2 * This file is part of libFirm.
3 * Copyright (C) 2012 University of Karlsruhe.
8 * @brief Performs lowering of perm nodes. Inserts copies to assure
9 * register constraints.
10 * @author Christian Wuerdig
20 #include "irnodeset.h"
21 #include "irnodehashmap.h"
23 #include "iredges_t.h"
33 #include "bessaconstr.h"
34 #include "beintlive_t.h"
36 #undef KEEP_ALIVE_COPYKEEP_HACK
38 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
39 DEBUG_ONLY(static firm_dbg_module_t *dbg_constr;)
40 DEBUG_ONLY(static firm_dbg_module_t *dbg_permmove;)
42 /** Associates an ir_node with its copy and CopyKeep. */
44 ir_nodeset_t copies; /**< all non-spillable copies of this irn */
45 const arch_register_class_t *cls;
48 /** Environment for constraints. */
51 ir_nodehashmap_t op_set;
55 /** Lowering walker environment. */
56 typedef struct lower_env_t {
61 /** Holds a Perm register pair. */
62 typedef struct reg_pair_t {
63 const arch_register_t *in_reg; /**< a perm IN register */
64 ir_node *in_node; /**< the in node to which the register belongs */
66 const arch_register_t *out_reg; /**< a perm OUT register */
67 ir_node *out_node; /**< the out node to which the register belongs */
69 int checked; /**< indicates whether the pair was check for cycle or not */
72 typedef enum perm_type_t {
77 /** Structure to represent cycles or chains in a Perm. */
78 typedef struct perm_cycle_t {
79 const arch_register_t **elems; /**< the registers in the cycle */
80 int n_elems; /**< number of elements in the cycle */
81 perm_type_t type; /**< type (CHAIN or CYCLE) */
84 /** returns the number register pairs marked as checked. */
85 static int get_n_unchecked_pairs(reg_pair_t const *const pairs, int const n)
90 for (i = 0; i < n; i++) {
91 if (!pairs[i].checked)
99 * Gets the node corresponding to an IN register from an array of register pairs.
100 * NOTE: The given registers pairs and the register to look for must belong
101 * to the same register class.
103 * @param pairs The array of register pairs
104 * @param n The number of pairs
105 * @param reg The register to look for
106 * @return The corresponding node or NULL if not found
108 static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg)
112 for (i = 0; i < n; i++) {
113 /* in register matches */
114 if (pairs[i].in_reg->index == reg->index)
115 return pairs[i].in_node;
122 * Gets the node corresponding to an OUT register from an array of register pairs.
123 * NOTE: The given registers pairs and the register to look for must belong
124 * to the same register class.
126 * @param pairs The array of register pairs
127 * @param n The number of pairs
128 * @param reg The register to look for
129 * @return The corresponding node or NULL if not found
131 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg)
135 for (i = 0; i < n; i++) {
136 /* out register matches */
137 if (pairs[i].out_reg->index == reg->index)
138 return pairs[i].out_node;
145 * Gets the index in the register pair array where the in register
146 * corresponds to reg_idx.
148 * @param pairs The array of register pairs
149 * @param n The number of pairs
150 * @param reg The register index to look for
152 * @return The corresponding index in pairs or -1 if not found
154 static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx)
158 for (i = 0; i < n; i++) {
159 /* in register matches */
160 if (pairs[i].in_reg->index == reg_idx)
167 * Gets the index in the register pair array where the out register
168 * corresponds to reg_idx.
170 * @param pairs The array of register pairs
171 * @param n The number of pairs
172 * @param reg The register index to look for
174 * @return The corresponding index in pairs or -1 if not found
176 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx)
180 for (i = 0; i < n; i++) {
181 /* out register matches */
182 if (pairs[i].out_reg->index == reg_idx)
189 * Gets an array of register pairs and tries to identify a cycle or chain
190 * starting at position start.
