4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
7 * NOTE: Comments my be (partially) wrong, since there was a major bug
8 * (spilling of phis, prespill) whose fixing changed a lot.
29 #include "iredges_t.h"
34 #include "bespillbelady.h"
36 #include "besched_t.h"
40 #include "bechordal_t.h"
49 #define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
50 static firm_dbg_module_t *dbg = NULL;
52 #define MIN(a,b) (((a)<(b))?(a):(b))
53 #undef SINGLE_START_PROJS
55 typedef struct _workset_t workset_t;
57 typedef struct _belady_env_t {
59 const arch_env_t *arch;
60 const arch_register_class_t *cls;
61 int n_regs; /** number of regs in this reg-class */
63 workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
64 be_uses_t *uses; /**< env for the next-use magic */
65 ir_node *instr; /**< current instruction */
66 unsigned instr_nr; /**< current instruction number (relative to block start) */
67 pset *used; /**< holds the values used (so far) in the current BB */
68 pset *copies; /**< holds all copies placed due to phi-spilling */
70 spill_env_t *senv; /* see bespill.h */
71 pset *reloads; /**< all reload nodes placed */
76 // int i; /**< used for iteration TODO remove this form the struct */
77 int len; /**< current length */
78 loc_t vals[1]; /**< inlined array of the values/distances in this working set */
81 typedef struct _block_info_t {
82 workset_t *ws_start, *ws_end;
86 * Alloc a new workset on obstack @p ob with maximum size @p max
88 static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) {
90 size_t size = sizeof(*res) + (bel->n_regs-1)*sizeof(res->vals[0]);
91 res = obstack_alloc(ob, size);
98 * Alloc a new instance on obstack and make it equal to @param ws
100 static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) {
102 size_t size = sizeof(*res) + (ws->bel->n_regs-1)*sizeof(res->vals[0]);
103 res = obstack_alloc(ob, size);
104 memcpy(res, ws, size);
109 * Do NOT alloc anything. Make @param tgt equal to @param src.
110 * returns @param tgt for convinience
112 static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) {
113 size_t size = sizeof(*src) + (src->bel->n_regs-1)*sizeof(src->vals[0]);
114 memcpy(tgt, src, size);
119 * Overwrites the current content array of @param ws with the
120 * @param count locations given at memory @param locs.
121 * Set the length of @param ws to count.
123 #define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=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(workset_t *ws, ir_node *val) {
132 /* check for current regclass */
133 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
134 DBG((dbg, DBG_DECIDE, "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 < ws->bel->n_regs && "Workset already full!");
145 ws->vals[ws->len++].irn = val;
149 * Inserts all values in array @p vals of length @p cnt
150 * into the workset. There must be enough space for the
153 static INLINE void workset_bulk_insert(workset_t *ws, int cnt, ir_node **vals) {
156 for(o=0; o<cnt; ++o) {
157 ir_node *val = vals[o];
158 DBG((dbg, DBG_TRACE, "Bulk insert %+F\n", val));
159 /* check for current regclass */
160 if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
161 DBG((dbg, DBG_TRACE, "Wrong reg class\n"));
165 /* check if val is already contained */
166 for(i=0; i<ws->len; ++i)
167 if (ws->vals[i].irn == val) {
168 DBG((dbg, DBG_TRACE, "Already contained\n"));
173 assert(ws->len < ws->bel->n_regs && "Workset does not have enough room!");
174 ws->vals[ws->len++].irn = val;
175 DBG((dbg, DBG_TRACE, "Inserted\n"));
178 /*epsilon statement :)*/;
183 * Removes all entries from this workset
185 #define workset_clear(ws) (ws)->len = 0;
188 * Removes the value @p val from the workset if present.
190 static INLINE void workset_remove(workset_t *ws, ir_node *val) {
192 for(i=0; i<ws->len; ++i)
193 if (ws->vals[i].irn == val) {
194 ws->vals[i] = ws->vals[--ws->len];
199 static INLINE int workset_contains(const workset_t *ws, const ir_node *val) {
201 for(i=0; i<ws->len; ++i)
202 if (ws->vals[i].irn == val)
208 * Iterates over all values in the working set.
