2 * Author: Daniel Grund, Matthias Braun
4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
15 #include "irprintf_t.h"
21 #include "iredges_t.h"
28 #include "bespillbelady.h"
30 #include "besched_t.h"
34 #include "bechordal_t.h"
35 #include "bespilloptions.h"
36 #include "beloopana.h"
45 #define DBG_WORKSET 128
46 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
49 * An association between a node and a point in time.
51 typedef struct _loc_t {
52 ir_node *irn; /**< A node. */
53 unsigned time; /**< A use time (see beuses.h). */
56 typedef struct _workset_t {
57 int len; /**< current length */
58 loc_t vals[0]; /**< inlined array of the values/distances in this working set */
61 typedef struct _belady_env_t {
63 const arch_env_t *arch;
64 const arch_register_class_t *cls;
66 be_loopana_t *loop_ana;
67 int n_regs; /** number of regs in this reg-class */
69 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
70 be_uses_t *uses; /**< env for the next-use magic */
71 ir_node *instr; /**< current instruction */
72 unsigned instr_nr; /**< current instruction number (relative to block start) */
75 spill_env_t *senv; /**< see bespill.h */
78 static int loc_compare(const void *a, const void *b)
82 return p->time - q->time;
85 static INLINE void workset_print(const workset_t *w)
89 for(i = 0; i < w->len; ++i) {
90 ir_fprintf(stderr, "%+F %d\n", w->vals[i].irn, w->vals[i].time);
95 * Alloc a new workset on obstack @p ob with maximum size @p max
97 static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) {
99 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
100 res = obstack_alloc(ob, size);
101 memset(res, 0, size);
106 * Alloc a new instance on obstack and make it equal to @param ws
108 static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) {
110 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
111 res = obstack_alloc(ob, size);
112 memcpy(res, ws, size);
117 * Do NOT alloc anything. Make @param tgt equal to @param src.
118 * returns @param tgt for convenience
120 static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src) {
121 size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]);
122 memcpy(tgt, src, size);
127 * Overwrites the current content array of @param ws with the
128 * @param count locations given at memory @param locs.
129 * Set the length of @param ws to count.
131 static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) {
132 workset->len = count;
133 memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0]));
137 * Inserts the value @p val into the workset, iff it is not
138 * already contained. The workset must not be full.
140 static INLINE void workset_insert(belady_env_t *env, workset_t *ws, ir_node *val) {
142 /* check for current regclass */
143 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
144 DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
148 /* check if val is already contained */
149 for(i=0; i<ws->len; ++i)
150 if (ws->vals[i].irn == val)
154 assert(ws->len < env->n_regs && "Workset already full!");
155 ws->vals[ws->len++].irn = val;
159 * Removes all entries from this workset
161 static INLINE void workset_clear(workset_t *ws) {
166 * Removes the value @p val from the workset if present.
168 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
170 for(i=0; i<ws->len; ++i) {
171 if (ws->vals[i].irn == val) {
172 ws->vals[i] = ws->vals[--ws->len];
178 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
180 for(i=0; i<ws->len; ++i) {
181 if (ws->vals[i].irn == val)
189 * Iterates over all values in the working set.