192 * @param cycle Variable to hold the cycle
193 * @param pairs Array of register pairs
194 * @param n length of the pairs array
195 * @param start Index to start
197 * @return The cycle or chain
199 static void get_perm_cycle(perm_cycle_t *const cycle,
200 reg_pair_t *const pairs,
204 int head = pairs[start].in_reg->index;
205 int cur_idx = pairs[start].out_reg->index;
206 int const n_pairs_todo = get_n_unchecked_pairs(pairs, n);
207 perm_type_t cycle_tp = PERM_CYCLE;
210 /* We could be right in the middle of a chain, so we need to find the start */
211 while (head != cur_idx) {
212 /* goto previous register in cycle or chain */
213 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
215 if (cur_pair_idx < 0) {
216 cycle_tp = PERM_CHAIN;
219 head = pairs[cur_pair_idx].in_reg->index;
220 start = cur_pair_idx;
224 /* assume worst case: all remaining pairs build a cycle or chain */
225 cycle->elems = XMALLOCNZ(const arch_register_t*, n_pairs_todo * 2);
226 cycle->n_elems = 2; /* initial number of elements is 2 */
227 cycle->elems[0] = pairs[start].in_reg;
228 cycle->elems[1] = pairs[start].out_reg;
229 cycle->type = cycle_tp;
230 cur_idx = pairs[start].out_reg->index;
233 /* check for cycle or end of a chain */
234 while (cur_idx != head) {
235 /* goto next register in cycle or chain */
236 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
238 if (cur_pair_idx < 0)
241 cur_idx = pairs[cur_pair_idx].out_reg->index;
243 /* it's not the first element: insert it */
244 if (cur_idx != head) {
245 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
248 /* we are there where we started -> CYCLE */
249 cycle->type = PERM_CYCLE;
253 /* mark all pairs having one in/out register with cycle in common as checked */
254 for (idx = 0; idx < cycle->n_elems; idx++) {
257 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
258 if (cur_pair_idx >= 0)
259 pairs[cur_pair_idx].checked = 1;
261 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
262 if (cur_pair_idx >= 0)
263 pairs[cur_pair_idx].checked = 1;
268 * Lowers a perm node. Resolves cycles and creates a bunch of
269 * copy and swap operations to permute registers.
270 * Note: The caller of this function has to make sure, that irn
273 * @param irn The perm node
274 * @param block The block the perm node belongs to
275 * @param env The lowerer environment
277 static void lower_perm_node(ir_node *irn, lower_env_t *env)
279 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
280 ir_node *const block = get_nodes_block(irn);
281 int const arity = get_irn_arity(irn);
282 reg_pair_t *const pairs = ALLOCAN(reg_pair_t, arity);
284 int do_copy = env->do_copy;
285 /* Get the schedule predecessor node to the perm.
286 * NOTE: This works with auto-magic. If we insert the new copy/exchange
287 * nodes after this node, everything should be ok. */
288 ir_node * sched_point = sched_prev(irn);
292 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
293 assert(sched_point && "Perm is not scheduled or has no predecessor");
295 assert(arity == get_irn_n_edges(irn) && "perm's in and out numbers different");
297 /* build the list of register pairs (in, out) */
299 foreach_out_edge_safe(irn, edge) {
300 ir_node *const out = get_edge_src_irn(edge);
301 long const pn = get_Proj_proj(out);
302 ir_node *const in = get_irn_n(irn, pn);
303 arch_register_t const *const in_reg = arch_get_irn_register(in);
304 arch_register_t const *const out_reg = arch_get_irn_register(out);
307 if (in_reg == out_reg) {
308 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
309 irn, in, out, out_reg->name));
316 pair->in_reg = in_reg;
317 pair->out_node = out;
318 pair->out_reg = out_reg;
322 DBG((dbg, LEVEL_1, "%+F has %d unresolved constraints\n", irn, n));
324 /* Set do_copy to 0 if it's on but we have no free register */
