2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Performs lowering of perm nodes. Inserts copies to assure register constraints.
23 * @author Christian Wuerdig
34 #include "irnodeset.h"
35 #include "irnodemap.h"
37 #include "iredges_t.h"
46 #include "bessaconstr.h"
47 #include "beintlive_t.h"
49 #undef KEEP_ALIVE_COPYKEEP_HACK
51 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
52 DEBUG_ONLY(static firm_dbg_module_t *dbg_constr;)
53 DEBUG_ONLY(static firm_dbg_module_t *dbg_permmove;)
55 /** Associates an ir_node with it's copy and CopyKeep. */
57 ir_nodeset_t copies; /**< all non-spillable copies of this irn */
58 const arch_register_class_t *cls;
61 /** Environment for constraints. */
68 /** Lowering walker environment. */
69 typedef struct lower_env_t {
74 /** Holds a Perm register pair. */
75 typedef struct reg_pair_t {
76 const arch_register_t *in_reg; /**< a perm IN register */
77 ir_node *in_node; /**< the in node to which the register belongs */
79 const arch_register_t *out_reg; /**< a perm OUT register */
80 ir_node *out_node; /**< the out node to which the register belongs */
82 int checked; /**< indicates whether the pair was check for cycle or not */
85 typedef enum perm_type_t {
92 /** Structure to represent cycles or chains in a Perm. */
93 typedef struct perm_cycle_t {
94 const arch_register_t **elems; /**< the registers in the cycle */
95 int n_elems; /**< number of elements in the cycle */
96 perm_type_t type; /**< type (CHAIN or CYCLE) */
99 /** returns the number register pairs marked as checked. */
100 static int get_n_unchecked_pairs(reg_pair_t const *const pairs, int const n)
105 for (i = 0; i < n; i++) {
106 if (!pairs[i].checked)
114 * Gets the node corresponding to an IN register from an array of register pairs.
115 * NOTE: The given registers pairs and the register to look for must belong
116 * to the same register class.
118 * @param pairs The array of register pairs
119 * @param n The number of pairs
120 * @param reg The register to look for
121 * @return The corresponding node or NULL if not found
123 static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg)
127 for (i = 0; i < n; i++) {
128 /* in register matches */
129 if (pairs[i].in_reg->index == reg->index)
130 return pairs[i].in_node;
137 * Gets the node corresponding to an OUT register from an array of register pairs.
138 * NOTE: The given registers pairs and the register to look for must belong
139 * to the same register class.
141 * @param pairs The array of register pairs
142 * @param n The number of pairs
143 * @param reg The register to look for
144 * @return The corresponding node or NULL if not found
146 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg)
150 for (i = 0; i < n; i++) {
151 /* out register matches */
152 if (pairs[i].out_reg->index == reg->index)
153 return pairs[i].out_node;
160 * Gets the index in the register pair array where the in register
161 * corresponds to reg_idx.
163 * @param pairs The array of register pairs
164 * @param n The number of pairs
165 * @param reg The register index to look for
167 * @return The corresponding index in pairs or -1 if not found
169 static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx)
173 for (i = 0; i < n; i++) {
174 /* in register matches */
175 if (pairs[i].in_reg->index == reg_idx)
182 * Gets the index in the register pair array where the out register
183 * corresponds to reg_idx.
185 * @param pairs The array of register pairs
186 * @param n The number of pairs
187 * @param reg The register index to look for
189 * @return The corresponding index in pairs or -1 if not found
191 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx)
195 for (i = 0; i < n; i++) {
196 /* out register matches */
197 if (pairs[i].out_reg->index == reg_idx)
204 * Gets an array of register pairs and tries to identify a cycle or chain
205 * starting at position start.