190 * @p ws The workset to iterate
191 * @p v A variable to put the current value in
192 * @p i An integer for internal use
194 #define workset_foreach(ws, v, i) for(i=0; \
195 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
198 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
199 #define workset_get_time(ws, i) (ws)->vals[i].time
200 #define workset_set_length(ws, length) (ws)->len = length
201 #define workset_get_length(ws) ((ws)->len)
202 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
203 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
205 typedef struct _block_info_t {
206 workset_t *ws_start, *ws_end;
211 static INLINE void *new_block_info(struct obstack *ob) {
212 block_info_t *res = obstack_alloc(ob, sizeof(*res));
213 res->ws_start = NULL;
220 #define get_block_info(block) ((block_info_t *)get_irn_link(block))
221 #define set_block_info(block, info) set_irn_link(block, info)
224 * @return The distance to the next use or 0 if irn has dont_spill flag set
226 static INLINE unsigned get_distance(belady_env_t *env, ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
229 int flags = arch_irn_get_flags(env->arch, def);
231 assert(! (flags & arch_irn_flags_ignore));
233 use = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
234 if(USES_IS_INFINITE(use.time))
235 return USES_INFINITY;
237 /* We have to keep nonspillable nodes in the workingset */
238 if(flags & arch_irn_flags_dont_spill)
245 * Performs the actions necessary to grant the request that:
246 * - new_vals can be held in registers
247 * - as few as possible other values are disposed
248 * - the worst values get disposed
250 * @p is_usage indicates that the values in new_vals are used (not defined)
251 * In this case reloads must be performed
253 static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) {
255 int i, len, max_allowed, demand, iter;
257 workset_t *ws = env->ws;
258 ir_node **to_insert = alloca(env->n_regs * sizeof(*to_insert));
261 1. Identify the number of needed slots and the values to reload
264 workset_foreach(new_vals, val, iter) {
265 /* mark value as used */
267 pset_insert_ptr(env->used, val);
269 if (! workset_contains(ws, val)) {
270 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
271 to_insert[demand++] = val;
273 DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr));
274 be_add_reload(env->senv, val, env->instr, env->cls, 1);
278 assert(is_usage || "Defined value already in workset?!?");
279 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
282 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
285 2. Make room for at least 'demand' slots
287 len = workset_get_length(ws);
288 max_allowed = env->n_regs - demand;
290 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
292 /* Only make more free room if we do not have enough */
293 if (len > max_allowed) {
294 /* get current next-use distance */
295 for (i = 0; i < ws->len; ++i) {
296 unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage);
297 workset_set_time(ws, i, dist);
300 /* sort entries by increasing nextuse-distance*/
304 Logic for not needed live-ins: If a value is disposed
305 before its first usage, remove it from start workset
306 We don't do this for phis though
308 for (i = max_allowed; i < ws->len; ++i) {
309 ir_node *irn = ws->vals[i].irn;
311 DBG((dbg, DBG_DECIDE, " disposing %+F (%u)\n", irn, workset_get_time(ws, i)));
316 if (! pset_find_ptr(env->used, irn)) {
317 ir_node *curr_bb = get_nodes_block(env->instr);
318 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
319 workset_remove(ws_start, irn);
321 DBG((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", irn));
325 /* kill the last 'demand' entries in the array */
326 workset_set_length(ws, max_allowed);
330 3. Insert the new values into the workset
332 for (i = 0; i < demand; ++i)
333 workset_insert(env, env->ws, to_insert[i]);
336 static void belady(ir_node *block, void *env);
338 /** Decides whether a specific node should be in the start workset or not
340 * @param env belady environment
342 * @param node the node to test
343 * @param block the block of the node
344 * @param loop the loop of the node
346 static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first,
347 ir_node *node, ir_node *block,
350 be_next_use_t next_use;
352 loc.time = USES_INFINITY;
355 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) {
356 loc.time = USES_INFINITY;
360 /* We have to keep nonspillable nodes in the workingset */
361 if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) {
363 DBG((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time));
367 next_use = be_get_next_use(env->uses, first, 0, node, 0);
368 if(USES_IS_INFINITE(next_use.time)) {
369 // the nodes marked as live in shouldn't be dead, so it must be a phi
370 assert(is_Phi(node));
371 loc.time = USES_INFINITY;
372 DBG((dbg, DBG_START, " %+F not taken (dead)\n", node));
374 be_spill_phi(env->senv, node);
379 loc.time = next_use.time;
381 if(next_use.outermost_loop >= get_loop_depth(loop)) {
382 DBG((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
384 loc.time = USES_PENDING;
385 DBG((dbg, DBG_START, " %+F delayed (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
