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 "bechordal_t.h"
28 typedef struct _reloader_t reloader_t;
29 typedef struct _spill_info_t spill_info_t;
36 struct _spill_info_t {
37 ir_node *spilled_node;
38 reloader_t *reloaders;
41 typedef struct _spill_ctx_t {
42 ir_node *spilled; /**< The spilled node. */
43 ir_node *user; /**< The node this spill is for. */
44 ir_node *spill; /**< The spill itself. */
48 firm_dbg_module_t *dbg;
49 const arch_register_class_t *cls;
50 const be_chordal_env_t *chordal_env;
53 set *spills; /**< all spill_info_t's, which must be placed */
54 pset *mem_phis; /**< set of all special spilled phis. allocated and freed seperately */
55 decide_irn_t is_mem_phi; /**< callback func to decide if a phi needs special spilling */
56 void *data; /**< data passed to all callbacks */
59 static int cmp_spillctx(const void *a, const void *b, size_t n) {
60 const spill_ctx_t *p = a;
61 const spill_ctx_t *q = b;
62 return !(p->user == q->user && p->spilled == q->spilled);
65 static int cmp_spillinfo(const void *x, const void *y, size_t size) {
66 const spill_info_t *xx = x;
67 const spill_info_t *yy = y;
68 return ! (xx->spilled_node == yy->spilled_node);
71 spill_env_t *be_new_spill_env(firm_dbg_module_t *dbg,
72 const be_chordal_env_t *chordal_env,
73 decide_irn_t is_mem_phi, void *data) {
75 spill_env_t *env = malloc(sizeof(env[0]));
76 env->spill_ctxs = new_set(cmp_spillctx, 1024);
77 env->spills = new_set(cmp_spillinfo, 1024);
78 env->cls = chordal_env->cls;
80 env->is_mem_phi = is_mem_phi;
82 env->chordal_env = chordal_env;
83 obstack_init(&env->obst);
87 void be_delete_spill_env(spill_env_t *senv) {
88 del_set(senv->spill_ctxs);
89 del_set(senv->spills);
90 obstack_free(&senv->obst, NULL);
94 static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_irn) {
97 templ.spilled = to_spill;
101 return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
104 static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn) {
106 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", irn, ctx_irn));
108 ctx = be_get_spill_ctx(senv->spill_ctxs, irn, ctx_irn);
110 const be_main_env_t *env = senv->chordal_env->birg->main_env;
111 ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
118 * If the first usage of a phi result would be out of memory
119 * there is no sense in allocating a register for it.
120 * Thus we spill it and all its operands to the same spill slot.
121 * Therefore the phi/dataB becomes a phi/Memory
123 static ir_node *be_spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn) {
124 int i, n = get_irn_arity(phi);
125 ir_node **ins, *bl = get_nodes_block(phi);
126 ir_graph *irg = senv->chordal_env->irg;
130 DBG((senv->dbg, LEVEL_1, "%+F in ctx %+F\n", phi, ctx_irn));
132 /* search an existing spill for this context */
133 ctx = be_get_spill_ctx(senv->spill_ctxs, phi, ctx_irn);
135 /* if not found spill the phi */
137 /* build a new PhiM with dummy in-array */
138 ins = malloc(n * sizeof(ins[0]));
140 ins[i] = new_r_Unknown(irg, mode_M);
141 ctx->spill = new_r_Phi(senv->chordal_env->irg, bl, n, ins, mode_M);
144 /* re-wire the phiM */
146 ir_node *arg = get_irn_n(phi, i);
149 if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg))
150 sub_res = be_spill_phi(senv, arg, ctx_irn);
152 sub_res = be_spill_irn(senv, arg, ctx_irn);
154 set_irn_n(ctx->spill, i, sub_res);
160 static ir_node *be_spill_node(spill_env_t *senv, ir_node *to_spill) {
162 if (pset_find_ptr(senv->mem_phis, to_spill))
163 res = be_spill_phi(senv, to_spill, to_spill);
165 res = be_spill_irn(senv, to_spill, to_spill);
170 static void phi_walker(ir_node *irn, void *env) {
171 spill_env_t *senv = env;
172 const arch_env_t *arch = senv->chordal_env->birg->main_env->arch_env;
174 if (is_Phi(irn) && arch_irn_has_reg_class(arch, irn, 0, senv->cls)
175 && senv->is_mem_phi(irn, senv->data)) {
176 DBG((senv->dbg, LEVEL_1, " %+F\n", irn));
177 pset_insert_ptr(senv->mem_phis, irn);
181 void be_insert_spills_reloads(spill_env_t *senv, pset *reload_set) {
182 const arch_env_t *aenv = senv->chordal_env->birg->main_env->arch_env;
183 ir_graph *irg = senv->chordal_env->irg;
190 /* get all special spilled phis */
191 DBG((senv->dbg, LEVEL_1, "Mem-phis:\n"));
192 senv->mem_phis = pset_new_ptr_default();