207 * @param cycle Variable to hold the cycle
208 * @param pairs Array of register pairs
209 * @param n length of the pairs array
210 * @param start Index to start
212 * @return The cycle or chain
214 static void get_perm_cycle(perm_cycle_t *const cycle,
215 reg_pair_t *const pairs,
219 int head = pairs[start].in_reg->index;
220 int cur_idx = pairs[start].out_reg->index;
221 int const n_pairs_todo = get_n_unchecked_pairs(pairs, n);
222 perm_type_t cycle_tp = PERM_CYCLE;
225 /* We could be right in the middle of a chain, so we need to find the start */
226 while (head != cur_idx) {
227 /* goto previous register in cycle or chain */
228 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
230 if (cur_pair_idx < 0) {
231 cycle_tp = PERM_CHAIN;
234 head = pairs[cur_pair_idx].in_reg->index;
235 start = cur_pair_idx;
239 /* assume worst case: all remaining pairs build a cycle or chain */
240 cycle->elems = XMALLOCNZ(const arch_register_t*, n_pairs_todo * 2);
241 cycle->n_elems = 2; /* initial number of elements is 2 */
242 cycle->elems[0] = pairs[start].in_reg;
243 cycle->elems[1] = pairs[start].out_reg;
244 cycle->type = cycle_tp;
245 cur_idx = pairs[start].out_reg->index;
248 /* check for cycle or end of a chain */
249 while (cur_idx != head) {
250 /* goto next register in cycle or chain */
251 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
253 if (cur_pair_idx < 0)
256 cur_idx = pairs[cur_pair_idx].out_reg->index;
258 /* it's not the first element: insert it */
259 if (cur_idx != head) {
260 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
263 /* we are there where we started -> CYCLE */
264 cycle->type = PERM_CYCLE;
268 /* mark all pairs having one in/out register with cycle in common as checked */
269 for (idx = 0; idx < cycle->n_elems; idx++) {
272 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
273 if (cur_pair_idx >= 0)
274 pairs[cur_pair_idx].checked = 1;
276 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
277 if (cur_pair_idx >= 0)
278 pairs[cur_pair_idx].checked = 1;
283 * Lowers a perm node. Resolves cycles and creates a bunch of
284 * copy and swap operations to permute registers.
285 * Note: The caller of this function has to make sure, that irn
288 * @param irn The perm node
289 * @param block The block the perm node belongs to
290 * @param env The lowerer environment
292 static void lower_perm_node(ir_node *irn, lower_env_t *env)
294 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
295 ir_node *const block = get_nodes_block(irn);
296 int const arity = get_irn_arity(irn);
297 reg_pair_t *const pairs = ALLOCAN(reg_pair_t, arity);
299 int do_copy = env->do_copy;
300 /* Get the schedule predecessor node to the perm.
301 * NOTE: This works with auto-magic. If we insert the new copy/exchange
302 * nodes after this node, everything should be ok. */
303 ir_node * sched_point = sched_prev(irn);
304 const ir_edge_t * edge;
305 const ir_edge_t * next;
309 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
310 assert(sched_point && "Perm is not scheduled or has no predecessor");
312 assert(arity == get_irn_n_edges(irn) && "perm's in and out numbers different");
314 /* build the list of register pairs (in, out) */
316 foreach_out_edge_safe(irn, edge, next) {
317 ir_node *const out = get_edge_src_irn(edge);
318 long const pn = get_Proj_proj(out);
319 ir_node *const in = get_irn_n(irn, pn);
320 arch_register_t const *const in_reg = arch_get_irn_register(in);
321 arch_register_t const *const out_reg = arch_get_irn_register(out);
324 if (in_reg == out_reg) {
325 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
326 irn, in, out, out_reg->name));
333 pair->in_reg = in_reg;
334 pair->out_node = out;
335 pair->out_reg = out_reg;
339 DBG((dbg, LEVEL_1, "%+F has %d unresolved constraints\n", irn, n));
341 /* Set do_copy to 0 if it's on but we have no free register */
342 /* TODO check for free register */
347 /* check for cycles and chains */