325 /* TODO check for free register */
330 /* check for cycles and chains */
331 while (get_n_unchecked_pairs(pairs, n) > 0) {
335 /* go to the first not-checked pair */
336 for (i = 0; pairs[i].checked; ++i) {}
337 get_perm_cycle(&cycle, pairs, n, i);
339 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle"));
340 for (j = 0; j < cycle.n_elems; j++) {
341 DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name));
343 DB((dbg, LEVEL_1, "\n"));
345 if (cycle.type == PERM_CYCLE && arity == 2) {
346 /* We don't need to do anything if we have a Perm with two elements
347 * which represents a cycle, because those nodes already represent
351 /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and
352 * insert copy to/from it before/after the copy cascade (this
353 * reduces the cycle into a chain) */
355 /* build copy/swap nodes from back to front */
356 for (i = cycle.n_elems - 2; i >= 0; i--) {
357 ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]);
358 ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]);
360 ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]);
361 ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]);
362 /* If we have a cycle and don't copy: we need to create exchange
364 * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
365 * IN_1 = in node with register i
366 * IN_2 = in node with register i + 1
367 * OUT_1 = out node with register i + 1
368 * OUT_2 = out node with register i */
370 if (cycle.type == PERM_CYCLE && !do_copy) {
376 /* At this point we have to handle the following problem:
378 * If we have a cycle with more than two elements, then this
379 * could correspond to the following Perm node:
381 * +----+ +----+ +----+
382 * | r1 | | r2 | | r3 |
383 * +-+--+ +-+--+ +--+-+
386 * +-+--------+---------+-+
388 * +-+--------+---------+-+
391 * +-+--+ +-+--+ +--+-+
392 * |Proj| |Proj| |Proj|
393 * | r2 | | r3 | | r1 |
394 * +----+ +----+ +----+
396 * This node is about to be split up into two 2x Perm's for
397 * which we need 4 Proj's and the one additional Proj of the
398 * first Perm has to be one IN of the second. So in general
399 * we need to create one additional Proj for each "middle"
400 * Perm and set this to one in node of the successor Perm. */
402 DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
403 irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name));
404 DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
405 irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
407 cpyxchg = be_new_Perm(reg_class, block, 2, in);
410 /* cycle is not done yet */
411 int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index);
413 /* create intermediate proj */
414 res1 = new_r_Proj(cpyxchg, get_irn_mode(res1), 0);
416 /* set as in for next Perm */
417 pairs[pidx].in_node = res1;
420 set_Proj_pred(res2, cpyxchg);
421 set_Proj_proj(res2, 0);
422 set_Proj_pred(res1, cpyxchg);
423 set_Proj_proj(res1, 1);
425 arch_set_irn_register(res2, cycle.elems[i + 1]);
426 arch_set_irn_register(res1, cycle.elems[i]);
428 DB((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
430 DB((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
431 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
433 cpyxchg = be_new_Copy(block, arg1);
434 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
436 /* exchange copy node and proj */
437 exchange(res2, cpyxchg);
440 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
441 sched_add_after(sched_point, cpyxchg);
443 /* set the new scheduling point */
444 sched_point = cpyxchg;
448 free((void*)cycle.elems);
451 /* remove the perm from schedule */
460 static int has_irn_users(const ir_node *irn)
462 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
465 static ir_node *find_copy(ir_node *irn, ir_node *op)