209 * @p ws The workset to iterate
210 * @p v A variable to put the current value in
211 * @p i An integer for internal use
213 #define workset_foreach(ws, v, i) for(i=0; \
214 v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \
217 #define workset_set_time(ws, i, t) (ws)->vals[i].time=t
218 #define workset_set_length(ws, length) (ws)->len = length
219 #define workset_get_length(ws) ((ws)->len)
220 #define workset_get_val(ws, i) ((ws)->vals[i].irn)
221 #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare);
224 static int is_mem_phi(const ir_node *irn, void *data) {
226 ir_node *blk = get_nodes_block(irn);
228 DBG((dbg, DBG_SPILL, "Is %+F a mem-phi?\n", irn));
229 sws = ((block_info_t *) get_irn_link(blk))->ws_start;
230 DBG((dbg, DBG_SPILL, " %d\n", !workset_contains(sws, irn)));
231 return !workset_contains(sws, irn);
235 * @return The distance to the next use
236 * Or 0 if irn is an ignore node
238 #define get_distance(bel, from, from_step, def, skip_from_uses) \
239 ((arch_irn_is_ignore(bel->arch, def) ) ? 0 : be_get_next_use(bel->uses, from, from_step, def, skip_from_uses))
243 * Collects all values live-in at block @p blk and all phi results in this block.
244 * Then it adds the best values (at most n_regs) to the blocks start_workset.
245 * The phis among the remaining values get spilled: Introduce psudo-copies of
246 * their args to break interference and make it possible to spill them to the
249 static void compute_block_start_info(ir_node *blk, void *env) {
250 belady_env_t *bel = env;
251 block_info_t *blk_info;
252 ir_node *irn, *first;
254 int i, count, ws_count;
255 loc_t loc, *starters;
256 ir_graph *irg = get_irn_irg(blk);
261 /* Get all values living at the block start */
262 DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
263 first = sched_first(blk);
265 sched_foreach(blk, irn)
266 if (is_Phi(irn) && arch_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) {
268 loc.time = get_distance(bel, first, 0, irn, 0);
269 obstack_grow(&ob, &loc, sizeof(loc));
270 DBG((dbg, DBG_START, " %+F:\n", irn));
275 live_foreach(blk, li)
276 if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) {
277 loc.irn = (ir_node *)li->irn;
278 loc.time = get_distance(bel, first, 0, li->irn, 0);
279 obstack_grow(&ob, &loc, sizeof(loc));
280 DBG((dbg, DBG_START, " %+F:\n", irn));
283 starters = obstack_finish(&ob);
285 /* Sort all values */
286 qsort(starters, count, sizeof(starters[0]), loc_compare);
288 /* Create the start workset for this block. Copy the best ones from starters */
289 blk_info = obstack_alloc(&bel->ob, sizeof(*blk_info));
290 set_irn_link(blk, blk_info);
292 ws_count = MIN(count, bel->n_regs);
293 blk_info->ws_start = new_workset(&bel->ob, bel);
294 workset_bulk_fill(blk_info->ws_start, ws_count, starters);
296 /* Spill the phis among the remaining values */
297 for (i=ws_count; i<count; ++i) {
300 irn = starters[i].irn;
301 if (!is_Phi(irn) || get_nodes_block(irn) != blk)
304 DBG((dbg, DBG_START, "For %+F:\n", irn));
306 for (max=get_irn_arity(irn), o=0; o<max; ++o) {
307 ir_node *arg = get_irn_n(irn, o);
308 ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
309 ir_node *cpy = be_new_Copy(bel->cls, irg, pred_block, arg);
310 pset_insert_ptr(bel->copies, cpy);
311 DBG((dbg, DBG_START, " place a %+F of %+F in %+F\n", cpy, arg, pred_block));
312 sched_add_before(pred_block, cpy);
313 set_irn_n(irn, o, cpy);
317 obstack_free(&ob, NULL);
321 * Performs the actions neccessary to grant the request that:
322 * - new_vals can be held in registers
323 * - as few as possible other values are disposed
324 * - the worst values get disposed
326 * @p is_usage indicates that the values in new_vals are used (not defined)
327 * In this case reloads must be performed
329 static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
331 int i, len, max_allowed, demand, iter;
332 workset_t *ws = bel->ws;
333 ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert));
