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"
30 /* This enables re-computation of values. Current state: Unfinished and buggy. */
33 typedef struct _reloader_t reloader_t;
34 typedef struct _spill_info_t spill_info_t;
41 struct _spill_info_t {
42 ir_node *spilled_node;
43 reloader_t *reloaders;
46 typedef struct _spill_ctx_t {
47 ir_node *spilled; /**< The spilled node. */
48 ir_node *user; /**< The node this spill is for. */
49 ir_node *spill; /**< The spill itself. */
53 firm_dbg_module_t *dbg;
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 */
64 static int cmp_spillctx(const void *a, const void *b, size_t n) {
65 const spill_ctx_t *p = a;
66 const spill_ctx_t *q = b;
67 return !(p->user == q->user && p->spilled == q->spilled);
70 static int cmp_spillinfo(const void *x, const void *y, size_t size) {
71 const spill_info_t *xx = x;
72 const spill_info_t *yy = y;
73 return ! (xx->spilled_node == yy->spilled_node);
76 spill_env_t *be_new_spill_env(firm_dbg_module_t *dbg,
77 const be_chordal_env_t *chordal_env,
78 decide_irn_t is_mem_phi, void *data) {
80 spill_env_t *env = xmalloc(sizeof(env[0]));
81 env->spill_ctxs = new_set(cmp_spillctx, 1024);
82 env->spills = new_set(cmp_spillinfo, 1024);
83 env->cls = chordal_env->cls;
85 env->is_mem_phi = is_mem_phi;
87 env->chordal_env = chordal_env;
88 obstack_init(&env->obst);
92 void be_delete_spill_env(spill_env_t *senv) {
93 del_set(senv->spill_ctxs);
94 del_set(senv->spills);
95 obstack_free(&senv->obst, NULL);
99 static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
102 templ.spilled = to_spill;
103 templ.user = ctx_irn;
106 return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
109 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
111 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
113 ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
115 const be_main_env_t *env = senv->chordal_env->birg->main_env;
116 ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
123 * If the first usage of a phi result would be out of memory
124 * there is no sense in allocating a register for it.
125 * Thus we spill it and all its operands to the same spill slot.
126 * Therefore the phi/dataB becomes a phi/Memory
128 static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
129 int i, n = get_irn_arity(phi);
130 ir_node **ins, *bl = get_nodes_block(phi);
131 ir_graph *irg = senv->chordal_env->irg;
135 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
137 /* search an existing spill for this context */
138 ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
140 /* if not found spill the phi */
142 /* build a new PhiM with dummy in-array */
143 NEW_ARR_A(ir_node *, ins, n);
145 ins[i] = new_r_Unknown(irg, mode_M);
146 ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
148 /* re-wire the phiM */
150 ir_node *arg = get_irn_n(phi, i);
153 if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
154 sub_res = be_spill_phi(senv, arg, ctx_irn);
156 sub_res = be_spill_irn(senv, arg, ctx_irn);
158 set_irn_n(ctx->spill, i, sub_res);
164 static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
166 if (pset_find_ptr(senv->mem_phis, to_spill))
167 res = be_spill_phi(senv, to_spill, to_spill);
169 res = be_spill_irn(senv, to_spill, to_spill);
174 static void phi_walker(ir_node *irn, void *env) {
175 spill_env_t *senv = env;
176 const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
178 if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
179 && senv->is_mem_phi(irn, senv->data)) {
180 DBG((senv->dbg, LEVEL_1, " %+F\n", irn));
181 pset_insert_ptr(senv->mem_phis, irn);
188 static int check_remat_conditions(spill_env_t *senv, ir_node *spill, ir_node *spilled, ir_node *reloader) {
191 /* check for 'normal' spill and general remat condition */
192 if (!be_is_Spill(spill) || !arch_irn_is(senv->chordal_env->birg->main_env->arch_env, spilled, rematerializable))
195 /* check availability of original arguments */
196 if (is_Block(reloader)) {
198 /* we want to remat at the end of a block.
