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 firm_dbg_module_t *dbg;
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 seperately */
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 */
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(firm_dbg_module_t *dbg,
78 const be_chordal_env_t *chordal_env,
79 decide_irn_t is_mem_phi, void *data) {
81 spill_env_t *env = xmalloc(sizeof(env[0]));
82 env->spill_ctxs = new_set(cmp_spillctx, 1024);
83 env->spills = new_set(cmp_spillinfo, 1024);
84 env->cls = chordal_env->cls;
86 env->is_mem_phi = is_mem_phi;
88 env->chordal_env = chordal_env;
89 obstack_init(&env->obst);
93 void be_delete_spill_env(spill_env_t *senv) {
94 del_set(senv->spill_ctxs);
95 del_set(senv->spills);
96 obstack_free(&senv->obst, NULL);
100 static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
103 templ.spilled = to_spill;
104 templ.user = ctx_irn;
107 return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
110 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
112 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
114 ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
116 const be_main_env_t *env = senv->chordal_env->birg->main_env;
117 ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
124 * If the first usage of a phi result would be out of memory
125 * there is no sense in allocating a register for it.
126 * Thus we spill it and all its operands to the same spill slot.
127 * Therefore the phi/dataB becomes a phi/Memory
129 static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
130 int i, n = get_irn_arity(phi);
131 ir_node **ins, *bl = get_nodes_block(phi);
132 ir_graph *irg = senv->chordal_env->irg;
136 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
138 /* search an existing spill for this context */
139 ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
141 /* if not found spill the phi */
143 /* build a new PhiM with dummy in-array */
144 NEW_ARR_A(ir_node *, ins, n);
146 ins[i] = new_r_Unknown(irg, mode_M);
147 ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
149 /* re-wire the phiM */
151 ir_node *arg = get_irn_n(phi, i);
154 if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
155 sub_res = be_spill_phi(senv, arg, ctx_irn);
157 sub_res = be_spill_irn(senv, arg, ctx_irn);
159 set_irn_n(ctx->spill, i, sub_res);
165 static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
167 if (pset_find_ptr(senv->mem_phis, to_spill))
168 res = be_spill_phi(senv, to_spill, to_spill);
170 res = be_spill_irn(senv, to_spill, to_spill);
175 static void phi_walker(ir_node *irn, void *env) {
176 spill_env_t *senv = env;
177 const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
179 if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
180 && senv->is_mem_phi(irn, senv->data)) {
181 DBG((senv->dbg, LEVEL_1, " %+F\n", irn));
182 pset_insert_ptr(senv->mem_phis, irn);
190 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
193 /* check for 'normal' spill and general remat condition */
194 if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
197 /* check availability of original arguments */
198 if (is_Block(reloader)) {
200 /* we want to remat at the end of a block.
201 * thus all arguments must be alive at the end of the block
203 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
204 ir_node *arg = get_irn_n(spilled, pos);
205 if (!is_live_end(reloader, arg))
211 /* we want to remat before the insn reloader
212 * thus an arguments is alive if
213 * - it interferes with the reloaders result
215 * - or it is (last-) used by reloader itself
217 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
218 ir_node *arg = get_irn_n(spilled, pos);
221 if (values_interfere(reloader, arg))
224 for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
225 ir_node *rel_arg = get_irn_n(reloader, i);
230 /* arg is not alive before reloader */
242 #else /* BUGGY_REMAT */
244 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
245 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
247 return get_irn_arity(spilled) == 0 &&
248 be_is_Spill(spill) &&
249 arch_irn_is(aenv, spilled, rematerializable);
252 #endif /* BUGGY_REMAT */
254 static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
256 ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
258 /* recompute the value */
259 res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
261 get_irn_mode(spilled),
262 get_irn_arity(spilled),
263 get_irn_in(spilled));
264 copy_node_attr(spilled, res);
266 DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
268 /* insert in schedule */
269 if (is_Block(reloader)) {
270 ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
271 sched_add_after(insert, res);
273 sched_add_before(reloader, res);
281 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
