2 * Copyright (C) 1995-2008 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"
44 #include "besched_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) {
126 for (i = 0; i < n; i++) {
127 /* in register matches */
128 if (pairs[i].in_reg->index == reg->index)
129 return pairs[i].in_node;
136 * Gets the node corresponding to an OUT register from an array of register pairs.
137 * NOTE: The given registers pairs and the register to look for must belong
138 * to the same register class.
140 * @param pairs The array of register pairs
141 * @param n The number of pairs
142 * @param reg The register to look for
143 * @return The corresponding node or NULL if not found
145 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
148 for (i = 0; i < n; i++) {
149 /* out register matches */
150 if (pairs[i].out_reg->index == reg->index)
151 return pairs[i].out_node;
158 * Gets the index in the register pair array where the in register
159 * corresponds to reg_idx.
161 * @param pairs The array of register pairs
162 * @param n The number of pairs
163 * @param reg The register index to look for
165 * @return The corresponding index in pairs or -1 if not found
167 static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
170 for (i = 0; i < n; i++) {
171 /* in register matches */
172 if (pairs[i].in_reg->index == reg_idx)
179 * Gets the index in the register pair array where the out register
180 * corresponds to reg_idx.
182 * @param pairs The array of register pairs
183 * @param n The number of pairs
184 * @param reg The register index to look for
186 * @return The corresponding index in pairs or -1 if not found
188 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
191 for (i = 0; i < n; i++) {
192 /* out register matches */
193 if (pairs[i].out_reg->index == reg_idx)
200 * Gets an array of register pairs and tries to identify a cycle or chain
201 * starting at position start.
203 * @param cycle Variable to hold the cycle
204 * @param pairs Array of register pairs
205 * @param n length of the pairs array
206 * @param start Index to start
208 * @return The cycle or chain
210 static void get_perm_cycle(perm_cycle_t *const cycle,
211 reg_pair_t *const pairs,
215 int head = pairs[start].in_reg->index;
216 int cur_idx = pairs[start].out_reg->index;
217 int const n_pairs_todo = get_n_unchecked_pairs(pairs, n);
218 perm_type_t cycle_tp = PERM_CYCLE;
221 /* We could be right in the middle of a chain, so we need to find the start */
222 while (head != cur_idx) {
223 /* goto previous register in cycle or chain */
224 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
226 if (cur_pair_idx < 0) {
227 cycle_tp = PERM_CHAIN;
230 head = pairs[cur_pair_idx].in_reg->index;
231 start = cur_pair_idx;
235 /* assume worst case: all remaining pairs build a cycle or chain */
236 cycle->elems = XMALLOCNZ(const arch_register_t*, n_pairs_todo * 2);
237 cycle->n_elems = 2; /* initial number of elements is 2 */
238 cycle->elems[0] = pairs[start].in_reg;
239 cycle->elems[1] = pairs[start].out_reg;
240 cycle->type = cycle_tp;
241 cur_idx = pairs[start].out_reg->index;
244 /* check for cycle or end of a chain */
245 while (cur_idx != head) {
246 /* goto next register in cycle or chain */
247 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
249 if (cur_pair_idx < 0)
252 cur_idx = pairs[cur_pair_idx].out_reg->index;
254 /* it's not the first element: insert it */
255 if (cur_idx != head) {
256 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
259 /* we are there where we started -> CYCLE */
260 cycle->type = PERM_CYCLE;
264 /* mark all pairs having one in/out register with cycle in common as checked */
265 for (idx = 0; idx < cycle->n_elems; idx++) {
268 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
269 if (cur_pair_idx >= 0)
270 pairs[cur_pair_idx].checked = 1;
272 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
273 if (cur_pair_idx >= 0)
274 pairs[cur_pair_idx].checked = 1;
279 * Lowers a perm node. Resolves cycles and creates a bunch of
280 * copy and swap operations to permute registers.