469 for (cur_node = irn;;) {
470 cur_node = sched_prev(cur_node);
471 if (! be_is_Copy(cur_node))
473 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
478 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env)
480 ir_nodehashmap_t *op_set;
482 const arch_register_class_t *cls;
484 op_copy_assoc_t *entry;
486 arch_register_req_t const *const req = arch_get_irn_register_req(other_different);
487 if (arch_register_req_is(req, ignore) ||
488 !mode_is_datab(get_irn_mode(other_different))) {
489 DB((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
493 op_set = &env->op_set;
494 block = get_nodes_block(irn);
497 /* Make a not spillable copy of the different node */
498 /* this is needed because the different irn could be */
499 /* in block far far away */
500 /* The copy is optimized later if not needed */
502 /* check if already exists such a copy in the schedule immediately before */
503 cpy = find_copy(skip_Proj(irn), other_different);
505 cpy = be_new_Copy(block, other_different);
506 arch_set_irn_flags(cpy, arch_irn_flags_dont_spill);
507 DB((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
509 DB((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
512 /* Add the Keep resp. CopyKeep and reroute the users */
513 /* of the other_different irn in case of CopyKeep. */
514 if (has_irn_users(other_different)) {
515 keep = be_new_CopyKeep_single(block, cpy, irn);
516 be_node_set_reg_class_in(keep, 1, cls);
522 keep = be_new_Keep(block, 2, in);
525 DB((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
527 /* insert copy and keep into schedule */
528 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
529 if (! sched_is_scheduled(cpy))
530 sched_add_before(skip_Proj(irn), cpy);
531 sched_add_after(skip_Proj(irn), keep);
533 /* insert the other different and its copies into the map */
534 entry = ir_nodehashmap_get(op_copy_assoc_t, op_set, other_different);
536 entry = OALLOC(&env->obst, op_copy_assoc_t);
538 ir_nodeset_init(&entry->copies);
540 ir_nodehashmap_insert(op_set, other_different, entry);
544 ir_nodeset_insert(&entry->copies, cpy);
546 /* insert keep in case of CopyKeep */
547 if (be_is_CopyKeep(keep))
548 ir_nodeset_insert(&entry->copies, keep);
552 * Checks if node has a must_be_different constraint in output and adds a Keep
553 * then to assure the constraint.
555 * @param irn the node to check
556 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
557 * @param env the constraint environment
559 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env)
561 const arch_register_req_t *req = arch_get_irn_register_req(irn);
563 if (arch_register_req_is(req, must_be_different)) {
564 const unsigned other = req->other_different;
567 if (arch_register_req_is(req, should_be_same)) {
568 const unsigned same = req->other_same;
570 if (is_po2(other) && is_po2(same)) {
571 int idx_other = ntz(other);
572 int idx_same = ntz(same);
575 * We can safely ignore a should_be_same x must_be_different y
576 * IFF both inputs are equal!
578 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
583 for (i = 0; 1U << i <= other; ++i) {
584 if (other & (1U << i)) {
585 ir_node *different_from = get_irn_n(skipped_irn, i);
586 gen_assure_different_pattern(irn, different_from, env);
593 * Calls the functions to assure register constraints.
595 * @param block The block to be checked
596 * @param walk_env The walker environment
598 static void assure_constraints_walker(ir_node *block, void *walk_env)
600 constraint_env_t *env = (constraint_env_t*)walk_env;
602 sched_foreach_reverse(block, irn) {
603 be_foreach_value(irn, value,
604 if (mode_is_datab(get_irn_mode(value)))
605 assure_different_constraints(value, irn, env);
611 * Melt all copykeeps pointing to the same node
612 * (or Projs of the same node), copying the same operand.