193 irg_walk_graph(senv->chordal_env->irg, phi_walker, NULL, senv);
195 /* Add reloads for mem_phis */
196 /* BETTER: These reloads (1) should only be inserted, if they are really needed */
197 DBG((senv->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
198 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
200 DBG((senv->dbg, LEVEL_1, " Mem-phi %+F\n", irn));
201 foreach_out_edge(irn, e) {
202 ir_node *user = e->src;
203 if (is_Phi(user) && !pset_find_ptr(senv->mem_phis, user)) {
204 ir_node *use_bl = get_nodes_block(user);
205 DBG((senv->dbg, LEVEL_1, " non-mem-phi user %+F\n", user));
206 be_add_reload_on_edge(senv, irn, use_bl, e->pos); /* (1) */
211 /* process each spilled node */
212 DBG((senv->dbg, LEVEL_1, "Insert spills and reloads:\n"));
213 for(si = set_first(senv->spills); si; si = set_next(senv->spills)) {
217 ir_mode *mode = get_irn_mode(si->spilled_node);
219 /* go through all reloads for this spill */
220 for(rld = si->reloaders; rld; rld = rld->next) {
221 /* the spill for this reloader */
222 ir_node *spill = be_spill_node(senv, si->spilled_node);
225 ir_node *reload = be_reload(aenv, senv->cls, rld->reloader, mode, spill);
227 DBG((senv->dbg, LEVEL_1, " %+F of %+F before %+F\n", reload, si->spilled_node, rld->reloader));
229 pset_insert_ptr(reload_set, reload);
231 /* remember the reload */
232 obstack_ptr_grow(&ob, reload);
236 assert(n_reloads > 0);
237 obstack_ptr_grow(&ob, si->spilled_node);
238 reloads = obstack_finish(&ob);
239 be_ssa_constr_ignore(senv->chordal_env->dom_front, n_reloads + 1, reloads, senv->mem_phis);
240 obstack_free(&ob, reloads);
243 obstack_free(&ob, NULL);
245 for(irn = pset_first(senv->mem_phis); irn; irn = pset_next(senv->mem_phis)) {
247 for(i = 0, n = get_irn_arity(irn); i < n; ++i)
248 set_irn_n(irn, i, new_r_Bad(senv->chordal_env->irg));
252 del_pset(senv->mem_phis);
255 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before) {
256 spill_info_t templ, *res;
259 templ.spilled_node = to_spill;
260 templ.reloaders = NULL;
261 res = set_insert(senv->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
263 rel = obstack_alloc(&senv->obst, sizeof(rel[0]));
264 rel->reloader = before;
265 rel->next = res->reloaders;
266 res->reloaders = rel;
269 void be_add_reload_on_edge(spill_env_t *senv, ir_node *to_spill, ir_node *bl, int pos) {
270 ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
271 be_add_reload(senv, to_spill, insert_bl);
276 /****************************************
278 SPILL SLOT MANAGEMENT AND OPTS
280 ****************************************/
282 typedef struct _spill_slot_t {
286 ir_mode *largest_mode; /* the mode of all members with largest size */
289 typedef struct _ss_env_t {
290 firm_dbg_module_t *dbg;
292 be_chordal_env_t *cenv;
293 pmap *slots; /* maps spill_contexts to spill_slots */
294 pmap *types; /* maps modes to types */
298 static void compute_spill_slots_walker(ir_node *spill, void *env) {
299 ss_env_t *ssenv = env;
304 if (!be_is_Spill(spill))
307 /* check, if this spill is for a context already known */
308 ctx = be_get_Spill_context(spill);
309 entry = pmap_find(ssenv->slots, ctx);
312 /* this is a new spill context */
313 ss = obstack_alloc(&ssenv->ob, sizeof(*ss));
314 ss->members = pset_new_ptr(8);
315 ss->largest_mode = get_irn_mode(get_irn_n(spill, 0));
316 ss->size = get_mode_size_bytes(ss->largest_mode);
317 ss->align = ss->size; /* TODO Assumed for now */
318 pmap_insert(ssenv->slots, ctx, ss);
321 /* values with the same spill_ctx must go into the same spill slot */
323 assert(ss->size == (unsigned)get_mode_size_bytes(get_irn_mode(get_irn_n(spill, 0))) && "Different sizes for the same spill slot are not allowed yet.");
324 for (irn = pset_first(ss->members); irn; irn = pset_next(ss->members)) {
325 /* use values_interfere here, because it uses the dominance check,
326 which does work for values in memory */
327 assert(!values_interfere(spill, irn) && "Spills for the same spill slot must not interfere!");
331 pset_insert_ptr(ss->members, spill);
334 static int ss_sorter(const void *v1, const void *v2) {
335 const spill_slot_t *ss1 = v1;
336 const spill_slot_t *ss2 = v2;
337 return ((int) ss2->size) - ((int) ss1->size);
342 * This function should optimize the spill slots.
