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"
42 #include "bespilloptions.h"
51 #define DBG_WORKSET 128
52 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
55 * An association between a node and a point in time.
57 typedef struct _loc_t {
58 ir_node *irn; /**< A node. */
59 unsigned time; /**< A use time (see beuses.h). */
62 typedef struct _workset_t {
63 int len; /**< current length */
64 loc_t vals[0]; /**< inlined array of the values/distances in this working set */
67 typedef struct _belady_env_t {
69 const arch_env_t *arch;
70 const arch_register_class_t *cls;
72 int n_regs; /** number of regs in this reg-class */
74 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
75 be_uses_t *uses; /**< env for the next-use magic */
76 ir_node *instr; /**< current instruction */
77 unsigned instr_nr; /**< current instruction number (relative to block start) */
80 spill_env_t *senv; /**< see bespill.h */
83 static int loc_compare(const void *a, const void *b)
87 return p->time - q->time;
90 void workset_print(const workset_t *w)
94 for(i = 0; i < w->len; ++i) {
95 ir_fprintf(stderr, "%+F %d\n", w->vals[i].irn, w->vals[i].time);
100 * Alloc a new workset on obstack @p ob with maximum size @p max
102 static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) {
104 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
105 res = obstack_alloc(ob, size);
106 memset(res, 0, size);
111 * Alloc a new instance on obstack and make it equal to @param ws
113 static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) {
115 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
116 res = obstack_alloc(ob, size);
117 memcpy(res, ws, size);
122 * Do NOT alloc anything. Make @param tgt equal to @param src.
123 * returns @param tgt for convenience
125 static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src) {
126 size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]);
127 memcpy(tgt, src, size);
132 * Overwrites the current content array of @param ws with the
133 * @param count locations given at memory @param locs.
134 * Set the length of @param ws to count.
136 static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) {
137 workset->len = count;
138 memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
142 * Inserts the value @p val into the workset, iff it is not
143 * already contained. The workset must not be full.
145 static INLINE void workset_insert(belady_env_t *env, workset_t *ws, ir_node *val) {
147 /* check for current regclass */
148 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
149 DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
153 /* check if val is already contained */
154 for(i=0; i<ws->len; ++i)
155 if (ws->vals[i].irn == val)
159 assert(ws->len < env->n_regs && "Workset already full!");
160 ws->vals[ws->len++].irn = val;
164 * Removes all entries from this workset
166 static INLINE void workset_clear(workset_t *ws) {
171 * Removes the value @p val from the workset if present.
173 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
175 for(i=0; i<ws->len; ++i) {
176 if (ws->vals[i].irn == val) {
177 ws->vals[i] = ws->vals[--ws->len];
183 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
185 for(i=0; i<ws->len; ++i) {
186 if (ws->vals[i].irn == val)
194 * Iterates over all values in the working set.
195 * @p ws The workset to iterate
196 * @p v A variable to put the current value in
197 * @p i An integer for internal use
199 #define workset_foreach(ws, v, i) for(i=0; \
200 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
203 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
204 #define workset_get_time(ws, i) (ws)->vals[i].time
205 #define workset_set_length(ws, length) (ws)->len = length
206 #define workset_get_length(ws) ((ws)->len)
207 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
208 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
210 typedef struct _block_info_t {
211 workset_t *ws_start, *ws_end;
216 static INLINE void *new_block_info(struct obstack *ob) {
217 block_info_t *res = obstack_alloc(ob, sizeof(*res));
218 res->ws_start = NULL;
225 #define get_block_info(block) ((block_info_t *)get_irn_link(block))
226 #define set_block_info(block, info) set_irn_link(block, info)
229 * @return The distance to the next use or 0 if irn has dont_spill flag set
231 static INLINE unsigned get_distance(belady_env_t *env, ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
234 int flags = arch_irn_get_flags(env->arch, def);
236 assert(! (flags & arch_irn_flags_ignore));
238 use = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
239 if(USES_IS_INFINITE(use.time))
240 return USES_INFINITY;
242 /* We have to keep nonspillable nodes in the workingset */
243 if(flags & arch_irn_flags_dont_spill)
251 * Fix to remove dead nodes (especially don't spill nodes) from workset.