281 * Note: The caller of this function has to make sure, that irn
284 * @param irn The perm node
285 * @param block The block the perm node belongs to
286 * @param env The lowerer environment
288 static void lower_perm_node(ir_node *irn, lower_env_t *env)
290 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
291 ir_graph *const irg = get_irn_irg(irn);
292 ir_node *const block = get_nodes_block(irn);
293 int const arity = get_irn_arity(irn);
294 reg_pair_t *const pairs = ALLOCAN(reg_pair_t, arity);
296 int do_copy = env->do_copy;
297 /* Get the schedule predecessor node to the perm.
298 * NOTE: This works with auto-magic. If we insert the new copy/exchange
299 * nodes after this node, everything should be ok. */
300 ir_node * sched_point = sched_prev(irn);
301 const ir_edge_t * edge;
302 const ir_edge_t * next;
306 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
307 assert(sched_point && "Perm is not scheduled or has no predecessor");
309 assert(arity == get_irn_n_edges(irn) && "perm's in and out numbers different");
311 /* build the list of register pairs (in, out) */
313 foreach_out_edge_safe(irn, edge, next) {
314 ir_node *const out = get_edge_src_irn(edge);
315 long const pn = get_Proj_proj(out);
316 ir_node *const in = get_irn_n(irn, pn);
317 arch_register_t const *const in_reg = arch_get_irn_register(in);
318 arch_register_t const *const out_reg = arch_get_irn_register(out);
321 if (in_reg == out_reg) {
322 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
323 irn, in, out, out_reg->name));
330 pair->in_reg = arch_get_irn_register(in);
331 pair->out_node = out;
332 pair->out_reg = arch_get_irn_register(out);
336 DBG((dbg, LEVEL_1, "%+F has %d unresolved constraints\n", irn, n));
338 /* Set do_copy to 0 if it's on but we have no free register */
339 /* TODO check for free register */
344 /* check for cycles and chains */
345 while (get_n_unchecked_pairs(pairs, n) > 0) {
349 /* go to the first not-checked pair */
350 for (i = 0; pairs[i].checked; ++i) {}
351 get_perm_cycle(&cycle, pairs, n, i);
353 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle"));
354 for (j = 0; j < cycle.n_elems; j++) {
355 DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name));
357 DB((dbg, LEVEL_1, "\n"));
359 if (cycle.type == PERM_CYCLE && arity == 2) {
360 /* We don't need to do anything if we have a Perm with two elements
361 * which represents a cycle, because those nodes already represent
365 /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and
366 * insert copy to/from it before/after the copy cascade (this
367 * reduces the cycle into a chain) */
369 /* build copy/swap nodes from back to front */
370 for (i = cycle.n_elems - 2; i >= 0; i--) {
371 ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]);
372 ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]);
374 ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]);
375 ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]);
376 /* If we have a cycle and don't copy: we need to create exchange
378 * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
379 * IN_1 = in node with register i
380 * IN_2 = in node with register i + 1
381 * OUT_1 = out node with register i + 1
382 * OUT_2 = out node with register i */
383 if (cycle.type == PERM_CYCLE && !do_copy) {
390 /* At this point we have to handle the following problem:
392 * If we have a cycle with more than two elements, then this
393 * could correspond to the following Perm node:
395 * +----+ +----+ +----+
396 * | r1 | | r2 | | r3 |
397 * +-+--+ +-+--+ +--+-+
400 * +-+--------+---------+-+
402 * +-+--------+---------+-+
405 * +-+--+ +-+--+ +--+-+
406 * |Proj| |Proj| |Proj|
407 * | r2 | | r3 | | r1 |
408 * +----+ +----+ +----+
410 * This node is about to be split up into two 2x Perm's for
411 * which we need 4 Proj's and the one additional Proj of the
412 * first Perm has to be one IN of the second. So in general
413 * we need to create one additional Proj for each "middle"
414 * Perm and set this to one in node of the successor Perm. */
416 DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
417 irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name));
418 DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
419 irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
421 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
424 /* cycle is not done yet */