614 static void melt_copykeeps(constraint_env_t *cenv)
616 ir_nodehashmap_iterator_t map_iter;
617 ir_nodehashmap_entry_t map_entry;
620 foreach_ir_nodehashmap(&cenv->op_set, map_entry, map_iter) {
621 op_copy_assoc_t *entry = (op_copy_assoc_t*)map_entry.data;
624 ir_node **ck_arr, **melt_arr;
628 /* collect all copykeeps */
630 foreach_ir_nodeset(&entry->copies, cp, iter) {
631 if (be_is_CopyKeep(cp)) {
632 obstack_grow(&obst, &cp, sizeof(cp));
635 #ifdef KEEP_ALIVE_COPYKEEP_HACK
637 set_irn_mode(cp, mode_ANY);
640 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
643 /* compare each copykeep with all other copykeeps */
644 ck_arr = (ir_node **)obstack_finish(&obst);
645 for (idx = 0; idx < num_ck; ++idx) {
646 ir_node *ref, *ref_mode_T;
651 ir_node *sched_pt = NULL;
655 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
656 obstack_grow(&obst, &ref, sizeof(ref));
658 DB((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
660 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
661 for (j = 0; j < num_ck; ++j) {
662 ir_node *cur_ck = ck_arr[j];
664 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
665 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
666 ir_nodeset_remove(&entry->copies, cur_ck);
667 DB((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
670 sched_remove(cur_ck);
675 /* check, if we found some candidates for melting */
677 DB((dbg_constr, LEVEL_1, "\tno candidate found\n"));
681 ir_nodeset_remove(&entry->copies, ref);
684 melt_arr = (ir_node **)obstack_finish(&obst);
685 /* melt all found copykeeps */
686 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
687 for (j = 0; j < n_melt; ++j) {
688 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
690 /* now, we can kill the melted keep, except the */
691 /* ref one, we still need some information */
692 if (melt_arr[j] != ref)
693 kill_node(melt_arr[j]);
696 ir_node *const new_ck = be_new_CopyKeep(get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in);
697 #ifdef KEEP_ALIVE_COPYKEEP_HACK
699 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
701 /* set register class for all kept inputs */
702 for (j = 1; j <= n_melt; ++j)
703 be_node_set_reg_class_in(new_ck, j, entry->cls);
705 ir_nodeset_insert(&entry->copies, new_ck);
707 /* find scheduling point */
708 sched_pt = ref_mode_T;
710 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
711 sched_pt = sched_next(sched_pt);
712 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
714 sched_add_before(sched_pt, new_ck);
715 DB((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
717 /* finally: kill the reference copykeep */
722 obstack_free(&obst, NULL);
726 void assure_constraints(ir_graph *irg)
728 constraint_env_t cenv;
729 ir_nodehashmap_iterator_t map_iter;
730 ir_nodehashmap_entry_t map_entry;
732 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
735 ir_nodehashmap_init(&cenv.op_set);
736 obstack_init(&cenv.obst);
738 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
740 /* melt copykeeps, pointing to projs of */
741 /* the same mode_T node and keeping the */
743 melt_copykeeps(&cenv);
746 foreach_ir_nodehashmap(&cenv.op_set, map_entry, map_iter) {
747 op_copy_assoc_t *entry = (op_copy_assoc_t*)map_entry.data;
748 size_t n = ir_nodeset_size(&entry->copies);
749 ir_node **nodes = ALLOCAN(ir_node*, n);
750 be_ssa_construction_env_t senv;
752 /* put the node in an array */
753 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
755 /* collect all copies */
757 foreach_ir_nodeset(&entry->copies, cp, iter) {
759 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
762 DB((dbg_constr, LEVEL_1, "\n"));
764 /* introduce the copies for the operand and its copies */
765 be_ssa_construction_init(&senv, irg);
766 be_ssa_construction_add_copy(&senv, map_entry.node);
767 be_ssa_construction_add_copies(&senv, nodes, n);
768 be_ssa_construction_fix_users(&senv, map_entry.node);
769 be_ssa_construction_destroy(&senv);
771 /* Could be that not all CopyKeeps are really needed, */
772 /* so we transform unnecessary ones into Keeps. */
773 foreach_ir_nodeset(&entry->copies, cp, iter) {
774 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
775 int n = get_irn_arity(cp);
778 keep = be_new_Keep(get_nodes_block(cp), n, get_irn_in(cp) + 1);
779 sched_replace(cp, keep);
781 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
786 ir_nodeset_destroy(&entry->copies);
789 ir_nodehashmap_destroy(&cenv.op_set);
790 obstack_free(&cenv.obst, NULL);
791 be_invalidate_live_sets(irg);
795 * Push nodes that do not need to be permed through the Perm.