253 static void fix_dead_values(workset_t *ws, ir_node *irn) {
256 ir_node *block = get_nodes_block(irn);
258 DBG((dbg, DBG_DECIDE, "fixing dead values at %+F:\n", irn));
260 workset_foreach(ws, node, idx) {
261 const ir_edge_t *edge;
264 /* skip already fixed nodes */
265 if (workset_get_time(ws, idx) == INT_MAX)
268 /* check all users */
269 foreach_out_edge(node, edge) {
270 ir_node *user = get_edge_src_irn(edge);
272 if ((get_nodes_block(user) != block) || /* user is in a different block */
273 (sched_is_scheduled(user) && sched_comes_after(irn, user)) || /* user is scheduled after irn */
274 user == irn) /* irn is the user */
275 { /* => don't fix distance */
281 /* all users scheduled prior to current irn in in same block as irn -> fix */
283 workset_set_time(ws, idx, INT_MAX);
284 DBG((dbg, DBG_DECIDE, "\tfixing time for %+F to INT_MAX\n", node));
292 * Performs the actions necessary to grant the request that:
293 * - new_vals can be held in registers
294 * - as few as possible other values are disposed
295 * - the worst values get disposed
297 * @p is_usage indicates that the values in new_vals are used (not defined)
298 * In this case reloads must be performed
300 static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) {
302 int i, len, max_allowed, demand, iter;
304 workset_t *ws = env->ws;
305 ir_node **to_insert = alloca(env->n_regs * sizeof(*to_insert));
308 1. Identify the number of needed slots and the values to reload
311 workset_foreach(new_vals, val, iter) {
312 /* mark value as used */
314 pset_insert_ptr(env->used, val);
316 if (! workset_contains(ws, val)) {
317 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
318 to_insert[demand++] = val;
320 DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr));
321 be_add_reload(env->senv, val, env->instr, env->cls, 1);
325 assert(is_usage || "Defined value already in workset?!?");
326 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
329 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
332 2. Make room for at least 'demand' slots
334 len = workset_get_length(ws);
335 max_allowed = env->n_regs - demand;
337 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
339 /* Only make more free room if we do not have enough */
340 if (len > max_allowed) {
341 /* get current next-use distance */
342 for (i = 0; i < ws->len; ++i) {
343 unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage);
344 workset_set_time(ws, i, dist);
349 FIX for don't spill nodes:
350 Problem is that get_distance always returns 0 for those nodes even if they are not
351 needed anymore (all their usages have already been visited).
352 Even if we change this behavior, get_distance doesn't distinguish between not
353 used anymore (dead) and live out of block.
354 Solution: Set distances of all nodes having all their usages in schedule prior to
355 current instruction to MAX_INT.
357 fix_dead_values(ws, env->instr);
360 /* sort entries by increasing nextuse-distance*/
364 Logic for not needed live-ins: If a value is disposed
365 before its first usage, remove it from start workset
366 We don't do this for phis though
368 for (i = max_allowed; i < ws->len; ++i) {
369 ir_node *irn = ws->vals[i].irn;
371 DBG((dbg, DBG_DECIDE, " disposing %+F (%u)\n", irn, workset_get_time(ws, i)));
376 if (! pset_find_ptr(env->used, irn)) {
377 ir_node *curr_bb = get_nodes_block(env->instr);
378 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
379 workset_remove(ws_start, irn);
381 DBG((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", irn));
385 /* kill the last 'demand' entries in the array */
386 workset_set_length(ws, max_allowed);
390 3. Insert the new values into the workset
392 for (i = 0; i < demand; ++i)
393 workset_insert(env, env->ws, to_insert[i]);
396 static void belady(ir_node *block, void *env);
398 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) {
399 be_next_use_t next_use;
401 loc.time = USES_INFINITY;
403 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) {
404 loc.time = USES_INFINITY;
410 /* We have to keep nonspillable nodes in the workingset */
411 if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) {
413 DBG((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time));
417 next_use = be_get_next_use(env->uses, first, 0, node, 0);
418 if(USES_IS_INFINITE(next_use.time)) {
419 // the nodes marked as live in shouldn't be dead, so it must be a phi
420 loc.time = USES_INFINITY;
421 DBG((dbg, DBG_START, " %+F not taken (dead)\n", node));
422 assert(is_Phi(node));
426 loc.time = next_use.time;
428 if(next_use.outermost_loop >= get_loop_depth(loop)) {
429 DBG((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
431 // ARR_APP1(loc_t, starters, loc);
433 loc.time = USES_INFINITY;
434 DBG((dbg, DBG_START, " %+F not taken (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
440 * Computes set of live-ins for each block with multiple predecessors
441 * and notifies spill algorithm which phis need to be spilled
443 static void compute_live_ins(ir_node *block, void *data) {
444 belady_env_t *env = data;
445 block_info_t *block_info;
446 ir_node *first, *irn;
447 loc_t loc, *starters;
448 int i, len, ws_count;
449 ir_loop *loop = get_irn_loop(block);
450 const be_lv_t *lv = env->lv;
452 if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
455 block_info = new_block_info(&env->ob);
456 set_block_info(block, block_info);
458 /* Collect all values living at start of block */
459 starters = NEW_ARR_F(loc_t, 0);
461 /* rebuild schedule time information, because it seems to be broken */
462 // Matze: is this still true?
