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"
25 #include "besched_t.h"
28 #include "bechordal_t.h"
31 /* This enables re-computation of values. Current state: Unfinished and buggy. */
34 typedef struct _reloader_t reloader_t;
35 typedef struct _spill_info_t spill_info_t;
42 struct _spill_info_t {
43 ir_node *spilled_node;
44 reloader_t *reloaders;
47 typedef struct _spill_ctx_t {
48 ir_node *spilled; /**< The spilled node. */
49 ir_node *user; /**< The node this spill is for. */
50 ir_node *spill; /**< The spill itself. */
54 const arch_register_class_t *cls;
55 const be_chordal_env_t *chordal_env;
58 set *spills; /**< all spill_info_t's, which must be placed */
59 pset *mem_phis; /**< set of all special spilled phis. allocated and freed separately */
60 decide_irn_t is_mem_phi; /**< callback func to decide if a phi needs special spilling */
61 void *data; /**< data passed to all callbacks */
62 DEBUG_ONLY(firm_dbg_module_t *dbg;)
65 static int cmp_spillctx(const void *a, const void *b, size_t n) {
66 const spill_ctx_t *p = a;
67 const spill_ctx_t *q = b;
68 return !(p->user == q->user && p->spilled == q->spilled);
71 static int cmp_spillinfo(const void *x, const void *y, size_t size) {
72 const spill_info_t *xx = x;
73 const spill_info_t *yy = y;
74 return ! (xx->spilled_node == yy->spilled_node);
78 void be_set_spill_env_dbg_module(spill_env_t *env, firm_dbg_module_t *dbg) {
83 spill_env_t *be_new_spill_env(const be_chordal_env_t *chordal_env, decide_irn_t is_mem_phi, void *data) {
84 spill_env_t *env = xmalloc(sizeof(env[0]));
85 env->spill_ctxs = new_set(cmp_spillctx, 1024);
86 env->spills = new_set(cmp_spillinfo, 1024);
87 env->cls = chordal_env->cls;
88 env->is_mem_phi = is_mem_phi;
90 env->chordal_env = chordal_env;
91 obstack_init(&env->obst);
95 void be_delete_spill_env(spill_env_t *senv) {
96 del_set(senv->spill_ctxs);
97 del_set(senv->spills);
98 obstack_free(&senv->obst, NULL);
102 static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
105 templ.spilled = to_spill;
106 templ.user = ctx_irn;
109 return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
112 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
114 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
116 ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
118 const be_main_env_t *env = senv->chordal_env->birg->main_env;
119 ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
126 * If the first usage of a phi result would be out of memory
127 * there is no sense in allocating a register for it.
128 * Thus we spill it and all its operands to the same spill slot.
129 * Therefore the phi/dataB becomes a phi/Memory
131 static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
132 int i, n = get_irn_arity(phi);
133 ir_node **ins, *bl = get_nodes_block(phi);
134 ir_graph *irg = senv->chordal_env->irg;
138 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
140 /* search an existing spill for this context */
141 ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
143 /* if not found spill the phi */
145 /* build a new PhiM with dummy in-array */
146 NEW_ARR_A(ir_node *, ins, n);
148 ins[i] = new_r_Unknown(irg, mode_M);
149 ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
151 /* re-wire the phiM */
153 ir_node *arg = get_irn_n(phi, i);
156 if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
157 sub_res = be_spill_phi(senv, arg, ctx_irn);
159 sub_res = be_spill_irn(senv, arg, ctx_irn);
161 set_irn_n(ctx->spill, i, sub_res);
167 static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
169 if (pset_find_ptr(senv->mem_phis, to_spill))
170 res = be_spill_phi(senv, to_spill, to_spill);
172 res = be_spill_irn(senv, to_spill, to_spill);
177 static void phi_walker(ir_node *irn, void *env) {
178 spill_env_t *senv = env;
179 const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
181 if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
182 && senv->is_mem_phi(irn, senv->data)) {
183 DBG((senv->dbg, LEVEL_1, " %+F\n", irn));
184 pset_insert_ptr(senv->mem_phis, irn);
192 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
195 /* check for 'normal' spill and general remat condition */
196 if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
199 /* check availability of original arguments */
200 if (is_Block(reloader)) {
202 /* we want to remat at the end of a block.