425 int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index);
427 /* create intermediate proj */
428 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
430 /* set as in for next Perm */
431 pairs[pidx].in_node = res1;
434 set_Proj_pred(res2, cpyxchg);
435 set_Proj_proj(res2, 0);
436 set_Proj_pred(res1, cpyxchg);
437 set_Proj_proj(res1, 1);
439 arch_set_irn_register(res2, cycle.elems[i + 1]);
440 arch_set_irn_register(res1, cycle.elems[i]);
442 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
443 sched_add_after(skip_Proj(sched_point), cpyxchg);
445 DB((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
447 /* set the new scheduling point */
452 DB((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
453 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
455 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
456 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
458 /* exchange copy node and proj */
459 exchange(res2, cpyxchg);
461 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
462 sched_add_after(skip_Proj(sched_point), cpyxchg);
464 /* set the new scheduling point */
465 sched_point = cpyxchg;
470 free((void*)cycle.elems);
473 /* remove the perm from schedule */
482 static int has_irn_users(const ir_node *irn) {
483 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
486 static ir_node *find_copy(ir_node *irn, ir_node *op)
490 for (cur_node = irn;;) {
491 cur_node = sched_prev(cur_node);
492 if (! be_is_Copy(cur_node))
494 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
499 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
501 ir_nodemap_t *op_set;
503 const arch_register_class_t *cls;
505 op_copy_assoc_t *entry;
507 if (arch_irn_is_ignore(other_different) ||
508 !mode_is_datab(get_irn_mode(other_different))) {
509 DB((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
513 irg = be_get_birg_irg(env->birg);
514 op_set = &env->op_set;
515 block = get_nodes_block(irn);
516 cls = arch_get_irn_reg_class_out(other_different);
518 /* Make a not spillable copy of the different node */
519 /* this is needed because the different irn could be */
520 /* in block far far away */
521 /* The copy is optimized later if not needed */
523 /* check if already exists such a copy in the schedule immediately before */
524 cpy = find_copy(skip_Proj(irn), other_different);
526 cpy = be_new_Copy(cls, irg, block, other_different);
527 arch_irn_set_flags(cpy, arch_irn_flags_dont_spill);
528 DB((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
530 DB((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
533 /* Add the Keep resp. CopyKeep and reroute the users */
534 /* of the other_different irn in case of CopyKeep. */
535 if (has_irn_users(other_different)) {
536 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
537 be_node_set_reg_class_in(keep, 1, cls);
543 keep = be_new_Keep(cls, irg, block, 2, in);
546 DB((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
548 /* insert copy and keep into schedule */
549 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
550 if (! sched_is_scheduled(cpy))
551 sched_add_before(skip_Proj(irn), cpy);
552 sched_add_after(skip_Proj(irn), keep);
554 /* insert the other different and it's copies into the map */
555 entry = ir_nodemap_get(op_set, other_different);
557 entry = obstack_alloc(&env->obst, sizeof(*entry));
559 ir_nodeset_init(&entry->copies);
561 ir_nodemap_insert(op_set, other_different, entry);
565 ir_nodeset_insert(&entry->copies, cpy);
567 /* insert keep in case of CopyKeep */
568 if (be_is_CopyKeep(keep))
569 ir_nodeset_insert(&entry->copies, keep);
573 * Checks if node has a must_be_different constraint in output and adds a Keep
574 * then to assure the constraint.
576 * @param irn the node to check
577 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
578 * @param env the constraint environment
580 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
581 const arch_register_req_t *req = arch_get_register_req_out(irn);
583 if (arch_register_req_is(req, must_be_different)) {
584 const unsigned other = req->other_different;
587 if (arch_register_req_is(req, should_be_same)) {
588 const unsigned same = req->other_same;
590 if (is_po2(other) && is_po2(same)) {
591 int idx_other = ntz(other);
592 int idx_same = ntz(same);
595 * We can safely ignore a should_be_same x must_be_different y
596 * IFF both inputs are equal!