199 * thus all arguments must be alive at the end of the block
201 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
202 ir_node *arg = get_irn_n(spilled, pos);
203 if (!is_live_end(reloader, arg))
209 /* we want to remat before the insn reloader
210 * thus an arguments is alive if
211 * - it interferes with the reloaders result
213 * - or it is (last-) used by reloader itself
215 for (pos=0, max=get_irn_arity(spilled); pos<max; ++pos) {
216 ir_node *arg = get_irn_n(spilled, pos);
219 if (values_interfere(reloader, arg))
222 for (i=0, m=get_irn_arity(reloader); i<m; ++i) {
223 ir_node *rel_arg = get_irn_n(reloader, i);
228 /* arg is not alive before reloader */
240 static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader) {
242 ir_node *bl = (is_Block(reloader)) ? reloader : get_nodes_block(reloader);
244 /* recompute the value */
245 res = new_ir_node(get_irn_dbg_info(spilled), senv->chordal_env->irg, bl,
247 get_irn_mode(spilled),
248 get_irn_arity(spilled),
249 get_irn_in(spilled));
251 DBG((senv->dbg, LEVEL_1, "Insert remat %+F before reloader %+F\n", res, reloader));
253 /* insert in schedule */
254 if (is_Block(reloader)) {
255 ir_node *insert = sched_skip(reloader, 0, sched_skip_cf_predicator, (void *) senv->chordal_env->birg->main_env->arch_env);
256 sched_add_after(insert, res);
258 sched_add_before(reloader, res);
266 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
267 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
268 ir_graph *irg = senv->chordal_env->irg;
272 /* get all special spilled phis */
273 DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
274 senv->mem_phis = pset_new_ptr_default();
275 irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
277 /* Add reloads for mem_phis */
278 /* BETTER: These reloads (1) should only be inserted, if they are really needed */
279 DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
280 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
282 DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
283 foreach_out_edge(irn, e) {
284 ir_node *user = e->src;
285 if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
286 ir_node *use_bl = get_nodes_block(user);
287 DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
288 be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
293 /* process each spilled node */
294 DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
295 for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
297 ir_mode *mode = get_irn_mode(si->spilled_node);
298 pset *values = pset_new_ptr(16);
300 /* go through all reloads for this spill */
301 for(rld = si->reloaders; rld; rld = rld->next) {
304 /* the spill for this reloader */
305 ir_node *spill = be_spill_node(senv, si->spilled_node);
308 if (check_remat_conditions(senv, spill, si->spilled_node, rld->reloader))
309 new_val = do_remat(senv, si->spilled_node, rld->reloader);
313 new_val = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
315 DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", new_val, si->spilled_node, rld->reloader));
316 pset_insert_ptr(values, new_val);
318 pset_insert_ptr(reload_set, new_val);
321 /* introduce copies, rewire the uses */
322 assert(pset_count(values) > 0 && "???");
323 pset_insert_ptr(values, si->spilled_node);
324 be_ssa_constr_set_ignore(senv->chordal_env->dom_front, values, senv->mem_phis);
329 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
331 for(i = 0, n = get_irn_arity(irn); i < n; ++i)
332 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
336 del_pset(senv->mem_phis);
339 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
340 spill_info_t templ, *res;
343 templ.spilled_node = to_spill;
344 templ.reloaders = NULL;
345 res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
347 rel = obstack_alloc(&senv->obst, sizeof(rel[0]));
348 rel->reloader = before;
349 rel->next = res->reloaders;
350 res->reloaders = rel;
353 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
354 ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
355 be_add_reload(senv, to_spill, insert_bl);
360 /****************************************
362 SPILL SLOT MANAGEMENT AND OPTS
364 ****************************************/
366 typedef struct _spill_slot_t {
370 ir_mode *largest_mode; /* the mode of all members with largest size */
373 typedef struct _ss_env_t {
374 firm_dbg_module_t *dbg;
376 be_chordal_env_t *cenv;
377 pmap *slots; /* maps spill_contexts to spill_slots */
378 pmap *types; /* maps modes to types */
382 static void compute_spill_slots_walker(ir_node *spill, void *env) {
383 ss_env_t *ssenv = env;
388 if (!be_is_Spill(spill))
391 /* check, if this spill is for a context already known */
392 ctx = be_get_Spill_context(spill);
393 entry = pmap_find(ssenv->slots, ctx);
396 /* this is a new spill context */
397 ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
398 ss->members = pset_new_ptr(8);
399 ss->largest_mode = get_irn_mode(get_irn_n(spill, be_pos_Spill_val));
400 ss->size = get_mode_size_bytes(ss->largest_mode);
401 ss->align = ss->size; /* TODO Assumed for now */
402 pmap_insert(ssenv->slots, ctx, ss);
405 /* values with the same spill_ctx must go into the same spill slot */
407 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.");
408 for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
409 /* use values_interfere here, because it uses the dominance check,
410 which does work for values in memory */
411 assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
415 pset_insert_ptr(ss->members, spill);
418 static int ss_sorter(const void *v1, const void *v2) {
419 const spill_slot_t *ss1 = v1;
420 const spill_slot_t *ss2 = v2;
421 return ((int) ss2->size) - ((int) ss1->size);
426 * This function should optimize the spill slots.
