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;)
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, "Dropped %+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(blk) ((block_info_t *)get_irn_link(blk))
225 #define set_block_info(blk, info) set_irn_link(blk, 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, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
232 int flags = arch_irn_get_flags(env->arch, def);
235 assert(! (flags & arch_irn_flags_ignore));
237 /* We have to keep nonspillable nodes in the workingset */
238 if(flags & arch_irn_flags_dont_spill)
241 dist = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
243 if(USES_IS_INFINITE(dist))
244 dist = USES_INFINITY;
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));
290 * Performs the actions necessary to grant the request that:
291 * - new_vals can be held in registers
292 * - as few as possible other values are disposed
293 * - the worst values get disposed
295 * @p is_usage indicates that the values in new_vals are used (not defined)
296 * In this case reloads must be performed
298 static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) {
300 int i, len, max_allowed, demand, iter;
302 workset_t *ws = env->ws;
303 ir_node **to_insert = alloca(env->n_regs * sizeof(*to_insert));
306 1. Identify the number of needed slots and the values to reload
309 workset_foreach(new_vals, val, iter) {
310 /* mark value as used */
312 pset_insert_ptr(env->used, val);
314 if (! workset_contains(ws, val)) {
315 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
316 to_insert[demand++] = val;
318 be_add_reload(env->senv, val, env->instr);
321 assert(is_usage || "Defined value already in workset?!?");
322 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
325 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
328 2. Make room for at least 'demand' slots
330 len = workset_get_length(ws);
331 max_allowed = env->n_regs - demand;
333 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
335 /* Only make more free room if we do not have enough */
336 if (len > max_allowed) {
337 /* get current next-use distance */
338 for (i = 0; i < ws->len; ++i) {
339 unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage);
340 workset_set_time(ws, i, dist);
344 FIX for don't spill nodes:
345 Problem is that get_distance always returns 0 for those nodes even if they are not
346 needed anymore (all their usages have already been visited).
347 Even if we change this behavior, get_distance doesn't distinguish between not
348 used anymore (dead) and live out of block.
349 Solution: Set distances of all nodes having all their usages in schedule prior to
350 current instruction to MAX_INT.
352 fix_dead_values(ws, env->instr);
354 /* sort entries by increasing nextuse-distance*/
358 Logic for not needed live-ins: If a value is disposed
359 before its first usage, remove it from start workset
360 We don't do this for phis though
362 for (i = max_allowed; i < ws->len; ++i) {
363 ir_node *irn = ws->vals[i].irn;
368 if (! pset_find_ptr(env->used, irn)) {
369 ir_node *curr_bb = get_nodes_block(env->instr);
370 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
371 workset_remove(ws_start, irn);
373 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
376 DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn));
380 /* kill the last 'demand' entries in the array */
381 workset_set_length(ws, max_allowed);
385 3. Insert the new values into the workset
387 for (i = 0; i < demand; ++i)
388 workset_insert(env, env->ws, to_insert[i]);
391 static void belady(ir_node *blk, void *env);
394 * Computes set of live-ins for each block with multiple predecessors
395 * and notifies spill algorithm which phis need to be spilled
397 static void spill_phi_walker(ir_node *block, void *data) {
398 belady_env_t *env = data;
399 block_info_t *block_info;
400 ir_node *first, *irn;
401 loc_t loc, *starters;
402 int i, len, ws_count;
404 if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
407 block_info = new_block_info(&env->ob);
408 set_block_info(block, block_info);
410 /* Collect all values living at start of block */
411 starters = NEW_ARR_F(loc_t, 0);
413 /* rebuild schedule time information, because it seems to be broken */
414 sched_renumber(block);
416 DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
417 first = sched_first(block);
418 sched_foreach(block, irn) {
421 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn))
425 loc.time = get_distance(env, first, 0, irn, 0);
426 ARR_APP1(loc_t, starters, loc);
427 DBG((dbg, DBG_START, " %+F:\n", irn));
430 be_lv_foreach(env->cenv->lv, block, be_lv_state_in, i) {
431 ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
432 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn))
436 loc.time = get_distance(env, first, 0, irn, 0);
437 ARR_APP1(loc_t, starters, loc);
438 DBG((dbg, DBG_START, " %+F:\n", irn));
441 // Sort start values by first use
442 qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
444 /* Copy the best ones from starters to start workset */
445 ws_count = MIN(ARR_LEN(starters), env->n_regs);
446 block_info->ws_start = new_workset(env, &env->ob);
447 workset_bulk_fill(block_info->ws_start, ws_count, starters);
449 /* The phis of this block which are not in the start set have to be spilled later. */
450 for (i = ws_count, len = ARR_LEN(starters); i < len; ++i) {
451 irn = starters[i].irn;
452 if (!is_Phi(irn) || get_nodes_block(irn) != block)
455 be_spill_phi(env->senv, irn);
462 * Collects all values live-in at block @p blk and all phi results in this block.
463 * Then it adds the best values (at most n_regs) to the blocks start_workset.