598 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
603 for (i = 0; 1U << i <= other; ++i) {
604 if (other & (1U << i)) {
605 ir_node *different_from = get_irn_n(skipped_irn, i);
606 gen_assure_different_pattern(irn, different_from, env);
613 * Calls the functions to assure register constraints.
615 * @param block The block to be checked
616 * @param walk_env The walker environment
618 static void assure_constraints_walker(ir_node *block, void *walk_env) {
621 sched_foreach_reverse(block, irn) {
622 ir_mode *mode = get_irn_mode(irn);
624 if (mode == mode_T) {
625 const ir_edge_t *edge;
627 foreach_out_edge(irn, edge) {
628 ir_node *proj = get_edge_src_irn(edge);
630 mode = get_irn_mode(proj);
631 if (mode_is_datab(mode))
632 assure_different_constraints(proj, irn, walk_env);
634 } else if (mode_is_datab(mode)) {
635 assure_different_constraints(irn, irn, walk_env);
641 * Melt all copykeeps pointing to the same node
642 * (or Projs of the same node), copying the same operand.
644 static void melt_copykeeps(constraint_env_t *cenv) {
645 be_irg_t *birg = cenv->birg;
646 ir_graph *irg = be_get_birg_irg(birg);
647 ir_nodemap_iterator_t map_iter;
648 ir_nodemap_entry_t map_entry;
651 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
652 op_copy_assoc_t *entry = map_entry.data;
656 ir_nodeset_iterator_t iter;
657 ir_node **ck_arr, **melt_arr;
661 /* collect all copykeeps */
663 foreach_ir_nodeset(&entry->copies, cp, iter) {
664 if (be_is_CopyKeep(cp)) {
665 obstack_grow(&obst, &cp, sizeof(cp));
668 #ifdef KEEP_ALIVE_COPYKEEP_HACK
670 set_irn_mode(cp, mode_ANY);
673 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
676 /* compare each copykeep with all other copykeeps */
677 ck_arr = (ir_node **)obstack_finish(&obst);
678 for (idx = 0; idx < num_ck; ++idx) {
679 ir_node *ref, *ref_mode_T;
685 ir_node *sched_pt = NULL;
689 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
690 obstack_grow(&obst, &ref, sizeof(ref));
692 DB((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
694 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
695 for (j = 0; j < num_ck; ++j) {
696 ir_node *cur_ck = ck_arr[j];
698 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
699 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
700 ir_nodeset_remove(&entry->copies, cur_ck);
701 DB((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
704 sched_remove(cur_ck);
709 /* check, if we found some candidates for melting */
711 DB((dbg_constr, LEVEL_1, "\tno candidate found\n"));
715 ir_nodeset_remove(&entry->copies, ref);
718 melt_arr = (ir_node **)obstack_finish(&obst);
719 /* melt all found copykeeps */
720 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
721 for (j = 0; j < n_melt; ++j) {
722 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
724 /* now, we can kill the melted keep, except the */
725 /* ref one, we still need some information */
726 if (melt_arr[j] != ref)
727 kill_node(melt_arr[j]);
730 #ifdef KEEP_ALIVE_COPYKEEP_HACK
731 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
734 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, get_irn_mode(ref));
735 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
737 /* set register class for all kept inputs */
738 for (j = 1; j <= n_melt; ++j)
739 be_node_set_reg_class_in(new_ck, j, entry->cls);
741 ir_nodeset_insert(&entry->copies, new_ck);
743 /* find scheduling point */
744 sched_pt = ref_mode_T;
746 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
747 sched_pt = sched_next(sched_pt);
748 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
750 sched_add_before(sched_pt, new_ck);
751 DB((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
753 /* finally: kill the reference copykeep */
758 obstack_free(&obst, NULL);
763 * Walks over all nodes to assure register constraints.