427 * - Coalescing of multiple slots
428 * - Ordering the slots
430 * Input slots are in @p ssenv->slots
431 * @p size The count of initial spill slots in @p ssenv->slots
432 * This also is the size of the preallocated array @p ass
434 * @return An array of spill slots @p ass in specific order
436 static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
437 int i, o, used_slots;
441 pmap_foreach(ssenv->slots, entr)
442 ass[i++] = entr->value;
444 /* Sort the array to minimize fragmentation and cache footprint.
445 Large slots come first */
446 qsort(ass, size, sizeof(ass[0]), ss_sorter);
448 /* For each spill slot:
449 - assign a new offset to this slot
450 - xor find another slot to coalesce with */
452 for (i=0; i<size; ++i) { /* for each spill slot */
456 DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
457 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
458 DBG((ssenv->dbg, LEVEL_1, " %+F\n", n1));
461 for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
462 /* check inter-slot-pairs for interference */
464 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
465 for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
466 if(values_interfere(n1, n2)) {
467 pset_break(ass[i]->members);
468 pset_break(ass[o]->members);
469 DBG((ssenv->dbg, LEVEL_1, " Interf %+F -- %+F\n", n1, n2));
470 goto interf_detected;
473 /* if we are here, there is no interference between ass[i] and ass[o] */
476 interf_detected: /*nothing*/ ;
479 /* now the members of ass[i] join the members of ass[tgt_slot] */
481 /* do we need a new slot? */
482 if (tgt_slot == -1) {
483 tgt_slot = used_slots;
488 ass[tgt_slot]->size = ass[i]->size;
489 del_pset(ass[tgt_slot]->members);
490 ass[tgt_slot]->members = pset_new_ptr(8);
494 /* copy the members to the target pset */
495 /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
497 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
498 pset_insert_ptr(ass[tgt_slot]->members, n1);
502 #define ALIGN_SPILL_AREA 16
503 #define pset_foreach(pset, elm) for(elm=pset_first(pset); elm; elm=pset_next(pset))
506 * Returns a spill type for a mode. Keep them in a map to reduce
507 * the number of types.
509 static ir_type *get_spill_type(pmap *types, ir_mode *mode) {
510 pmap_entry *e = pmap_find(types, mode);
515 snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(mode));
516 res = new_type_primitive(new_id_from_str(buf), mode);
517 pmap_insert(types, mode, res);
524 static void assign_entities(ss_env_t *ssenv, int n, spill_slot_t **ss) {
526 ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
528 /* aligning by increasing frame size */
529 offset = get_type_size_bits(frame) / 8;
530 offset = round_up2(offset, ALIGN_SPILL_AREA);
531 set_type_size_bytes(frame, -1);
533 /* create entities and assign offsets according to size and alignment*/
534 for (i=0; i<n; ++i) {
541 snprintf(buf, sizeof(buf), "spill_slot_%d", i);
542 name = new_id_from_str(buf);
544 spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]->largest_mode));
547 offset = round_up2(offset, ss[i]->align);
549 set_entity_offset_bytes(spill_ent, offset);
550 /* next possible offset */
551 offset += ss[i]->size;
553 pset_foreach(ss[i]->members, irn)
554 be_set_Spill_entity(irn, spill_ent);
557 /* set final size of stack frame */
558 set_type_size_bytes(frame, offset);
561 void be_compute_spill_offsets(be_chordal_env_t *cenv) {
567 obstack_init(&ssenv.ob);
569 ssenv.slots = pmap_create();
570 ssenv.types = pmap_create();
571 ssenv.dbg = firm_dbg_register("ir.be.spillslots");
573 /* Get initial spill slots */
574 irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
576 /* Build an empty array for optimized spill slots */
577 ss_size = pmap_count(ssenv.slots);
578 ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
579 optimize_slots(&ssenv, ss_size, ss);
581 /* Integrate slots into the stack frame entity */
582 assign_entities(&ssenv, ss_size, ss);
585 pmap_foreach(ssenv.slots, pme)
586 del_pset(((spill_slot_t *)pme->value)->members);
587 pmap_destroy(ssenv.slots);
588 pmap_destroy(ssenv.types);
589 obstack_free(&ssenv.ob, NULL);
591 be_copy_entities_to_reloads(cenv->irg);