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
35 #include "irnodeset.h"
36 #include "irnodemap.h"
38 #include "iredges_t.h"
45 #include "besched_t.h"
47 #include "bessaconstr.h"
48 #include "beintlive_t.h"
50 #undef KEEP_ALIVE_COPYKEEP_HACK
52 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
53 DEBUG_ONLY(static firm_dbg_module_t *dbg_constr;)
54 DEBUG_ONLY(static firm_dbg_module_t *dbg_permmove;)
56 /** Associates an ir_node with it's copy and CopyKeep. */
58 ir_nodeset_t copies; /**< all non-spillable copies of this irn */
59 const arch_register_class_t *cls;
62 /** Environment for constraints. */
69 /** Lowering walker environment. */
70 typedef struct _lower_env_t {
75 /** Holds a Perm register pair. */
76 typedef struct _reg_pair_t {
77 const arch_register_t *in_reg; /**< a perm IN register */
78 ir_node *in_node; /**< the in node to which the register belongs */
80 const arch_register_t *out_reg; /**< a perm OUT register */
81 ir_node *out_node; /**< the out node to which the register belongs */
83 int checked; /**< indicates whether the pair was check for cycle or not */
86 typedef enum _perm_type_t {
93 /** Structure to represent cycles or chains in a Perm. */
94 typedef struct _perm_cycle_t {
95 const arch_register_t **elems; /**< the registers in the cycle */
96 int n_elems; /**< number of elements in the cycle */
97 perm_type_t type; /**< type (CHAIN or CYCLE) */
100 /** returns the number register pairs marked as checked. */
101 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
102 int i, n_checked = 0;
104 for (i = 0; i < n; i++) {
105 if (pairs[i].checked)
113 * Gets the node corresponding to an IN register from an array of register pairs.
114 * NOTE: The given registers pairs and the register to look for must belong
115 * to the same register class.
117 * @param pairs The array of register pairs
118 * @param n The number of pairs
119 * @param reg The register to look for
120 * @return The corresponding node or NULL if not found
122 static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
125 for (i = 0; i < n; i++) {
126 /* in register matches */
127 if (pairs[i].in_reg->index == reg->index)
128 return pairs[i].in_node;
135 * Gets the node corresponding to an OUT register from an array of register pairs.
136 * NOTE: The given registers pairs and the register to look for must belong
137 * to the same register class.
139 * @param pairs The array of register pairs
140 * @param n The number of pairs
141 * @param reg The register to look for
142 * @return The corresponding node or NULL if not found
144 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
147 for (i = 0; i < n; i++) {
148 /* out register matches */
149 if (pairs[i].out_reg->index == reg->index)
150 return pairs[i].out_node;
157 * Gets the index in the register pair array where the in register
158 * corresponds to reg_idx.
160 * @param pairs The array of register pairs
161 * @param n The number of pairs
162 * @param reg The register index to look for
164 * @return The corresponding index in pairs or -1 if not found
166 static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
169 for (i = 0; i < n; i++) {
170 /* in register matches */
171 if (pairs[i].in_reg->index == reg_idx)
178 * Gets the index in the register pair array where the out register
179 * corresponds to reg_idx.
181 * @param pairs The array of register pairs
182 * @param n The number of pairs
183 * @param reg The register index to look for
185 * @return The corresponding index in pairs or -1 if not found
187 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
190 for (i = 0; i < n; i++) {
191 /* out register matches */
192 if (pairs[i].out_reg->index == reg_idx)
199 * Gets an array of register pairs and tries to identify a cycle or chain
200 * starting at position start.