765 * @param birg The birg structure containing the irg
767 void assure_constraints(be_irg_t *birg) {
768 ir_graph *irg = be_get_birg_irg(birg);
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 = map_entry.data;
789 int 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, birg);
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 const arch_register_class_t *cls = arch_get_irn_reg_class_out(cp);
819 int n = get_irn_arity(cp);
822 keep = be_new_Keep(cls, irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
823 sched_add_before(cp, keep);
825 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
831 ir_nodeset_destroy(&entry->copies);
834 ir_nodemap_destroy(&cenv.op_set);
835 obstack_free(&cenv.obst, NULL);
836 be_liveness_invalidate(be_get_birg_liveness(birg));
841 * Push nodes that do not need to be permed through the Perm.
842 * This is commonly a reload cascade at block ends.
843 * @note This routine needs interference.
844 * @note Probably, we can implement it a little more efficient.
845 * Especially searching the frontier lazily might be better.
847 * @param perm The perm
848 * @param env The lowerer environment
850 * @return 1, if there is something left to perm over.
851 * 0, if removed the complete perm.
853 static int push_through_perm(ir_node *perm, lower_env_t *env)
855 ir_graph *irg = get_irn_irg(perm);
856 ir_node *bl = get_nodes_block(perm);
858 int arity = get_irn_arity(perm);
861 bitset_t *moved = bitset_alloca(arity);
864 ir_node *frontier = bl;
868 /* get some Proj and find out the register class of that Proj. */
869 const ir_edge_t *edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
870 ir_node *one_proj = get_edge_src_irn(edge);
871 const arch_register_class_t *cls = arch_get_irn_reg_class_out(one_proj);
872 assert(is_Proj(one_proj));
874 DB((dbg_permmove, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
876 /* Find the point in the schedule after which the
877 * potentially movable nodes must be defined.
878 * A Perm will only be pushed up to first instruction
879 * which lets an operand of itself die.
880 * If we would allow to move the Perm above this instruction,
881 * the former dead operand would be live now at the point of
882 * the Perm, increasing the register pressure by one.
884 sched_foreach_reverse_from(sched_prev(perm), irn) {
885 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
886 ir_node *op = get_irn_n(irn, i);
887 if (arch_irn_consider_in_reg_alloc(cls, op) &&
888 !values_interfere(env->birg, 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 edges_reroute(proj, node, irg);
948 set_Proj_pred(proj, new_Bad());
951 bitset_set(moved, input);
954 node = sched_prev(node);
957 /* well, we could not push anything through the perm */
961 new_size = arity - n_moved;
968 map = ALLOCAN(int, new_size);
969 proj_map = ALLOCAN(int, arity);
970 memset(proj_map, -1, sizeof(proj_map[0]));
972 for (i = 0; i < arity; ++i) {
973 if (bitset_is_set(moved, i))
979 assert(n == new_size);
980 foreach_out_edge(perm, edge) {
981 ir_node *proj = get_edge_src_irn(edge);
982 int pn = get_Proj_proj(proj);
985 set_Proj_proj(proj, pn);
988 be_Perm_reduce(perm, new_size, map);
993 * Calls the corresponding lowering function for the node.
995 * @param irn The node to be checked for lowering
996 * @param walk_env The walker environment
998 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
1002 if (!be_is_Perm(irn))
1005 perm_stayed = push_through_perm(irn, walk_env);
1007 lower_perm_node(irn, walk_env);
1011 * Walks over all blocks in an irg and performs lowering need to be
1012 * done after register allocation (e.g. perm lowering).
1014 * @param birg The birg object
1015 * @param do_copy 1 == resolve cycles with a free reg if available
1017 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1021 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1022 FIRM_DBG_REGISTER(dbg_permmove, "firm.be.lower.permmove");
1025 env.do_copy = do_copy;
1027 /* we will need interference */
1028 be_liveness_assure_chk(be_get_birg_liveness(birg));
1030 irg = be_get_birg_irg(birg);
1031 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);