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) */
67 spill_env_t *senv; /**< see bespill.h */
71 int len; /**< current length */
72 loc_t vals[0]; /**< inlined array of the values/distances in this working set */
75 void workset_print(const workset_t *w)
79 for(i = 0; i < w->len; ++i) {
80 ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time);
85 * Alloc a new workset on obstack @p ob with maximum size @p max
87 static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) {
89 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
90 res = obstack_alloc(ob, size);
96 * Alloc a new instance on obstack and make it equal to @param ws
98 static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) {
100 size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]);
101 res = obstack_alloc(ob, size);
102 memcpy(res, ws, size);
107 * Do NOT alloc anything. Make @param tgt equal to @param src.
108 * returns @param tgt for convinience
110 static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src) {
111 size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]);
112 memcpy(tgt, src, size);
117 * Overwrites the current content array of @param ws with the
118 * @param count locations given at memory @param locs.
119 * Set the length of @param ws to count.
121 static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) {
122 workset->len = count;
123 memcpy(&(workset->vals[0]), locs, 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(belady_env_t *env, workset_t *ws, ir_node *val) {
132 /* check for current regclass */
133 if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
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 < env->n_regs && "Workset already full!");
145 ws->vals[ws->len++].irn = val;
149 * Removes all entries from this workset
151 static INLINE void workset_clear(workset_t *ws) {
156 * Removes the value @p val from the workset if present.
158 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
160 for(i=0; i<ws->len; ++i) {
161 if (ws->vals[i].irn == val) {
162 ws->vals[i] = ws->vals[--ws->len];
168 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
170 for(i=0; i<ws->len; ++i) {
171 if (ws->vals[i].irn == val)
179 * Iterates over all values in the working set.
180 * @p ws The workset to iterate
181 * @p v A variable to put the current value in
182 * @p i An integer for internal use
184 #define workset_foreach(ws, v, i) for(i=0; \
185 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
188 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
189 #define workset_set_length(ws, length) (ws)->len = length
190 #define workset_get_length(ws) ((ws)->len)
191 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
192 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
194 typedef struct _block_info_t {
195 workset_t *ws_start, *ws_end;
200 static INLINE void *new_block_info(struct obstack *ob) {
201 block_info_t *res = obstack_alloc(ob, sizeof(*res));
202 res->ws_start = NULL;
209 #define get_block_info(blk) ((block_info_t *)get_irn_link(blk))
210 #define set_block_info(blk, info) set_irn_link(blk, info)
213 * @return The distance to the next use or 0 if irn has dont_spill flag set
215 static INLINE unsigned get_distance(belady_env_t *env, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses)
217 int flags = arch_irn_get_flags(env->arch, def);
218 unsigned dist = be_get_next_use(env->uses, from, from_step, def, skip_from_uses);
220 assert(! (flags & arch_irn_flags_ignore));
221 // we have to keep nonspillable nodes in the workingset
222 if(flags & arch_irn_flags_dont_spill)
229 * Performs the actions necessary to grant the request that:
230 * - new_vals can be held in registers
231 * - as few as possible other values are disposed
232 * - the worst values get disposed
234 * @p is_usage indicates that the values in new_vals are used (not defined)
235 * In this case reloads must be performed
237 static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) {
239 int i, len, max_allowed, demand, iter;
240 workset_t *ws = env->ws;
241 ir_node **to_insert = alloca(env->n_regs * sizeof(*to_insert));
244 * 1. Identify the number of needed slots and the values to reload
247 workset_foreach(new_vals, val, iter) {
248 /* mark value as used */
250 pset_insert_ptr(env->used, val);
252 if (!workset_contains(ws, val)) {
253 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
254 to_insert[demand++] = val;
256 be_add_reload(env->senv, val, env->instr);
258 assert(is_usage || "Defined value already in workset?!?");
259 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
262 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
265 * 2. Make room for at least 'demand' slots
267 len = workset_get_length(ws);
268 max_allowed = env->n_regs - demand;
270 DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
272 /* Only make more free room if we do not have enough */
273 if (len > max_allowed) {
274 /* get current next-use distance */
275 for (i=0; i<ws->len; ++i)
276 workset_set_time(ws, i, get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage));