391 * Computes set of live-ins for each block with multiple predecessors
392 * and notifies spill algorithm which phis need to be spilled
394 static void compute_live_ins(ir_node *block, void *data) {
395 belady_env_t *env = data;
396 ir_loop *loop = get_irn_loop(block);
397 const be_lv_t *lv = env->lv;
398 block_info_t *block_info;
399 ir_node *first, *irn;
400 loc_t loc, *starters, *delayed;
401 int i, len, ws_count;
402 int free_slots, free_pressure_slots;
405 if (get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
408 block_info = new_block_info(&env->ob);
409 set_block_info(block, block_info);
411 /* Collect all values living at start of block */
412 starters = NEW_ARR_F(loc_t, 0);
413 delayed = NEW_ARR_F(loc_t, 0);
415 DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
416 first = sched_first(block);
418 /* check all Phis first */
419 sched_foreach(block, irn) {
423 loc = to_take_or_not_to_take(env, first, irn, block, loop);
425 if (! USES_IS_INFINITE(loc.time)) {
426 if (USES_IS_PENDING(loc.time))
427 ARR_APP1(loc_t, delayed, loc);
429 ARR_APP1(loc_t, starters, loc);
433 /* check all Live-Ins */
434 be_lv_foreach(lv, block, be_lv_state_in, i) {
435 ir_node *node = be_lv_get_irn(lv, block, i);
437 loc = to_take_or_not_to_take(env, first, node, block, loop);
439 if (! USES_IS_INFINITE(loc.time)) {
440 if (USES_IS_PENDING(loc.time))
441 ARR_APP1(loc_t, delayed, loc);
443 ARR_APP1(loc_t, starters, loc);
447 pressure = be_get_loop_pressure(env->loop_ana, env->cls, loop);
448 assert(ARR_LEN(delayed) <= (signed)pressure);
449 free_slots = env->n_regs - ARR_LEN(starters);
450 free_pressure_slots = env->n_regs - (pressure - ARR_LEN(delayed));
451 free_slots = MIN(free_slots, free_pressure_slots);
452 /* append nodes delayed due to loop structure until start set is full */
453 for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
454 DBG((dbg, DBG_START, " delayed %+F taken\n", delayed[i].irn));
455 ARR_APP1(loc_t, starters, delayed[i]);
456 delayed[i].irn = NULL;
459 /* spill all delayed phis which didn't make it into start workset */
460 for (i = ARR_LEN(delayed) - 1; i >= 0; --i) {
461 ir_node *irn = delayed[i].irn;
462 if (irn && is_Phi(irn)) {
463 DBG((dbg, DBG_START, " spilling delayed phi %+F\n", irn));
464 be_spill_phi(env->senv, irn);
469 /* Sort start values by first use */
470 qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
472 /* Copy the best ones from starters to start workset */
473 ws_count = MIN(ARR_LEN(starters), env->n_regs);
474 block_info->ws_start = new_workset(env, &env->ob);
475 workset_bulk_fill(block_info->ws_start, ws_count, starters);
477 /* The phis of this block which are not in the start set have to be spilled later. */
478 len = ARR_LEN(starters);
479 for (i = ws_count; i < len; ++i) {
480 irn = starters[i].irn;
481 if (! is_Phi(irn) || get_nodes_block(irn) != block)
484 be_spill_phi(env->senv, irn);
491 * Collects all values live-in at block @p block and all phi results in this block.
492 * Then it adds the best values (at most n_regs) to the blocks start_workset.
493 * The phis among the remaining values get spilled: Introduce psudo-copies of
494 * their args to break interference and make it possible to spill them to the
497 static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) {
499 block_info_t *res, *pred_info;
501 /* Have we seen this block before? */
502 res = get_block_info(block);
506 /* Create the block info for this block. */
507 res = new_block_info(&env->ob);
508 set_block_info(block, res);
510 /* Use endset of predecessor block as startset */
511 assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block)));
512 pred_block = get_Block_cfgpred_block(block, 0);
513 pred_info = get_block_info(pred_block);
515 /* if pred block has not been processed yet, do it now */
516 if (pred_info == NULL || pred_info->processed == 0) {
517 belady(pred_block, env);
518 pred_info = get_block_info(pred_block);
521 /* now we have an end_set of pred */
522 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
523 res->ws_start = workset_clone(env, &env->ob, pred_info->ws_end);
530 * For the given block @p block, decide for each values
531 * whether it is used from a register or is reloaded
534 static void belady(ir_node *block, void *data) {
535 belady_env_t *env = data;
539 block_info_t *block_info;
541 /* make sure we have blockinfo (with startset) */
542 block_info = get_block_info(block);
543 if (block_info == NULL)
544 block_info = compute_block_start_info(env, block);
546 /* Don't do a block twice */
547 if(block_info->processed)
550 /* get the starting workset for this block */
551 DBG((dbg, DBG_DECIDE, "\n"));
552 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
554 workset_copy(env, env->ws, block_info->ws_start);
555 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
556 workset_foreach(env->ws, irn, iter)
557 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(env->ws, iter)));
559 /* process the block from start to end */
560 DBG((dbg, DBG_WSETS, "Processing...\n"));
561 env->used = pset_new_ptr_default();
563 new_vals = new_workset(env, &env->ob);
564 sched_foreach(block, irn) {
566 assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
568 /* projs are handled with the tuple value.