463 //sched_renumber(block);
465 DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
466 first = sched_first(block);
468 sched_foreach(block, irn) {
472 loc = to_take_or_not_to_take(env, first, irn, block, loop);
474 if(!USES_IS_INFINITE(loc.time)) {
475 ARR_APP1(loc_t, starters, loc);
477 be_spill_phi(env->senv, irn);
481 be_lv_foreach(lv, block, be_lv_state_in, i) {
482 ir_node *node = be_lv_get_irn(lv, block, i);
484 loc = to_take_or_not_to_take(env, first, node, block, loop);
486 if(!USES_IS_INFINITE(loc.time)) {
487 ARR_APP1(loc_t, starters, loc);
491 // Sort start values by first use
492 qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
494 /* Copy the best ones from starters to start workset */
495 ws_count = MIN(ARR_LEN(starters), env->n_regs);
496 block_info->ws_start = new_workset(env, &env->ob);
497 workset_bulk_fill(block_info->ws_start, ws_count, starters);
499 /* The phis of this block which are not in the start set have to be spilled later. */
500 len = ARR_LEN(starters);
501 for (i = ws_count; i < len; ++i) {
502 irn = starters[i].irn;
503 if (!is_Phi(irn) || get_nodes_block(irn) != block)
506 be_spill_phi(env->senv, irn);
513 * Collects all values live-in at block @p block and all phi results in this block.
514 * Then it adds the best values (at most n_regs) to the blocks start_workset.
515 * The phis among the remaining values get spilled: Introduce psudo-copies of
516 * their args to break interference and make it possible to spill them to the
519 static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) {
521 block_info_t *res, *pred_info;
523 /* Have we seen this block before? */
524 res = get_block_info(block);
528 /* Create the block info for this block. */
529 res = new_block_info(&env->ob);
530 set_block_info(block, res);
532 /* Use endset of predecessor block as startset */
533 assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block)));
534 pred_block = get_Block_cfgpred_block(block, 0);
535 pred_info = get_block_info(pred_block);
537 /* if pred block has not been processed yet, do it now */
538 if (pred_info == NULL || pred_info->processed == 0) {
539 belady(pred_block, env);
540 pred_info = get_block_info(pred_block);
543 /* now we have an end_set of pred */
544 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
545 res->ws_start = workset_clone(env, &env->ob, pred_info->ws_end);
552 * For the given block @p block, decide for each values
553 * whether it is used from a register or is reloaded
556 static void belady(ir_node *block, void *data) {
557 belady_env_t *env = data;
561 block_info_t *block_info;
563 /* make sure we have blockinfo (with startset) */
564 block_info = get_block_info(block);
565 if (block_info == NULL)
566 block_info = compute_block_start_info(env, block);
568 /* Don't do a block twice */
569 if(block_info->processed)
572 /* get the starting workset for this block */
573 DBG((dbg, DBG_DECIDE, "\n"));
574 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
576 workset_copy(env, env->ws, block_info->ws_start);
577 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
578 workset_foreach(env->ws, irn, iter)
579 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(env->ws, iter)));
581 /* process the block from start to end */
582 DBG((dbg, DBG_WSETS, "Processing...\n"));
583 env->used = pset_new_ptr_default();
585 new_vals = new_workset(env, &env->ob);
586 sched_foreach(block, irn) {
588 assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
590 /* projs are handled with the tuple value.
