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 seperately */
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);
77 spill_env_t *be_new_spill_env(const be_chordal_env_t *chordal_env, decide_irn_t is_mem_phi, void *data) {
78 spill_env_t *env = xmalloc(sizeof(env[0]));
79 env->spill_ctxs = new_set(cmp_spillctx, 1024);
80 env->spills = new_set(cmp_spillinfo, 1024);
81 env->cls = chordal_env->cls;
82 env->is_mem_phi = is_mem_phi;
84 env->chordal_env = chordal_env;
85 obstack_init(&env->obst);
89 void be_delete_spill_env(spill_env_t *senv) {
90 del_set(senv->spill_ctxs);
91 del_set(senv->spills);
92 obstack_free(&senv->obst, NULL);
96 static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
99 templ.spilled = to_spill;
100 templ.user = ctx_irn;
103 return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
106 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
108 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
110 ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
112 const be_main_env_t *env = senv->chordal_env->birg->main_env;
113 ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
120 * If the first usage of a phi result would be out of memory
121 * there is no sense in allocating a register for it.
122 * Thus we spill it and all its operands to the same spill slot.
123 * Therefore the phi/dataB becomes a phi/Memory
125 static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
126 int i, n = get_irn_arity(phi);
127 ir_node **ins, *bl = get_nodes_block(phi);
128 ir_graph *irg = senv->chordal_env->irg;
132 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
134 /* search an existing spill for this context */
135 ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
137 /* if not found spill the phi */
139 /* build a new PhiM with dummy in-array */
140 NEW_ARR_A(ir_node *, ins, n);
142 ins[i] = new_r_Unknown(irg, mode_M);
143 ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
145 /* re-wire the phiM */
147 ir_node *arg = get_irn_n(phi, i);
150 if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
151 sub_res = be_spill_phi(senv, arg, ctx_irn);
153 sub_res = be_spill_irn(senv, arg, ctx_irn);
155 set_irn_n(ctx->spill, i, sub_res);
161 static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
163 if (pset_find_ptr(senv->mem_phis, to_spill))
164 res = be_spill_phi(senv, to_spill, to_spill);
166 res = be_spill_irn(senv, to_spill, to_spill);
171 static void phi_walker(ir_node *irn, void *env) {
172 spill_env_t *senv = env;
173 const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
175 if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
176 && senv->is_mem_phi(irn, senv->data)) {
177 DBG((senv->dbg, LEVEL_1, " %+F\n", irn));
178 pset_insert_ptr(senv->mem_phis, irn);
186 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
189 /* check for 'normal' spill and general remat condition */
190 if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
193 /* check availability of original arguments */
194 if (is_Block(reloader)) {
196 /* we want to remat at the end of a block.
197 * thus all arguments must be alive at the end of the block
199 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
200 ir_node *arg = get_irn_n(spilled, pos);
201 if (!is_live_end(reloader, arg))
207 /* we want to remat before the insn reloader
208 * thus an arguments is alive if
209 * - it interferes with the reloaders result
211 * - or it is (last-) used by reloader itself
213 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
214 ir_node *arg = get_irn_n(spilled, pos);
217 if (values_interfere(reloader, arg))
220 for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
221 ir_node *rel_arg = get_irn_n(reloader, i);
226 /* arg is not alive before reloader */
238 #else /* BUGGY_REMAT */
240 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
241 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
243 return get_irn_arity(spilled) == 0 &&
244 be_is_Spill(spill) &&
245 arch_irn_is(aenv, spilled, rematerializable);
248 #endif /* BUGGY_REMAT */
250 static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
252 ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
254 /* recompute the value */
255 res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
257 get_irn_mode(spilled),
258 get_irn_arity(spilled),
259 get_irn_in(spilled));
260 copy_node_attr(spilled, res);
262 DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
264 /* insert in schedule */
265 if (is_Block(reloader)) {
266 ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