336 * 1. Identify the number of needed slots and the values to reload
339 workset_foreach(new_vals, val, iter) {
340 /* mark value as used */
342 pset_insert_ptr(bel->used, val);
344 if (!workset_contains(ws, val)) {
345 DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
346 to_insert[demand++] = val;
348 be_add_reload(bel->senv, val, bel->instr);
350 DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
352 DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
356 * 2. Make room for at least 'demand' slots
358 len = workset_get_length(ws);
359 max_allowed = bel->n_regs - demand;
361 /* Only make more free room if we do not have enough */
362 if (len > max_allowed) {
363 /* get current next-use distance */
364 for (i=0; i<ws->len; ++i)
365 workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage));
367 /* sort entries by increasing nextuse-distance*/
370 /* Logic for not needed live-ins: If a value is disposed
371 before its first usage, remove it from start workset */
372 for (i=max_allowed; i<ws->len; ++i) {
373 ir_node *irn = ws->vals[i].irn;
374 if (!pset_find_ptr(bel->used, irn)) {
375 ir_node *curr_bb = get_nodes_block(bel->instr);
376 workset_t *ws_start = ((block_info_t *) get_irn_link(curr_bb))->ws_start;
377 workset_remove(ws_start, irn);
379 DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn));
381 DBG((dbg, DBG_DECIDE, " dispose %+F\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 workset_bulk_insert(bel->ws, demand, to_insert);
395 * For the given block @p blk, decide for each values
396 * whether it is used from a register or is reloaded
399 static void belady(ir_node *blk, void *env) {
400 belady_env_t *bel = env;
405 #ifdef SINGLE_START_PROJS
406 ir_node *start_blk = get_irg_start_block(get_irn_irg(blk));
408 block_info_t *blk_info = get_irn_link(blk);
410 DBG((dbg, DBG_DECIDE, "\n"));
411 DBG((dbg, DBG_DECIDE, "Decide for %+F\n", blk));
413 workset_copy(bel->ws, blk_info->ws_start);
414 DBG((dbg, DBG_WSETS, "Initial start workset for %+F:\n", blk));
415 workset_foreach(bel->ws, irn, iter)
416 DBG((dbg, DBG_WSETS, " %+F\n", irn));
418 /* process the block from start to end */
419 DBG((dbg, DBG_WSETS, "Processing...\n"));
420 bel->used = pset_new_ptr(32);
422 new_vals = new_workset(&bel->ob, bel);
423 sched_foreach(blk, irn) {
424 assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!");
427 #ifdef SINGLE_START_PROJS
429 (is_Proj(irn) && blk!=start_blk) ||
430 (get_irn_mode(irn) == mode_T && blk==start_blk)) {
431 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
435 /* projs are handled with the tuple value.
436 * Phis are no real instr (see insert_starters)
437 * instr_nr does not increase */
438 if (is_Proj(irn) || is_Phi(irn)) {
439 DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn));
443 DBG((dbg, DBG_DECIDE, " ...%+F\n", irn));
445 /* set instruction in the workset */
448 /* allocate all values _used_ by this instruction */
449 workset_clear(new_vals);
450 workset_bulk_insert(new_vals, get_irn_arity(irn)+1, get_irn_in(irn));
451 displace(bel, new_vals, 1);
453 /* allocate all values _defined_ by this instruction */
454 workset_clear(new_vals);
455 if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */
457 for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj))
458 workset_insert(new_vals, proj);
460 workset_insert(new_vals, irn);
462 displace(bel, new_vals, 0);
468 /* Remember end-workset for this block */
469 blk_info->ws_end = workset_clone(&bel->ob, bel->ws);
470 DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk));
471 workset_foreach(blk_info->ws_start, irn, iter)
472 DBG((dbg, DBG_WSETS, " %+F\n", irn));
473 DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk));
474 workset_foreach(blk_info->ws_end, irn, iter)
475 DBG((dbg, DBG_WSETS, " %+F\n", irn));
479 * 'decide' is block-local and makes assumptions
480 * about the set of live-ins. Thus we must adapt the
481 * live-outs to the live-ins at each block-border.
