2 * Author: Daniel Grund, Sebastian Hack
4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
16 #include "iredges_t.h"
26 #include "besched_t.h"
29 #include "bechordal_t.h"
32 /* This enables re-computation of values. Current state: Unfinished and buggy. */
35 typedef struct _reloader_t reloader_t;
36 typedef struct _spill_info_t spill_info_t;
43 struct _spill_info_t {
44 ir_node *spilled_node;
45 reloader_t *reloaders;
48 typedef struct _spill_ctx_t {
49 ir_node *spilled; /**< The spilled node. */
50 ir_node *user; /**< The node this spill is for. */
51 ir_node *spill; /**< The spill itself. */
55 const arch_register_class_t *cls;
56 const be_chordal_env_t *chordal_env;
59 set *spills; /**< all spill_info_t's, which must be placed */
60 pset *mem_phis; /**< set of all special spilled phis. allocated and freed separately */
61 decide_irn_t is_mem_phi; /**< callback func to decide if a phi needs special spilling */
62 void *data; /**< data passed to all callbacks */
63 DEBUG_ONLY(firm_dbg_module_t *dbg;)
66 static int cmp_spillctx(const void *a, const void *b, size_t n) {
67 const spill_ctx_t *p = a;
68 const spill_ctx_t *q = b;
69 return !(p->user == q->user && p->spilled == q->spilled);
72 static int cmp_spillinfo(const void *x, const void *y, size_t size) {
73 const spill_info_t *xx = x;
74 const spill_info_t *yy = y;
75 return ! (xx->spilled_node == yy->spilled_node);
79 void be_set_spill_env_dbg_module(spill_env_t *env, firm_dbg_module_t *dbg) {
84 spill_env_t *be_new_spill_env(const be_chordal_env_t *chordal_env, decide_irn_t is_mem_phi, void *data) {
85 spill_env_t *env = xmalloc(sizeof(env[0]));
86 env->spill_ctxs = new_set(cmp_spillctx, 1024);
87 env->spills = new_set(cmp_spillinfo, 1024);
88 env->cls = chordal_env->cls;
89 env->is_mem_phi = is_mem_phi;
91 env->chordal_env = chordal_env;
92 obstack_init(&env->obst);
96 void be_delete_spill_env(spill_env_t *senv) {
97 del_set(senv->spill_ctxs);
98 del_set(senv->spills);
99 obstack_free(&senv->obst, NULL);
103 static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
106 templ.spilled = to_spill;
107 templ.user = ctx_irn;
110 return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
113 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
115 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
117 ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
119 const be_main_env_t *env = senv->chordal_env->birg->main_env;
120 ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
127 * If the first usage of a phi result would be out of memory
128 * there is no sense in allocating a register for it.
129 * Thus we spill it and all its operands to the same spill slot.
130 * Therefore the phi/dataB becomes a phi/Memory
132 static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
133 int i, n = get_irn_arity(phi);
134 ir_node **ins, *bl = get_nodes_block(phi);
135 ir_graph *irg = senv->chordal_env->irg;
139 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
141 /* search an existing spill for this context */
142 ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
144 /* if not found spill the phi */
146 /* build a new PhiM with dummy in-array */
147 NEW_ARR_A(ir_node *, ins, n);
149 ins[i] = new_r_Unknown(irg, mode_M);
150 ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
152 /* re-wire the phiM */
154 ir_node *arg = get_irn_n(phi, i);
157 if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
158 sub_res = be_spill_phi(senv, arg, ctx_irn);
160 sub_res = be_spill_irn(senv, arg, ctx_irn);
162 set_irn_n(ctx->spill, i, sub_res);
168 static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
170 if (pset_find_ptr(senv->mem_phis, to_spill))
171 res = be_spill_phi(senv, to_spill, to_spill);
173 res = be_spill_irn(senv, to_spill, to_spill);
178 static void phi_walker(ir_node *irn, void *env) {
179 spill_env_t *senv = env;
180 const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
182 if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
183 && senv->is_mem_phi(irn, senv->data)) {
184 DBG((senv->dbg, LEVEL_1, " %+F\n", irn));
185 pset_insert_ptr(senv->mem_phis, irn);
194 * Check if a spilled node could be rematerialized.