348 while (get_n_unchecked_pairs(pairs, n) > 0) {
352 /* go to the first not-checked pair */
353 for (i = 0; pairs[i].checked; ++i) {}
354 get_perm_cycle(&cycle, pairs, n, i);
356 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle"));
357 for (j = 0; j < cycle.n_elems; j++) {
358 DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name));
360 DB((dbg, LEVEL_1, "\n"));
362 if (cycle.type == PERM_CYCLE && arity == 2) {
363 /* We don't need to do anything if we have a Perm with two elements
364 * which represents a cycle, because those nodes already represent
368 /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and
369 * insert copy to/from it before/after the copy cascade (this
370 * reduces the cycle into a chain) */
372 /* build copy/swap nodes from back to front */
373 for (i = cycle.n_elems - 2; i >= 0; i--) {
374 ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]);
375 ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]);
377 ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]);
378 ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]);
379 /* If we have a cycle and don't copy: we need to create exchange
381 * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
382 * IN_1 = in node with register i
383 * IN_2 = in node with register i + 1
384 * OUT_1 = out node with register i + 1
385 * OUT_2 = out node with register i */
386 if (cycle.type == PERM_CYCLE && !do_copy) {
393 /* At this point we have to handle the following problem:
395 * If we have a cycle with more than two elements, then this
396 * could correspond to the following Perm node:
398 * +----+ +----+ +----+
399 * | r1 | | r2 | | r3 |
400 * +-+--+ +-+--+ +--+-+
403 * +-+--------+---------+-+
405 * +-+--------+---------+-+
408 * +-+--+ +-+--+ +--+-+
409 * |Proj| |Proj| |Proj|
410 * | r2 | | r3 | | r1 |
411 * +----+ +----+ +----+
413 * This node is about to be split up into two 2x Perm's for
414 * which we need 4 Proj's and the one additional Proj of the
415 * first Perm has to be one IN of the second. So in general
416 * we need to create one additional Proj for each "middle"
417 * Perm and set this to one in node of the successor Perm. */
419 DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
420 irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name));
421 DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
422 irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
424 cpyxchg = be_new_Perm(reg_class, block, 2, in);
427 /* cycle is not done yet */
428 int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index);
430 /* create intermediate proj */
431 res1 = new_r_Proj(cpyxchg, get_irn_mode(res1), 0);
433 /* set as in for next Perm */
434 pairs[pidx].in_node = res1;
437 set_Proj_pred(res2, cpyxchg);
438 set_Proj_proj(res2, 0);
439 set_Proj_pred(res1, cpyxchg);
440 set_Proj_proj(res1, 1);
442 arch_set_irn_register(res2, cycle.elems[i + 1]);
443 arch_set_irn_register(res1, cycle.elems[i]);
445 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
446 sched_add_after(skip_Proj(sched_point), cpyxchg);
448 DB((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
450 /* set the new scheduling point */
455 DB((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
456 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
458 cpyxchg = be_new_Copy(reg_class, block, arg1);
459 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
461 /* exchange copy node and proj */
462 exchange(res2, cpyxchg);
464 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
465 sched_add_after(skip_Proj(sched_point), cpyxchg);
467 /* set the new scheduling point */
468 sched_point = cpyxchg;
473 free((void*)cycle.elems);
476 /* remove the perm from schedule */
485 static int has_irn_users(const ir_node *irn)
487 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
490 static ir_node *find_copy(ir_node *irn, ir_node *op)
494 for (cur_node = irn;;) {
495 cur_node = sched_prev(cur_node);
496 if (! be_is_Copy(cur_node))