282 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
283 ir_graph *irg = senv->chordal_env->irg;
287 /* get all special spilled phis */
288 DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
289 senv->mem_phis = pset_new_ptr_default();
290 irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
292 /* Add reloads for mem_phis */
293 /* BETTER: These reloads (1) should only be inserted, if they are really needed */
294 DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
295 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
297 DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
298 foreach_out_edge(irn, e) {
299 ir_node *user = e->src;
300 if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
301 ir_node *use_bl = get_nodes_block(user);
302 DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
303 be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
308 /* process each spilled node */
309 DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
310 for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
312 ir_mode *mode = get_irn_mode(si->spilled_node);
313 pset *values = pset_new_ptr(16);
315 /* go through all reloads for this spill */
316 for(rld = si->reloaders; rld; rld = rld->next) {
319 /* the spill for this reloader */
320 ir_node *spill = be_spill_node(senv, si->spilled_node);
323 if (check_remat_conditions(senv, spill, si->spilled_node, rld->reloader))
324 new_val = do_remat(senv, si->spilled_node, rld->reloader);
328 new_val = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
330 DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
331 pset_insert_ptr(values, new_val);
333 pset_insert_ptr(reload_set, new_val);
336 /* introduce copies, rewire the uses */
337 assert(pset_count(values) > 0 && "???");
338 pset_insert_ptr(values, si->spilled_node);
339 be_ssa_constr_set_ignore(senv->chordal_env->dom_front, values, senv->mem_phis);
344 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
346 for(i = 0, n = get_irn_arity(irn); i < n; ++i)
347 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
351 del_pset(senv->mem_phis);
354 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
355 spill_info_t templ, *res;
358 templ.spilled_node = to_spill;
359 templ.reloaders = NULL;
360 res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
362 rel = obstack_alloc(&senv->obst, sizeof(rel[0]));
363 rel->reloader = before;
364 rel->next = res->reloaders;
365 res->reloaders = rel;
368 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
369 ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
370 be_add_reload(senv, to_spill, insert_bl);
375 /****************************************
377 SPILL SLOT MANAGEMENT AND OPTS
379 ****************************************/
381 typedef struct _spill_slot_t {
385 ir_mode *largest_mode; /* the mode of all members with largest size */
388 typedef struct _ss_env_t {
389 firm_dbg_module_t *dbg;
391 be_chordal_env_t *cenv;
392 pmap *slots; /* maps spill_contexts to spill_slots */
393 pmap *types; /* maps modes to types */
397 static void compute_spill_slots_walker(ir_node *spill, void *env) {
398 ss_env_t *ssenv = env;
403 if (!be_is_Spill(spill))
406 /* check, if this spill is for a context already known */
407 ctx = be_get_Spill_context(spill);
408 entry = pmap_find(ssenv->slots, ctx);
411 /* this is a new spill context */
412 ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
413 ss->members = pset_new_ptr(8);
414 ss->largest_mode = get_irn_mode(get_irn_n(spill, be_pos_Spill_val));
415 ss->size = get_mode_size_bytes(ss->largest_mode);
416 ss->align = ss->size; /* TODO Assumed for now */
417 pmap_insert(ssenv->slots, ctx, ss);
420 /* values with the same spill_ctx must go into the same spill slot */
422 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.");
423 for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
424 /* use values_interfere here, because it uses the dominance check,
425 which does work for values in memory */
426 assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
430 pset_insert_ptr(ss->members, spill);
433 static int ss_sorter(const void *v1, const void *v2) {
434 const spill_slot_t *ss1 = v1;
435 const spill_slot_t *ss2 = v2;
436 return ((int) ss2->size) - ((int) ss1->size);
441 * This function should optimize the spill slots.
442 * - Coalescing of multiple slots
443 * - Ordering the slots
445 * Input slots are in @p ssenv->slots
446 * @p size The count of initial spill slots in @p ssenv->slots
447 * This also is the size of the preallocated array @p ass
449 * @return An array of spill slots @p ass in specific order
451 static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
452 int i, o, used_slots;
456 pmap_foreach(ssenv->slots, entr)
457 ass[i++] = entr->value;
459 /* Sort the array to minimize fragmentation and cache footprint.