343 * - Coalescing of multiple slots
344 * - Ordering the slots
346 * Input slots are in @p ssenv->slots
347 * @p size The count of initial spill slots in @p ssenv->slots
348 * This also is the size of the preallocated array @p ass
350 * @return An array of spill slots @p ass in specific order
352 static void optimize_slots(ss_env_t *ssenv, int size, spill_slot_t **ass) {
353 int i, o, used_slots;
357 pmap_foreach(ssenv->slots, entr)
358 ass[i++] = entr->value;
360 /* Sort the array to minimize fragmentation and cache footprint.
361 Large slots come first */
362 qsort(ass, size, sizeof(ass[0]), ss_sorter);
364 /* For each spill slot:
365 - assign a new offset to this slot
366 - xor find another slot to coalesce with */
368 for (i=0; i<size; ++i) { /* for each spill slot */
372 DBG((ssenv->dbg, LEVEL_1, "Spill slot %d members:\n", i));
373 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
374 DBG((ssenv->dbg, LEVEL_1, " %+F\n", n1));
377 for (o=0; o < used_slots && tgt_slot == -1; ++o) { /* for each offset-assigned spill slot */
378 /* check inter-slot-pairs for interference */
380 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
381 for(n2 = pset_first(ass[o]->members); n2; n2 = pset_next(ass[o]->members))
382 if(values_interfere(n1, n2)) {
383 pset_break(ass[i]->members);
384 pset_break(ass[o]->members);
385 DBG((ssenv->dbg, LEVEL_1, " Interf %+F -- %+F\n", n1, n2));
386 goto interf_detected;
389 /* if we are here, there is no interference between ass[i] and ass[o] */
392 interf_detected: /*nothing*/ ;
395 /* now the members of ass[i] join the members of ass[tgt_slot] */
397 /* do we need a new slot? */
398 if (tgt_slot == -1) {
399 tgt_slot = used_slots;
404 ass[tgt_slot]->size = ass[i]->size;
405 del_pset(ass[tgt_slot]->members);
406 ass[tgt_slot]->members = pset_new_ptr(8);
410 /* copy the members to the target pset */
411 /* NOTE: If src and tgt pset are the same, inserting while iterating is not allowed */
413 for(n1 = pset_first(ass[i]->members); n1; n1 = pset_next(ass[i]->members))
414 pset_insert_ptr(ass[tgt_slot]->members, n1);
418 #define ALIGN_SPILL_AREA 16
419 #define pset_foreach(pset, elm) for(elm=pset_first(pset); elm; elm=pset_next(pset))
422 * Returns a spill type for a mode. Keep them in a map to reduce
423 * the number of types.
425 static ir_type *get_spill_type(pmap *types, ir_mode *mode) {
426 pmap_entry *e = pmap_find(types, mode);
431 snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(mode));
432 res = new_type_primitive(new_id_from_str(buf), mode);
433 pmap_insert(types, mode, res);
440 static void assign_entities(ss_env_t *ssenv, int n, spill_slot_t **ss) {
442 ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
444 /* aligning by increasing frame size */
445 offset = get_type_size_bits(frame) / 8;
446 offset = round_up2(offset, ALIGN_SPILL_AREA);
447 set_type_size_bytes(frame, -1);
449 /* create entities and assign offsets according to size and alignment*/
450 for (i=0; i<n; ++i) {
457 snprintf(buf, sizeof(buf), "spill_slot_%d", i);
458 name = new_id_from_str(buf);
460 spill_ent = new_entity(frame, name, get_spill_type(ssenv->types, ss[i]->largest_mode));
463 offset = round_up2(offset, ss[i]->align);
465 set_entity_offset_bytes(spill_ent, offset);
466 /* next possible offset */
467 offset += ss[i]->size;
469 pset_foreach(ss[i]->members, irn)
470 be_set_Spill_entity(irn, spill_ent);
473 /* set final size of stack frame */
474 set_type_size_bytes(frame, offset);
477 void be_compute_spill_offsets(be_chordal_env_t *cenv) {
483 obstack_init(&ssenv.ob);
485 ssenv.slots = pmap_create();
486 ssenv.types = pmap_create();
487 ssenv.dbg = firm_dbg_register("ir.be.spillslots");
489 /* Get initial spill slots */
490 irg_walk_graph(cenv->irg, NULL, compute_spill_slots_walker, &ssenv);
492 /* Build an empty array for optimized spill slots */
493 ss_size = pmap_count(ssenv.slots);
494 ss = obstack_alloc(&ssenv.ob, ss_size * sizeof(*ss));
495 optimize_slots(&ssenv, ss_size, ss);
497 /* Integrate slots into the stack frame entity */
498 assign_entities(&ssenv, ss_size, ss);
501 pmap_foreach(ssenv.slots, pme)
502 del_pset(((spill_slot_t *)pme->value)->members);
503 pmap_destroy(ssenv.slots);
504 pmap_destroy(ssenv.types);
505 obstack_free(&ssenv.ob, NULL);
507 be_copy_entities_to_reloads(cenv->irg);