498 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
503 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env)
505 ir_nodemap_t *op_set;
507 const arch_register_class_t *cls;
509 op_copy_assoc_t *entry;
511 if (arch_irn_is_ignore(other_different) ||
512 !mode_is_datab(get_irn_mode(other_different))) {
513 DB((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
517 op_set = &env->op_set;
518 block = get_nodes_block(irn);
519 cls = arch_get_irn_reg_class_out(other_different);
521 /* Make a not spillable copy of the different node */
522 /* this is needed because the different irn could be */
523 /* in block far far away */
524 /* The copy is optimized later if not needed */
526 /* check if already exists such a copy in the schedule immediately before */
527 cpy = find_copy(skip_Proj(irn), other_different);
529 cpy = be_new_Copy(cls, block, other_different);
530 arch_irn_set_flags(cpy, arch_irn_flags_dont_spill);
531 DB((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
533 DB((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
536 /* Add the Keep resp. CopyKeep and reroute the users */
537 /* of the other_different irn in case of CopyKeep. */
538 if (has_irn_users(other_different)) {
539 keep = be_new_CopyKeep_single(cls, block, cpy, irn, get_irn_mode(other_different));
540 be_node_set_reg_class_in(keep, 1, cls);
546 keep = be_new_Keep(block, 2, in);
549 DB((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
551 /* insert copy and keep into schedule */
552 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
553 if (! sched_is_scheduled(cpy))
554 sched_add_before(skip_Proj(irn), cpy);
555 sched_add_after(skip_Proj(irn), keep);
557 /* insert the other different and it's copies into the map */
558 entry = (op_copy_assoc_t*)ir_nodemap_get(op_set, other_different);
560 entry = OALLOC(&env->obst, op_copy_assoc_t);
562 ir_nodeset_init(&entry->copies);
564 ir_nodemap_insert(op_set, other_different, entry);
568 ir_nodeset_insert(&entry->copies, cpy);
570 /* insert keep in case of CopyKeep */
571 if (be_is_CopyKeep(keep))
572 ir_nodeset_insert(&entry->copies, keep);
576 * Checks if node has a must_be_different constraint in output and adds a Keep
577 * then to assure the constraint.
579 * @param irn the node to check
580 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
581 * @param env the constraint environment
583 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env)
585 const arch_register_req_t *req = arch_get_register_req_out(irn);
587 if (arch_register_req_is(req, must_be_different)) {
588 const unsigned other = req->other_different;
591 if (arch_register_req_is(req, should_be_same)) {
592 const unsigned same = req->other_same;
594 if (is_po2(other) && is_po2(same)) {
595 int idx_other = ntz(other);
596 int idx_same = ntz(same);
599 * We can safely ignore a should_be_same x must_be_different y
600 * IFF both inputs are equal!
602 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
607 for (i = 0; 1U << i <= other; ++i) {
608 if (other & (1U << i)) {
609 ir_node *different_from = get_irn_n(skipped_irn, i);
610 gen_assure_different_pattern(irn, different_from, env);
617 * Calls the functions to assure register constraints.
619 * @param block The block to be checked
620 * @param walk_env The walker environment
622 static void assure_constraints_walker(ir_node *block, void *walk_env)
625 constraint_env_t *env = (constraint_env_t*)walk_env;
627 sched_foreach_reverse(block, irn) {
628 ir_mode *mode = get_irn_mode(irn);
630 if (mode == mode_T) {
631 const ir_edge_t *edge;
633 foreach_out_edge(irn, edge) {
634 ir_node *proj = get_edge_src_irn(edge);
636 mode = get_irn_mode(proj);
637 if (mode_is_datab(mode))
638 assure_different_constraints(proj, irn, env);
640 } else if (mode_is_datab(mode)) {
641 assure_different_constraints(irn, irn, env);
647 * Melt all copykeeps pointing to the same node
648 * (or Projs of the same node), copying the same operand.