202 * @param cycle Variable to hold the cycle
203 * @param pairs Array of register pairs
204 * @param n length of the pairs array
205 * @param start Index to start
207 * @return The cycle or chain
209 static void get_perm_cycle(perm_cycle_t *const cycle,
210 reg_pair_t *const pairs,
214 int head = pairs[start].in_reg->index;
215 int cur_idx = pairs[start].out_reg->index;
216 int const n_pairs_done = get_n_checked_pairs(pairs, n);
217 perm_type_t cycle_tp = PERM_CYCLE;
220 /* We could be right in the middle of a chain, so we need to find the start */
221 while (head != cur_idx) {
222 /* goto previous register in cycle or chain */
223 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
225 if (cur_pair_idx < 0) {
226 cycle_tp = PERM_CHAIN;
229 head = pairs[cur_pair_idx].in_reg->index;
230 start = cur_pair_idx;
234 /* assume worst case: all remaining pairs build a cycle or chain */
235 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
236 cycle->n_elems = 2; /* initial number of elements is 2 */
237 cycle->elems[0] = pairs[start].in_reg;
238 cycle->elems[1] = pairs[start].out_reg;
239 cycle->type = cycle_tp;
240 cur_idx = pairs[start].out_reg->index;
243 /* check for cycle or end of a chain */
244 while (cur_idx != head) {
245 /* goto next register in cycle or chain */
246 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
248 if (cur_pair_idx < 0)
251 cur_idx = pairs[cur_pair_idx].out_reg->index;
253 /* it's not the first element: insert it */
254 if (cur_idx != head) {
255 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
258 /* we are there where we started -> CYCLE */
259 cycle->type = PERM_CYCLE;
263 /* mark all pairs having one in/out register with cycle in common as checked */
264 for (idx = 0; idx < cycle->n_elems; idx++) {
267 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
268 if (cur_pair_idx >= 0)
269 pairs[cur_pair_idx].checked = 1;
271 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
272 if (cur_pair_idx >= 0)
273 pairs[cur_pair_idx].checked = 1;
278 * Lowers a perm node. Resolves cycles and creates a bunch of
279 * copy and swap operations to permute registers.
280 * Note: The caller of this function has to make sure, that irn
283 * @param irn The perm node
284 * @param block The block the perm node belongs to
285 * @param env The lowerer environment
287 static void lower_perm_node(ir_node *irn, lower_env_t *env)
289 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
290 ir_graph *const irg = get_irn_irg(irn);
291 ir_node *const block = get_nodes_block(irn);
292 int const n = get_irn_arity(irn);
293 reg_pair_t *const pairs = alloca(n * sizeof(pairs[0]));
295 int do_copy = env->do_copy;
296 /* Get the schedule predecessor node to the perm.
297 * NOTE: This works with auto-magic. If we insert the new copy/exchange
298 * nodes after this node, everything should be ok. */
299 ir_node * sched_point = sched_prev(irn);
300 const ir_edge_t * edge;
303 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
304 assert(sched_point && "Perm is not scheduled or has no predecessor");
306 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
308 /* build the list of register pairs (in, out) */
310 foreach_out_edge(irn, edge) {
311 reg_pair_t *const pair = &pairs[i++];
312 ir_node *const out = get_edge_src_irn(edge);
313 long const pn = get_Proj_proj(out);
314 ir_node *const in = get_irn_n(irn, pn);
317 pair->in_reg = arch_get_irn_register(in);
318 pair->out_node = out;
319 pair->out_reg = arch_get_irn_register(out);
323 /* Mark all equal pairs as checked, and exchange the OUT proj with the IN
325 for (i = 0; i < n; i++) {
326 reg_pair_t *const pair = &pairs[i];
328 if (pair->in_reg->index != pair->out_reg->index)
331 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
332 irn, pair->in_node, pair->out_node, pair->out_reg->name));
334 /* reroute the edges from the proj to the argument */
335 exchange(pair->out_node, pair->in_node);
340 /* Set do_copy to 0 if it's on but we have no free register */
341 /* TODO check for free register */
346 /* check for cycles and chains */
347 while (get_n_checked_pairs(pairs, n) < n) {
351 /* go to the first not-checked pair */
352 for (i = 0; pairs[i].checked; ++i) {}
353 get_perm_cycle(&cycle, pairs, n, i);
355 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle"));
356 for (j = 0; j < cycle.n_elems; j++) {
357 DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name));
359 DB((dbg, LEVEL_1, "\n"));
361 if (n == 2 && cycle.type == PERM_CYCLE) {
362 /* We don't need to do anything if we have a Perm with two elements
363 * which represents a cycle, because those nodes already represent
367 /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and
368 * insert copy to/from it before/after the copy cascade (this
369 * reduces the cycle into a chain) */