203 * thus all arguments must be alive at the end of the block
205 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
206 ir_node *arg = get_irn_n(spilled, pos);
207 if (!is_live_end(reloader, arg))
213 /* we want to remat before the insn reloader
214 * thus an arguments is alive if
215 * - it interferes with the reloaders result
217 * - or it is (last-) used by reloader itself
219 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
220 ir_node *arg = get_irn_n(spilled, pos);
223 if (values_interfere(reloader, arg))
226 for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
227 ir_node *rel_arg = get_irn_n(reloader, i);
232 /* arg is not alive before reloader */
244 #else /* BUGGY_REMAT */
246 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
247 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
249 return get_irn_arity(spilled) == 0 &&
250 be_is_Spill(spill) &&
251 arch_irn_is(aenv, spilled, rematerializable);
254 #endif /* BUGGY_REMAT */
256 static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
258 ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
260 /* recompute the value */
261 res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
263 get_irn_mode(spilled),
264 get_irn_arity(spilled),
265 get_irn_in(spilled));
266 copy_node_attr(spilled, res);
268 DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
270 /* insert in schedule */
271 if (is_Block(reloader)) {
272 ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
273 sched_add_after(insert, res);
275 sched_add_before(reloader, res);
283 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
284 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
285 ir_graph *irg = senv->chordal_env->irg;
289 /* get all special spilled phis */
290 DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
291 senv->mem_phis = pset_new_ptr_default();
292 irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
294 /* Add reloads for mem_phis */
295 /* BETTER: These reloads (1) should only be inserted, if they are really needed */
296 DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
297 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
299 DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
300 foreach_out_edge(irn, e) {
301 ir_node *user = e->src;
302 if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
303 ir_node *use_bl = get_nodes_block(user);
304 DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
305 be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
310 /* process each spilled node */
311 DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
312 for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
314 ir_mode *mode = get_irn_mode(si->spilled_node);
315 pset *values = pset_new_ptr(16);
317 /* go through all reloads for this spill */
318 for(rld = si->reloaders; rld; rld = rld->next) {
321 /* the spill for this reloader */
322 ir_node *spill = be_spill_node(senv, si->spilled_node);
325 if (check_remat_conditions(senv, spill, si->spilled_node, rld->reloader))
326 new_val = do_remat(senv, si->spilled_node, rld->reloader);
330 new_val = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
332 DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
333 pset_insert_ptr(values, new_val);
335 pset_insert_ptr(reload_set, new_val);
338 /* introduce copies, rewire the uses */
339 assert(pset_count(values) > 0 && "???");
340 pset_insert_ptr(values, si->spilled_node);
341 be_ssa_constr_set_ignore(senv->chordal_env->dom_front, values, senv->mem_phis);
346 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
348 for(i = 0, n = get_irn_arity(irn); i < n; ++i)
349 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
353 del_pset(senv->mem_phis);
356 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
357 spill_info_t templ, *res;
360 templ.spilled_node = to_spill;
361 templ.reloaders = NULL;
362 res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
364 rel = obstack_alloc(&senv->obst, sizeof(rel[0]));
365 rel->reloader = before;
366 rel->next = res->reloaders;
367 res->reloaders = rel;
370 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
371 ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
372 be_add_reload(senv, to_spill, insert_bl);
377 /****************************************
379 SPILL SLOT MANAGEMENT AND OPTS
381 ****************************************/
383 typedef struct _spill_slot_t {
387 ir_mode *largest_mode; /* the mode of all members with largest size */
390 typedef struct _ss_env_t {
392 be_chordal_env_t *cenv;
393 pmap *slots; /* maps spill_contexts to spill_slots */
394 pmap *types; /* maps modes to types */
395 DEBUG_ONLY(firm_dbg_module_t *dbg;)
400 * Walker: compute the spill slots
402 static void compute_spill_slots_walker(ir_node *spill, void *env) {
403 ss_env_t *ssenv = env;
408 if (!be_is_Spill(spill))
411 /* check, if this spill is for a context already known */
412 ctx = be_get_Spill_context(spill);
413 entry = pmap_find(ssenv->slots, ctx);
416 struct _arch_env_t *arch_env = ssenv->cenv->birg->main_env->arch_env;
417 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
418 ir_mode *largest_mode = arch_register_class_mode(cls);
420 /* this is a new spill context */
421 ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
422 ss->members = pset_new_ptr(8);
423 ss->largest_mode = largest_mode;
424 ss->size = get_mode_size_bytes(ss->largest_mode);
425 ss->align = arch_isa_get_reg_class_alignment(arch_env->isa, cls);
426 pmap_insert(ssenv->slots, ctx, ss);
428 /* values with the same spill_ctx must go into the same spill slot */
432 /* ugly mega assert :-) */
435 struct _arch_env_t *arch_env = ssenv->cenv->birg->main_env->arch_env;
436 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, spill, be_pos_Spill_val);
437 int size = get_mode_size_bytes(arch_register_class_mode(cls));
438 assert(ss->size == size && "Different sizes for the same spill slot are not allowed.");
439 for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
440 /* use values_interfere here, because it uses the dominance check,
441 which does work for values in memory */
442 assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
448 pset_insert_ptr(ss->members, spill);
452 * qsort compare function, sort spill slots by size.
454 static int ss_sorter(const void *v1, const void *v2) {
455 const spill_slot_t *ss1 = v1;
456 const spill_slot_t *ss2 = v2;
457 return ((int) ss2->size) - ((int) ss1->size);
462 * This function should optimize the spill slots.