650 static void melt_copykeeps(constraint_env_t *cenv)
652 ir_nodemap_iterator_t map_iter;
653 ir_nodemap_entry_t map_entry;
656 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
657 op_copy_assoc_t *entry = (op_copy_assoc_t*)map_entry.data;
661 ir_nodeset_iterator_t iter;
662 ir_node **ck_arr, **melt_arr;
666 /* collect all copykeeps */
668 foreach_ir_nodeset(&entry->copies, cp, iter) {
669 if (be_is_CopyKeep(cp)) {
670 obstack_grow(&obst, &cp, sizeof(cp));
673 #ifdef KEEP_ALIVE_COPYKEEP_HACK
675 set_irn_mode(cp, mode_ANY);
678 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
681 /* compare each copykeep with all other copykeeps */
682 ck_arr = (ir_node **)obstack_finish(&obst);
683 for (idx = 0; idx < num_ck; ++idx) {
684 ir_node *ref, *ref_mode_T;
690 ir_node *sched_pt = NULL;
694 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
695 obstack_grow(&obst, &ref, sizeof(ref));
697 DB((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
699 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
700 for (j = 0; j < num_ck; ++j) {
701 ir_node *cur_ck = ck_arr[j];
703 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
704 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
705 ir_nodeset_remove(&entry->copies, cur_ck);
706 DB((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
709 sched_remove(cur_ck);
714 /* check, if we found some candidates for melting */
716 DB((dbg_constr, LEVEL_1, "\tno candidate found\n"));
720 ir_nodeset_remove(&entry->copies, ref);
723 melt_arr = (ir_node **)obstack_finish(&obst);
724 /* melt all found copykeeps */
725 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
726 for (j = 0; j < n_melt; ++j) {
727 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
729 /* now, we can kill the melted keep, except the */
730 /* ref one, we still need some information */
731 if (melt_arr[j] != ref)
732 kill_node(melt_arr[j]);
735 #ifdef KEEP_ALIVE_COPYKEEP_HACK
736 new_ck = be_new_CopyKeep(entry->cls, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
739 new_ck = be_new_CopyKeep(entry->cls, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, get_irn_mode(ref));
740 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
742 /* set register class for all kept inputs */
743 for (j = 1; j <= n_melt; ++j)
744 be_node_set_reg_class_in(new_ck, j, entry->cls);
746 ir_nodeset_insert(&entry->copies, new_ck);
748 /* find scheduling point */
749 sched_pt = ref_mode_T;
751 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
752 sched_pt = sched_next(sched_pt);
753 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
755 sched_add_before(sched_pt, new_ck);
756 DB((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
758 /* finally: kill the reference copykeep */
763 obstack_free(&obst, NULL);
767 void assure_constraints(ir_graph *irg)
769 constraint_env_t cenv;
770 ir_nodemap_iterator_t map_iter;
771 ir_nodemap_entry_t map_entry;
773 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
776 ir_nodemap_init(&cenv.op_set);
777 obstack_init(&cenv.obst);
779 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
781 /* melt copykeeps, pointing to projs of */
782 /* the same mode_T node and keeping the */
784 melt_copykeeps(&cenv);
787 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
788 op_copy_assoc_t *entry = (op_copy_assoc_t*)map_entry.data;
789 size_t n = ir_nodeset_size(&entry->copies);
790 ir_node **nodes = ALLOCAN(ir_node*, n);
792 ir_nodeset_iterator_t iter;
793 be_ssa_construction_env_t senv;
795 /* put the node in an array */
796 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
798 /* collect all copies */
800 foreach_ir_nodeset(&entry->copies, cp, iter) {
802 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
805 DB((dbg_constr, LEVEL_1, "\n"));
807 /* introduce the copies for the operand and it's copies */
808 be_ssa_construction_init(&senv, irg);
809 be_ssa_construction_add_copy(&senv, map_entry.node);
810 be_ssa_construction_add_copies(&senv, nodes, n);
811 be_ssa_construction_fix_users(&senv, map_entry.node);
812 be_ssa_construction_destroy(&senv);
814 /* Could be that not all CopyKeeps are really needed, */
815 /* so we transform unnecessary ones into Keeps. */
816 foreach_ir_nodeset(&entry->copies, cp, iter) {
817 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
818 int n = get_irn_arity(cp);
821 keep = be_new_Keep(get_nodes_block(cp), n, get_irn_in(cp) + 1);
822 sched_add_before(cp, keep);
824 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
830 ir_nodeset_destroy(&entry->copies);
833 ir_nodemap_destroy(&cenv.op_set);
834 obstack_free(&cenv.obst, NULL);
835 be_liveness_invalidate(be_get_irg_liveness(irg));
840 * Push nodes that do not need to be permed through the Perm.
841 * This is commonly a reload cascade at block ends.
842 * @note This routine needs interference.