267 sched_add_after(insert, res);
269 sched_add_before(reloader, res);
277 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
278 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
279 ir_graph *irg = senv->chordal_env->irg;
283 /* get all special spilled phis */
284 DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
285 senv->mem_phis = pset_new_ptr_default();
286 irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
288 /* Add reloads for mem_phis */
289 /* BETTER: These reloads (1) should only be inserted, if they are really needed */
290 DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
291 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
293 DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
294 foreach_out_edge(irn, e) {
295 ir_node *user = e->src;
296 if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
297 ir_node *use_bl = get_nodes_block(user);
298 DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
299 be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
304 /* process each spilled node */
305 DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
306 for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
308 ir_mode *mode = get_irn_mode(si->spilled_node);
309 pset *values = pset_new_ptr(16);
311 /* go through all reloads for this spill */
312 for(rld = si->reloaders; rld; rld = rld->next) {
315 /* the spill for this reloader */
316 ir_node *spill = be_spill_node(senv, si->spilled_node);
319 if (check_remat_conditions(senv, spill, si->spilled_node, rld->reloader))
320 new_val = do_remat(senv, si->spilled_node, rld->reloader);
324 new_val = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
326 DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
327 pset_insert_ptr(values, new_val);
329 pset_insert_ptr(reload_set, new_val);
332 /* introduce copies, rewire the uses */
333 assert(pset_count(values) > 0 && "???");
334 pset_insert_ptr(values, si->spilled_node);
335 be_ssa_constr_set_ignore(senv->chordal_env->dom_front, values, senv->mem_phis);
340 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
342 for(i = 0, n = get_irn_arity(irn); i < n; ++i)
343 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
347 del_pset(senv->mem_phis);
350 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
351 spill_info_t templ, *res;
354 templ.spilled_node = to_spill;
355 templ.reloaders = NULL;
356 res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
358 rel = obstack_alloc(&senv->obst, sizeof(rel[0]));
359 rel->reloader = before;
360 rel->next = res->reloaders;
361 res->reloaders = rel;
364 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
365 ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
366 be_add_reload(senv, to_spill, insert_bl);
371 /****************************************
373 SPILL SLOT MANAGEMENT AND OPTS
375 ****************************************/
377 typedef struct _spill_slot_t {
381 ir_mode *largest_mode; /* the mode of all members with largest size */
384 typedef struct _ss_env_t {
386 be_chordal_env_t *cenv;
387 pmap *slots; /* maps spill_contexts to spill_slots */
388 pmap *types; /* maps modes to types */
389 DEBUG_ONLY(firm_dbg_module_t *dbg;)
393 static void compute_spill_slots_walker(ir_node *spill, void *env) {
394 ss_env_t *ssenv = env;
399 if (!be_is_Spill(spill))
402 /* check, if this spill is for a context already known */
403 ctx = be_get_Spill_context(spill);
404 entry = pmap_find(ssenv->slots, ctx);
407 /* this is a new spill context */
408 ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
409 ss->members = pset_new_ptr(8);
410 ss->largest_mode = get_irn_mode(get_irn_n(spill, be_pos_Spill_val));
411 ss->size = get_mode_size_bytes(ss->largest_mode);
412 ss->align = ss->size; /* TODO Assumed for now */
413 pmap_insert(ssenv->slots, ctx, ss);
416 /* values with the same spill_ctx must go into the same spill slot */
418 assert(ss->size == (unsigned)get_mode_size_bytes(get_irn_mode(get_irn_n(spill, be_pos_Spill_val))) && "Different sizes for the same spill slot are not allowed yet.");
419 for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
420 /* use values_interfere here, because it uses the dominance check,
421 which does work for values in memory */
422 assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
426 pset_insert_ptr(ss->members, spill);
429 static int ss_sorter(const void *v1, const void *v2) {
430 const spill_slot_t *ss1 = v1;
431 const spill_slot_t *ss2 = v2;
432 return ((int) ss2->size) - ((int) ss1->size);
437 * This function should optimize the spill slots.