463 * - Coalescing of multiple slots
464 * - Ordering the slots
466 * Input slots are in @p ssenv->slots
467 * @p size The count of initial spill slots in @p ssenv->slots
468 * This also is the size of the preallocated array @p ass
470 * @return An array of spill slots @p ass in specific order
472 static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
473 int i, o, used_slots;
477 pmap_foreach(ssenv->slots, entr)
478 ass[i++] = entr->value;
480 /* Sort the array to minimize fragmentation and cache footprint.
481 Large slots come first */
482 qsort(ass, size, sizeof(ass[0]), ss_sorter);
484 /* For each spill slot:
485 - assign a new offset to this slot
486 - xor find another slot to coalesce with */
488 for (i=0; i<size; ++i) { /* for each spill slot */
492 DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
493 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
494 DBG((ssenv->dbg, LEVEL_1, " %+F\n", n1));
497 for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
498 /* check inter-slot-pairs for interference */
500 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
501 for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
502 if(values_interfere(n1, n2)) {
503 pset_break(ass[i]->members);
504 pset_break(ass[o]->members);
505 DBG((ssenv->dbg, LEVEL_1, " Interf %+F -- %+F\n", n1, n2));
506 goto interf_detected;
509 /* if we are here, there is no interference between ass[i] and ass[o] */
512 interf_detected: /*nothing*/ ;
515 /* now the members of ass[i] join the members of ass[tgt_slot] */
517 /* do we need a new slot? */
518 if (tgt_slot == -1) {
519 tgt_slot = used_slots;
524 ass[tgt_slot]->size = ass[i]->size;
525 del_pset(ass[tgt_slot]->members);
526 ass[tgt_slot]->members = pset_new_ptr(8);
530 /* copy the members to the target pset */
531 /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
533 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
534 pset_insert_ptr(ass[tgt_slot]->members, n1);
538 #define ALIGN_SPILL_AREA 16
539 #define pset_foreach(pset, elm) for(elm=pset_first(pset); elm; elm=pset_next(pset))
542 * Returns a spill type for a mode. Keep them in a map to reduce
543 * the number of types.
545 * @param types a map containing all created types
546 * @param ss the spill slot
548 * Note that type types should are identical for every mode.
549 * This rule might break if two different register classes return the same
550 * mode but different alignments.
552 static ir_type *get_spill_type(pmap *types, spill_slot_t *ss) {
553 pmap_entry *e = pmap_find(types, ss->largest_mode);
558 snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(ss->largest_mode));
559 res = new_type_primitive(new_id_from_str(buf), ss->largest_mode);
560 set_type_alignment_bytes(res, ss->align);
561 pmap_insert(types, ss->largest_mode, res);
565 assert(get_type_alignment_bytes(res) == (int)ss->align);
571 * Create spill slot entities on the frame type.
573 * @param ssenv the spill environment
574 * @param n number of spill slots
575 * @param ss array of spill slots
577 static void assign_entities(ss_env_t *ssenv, int n_slots, spill_slot_t *ss[]) {
578 int i, offset, frame_align;
579 ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
581 /* aligning by increasing frame size */
582 offset = get_type_size_bits(frame) / 8;
583 offset = round_up2(offset, ALIGN_SPILL_AREA);
584 set_type_size_bytes(frame, -1);
586 /* create entities and assign offsets according to size and alignment*/
587 for (i = 0; i < n_slots; ++i) {
594 snprintf(buf, sizeof(buf), "spill_slot_%d", i);
595 name = new_id_from_str(buf);
597 spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]));
600 offset = round_up2(offset, ss[i]->align);
602 set_entity_offset_bytes(spill_ent, offset);
603 /* next possible offset */
604 offset += round_up2(ss[i]->size, ss[i]->align);
606 pset_foreach(ss[i]->members, irn)
607 be_set_Spill_entity(irn, spill_ent);
610 /* set final size of stack frame */
611 frame_align = get_type_alignment_bytes(frame);
612 set_type_size_bytes(frame, round_up2(offset, frame_align));
615 void be_compute_spill_offsets(be_chordal_env_t *cenv) {
621 obstack_init(&ssenv.ob);
623 ssenv.slots = pmap_create();
624 ssenv.types = pmap_create();
625 FIRM_DBG_REGISTER(ssenv.dbg, "ir.be.spillslots");
627 /* Get initial spill slots */
628 irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
630 /* Build an empty array for optimized spill slots */
631 ss_size = pmap_count(ssenv.slots);
632 ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
633 optimize_slots(&ssenv, ss_size, ss);
635 /* Integrate slots into the stack frame entity */
636 assign_entities(&ssenv, ss_size, ss);
639 pmap_foreach(ssenv.slots, pme)
640 del_pset(((spill_slot_t *)pme->value)->members);
641 pmap_destroy(ssenv.slots);
642 pmap_destroy(ssenv.types);
643 obstack_free(&ssenv.ob, NULL);
645 be_copy_entities_to_reloads(cenv->irg);