196 * @param senv the spill environment
197 * @param spill the Spill node
198 * @param spilled the node that was spilled
199 * @param reloader a irn that requires a reload
201 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
204 /* check for 'normal' spill and general remat condition */
205 if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
208 /* check availability of original arguments */
209 if (is_Block(reloader)) {
211 /* we want to remat at the end of a block.
212 * thus all arguments must be alive at the end of the block
214 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
215 ir_node *arg = get_irn_n(spilled, pos);
216 if (!is_live_end(reloader, arg))
222 /* we want to remat before the insn reloader
223 * thus an arguments is alive if
224 * - it interferes with the reloaders result
226 * - or it is (last-) used by reloader itself
228 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
229 ir_node *arg = get_irn_n(spilled, pos);
232 if (values_interfere(reloader, arg))
235 for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
236 ir_node *rel_arg = get_irn_n(reloader, i);
241 /* arg is not alive before reloader */
253 #else /* BUGGY_REMAT */
256 * A very simple rematerialization checker.
258 * @param senv the spill environment
259 * @param spill the Spill node
260 * @param spilled the node that was spilled
261 * @param reloader a irn that requires a reload
263 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
264 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
266 return get_irn_arity(spilled) == 0 &&
267 be_is_Spill(spill) &&
268 arch_irn_is(aenv, spilled, rematerializable);
271 #endif /* BUGGY_REMAT */
276 * Rematerialize a node.
278 * @param senv the spill environment
279 * @param spilled the node that was spilled
280 * @param reloader a irn that requires a reload
282 static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
284 ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
286 /* recompute the value */
287 res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
289 get_irn_mode(spilled),
290 get_irn_arity(spilled),
291 get_irn_in(spilled) + 1);
292 copy_node_attr(spilled, res);
294 DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
296 /* insert in schedule */
297 if (is_Block(reloader)) {
298 ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
299 sched_add_after(insert, res);
301 sched_add_before(reloader, res);
307 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
308 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
309 ir_graph *irg = senv->chordal_env->irg;
314 /* get all special spilled phis */
315 DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
316 senv->mem_phis = pset_new_ptr_default();
317 irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
319 /* Add reloads for mem_phis */
320 /* BETTER: These reloads (1) should only be inserted, if they are really needed */
321 DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
322 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
324 DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
325 foreach_out_edge(irn, e) {
326 ir_node *user = e->src;
327 if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
328 ir_node *use_bl = get_nodes_block(user);
329 DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
330 be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
335 /* process each spilled node */
336 DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
337 possibly_dead = new_pdeq();
338 for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
340 ir_mode *mode = get_irn_mode(si->spilled_node);
341 pset *values = pset_new_ptr(16);
343 /* go through all reloads for this spill */
344 for(rld = si->reloaders; rld; rld = rld->next) {
347 /* the spill for this reloader */
348 ir_node *spill = be_spill_node(senv, si->spilled_node);
351 if (check_remat_conditions(senv, spill, si->spilled_node, rld->reloader)) {
352 new_val = do_remat(senv, si->spilled_node, rld->reloader);
353 pdeq_putl(possibly_dead, spill);
358 new_val = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
360 DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
361 pset_insert_ptr(values, new_val);