483 static void fix_block_borders(ir_node *blk, void *env) {
485 belady_env_t *bel = env;
488 DBG((dbg, DBG_FIX, "\n"));
489 DBG((dbg, DBG_FIX, "Fixing %+F\n", blk));
491 wsb = ((block_info_t *)get_irn_link(blk))->ws_start;
493 /* process all pred blocks */
494 for (i=0, max=get_irn_arity(blk); i<max; ++i) {
495 ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(blk, i);
496 workset_t *wsp = ((block_info_t *)get_irn_link(pred))->ws_end;
498 DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
500 workset_foreach(wsb, irnb, iter) {
501 /* if irnb is a phi of the current block we reload
502 * the corresponding argument, else irnb itself */
503 if(is_Phi(irnb) && blk == get_nodes_block(irnb))
504 irnb = get_irn_n(irnb, i);
506 /* Unknowns are available everywhere */
507 if(get_irn_opcode(irnb) == iro_Unknown)
510 /* check if irnb is in a register at end of pred */
511 workset_foreach(wsp, irnp, iter)
515 /* irnb is in memory at the end of pred, so we have to reload it */
516 DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
517 be_add_reload_on_edge(bel->senv, irnb, blk, i);
520 /*epsilon statement :)*/;
526 * Removes all used reloads from bel->reloads.
527 * The remaining nodes in bel->reloads will be removed from the graph.
529 static void rescue_used_reloads(ir_node *irn, void *env) {
530 pset *rlds = (pset *)env;
531 if (pset_find_ptr(rlds, irn))
532 pset_remove_ptr(rlds, irn);
536 * Removes all copies introduced for phi-spills
538 static void remove_copies(belady_env_t *bel) {
541 for (irn = pset_first(bel->copies); irn; irn = pset_next(bel->copies)) {
542 ir_node *src, *spill;
544 assert(be_is_Copy(irn));
545 assert(get_irn_n_edges(irn) == 1 && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
547 spill = get_irn_edge(get_irn_irg(irn), irn, 0)->src;
548 assert(be_is_Spill(spill) && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
550 src = get_irn_n(irn, 0);
551 set_irn_n(spill, 0, src);
556 * Finds all unused reloads and remove them from the schedule
557 * Also removes spills if they are not used anymore after removing reloads
559 static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) {
562 irg_walk_graph(irg, rescue_used_reloads, NULL, bel->reloads);
563 for(irn = pset_first(bel->reloads); irn; irn = pset_next(bel->reloads)) {
565 DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn)));
567 spill = get_irn_n(irn, 0);
570 set_irn_n(irn, 0, new_Bad());
573 /* if spill not used anymore, remove it too
574 * test of regclass is necessary since spill may be a phi-M */
575 if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) {
576 set_irn_n(spill, 0, new_Bad());
582 void be_spill_belady(const be_chordal_env_t *chordal_env) {
585 dbg = firm_dbg_register("ir.be.spillbelady");
587 /* init belady env */
588 obstack_init(&bel.ob);
589 bel.arch = chordal_env->main_env->arch_env;
590 bel.cls = chordal_env->cls;
591 bel.n_regs = arch_register_class_n_regs(bel.cls);
592 bel.ws = new_workset(&bel.ob, &bel);
593 bel.uses = be_begin_uses(chordal_env->irg, chordal_env->main_env->arch_env, bel.cls);
594 bel.senv = be_new_spill_env(dbg, chordal_env, is_mem_phi, NULL);
595 bel.reloads = pset_new_ptr_default();
596 bel.copies = pset_new_ptr_default();
598 DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
601 irg_block_walk_graph(chordal_env->irg, compute_block_start_info, NULL, &bel);
602 irg_block_walk_graph(chordal_env->irg, belady, NULL, &bel);
603 irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
604 be_insert_spills_reloads(bel.senv, bel.reloads);
605 remove_unused_reloads(chordal_env->irg, &bel);
611 del_pset(bel.reloads);
612 be_delete_spill_env(bel.senv);
613 be_end_uses(bel.uses);
614 obstack_free(&bel.ob, NULL);