278 /* sort entries by increasing nextuse-distance*/
281 /* Logic for not needed live-ins: If a value is disposed
282 * before its first usage, remove it from start workset
283 * We don't do this for phis though
285 for (i=max_allowed; i<ws->len; ++i) {
286 ir_node *irn = ws->vals[i].irn;
291 if (!pset_find_ptr(env->used, irn)) {
292 ir_node *curr_bb = get_nodes_block(env->instr);
293 workset_t *ws_start = get_block_info(curr_bb)->ws_start;
294 workset_remove(ws_start, irn);
296 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
298 DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn));
302 /* kill the last 'demand' entries in the array */
303 workset_set_length(ws, max_allowed);
307 * 3. Insert the new values into the workset
309 for(i = 0; i < demand; ++i)
310 workset_insert(env, env->ws, to_insert[i]);
313 static void belady(ir_node *blk, void *env);
316 * Computes set of live-ins for each block with multiple predecessors and
317 * places copies in the predecessors when phis get spilled
319 static void place_copy_walker(ir_node *block, void *data) {
320 belady_env_t *env = data;
321 block_info_t *block_info;
323 ir_node *first, *irn;
324 loc_t loc, *starters;
325 int i, len, ws_count;
327 if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
330 block_info = new_block_info(&env->ob);
331 set_block_info(block, block_info);
333 /* Collect all values living at start of block */
334 starters = NEW_ARR_F(loc_t, 0);
336 DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
337 first = sched_first(block);
338 sched_foreach(block, irn) {
341 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn))
345 loc.time = get_distance(env, first, 0, irn, 0);
346 ARR_APP1(loc_t, starters, loc);
347 DBG((dbg, DBG_START, " %+F:\n", irn));
350 live_foreach(block, li) {
351 if (!live_is_in(li) || !arch_irn_consider_in_reg_alloc(env->arch, env->cls, li->irn))
354 loc.irn = (ir_node *)li->irn;
355 loc.time = get_distance(env, first, 0, li->irn, 0);
356 ARR_APP1(loc_t, starters, loc);
357 DBG((dbg, DBG_START, " %+F:\n", li->irn));
360 // Sort start values by first use
361 qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
363 /* Copy the best ones from starters to start workset */
364 ws_count = MIN(ARR_LEN(starters), env->n_regs);
365 block_info->ws_start = new_workset(env, &env->ob);
366 workset_bulk_fill(block_info->ws_start, ws_count, starters);
368 /* The phis of this block which are not in the start set have to be spilled later. */
369 for (i = ws_count, len = ARR_LEN(starters); i < len; ++i) {
370 irn = starters[i].irn;
371 if (!is_Phi(irn) || get_nodes_block(irn) != block)
374 be_spill_phi(env->senv, irn);
381 * Collects all values live-in at block @p blk and all phi results in this block.
382 * Then it adds the best values (at most n_regs) to the blocks start_workset.
383 * The phis among the remaining values get spilled: Introduce psudo-copies of
384 * their args to break interference and make it possible to spill them to the
387 static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) {
389 block_info_t *res, *pred_info;
391 /* Have we seen this block before? */
392 res = get_block_info(block);
396 /* Create the block info for this block. */
397 res = new_block_info(&env->ob);
398 set_block_info(block, res);
400 /* Use endset of predecessor block as startset */
401 assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block)));
402 pred_block = get_Block_cfgpred_block(block, 0);
403 pred_info = get_block_info(pred_block);
405 /* if pred block has not been processed yet, do it now */
406 if (pred_info == NULL || pred_info->processed == 0) {
407 belady(pred_block, env);
408 pred_info = get_block_info(pred_block);
411 /* now we have an end_set of pred */
412 assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
413 res->ws_start = workset_clone(env, &env->ob, pred_info->ws_end);
420 * For the given block @p blk, decide for each values
421 * whether it is used from a register or is reloaded
424 static void belady(ir_node *block, void *data) {
425 belady_env_t *env = data;
429 block_info_t *block_info;
431 /* make sure we have blockinfo (with startset) */
432 block_info = get_block_info(block);
433 if (block_info == NULL)
434 block_info = compute_block_start_info(env, block);
436 /* Don't do a block twice */
437 if(block_info->processed)
440 /* get the starting workset for this block */
441 DBG((dbg, DBG_DECIDE, "\n"));
442 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block));
444 workset_copy(env, env->ws, block_info->ws_start);
445 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block));
446 workset_foreach(env->ws, irn, iter)
447 DBG((dbg, DBG_WSETS, " %+F\n", irn));
449 /* process the block from start to end */
450 DBG((dbg, DBG_WSETS, "Processing...\n"));
451 env->used = pset_new_ptr_default();
453 new_vals = new_workset(env, &env->ob);
454 sched_foreach(block, irn) {
456 assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!");
458 /* projs are handled with the tuple value.