460 Large slots come first */
461 qsort(ass, size, sizeof(ass[0]), ss_sorter);
463 /* For each spill slot:
464 - assign a new offset to this slot
465 - xor find another slot to coalesce with */
467 for (i=0; i<size; ++i) { /* for each spill slot */
471 DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
472 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
473 DBG((ssenv->dbg, LEVEL_1, " %+F\n", n1));
476 for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
477 /* check inter-slot-pairs for interference */
479 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
480 for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
481 if(values_interfere(n1, n2)) {
482 pset_break(ass[i]->members);
483 pset_break(ass[o]->members);
484 DBG((ssenv->dbg, LEVEL_1, " Interf %+F -- %+F\n", n1, n2));
485 goto interf_detected;
488 /* if we are here, there is no interference between ass[i] and ass[o] */
491 interf_detected: /*nothing*/ ;
494 /* now the members of ass[i] join the members of ass[tgt_slot] */
496 /* do we need a new slot? */
497 if (tgt_slot == -1) {
498 tgt_slot = used_slots;
503 ass[tgt_slot]->size = ass[i]->size;
504 del_pset(ass[tgt_slot]->members);
505 ass[tgt_slot]->members = pset_new_ptr(8);
509 /* copy the members to the target pset */
510 /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
512 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
513 pset_insert_ptr(ass[tgt_slot]->members, n1);
517 #define ALIGN_SPILL_AREA 16
518 #define pset_foreach(pset, elm) for(elm=pset_first(pset); elm; elm=pset_next(pset))
521 * Returns a spill type for a mode. Keep them in a map to reduce
522 * the number of types.
524 static ir_type *get_spill_type(pmap *types, ir_mode *mode) {
525 pmap_entry *e = pmap_find(types, mode);
530 snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(mode));
531 res = new_type_primitive(new_id_from_str(buf), mode);
532 pmap_insert(types, mode, res);
539 static void assign_entities(ss_env_t *ssenv, int n, spill_slot_t **ss) {
541 ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
543 /* aligning by increasing frame size */
544 offset = get_type_size_bits(frame) / 8;
545 offset = round_up2(offset, ALIGN_SPILL_AREA);
546 set_type_size_bytes(frame, -1);
548 /* create entities and assign offsets according to size and alignment*/
549 for (i=0; i<n; ++i) {
556 snprintf(buf, sizeof(buf), "spill_slot_%d", i);
557 name = new_id_from_str(buf);
559 spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]->largest_mode));
562 offset = round_up2(offset, ss[i]->align);
564 set_entity_offset_bytes(spill_ent, offset);
565 /* next possible offset */
566 offset += ss[i]->size;
568 pset_foreach(ss[i]->members, irn)
569 be_set_Spill_entity(irn, spill_ent);
572 /* set final size of stack frame */
573 set_type_size_bytes(frame, offset);
576 void be_compute_spill_offsets(be_chordal_env_t *cenv) {
582 obstack_init(&ssenv.ob);
584 ssenv.slots = pmap_create();
585 ssenv.types = pmap_create();
586 FIRM_DBG_REGISTER(ssenv.dbg, "ir.be.spillslots");
588 /* Get initial spill slots */
589 irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
591 /* Build an empty array for optimized spill slots */
592 ss_size = pmap_count(ssenv.slots);
593 ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
594 optimize_slots(&ssenv, ss_size, ss);
596 /* Integrate slots into the stack frame entity */
597 assign_entities(&ssenv, ss_size, ss);
600 pmap_foreach(ssenv.slots, pme)
601 del_pset(((spill_slot_t *)pme->value)->members);
602 pmap_destroy(ssenv.slots);
603 pmap_destroy(ssenv.types);
604 obstack_free(&ssenv.ob, NULL);
606 be_copy_entities_to_reloads(cenv->irg);