env->used = pset_new_ptr_default();
584 new_vals = new_workset(env, &env->ob);
585 sched_foreach(block, irn) {
587 assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
589 /* projs are handled with the tuple value.
590 * Phis are no real instr (see insert_starters())
591 * instr_nr does not increase */
592 if (is_Proj(irn) || is_Phi(irn)) {
593 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
596 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
598 /* set instruction in the workset */
601 /* allocate all values _used_ by this instruction */
602 workset_clear(new_vals);
603 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
604 workset_insert(env, new_vals, get_irn_n(irn, i));
606 displace(env, new_vals, 1);
608 /* allocate all values _defined_ by this instruction */
609 workset_clear(new_vals);
610 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
612 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
613 workset_insert(env, new_vals, proj);
615 workset_insert(env, new_vals, irn);
617 displace(env, new_vals, 0);
623 /* Remember end-workset for this block */
624 block_info->ws_end = workset_clone(env, &env->ob, env->ws);
625 block_info->processed = 1;
626 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
627 workset_foreach(block_info->ws_end, irn, iter)
628 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(block_info->ws_end, iter)));
632 * 'decide' is block-local and makes assumptions
633 * about the set of live-ins. Thus we must adapt the
634 * live-outs to the live-ins at each block-border.
636 static void fix_block_borders(ir_node *block, void *data) {
637 belady_env_t *env = data;
639 ir_graph *irg = get_irn_irg(block);
640 ir_node *startblock = get_irg_start_block(irg);
641 int i, max, iter, iter2;
643 if(block == startblock)
646 DBG((dbg, DBG_FIX, "\n"));
647 DBG((dbg, DBG_FIX, "Fixing %+F\n", block));
649 wsb = get_block_info(block)->ws_start;
651 /* process all pred blocks */
652 for (i=0, max=get_irn_arity(block); i<max; ++i) {
653 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(block, i);
654 workset_t *wsp = get_block_info(pred)->ws_end;
656 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
658 workset_foreach(wsb, irnb, iter) {
659 /* if irnb is a phi of the current block we reload
660 * the corresponding argument, else irnb itself */
661 if(is_Phi(irnb) && block == get_nodes_block(irnb)) {
662 irnb = get_irn_n(irnb, i);
664 // we might have unknowns as argument for the phi
665 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
669 /* Unknowns are available everywhere */
670 if(get_irn_opcode(irnb) == iro_Unknown)
673 /* check if irnb is in a register at end of pred */
674 workset_foreach(wsp, irnp, iter2) {
679 /* irnb is not in memory at the end of pred, so we have to reload it */
680 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
681 DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", irnb, block, i));
682 be_add_reload_on_edge(env->senv, irnb, block, i, env->cls);
685 /*epsilon statement :)*/;
690 void be_spill_belady(const be_chordal_env_t *chordal_env) {
691 be_spill_belady_spill_env(chordal_env, NULL);
694 void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
697 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
698 //firm_dbg_set_mask(dbg, DBG_SPILL);
700 /* init belady env */
701 obstack_init(&env.ob);
702 env.cenv = chordal_env;
703 env.arch = chordal_env->birg->main_env->arch_env;
704 env.cls = chordal_env->cls;
705 env.n_regs = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, chordal_env->cls, NULL);
706 env.ws = new_workset(&env, &env.ob);
707 env.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->lv);
708 if(spill_env == NULL) {
709 env.senv = be_new_spill_env(chordal_env);
711 env.senv = spill_env;
713 DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
715 be_clear_links(chordal_env->irg);
716 /* Decide which phi nodes will be spilled and place copies for them into the graph */
717 irg_block_walk_graph(chordal_env->irg, compute_live_ins, NULL, &env);
718 /* Fix high register pressure with belady algorithm */
719 irg_block_walk_graph(chordal_env->irg, NULL, belady, &env);
720 /* belady was block-local, fix the global flow by adding reloads on the edges */
721 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &env);
722 /* Insert spill/reload nodes into the graph and fix usages */
723 be_insert_spills_reloads(env.senv);
726 if(spill_env == NULL)
727 be_delete_spill_env(env.senv);
728 be_end_uses(env.uses);
729 obstack_free(&env.ob, NULL);