438 * - Coalescing of multiple slots
439 * - Ordering the slots
441 * Input slots are in @p ssenv->slots
442 * @p size The count of initial spill slots in @p ssenv->slots
443 * This also is the size of the preallocated array @p ass
445 * @return An array of spill slots @p ass in specific order
447 static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
448 int i, o, used_slots;
452 pmap_foreach(ssenv->slots, entr)
453 ass[i++] = entr->value;
455 /* Sort the array to minimize fragmentation and cache footprint.
456 Large slots come first */
457 qsort(ass, size, sizeof(ass[0]), ss_sorter);
459 /* For each spill slot:
460 - assign a new offset to this slot
461 - xor find another slot to coalesce with */
463 for (i=0; i<size; ++i) { /* for each spill slot */
467 DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
468 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
469 DBG((ssenv->dbg, LEVEL_1, " %+F\n", n1));
472 for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
473 /* check inter-slot-pairs for interference */
475 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
476 for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
477 if(values_interfere(n1, n2)) {
478 pset_break(ass[i]->members);
479 pset_break(ass[o]->members);
480 DBG((ssenv->dbg, LEVEL_1, " Interf %+F -- %+F\n", n1, n2));
481 goto interf_detected;
484 /* if we are here, there is no interference between ass[i] and ass[o] */
487 interf_detected: /*nothing*/ ;
490 /* now the members of ass[i] join the members of ass[tgt_slot] */
492 /* do we need a new slot? */
493 if (tgt_slot == -1) {
494 tgt_slot = used_slots;
499 ass[tgt_slot]->size = ass[i]->size;
500 del_pset(ass[tgt_slot]->members);
501 ass[tgt_slot]->members = pset_new_ptr(8);
505 /* copy the members to the target pset */
506 /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
508 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
509 pset_insert_ptr(ass[tgt_slot]->members, n1);
513 #define ALIGN_SPILL_AREA 16
514 #define pset_foreach(pset, elm) for(elm=pset_first(pset); elm; elm=pset_next(pset))
517 * Returns a spill type for a mode. Keep them in a map to reduce
518 * the number of types.
520 static ir_type *get_spill_type(pmap *types, ir_mode *mode) {
521 pmap_entry *e = pmap_find(types, mode);
526 snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(mode));
527 res = new_type_primitive(new_id_from_str(buf), mode);
528 pmap_insert(types, mode, res);
535 static void assign_entities(ss_env_t *ssenv, int n, spill_slot_t **ss) {
537 ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
539 /* aligning by increasing frame size */
540 offset = get_type_size_bits(frame) / 8;
541 offset = round_up2(offset, ALIGN_SPILL_AREA);
542 set_type_size_bytes(frame, -1);
544 /* create entities and assign offsets according to size and alignment*/
545 for (i=0; i<n; ++i) {
552 snprintf(buf, sizeof(buf), "spill_slot_%d", i);
553 name = new_id_from_str(buf);
555 spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]->largest_mode));
558 offset = round_up2(offset, ss[i]->align);
560 set_entity_offset_bytes(spill_ent, offset);
561 /* next possible offset */
562 offset += ss[i]->size;
564 pset_foreach(ss[i]->members, irn)
565 be_set_Spill_entity(irn, spill_ent);
568 /* set final size of stack frame */
569 set_type_size_bytes(frame, offset);
572 void be_compute_spill_offsets(be_chordal_env_t *cenv) {
578 obstack_init(&ssenv.ob);
580 ssenv.slots = pmap_create();
581 ssenv.types = pmap_create();
582 FIRM_DBG_REGISTER(ssenv.dbg, "ir.be.spillslots");
584 /* Get initial spill slots */
585 irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
587 /* Build an empty array for optimized spill slots */
588 ss_size = pmap_count(ssenv.slots);
589 ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
590 optimize_slots(&ssenv, ss_size, ss);
592 /* Integrate slots into the stack frame entity */
593 assign_entities(&ssenv, ss_size, ss);
596 pmap_foreach(ssenv.slots, pme)
597 del_pset(((spill_slot_t *)pme->value)->members);
598 pmap_destroy(ssenv.slots);
599 pmap_destroy(ssenv.types);
600 obstack_free(&ssenv.ob, NULL);
602 be_copy_entities_to_reloads(cenv->irg);