363 pset_insert_ptr(reload_set, new_val);
366 /* introduce copies, rewire the uses */
367 assert(pset_count(values) > 0 && "???");
368 pset_insert_ptr(values, si->spilled_node);
369 be_ssa_constr_set_ignore(senv->chordal_env->dom_front, values, senv->mem_phis);
374 foreach_pset(senv->mem_phis, irn) {
376 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
377 pdeq_putl(possibly_dead, get_irn_n(irn, i));
378 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
383 /* check if possibly dead nodes are really dead yet */
384 while (! pdeq_empty(possibly_dead)) {
385 ir_node *irn = pdeq_getr(possibly_dead);
386 const ir_edge_t *edge = get_irn_out_edge_first(irn);
390 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
391 pdeq_putl(possibly_dead, get_irn_n(irn, i));
392 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
397 del_pdeq(possibly_dead);
398 del_pset(senv->mem_phis);
401 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
402 spill_info_t templ, *res;
405 templ.spilled_node = to_spill;
406 templ.reloaders = NULL;
407 res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
409 rel = obstack_alloc(&senv->obst, sizeof(rel[0]));
410 rel->reloader = before;
411 rel->next = res->reloaders;
412 res->reloaders = rel;
415 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
416 ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
417 be_add_reload(senv, to_spill, insert_bl);
422 /****************************************
424 SPILL SLOT MANAGEMENT AND OPTS
426 ****************************************/
428 typedef struct _spill_slot_t {
432 ir_mode *largest_mode; /* the mode of all members with largest size */
435 typedef struct _ss_env_t {
437 be_chordal_env_t *cenv;
438 pmap *slots; /* maps spill_contexts to spill_slots */
439 pmap *types; /* maps modes to types */
440 DEBUG_ONLY(firm_dbg_module_t *dbg;)
445 * Walker: compute the spill slots
447 static void compute_spill_slots_walker(ir_node *spill, void *env) {
448 ss_env_t *ssenv = env;
453 if (!be_is_Spill(spill))
456 /* check, if this spill is for a context already known */
457 ctx = be_get_Spill_context(spill);
458 entry = pmap_find(ssenv->slots, ctx);
461 struct _arch_env_t *arch_env = ssenv->cenv->birg->main_env->arch_env;
462 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
463 ir_mode *largest_mode = arch_register_class_mode(cls);
465 /* this is a new spill context */
466 ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
467 ss->members = pset_new_ptr(8);
468 ss->largest_mode = largest_mode;
469 ss->size = get_mode_size_bytes(ss->largest_mode);
470 ss->align = arch_isa_get_reg_class_alignment(arch_env->isa, cls);
471 pmap_insert(ssenv->slots, ctx, ss);
473 /* values with the same spill_ctx must go into the same spill slot */
477 /* ugly mega assert :-) */
480 struct _arch_env_t *arch_env = ssenv->cenv->birg->main_env->arch_env;
481 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
482 int size = get_mode_size_bytes(arch_register_class_mode(cls));
483 assert(ss->size == size && "Different sizes for the same spill slot are not allowed.");
484 for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
485 /* use values_interfere here, because it uses the dominance check,
486 which does work for values in memory */
487 assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
493 pset_insert_ptr(ss->members, spill);
497 * qsort compare function, sort spill slots by size.
499 static int ss_sorter(const void *v1, const void *v2) {
500 const spill_slot_t *ss1 = v1;
501 const spill_slot_t *ss2 = v2;
502 return ((int) ss2->size) - ((int) ss1->size);
507 * This function should optimize the spill slots.
508 * - Coalescing of multiple slots
509 * - Ordering the slots
511 * Input slots are in @p ssenv->slots
512 * @p size The count of initial spill slots in @p ssenv->slots
513 * This also is the size of the preallocated array @p ass
515 * @return An array of spill slots @p ass in specific order
517 static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
518 int i, o, used_slots;
522 pmap_foreach(ssenv->slots, entr)
523 ass[i++] = entr->value;
525 /* Sort the array to minimize fragmentation and cache footprint.