843 * @note Probably, we can implement it a little more efficient.
844 * Especially searching the frontier lazily might be better.
846 * @param perm The perm
847 * @param env The lowerer environment
849 * @return 1, if there is something left to perm over.
850 * 0, if removed the complete perm.
852 static int push_through_perm(ir_node *perm)
854 ir_graph *irg = get_irn_irg(perm);
855 ir_node *bl = get_nodes_block(perm);
857 int arity = get_irn_arity(perm);
860 bitset_t *moved = bitset_alloca(arity);
863 ir_node *frontier = bl;
867 /* get some Proj and find out the register class of that Proj. */
868 const ir_edge_t *edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
869 ir_node *one_proj = get_edge_src_irn(edge);
870 const arch_register_class_t *cls = arch_get_irn_reg_class_out(one_proj);
871 assert(is_Proj(one_proj));
873 DB((dbg_permmove, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
875 /* Find the point in the schedule after which the
876 * potentially movable nodes must be defined.
877 * A Perm will only be pushed up to first instruction
878 * which lets an operand of itself die.
879 * If we would allow to move the Perm above this instruction,
880 * the former dead operand would be live now at the point of
881 * the Perm, increasing the register pressure by one.
883 sched_foreach_reverse_from(sched_prev(perm), irn) {
884 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
885 ir_node *op = get_irn_n(irn, i);
886 be_lv_t *lv = be_get_irg_liveness(irg);
887 if (arch_irn_consider_in_reg_alloc(cls, op) &&
888 !be_values_interfere(lv, op, one_proj)) {
896 DB((dbg_permmove, LEVEL_2, "\tfrontier: %+F\n", frontier));
898 node = sched_prev(perm);
900 while (!sched_is_begin(node)) {
901 const arch_register_req_t *req;
905 /* search if node is a INPUT of Perm */
906 foreach_out_edge(perm, edge) {
907 ir_node *out = get_edge_src_irn(edge);
908 int pn = get_Proj_proj(out);
909 ir_node *in = get_irn_n(perm, pn);
916 /* it wasn't an input to the perm, we can't do anything more */
919 if (!sched_comes_after(frontier, node))
921 if (arch_irn_is(node, modify_flags))
923 req = arch_get_register_req_out(node);
924 if (req->type != arch_register_req_type_normal)
926 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
927 ir_node *opop = get_irn_n(node, i);
928 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
935 DBG((dbg_permmove, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
937 /* move the movable node in front of the Perm */
939 sched_add_after(perm, node);
941 /* give it the proj's register */
942 arch_set_irn_register(node, arch_get_irn_register(proj));
944 /* reroute all users of the proj to the moved node. */
945 exchange(proj, node);
947 bitset_set(moved, input);
950 node = sched_prev(node);
953 /* well, we could not push anything through the perm */
957 new_size = arity - n_moved;
964 map = ALLOCAN(int, new_size);
965 proj_map = ALLOCAN(int, arity);
966 memset(proj_map, -1, sizeof(proj_map[0]));
968 for (i = 0; i < arity; ++i) {
969 if (bitset_is_set(moved, i))
975 assert(n == new_size);
976 foreach_out_edge(perm, edge) {
977 ir_node *proj = get_edge_src_irn(edge);
978 int pn = get_Proj_proj(proj);
981 set_Proj_proj(proj, pn);
984 be_Perm_reduce(perm, new_size, map);
989 * Calls the corresponding lowering function for the node.
991 * @param irn The node to be checked for lowering
992 * @param walk_env The walker environment
994 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
998 if (!be_is_Perm(irn))
1001 perm_stayed = push_through_perm(irn);
1003 lower_perm_node(irn, (lower_env_t*)walk_env);
1006 void lower_nodes_after_ra(ir_graph *irg, int do_copy)
1010 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1011 FIRM_DBG_REGISTER(dbg_permmove, "firm.be.lower.permmove");
1014 env.do_copy = do_copy;
1016 /* we will need interference */
1017 be_liveness_assure_chk(be_get_irg_liveness(irg));
1019 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);