796 * This is commonly a reload cascade at block ends.
797 * @note This routine needs interference.
798 * @note Probably, we can implement it a little more efficient.
799 * Especially searching the frontier lazily might be better.
801 * @param perm The perm
802 * @param env The lowerer environment
804 * @return 1, if there is something left to perm over.
805 * 0, if removed the complete perm.
807 static int push_through_perm(ir_node *perm)
809 ir_graph *irg = get_irn_irg(perm);
810 ir_node *bl = get_nodes_block(perm);
811 int arity = get_irn_arity(perm);
814 bitset_t *moved = bitset_alloca(arity);
817 ir_node *frontier = bl;
819 be_lv_t *lv = be_get_irg_liveness(irg);
821 /* get some Proj and find out the register class of that Proj. */
822 ir_node *one_proj = get_edge_src_irn(get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL));
823 const arch_register_class_t *cls = arch_get_irn_reg_class(one_proj);
824 assert(is_Proj(one_proj));
826 DB((dbg_permmove, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
828 /* Find the point in the schedule after which the
829 * potentially movable nodes must be defined.
830 * A Perm will only be pushed up to first instruction
831 * which lets an operand of itself die.
832 * If we would allow to move the Perm above this instruction,
833 * the former dead operand would be live now at the point of
834 * the Perm, increasing the register pressure by one.
836 sched_foreach_reverse_before(perm, irn) {
837 be_foreach_use(irn, cls, in_req_, op, op_req_,
838 if (!be_values_interfere(lv, op, one_proj)) {
846 DB((dbg_permmove, LEVEL_2, "\tfrontier: %+F\n", frontier));
850 ir_node *const node = sched_prev(perm);
851 if (node == frontier)
854 const arch_register_req_t *req;
856 ir_node *proj = NULL;
858 /* search if node is a INPUT of Perm */
859 foreach_out_edge(perm, edge) {
860 ir_node *out = get_edge_src_irn(edge);
861 int pn = get_Proj_proj(out);
862 ir_node *in = get_irn_n(perm, pn);
869 /* it wasn't an input to the perm, we can't do anything more */
872 if (arch_irn_is(node, modify_flags))
874 req = arch_get_irn_register_req(node);
875 if (req->type != arch_register_req_type_normal)
877 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
878 ir_node *opop = get_irn_n(node, i);
879 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
886 DBG((dbg_permmove, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
888 /* move the movable node in front of the Perm */
890 sched_add_after(perm, node);
892 /* give it the proj's register */
893 arch_set_irn_register(node, arch_get_irn_register(proj));
895 /* reroute all users of the proj to the moved node. */
896 exchange(proj, node);
898 bitset_set(moved, input);
902 /* well, we could not push anything through the perm */
906 new_size = arity - n_moved;
913 map = ALLOCAN(int, new_size);
914 proj_map = ALLOCAN(int, arity);
915 memset(proj_map, -1, sizeof(proj_map[0]));
917 for (i = 0; i < arity; ++i) {
918 if (bitset_is_set(moved, i))
924 assert(n == new_size);
925 foreach_out_edge(perm, edge) {
926 ir_node *proj = get_edge_src_irn(edge);
927 int pn = get_Proj_proj(proj);
930 set_Proj_proj(proj, pn);
933 be_Perm_reduce(perm, new_size, map);
938 * Calls the corresponding lowering function for the node.
940 * @param irn The node to be checked for lowering
941 * @param walk_env The walker environment
943 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
947 if (!be_is_Perm(irn))
950 perm_stayed = push_through_perm(irn);
952 lower_perm_node(irn, (lower_env_t*)walk_env);
955 void lower_nodes_after_ra(ir_graph *irg, int do_copy)
959 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
960 FIRM_DBG_REGISTER(dbg_permmove, "firm.be.lower.permmove");
963 env.do_copy = do_copy;
965 /* we will need interference */
966 be_assure_live_chk(irg);
968 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);