591 * Phis are no real instr (see insert_starters())
592 * instr_nr does not increase */
593 if (is_Proj(irn) || is_Phi(irn)) {
594 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
597 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
599 /* set instruction in the workset */
602 /* allocate all values _used_ by this instruction */
603 workset_clear(new_vals);
604 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
605 workset_insert(env, new_vals, get_irn_n(irn, i));
607 displace(env, new_vals, 1);
609 /* allocate all values _defined_ by this instruction */
610 workset_clear(new_vals);
611 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
613 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
614 workset_insert(env, new_vals, proj);
616 workset_insert(env, new_vals, irn);
618 displace(env, new_vals, 0);
624 /* Remember end-workset for this block */
625 block_info->ws_end = workset_clone(env, &env->ob, env->ws);
626 block_info->processed = 1;
627 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
628 workset_foreach(block_info->ws_end, irn, iter)
629 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(block_info->ws_end, iter)));
633 * 'decide' is block-local and makes assumptions
634 * about the set of live-ins. Thus we must adapt the
635 * live-outs to the live-ins at each block-border.
637 static void fix_block_borders(ir_node *block, void *data) {
638 belady_env_t *env = data;
640 ir_graph *irg = get_irn_irg(block);
641 ir_node *startblock = get_irg_start_block(irg);
642 int i, max, iter, iter2;
644 if(block == startblock)
647 DBG((dbg, DBG_FIX, "\n"));
648 DBG((dbg, DBG_FIX, "Fixing %+F\n", block));
650 wsb = get_block_info(block)->ws_start;
652 /* process all pred blocks */
653 for (i=0, max=get_irn_arity(block); i<max; ++i) {
654 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(block, i);
655 workset_t *wsp = get_block_info(pred)->ws_end;
657 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
659 workset_foreach(wsb, irnb, iter) {
660 /* if irnb is a phi of the current block we reload
661 * the corresponding argument, else irnb itself */
662 if(is_Phi(irnb) && block == get_nodes_block(irnb)) {
663 irnb = get_irn_n(irnb, i);
665 // we might have unknowns as argument for the phi
666 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
670 /* Unknowns are available everywhere */
671 if(get_irn_opcode(irnb) == iro_Unknown)
674 /* check if irnb is in a register at end of pred */
675 workset_foreach(wsp, irnp, iter2) {
680 /* irnb is not in memory at the end of pred, so we have to reload it */
681 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
682 DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", irnb, block, i));
683 be_add_reload_on_edge(env->senv, irnb, block, i, env->cls, 1);
686 /*epsilon statement :)*/;
691 void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) {
692 be_spill_belady_spill_env(birg, cls, NULL);
695 void be_spill_belady_spill_env(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) {
697 ir_graph *irg = be_get_birg_irg(birg);
699 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
700 //firm_dbg_set_mask(dbg, DBG_SPILL);
702 be_assure_liveness(birg);
703 /* construct control flow loop tree */
704 if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
705 construct_cf_backedges(irg);
708 /* init belady env */
709 obstack_init(&env.ob);
710 env.arch = birg->main_env->arch_env;
712 env.lv = be_get_birg_liveness(birg);
713 env.n_regs = env.cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
714 env.ws = new_workset(&env, &env.ob);
715 env.uses = be_begin_uses(irg, env.lv);
716 if(spill_env == NULL) {
717 env.senv = be_new_spill_env(birg);
719 env.senv = spill_env;
721 DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
724 /* Decide which phi nodes will be spilled and place copies for them into the graph */
725 irg_block_walk_graph(irg, compute_live_ins, NULL, &env);
726 /* Fix high register pressure with belady algorithm */
727 irg_block_walk_graph(irg, NULL, belady, &env);
728 /* belady was block-local, fix the global flow by adding reloads on the edges */
729 irg_block_walk_graph(irg, fix_block_borders, NULL, &env);
730 /* Insert spill/reload nodes into the graph and fix usages */
731 be_insert_spills_reloads(env.senv);
734 if(spill_env == NULL)
735 be_delete_spill_env(env.senv);
736 be_end_uses(env.uses);
737 obstack_free(&env.ob, NULL);
740 void be_init_spillbelady(void)
742 static be_spiller_t belady_spiller = {
746 be_register_spiller("belady", &belady_spiller);
749 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);