526 Large slots come first */
527 qsort(ass, size, sizeof(ass[0]), ss_sorter);
529 /* For each spill slot:
530 - assign a new offset to this slot
531 - xor find another slot to coalesce with */
533 for (i=0; i<size; ++i) { /* for each spill slot */
537 DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
538 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
539 DBG((ssenv->dbg, LEVEL_1, " %+F\n", n1));
542 for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
543 /* check inter-slot-pairs for interference */
545 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
546 for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
547 if(values_interfere(n1, n2)) {
548 pset_break(ass[i]->members);
549 pset_break(ass[o]->members);
550 DBG((ssenv->dbg, LEVEL_1, " Interf %+F -- %+F\n", n1, n2));
551 goto interf_detected;
554 /* if we are here, there is no interference between ass[i] and ass[o] */
557 interf_detected: /*nothing*/ ;
560 /* now the members of ass[i] join the members of ass[tgt_slot] */
562 /* do we need a new slot? */
563 if (tgt_slot == -1) {
564 tgt_slot = used_slots;
569 ass[tgt_slot]->size = ass[i]->size;
570 del_pset(ass[tgt_slot]->members);
571 ass[tgt_slot]->members = pset_new_ptr(8);
575 /* copy the members to the target pset */
576 /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
578 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
579 pset_insert_ptr(ass[tgt_slot]->members, n1);
583 #define ALIGN_SPILL_AREA 16
584 #define pset_foreach(pset, elm) for(elm=pset_first(pset); elm; elm=pset_next(pset))
587 * Returns a spill type for a mode. Keep them in a map to reduce
588 * the number of types.
590 * @param types a map containing all created types
591 * @param ss the spill slot
593 * Note that type types should are identical for every mode.
594 * This rule might break if two different register classes return the same
595 * mode but different alignments.
597 static ir_type *get_spill_type(pmap *types, spill_slot_t *ss) {
598 pmap_entry *e = pmap_find(types, ss->largest_mode);
603 snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(ss->largest_mode));
604 res = new_type_primitive(new_id_from_str(buf), ss->largest_mode);
605 set_type_alignment_bytes(res, ss->align);
606 pmap_insert(types, ss->largest_mode, res);
610 assert(get_type_alignment_bytes(res) == (int)ss->align);
616 * Create spill slot entities on the frame type.
618 * @param ssenv the spill environment
619 * @param n number of spill slots
620 * @param ss array of spill slots
622 static void assign_entities(ss_env_t *ssenv, int n_slots, spill_slot_t *ss[]) {
623 int i, offset, frame_align;
624 ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
626 /* aligning by increasing frame size */
627 offset = get_type_size_bits(frame) / 8;
628 offset = round_up2(offset, ALIGN_SPILL_AREA);
629 set_type_size_bytes(frame, -1);
631 /* create entities and assign offsets according to size and alignment*/
632 for (i = 0; i < n_slots; ++i) {
639 snprintf(buf, sizeof(buf), "spill_slot_%d", i);
640 name = new_id_from_str(buf);
642 spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]));
645 offset = round_up2(offset, ss[i]->align);
647 set_entity_offset_bytes(spill_ent, offset);
648 /* next possible offset */
649 offset += round_up2(ss[i]->size, ss[i]->align);
651 pset_foreach(ss[i]->members, irn)
652 be_set_Spill_entity(irn, spill_ent);
655 /* set final size of stack frame */
656 frame_align = get_type_alignment_bytes(frame);
657 set_type_size_bytes(frame, round_up2(offset, frame_align));
660 void be_compute_spill_offsets(be_chordal_env_t *cenv) {
666 obstack_init(&ssenv.ob);
668 ssenv.slots = pmap_create();
669 ssenv.types = pmap_create();
670 FIRM_DBG_REGISTER(ssenv.dbg, "ir.be.spillslots");
672 /* Get initial spill slots */
673 irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
675 /* Build an empty array for optimized spill slots */
676 ss_size = pmap_count(ssenv.slots);
677 ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
678 optimize_slots(&ssenv, ss_size, ss);
680 /* Integrate slots into the stack frame entity */
681 assign_entities(&ssenv, ss_size, ss);
684 pmap_foreach(ssenv.slots, pme)
685 del_pset(((spill_slot_t *)pme->value)->members);
686 pmap_destroy(ssenv.slots);
687 pmap_destroy(ssenv.types);
688 obstack_free(&ssenv.ob, NULL);
690 be_copy_entities_to_reloads(cenv->irg);