459 * Phis are no real instr (see insert_starters())
460 * instr_nr does not increase */
461 if (is_Proj(irn) || is_Phi(irn)) {
462 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
465 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
467 /* set instruction in the workset */
470 /* allocate all values _used_ by this instruction */
471 workset_clear(new_vals);
472 for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
473 workset_insert(env, new_vals, get_irn_n(irn, i));
475 displace(env, new_vals, 1);
477 /* allocate all values _defined_ by this instruction */
478 workset_clear(new_vals);
479 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
481 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
482 workset_insert(env, new_vals, proj);
484 workset_insert(env, new_vals, irn);
486 displace(env, new_vals, 0);
492 /* Remember end-workset for this block */
493 block_info->ws_end = workset_clone(env, &env->ob, env->ws);
494 block_info->processed = 1;
495 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block));
496 workset_foreach(block_info->ws_end, irn, iter)
497 DBG((dbg, DBG_WSETS, " %+F\n", irn));
501 * 'decide' is block-local and makes assumptions
502 * about the set of live-ins. Thus we must adapt the
503 * live-outs to the live-ins at each block-border.
505 static void fix_block_borders(ir_node *blk, void *data) {
506 belady_env_t *env = data;
508 int i, max, iter, iter2;
510 DBG((dbg, DBG_FIX, "\n"));
511 DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
513 wsb = get_block_info(blk)->ws_start;
515 /* process all pred blocks */
516 for (i=0, max=get_irn_arity(blk); i<max; ++i) {
517 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
518 workset_t *wsp = get_block_info(pred)->ws_end;
520 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
522 workset_foreach(wsb, irnb, iter) {
523 /* if irnb is a phi of the current block we reload
524 * the corresponding argument, else irnb itself */
525 if(is_Phi(irnb) && blk == get_nodes_block(irnb)) {
526 irnb = get_irn_n(irnb, i);
528 // we might have unknowns as argument for the phi
529 if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
533 /* Unknowns are available everywhere */
534 if(get_irn_opcode(irnb) == iro_Unknown)
537 /* check if irnb is in a register at end of pred */
538 workset_foreach(wsp, irnp, iter2) {
543 /* irnb is not in memory at the end of pred, so we have to reload it */
544 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
545 be_add_reload_on_edge(env->senv, irnb, blk, i);
548 /*epsilon statement :)*/;
553 void be_spill_belady(const be_chordal_env_t *chordal_env) {
554 be_spill_belady_spill_env(chordal_env, NULL);
557 void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
560 FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
562 /* init belady env */
563 obstack_init(&env.ob);
564 env.arch = chordal_env->birg->main_env->arch_env;
565 env.cls = chordal_env->cls;
566 env.n_regs = arch_count_non_ignore_regs(env.arch, env.cls);
567 env.ws = new_workset(&env, &env.ob);
568 env.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, env.cls);
569 if(spill_env == NULL) {
570 env.senv = be_new_spill_env(chordal_env);
572 env.senv = spill_env;
574 DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
576 DBG((dbg, LEVEL_1, "running on register class: %s\n", env.cls->name));
579 be_clear_links(chordal_env->irg);
580 irg_block_walk_graph(chordal_env->irg, place_copy_walker, NULL, &env);
581 irg_block_walk_graph(chordal_env->irg, NULL, belady, &env);
582 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &env);
583 be_insert_spills_reloads(env.senv);
585 be_remove_dead_nodes_from_schedule(chordal_env->irg);
588 if(spill_env == NULL)
589 be_delete_spill_env(env.senv);
590 be_end_uses(env.uses);
591 obstack_free(&env.ob, NULL);