464 * The phis among the remaining values get spilled: Introduce psudo-copies of
465 * their args to break interference and make it possible to spill them to the
468 static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) {
470 block_info_t *res, *pred_info;
472 /* Have we seen this block before? */
473 res = get_block_info(block);
477 /* Create the block info for this block. */
478 res = new_block_info(&env->ob);
479 set_block_info(block, res);
481 /* Use endset of predecessor block as startset */
482 assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block)));
483 pred_block = get_Block_cfgpred_block(block, 0);
484 pred_info = get_block_info(pred_block);
486 /* if pred block has not been processed yet, do it now */
487 if (pred_info == NULL || pred_info->processed == 0) {
488 belady(pred_block, env);
489 pred_info = get_block_info(pred_block);
492 /* now we have an end_set of pred */
493 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
494 res->ws_start = workset_clone(env, &env->ob, pred_info->ws_end);
501 * For the given block @p blk, decide for each values
502 * whether it is used from a register or is reloaded
505 static void belady(ir_node *block, void *data) {
506 belady_env_t *env = data;
510 block_info_t *block_info;
512 /* make sure we have blockinfo (with startset) */
513 block_info = get_block_info(block);
514 if (block_info == NULL)
515 block_info = compute_block_start_info(env, block);
517 /* Don't do a block twice */
518 if(block_info->processed)
521 /* get the starting workset for this block */
522 DBG((dbg, DBG_DECIDE, "\n"));
523 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
525 workset_copy(env, env->ws, block_info->ws_start);
526 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
527 workset_foreach(env->ws, irn, iter)
528 DBG((dbg, DBG_WSETS, " %+F\n", irn));
530 /* process the block from start to end */
531 DBG((dbg, DBG_WSETS, "Processing...\n"));
532 env->used = pset_new_ptr_default();
534 new_vals = new_workset(env, &env->ob);
535 sched_foreach(block, irn) {
537 assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
539 /* projs are handled with the tuple value.
540 * Phis are no real instr (see insert_starters())
541 * instr_nr does not increase */
542 if (is_Proj(irn) || is_Phi(irn)) {
543 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
546 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
548 /* set instruction in the workset */
551 /* allocate all values _used_ by this instruction */
552 workset_clear(new_vals);
553 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
554 workset_insert(env, new_vals, get_irn_n(irn, i));
556 displace(env, new_vals, 1);
558 /* allocate all values _defined_ by this instruction */
559 workset_clear(new_vals);
560 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
562 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
563 workset_insert(env, new_vals, proj);
565 workset_insert(env, new_vals, irn);
567 displace(env, new_vals, 0);
573 /* Remember end-workset for this block */
574 block_info->ws_end = workset_clone(env, &env->ob, env->ws);
575 block_info->processed = 1;
576 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
577 workset_foreach(block_info->ws_end, irn, iter)
578 DBG((dbg, DBG_WSETS, " %+F\n", irn));
582 * 'decide' is block-local and makes assumptions
583 * about the set of live-ins. Thus we must adapt the
584 * live-outs to the live-ins at each block-border.
586 static void fix_block_borders(ir_node *blk, void *data) {
587 belady_env_t *env = data;
589 int i, max, iter, iter2;
591 DBG((dbg, DBG_FIX, "\n"));
592 DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
594 wsb = get_block_info(blk)->ws_start;
596 /* process all pred blocks */
597 for (i=0, max=get_irn_arity(blk); i<max; ++i) {
598 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
599 workset_t *wsp = get_block_info(pred)->ws_end;
601 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
603 workset_foreach(wsb, irnb, iter) {
604 /* if irnb is a phi of the current block we reload
605 * the corresponding argument, else irnb itself */
606 if(is_Phi(irnb) && blk == get_nodes_block(irnb)) {
607 irnb = get_irn_n(irnb, i);
609 // we might have unknowns as argument for the phi
610 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
614 /* Unknowns are available everywhere */
615 if(get_irn_opcode(irnb) == iro_Unknown)
618 /* check if irnb is in a register at end of pred */
619 workset_foreach(wsp, irnp, iter2) {
624 /* irnb is not in memory at the end of pred, so we have to reload it */
625 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
626 be_add_reload_on_edge(env->senv, irnb, blk, i);
629 /*epsilon statement :)*/;
634 void be_spill_belady(const be_chordal_env_t *chordal_env) {
635 be_spill_belady_spill_env(chordal_env, NULL);
638 void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
641 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
642 //firm_dbg_set_mask(dbg, DBG_START);
644 /* init belady env */
645 obstack_init(&env.ob);
646 env.cenv = chordal_env;
647 env.arch = chordal_env->birg->main_env->arch_env;
648 env.cls = chordal_env->cls;
649 env.n_regs = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, chordal_env->cls, NULL);
650 env.ws = new_workset(&env, &env.ob);
651 env.uses = be_begin_uses(chordal_env->irg, chordal_env->exec_freq, chordal_env->lv);
652 if(spill_env == NULL) {
653 env.senv = be_new_spill_env(chordal_env);
655 env.senv = spill_env;
657 DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
659 DBG((dbg, LEVEL_1, "running on register class: %s\n", env.cls->name));
661 be_clear_links(chordal_env->irg);
662 /* Decide which phi nodes will be spilled and place copies for them into the graph */
663 irg_block_walk_graph(chordal_env->irg, spill_phi_walker, NULL, &env);
664 /* Fix high register pressure with belady algorithm */
665 irg_block_walk_graph(chordal_env->irg, NULL, belady, &env);
666 /* belady was block-local, fix the global flow by adding reloads on the edges */
667 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &env);
668 /* Insert spill/reload nodes into the graph and fix usages */
669 be_insert_spills_reloads(env.senv);
672 if(spill_env == NULL)
673 be_delete_spill_env(env.senv);
674 be_end_uses(env.uses);
675 obstack_free(&env.ob, NULL);