569 * Phis are no real instr (see insert_starters())
570 * instr_nr does not increase */
571 if (is_Proj(irn) || is_Phi(irn)) {
572 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
575 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
577 /* set instruction in the workset */
580 /* allocate all values _used_ by this instruction */
581 workset_clear(new_vals);
582 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
583 workset_insert(env, new_vals, get_irn_n(irn, i));
585 displace(env, new_vals, 1);
587 /* allocate all values _defined_ by this instruction */
588 workset_clear(new_vals);
589 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
591 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
592 workset_insert(env, new_vals, proj);
594 workset_insert(env, new_vals, irn);
596 displace(env, new_vals, 0);
602 /* Remember end-workset for this block */
603 block_info->ws_end = workset_clone(env, &env->ob, env->ws);
604 block_info->processed = 1;
605 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
606 workset_foreach(block_info->ws_end, irn, iter)
607 DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(block_info->ws_end, iter)));
611 * 'decide' is block-local and makes assumptions
612 * about the set of live-ins. Thus we must adapt the
613 * live-outs to the live-ins at each block-border.
615 static void fix_block_borders(ir_node *block, void *data) {
616 belady_env_t *env = data;
618 ir_graph *irg = get_irn_irg(block);
619 ir_node *startblock = get_irg_start_block(irg);
620 int i, max, iter, iter2;
622 if(block == startblock)
625 DBG((dbg, DBG_FIX, "\n"));
626 DBG((dbg, DBG_FIX, "Fixing %+F\n", block));
628 wsb = get_block_info(block)->ws_start;
630 /* process all pred blocks */
631 for (i=0, max=get_irn_arity(block); i<max; ++i) {
632 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(block, i);
633 workset_t *wsp = get_block_info(pred)->ws_end;
635 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
637 workset_foreach(wsb, irnb, iter) {
638 /* if irnb is a phi of the current block we reload
639 * the corresponding argument, else irnb itself */
640 if(is_Phi(irnb) && block == get_nodes_block(irnb)) {
641 irnb = get_irn_n(irnb, i);
643 // we might have unknowns as argument for the phi
644 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
648 /* Unknowns are available everywhere */
649 if(get_irn_opcode(irnb) == iro_Unknown)
652 /* check if irnb is in a register at end of pred */
653 workset_foreach(wsp, irnp, iter2) {
658 /* irnb is not in memory at the end of pred, so we have to reload it */
659 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
660 DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", irnb, block, i));
661 be_add_reload_on_edge(env->senv, irnb, block, i, env->cls, 1);
664 /*epsilon statement :)*/;
670 * Do spilling for a register class on a graph using the belady heuristic.
671 * In the transformed graph, the register pressure never exceeds the number
672 * of available registers.
674 * @param birg The backend graph
675 * @param cls The register class to spill
677 static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) {
678 be_spill_belady_spill_env(birg, cls, NULL);
681 void be_spill_belady_spill_env(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) {
683 ir_graph *irg = be_get_birg_irg(birg);
686 /* some special classes contain only ignore regs, nothing to do then */
687 n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
691 be_invalidate_liveness(birg);
692 be_assure_liveness(birg);
693 /* construct control flow loop tree */
694 if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
695 construct_cf_backedges(irg);
698 /* init belady env */
699 obstack_init(&env.ob);
700 env.arch = birg->main_env->arch_env;
702 env.lv = be_get_birg_liveness(birg);
704 env.ws = new_workset(&env, &env.ob);
705 env.uses = be_begin_uses(irg, env.lv);
706 env.loop_ana = be_new_loop_pressure(birg);
707 if(spill_env == NULL) {
708 env.senv = be_new_spill_env(birg);
710 env.senv = spill_env;
712 DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
715 /* Decide which phi nodes will be spilled and place copies for them into the graph */
716 irg_block_walk_graph(irg, compute_live_ins, NULL, &env);
717 /* Fix high register pressure with belady algorithm */
718 irg_block_walk_graph(irg, NULL, belady, &env);
719 /* belady was block-local, fix the global flow by adding reloads on the edges */
720 irg_block_walk_graph(irg, fix_block_borders, NULL, &env);
721 /* Insert spill/reload nodes into the graph and fix usages */
722 be_insert_spills_reloads(env.senv);
725 if(spill_env == NULL)
726 be_delete_spill_env(env.senv);
727 be_end_uses(env.uses);
728 be_free_loop_pressure(env.loop_ana);
729 obstack_free(&env.ob, NULL);
732 void be_init_spillbelady(void)
734 static be_spiller_t belady_spiller = {
738 be_register_spiller("belady", &belady_spiller);
739 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
742 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);