371 /* build copy/swap nodes from back to front */
372 for (i = cycle.n_elems - 2; i >= 0; i--) {
373 ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]);
374 ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]);
376 ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]);
377 ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]);
378 /* If we have a cycle and don't copy: we need to create exchange
380 * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
381 * IN_1 = in node with register i
382 * IN_2 = in node with register i + 1
383 * OUT_1 = out node with register i + 1
384 * OUT_2 = out node with register i */
385 if (cycle.type == PERM_CYCLE && !do_copy) {
392 /* At this point we have to handle the following problem:
394 * If we have a cycle with more than two elements, then this
395 * could correspond to the following Perm node:
397 * +----+ +----+ +----+
398 * | r1 | | r2 | | r3 |
399 * +-+--+ +-+--+ +--+-+
402 * +-+--------+---------+-+
404 * +-+--------+---------+-+
407 * +-+--+ +-+--+ +--+-+
408 * |Proj| |Proj| |Proj|
409 * | r2 | | r3 | | r1 |
410 * +----+ +----+ +----+
412 * This node is about to be split up into two 2x Perm's for
413 * which we need 4 Proj's and the one additional Proj of the
414 * first Perm has to be one IN of the second. So in general
415 * we need to create one additional Proj for each "middle"
416 * Perm and set this to one in node of the successor Perm. */
418 DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
419 irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name));
420 DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
421 irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
423 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
426 /* cycle is not done yet */
427 int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index);
429 /* create intermediate proj */
430 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
432 /* set as in for next Perm */
433 pairs[pidx].in_node = res1;
436 set_Proj_pred(res2, cpyxchg);
437 set_Proj_proj(res2, 0);
438 set_Proj_pred(res1, cpyxchg);
439 set_Proj_proj(res1, 1);
441 arch_set_irn_register(res2, cycle.elems[i + 1]);
442 arch_set_irn_register(res1, cycle.elems[i]);
444 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
445 sched_add_after(sched_point, cpyxchg);
447 DBG((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
449 /* set the new scheduling point */
454 DBG((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
455 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
457 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
458 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
460 /* exchange copy node and proj */
461 exchange(res2, cpyxchg);
463 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
464 sched_add_after(sched_point, cpyxchg);
466 /* set the new scheduling point */
467 sched_point = cpyxchg;
472 free((void*)cycle.elems);
475 /* remove the perm from schedule */
484 static int has_irn_users(const ir_node *irn) {
485 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
488 static ir_node *find_copy(ir_node *irn, ir_node *op)
492 for (cur_node = irn;;) {
493 cur_node = sched_prev(cur_node);
494 if (! be_is_Copy(cur_node))
496 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
501 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
503 ir_nodemap_t *op_set;
505 const arch_register_class_t *cls;
507 op_copy_assoc_t *entry;
509 if (arch_irn_is(other_different, ignore) ||
510 !mode_is_datab(get_irn_mode(other_different))) {
511 DBG((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
515 irg = be_get_birg_irg(env->birg);
516 op_set = &env->op_set;
517 block = get_nodes_block(irn);
518 cls = arch_get_irn_reg_class(other_different, -1);
520 /* Make a not spillable copy of the different node */
521 /* this is needed because the different irn could be */
522 /* in block far far away */
523 /* The copy is optimized later if not needed */
525 /* check if already exists such a copy in the schedule immediately before */
526 cpy = find_copy(skip_Proj(irn), other_different);
528 cpy = be_new_Copy(cls, irg, block, other_different);
529 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
530 DBG((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
532 DBG((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
535 /* Add the Keep resp. CopyKeep and reroute the users */
536 /* of the other_different irn in case of CopyKeep. */
537 if (has_irn_users(other_different)) {
538 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
539 be_node_set_reg_class(keep, 1, cls);
545 keep = be_new_Keep(cls, irg, block, 2, in);
548 DBG((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
550 /* insert copy and keep into schedule */
551 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
552 if (! sched_is_scheduled(cpy))
553 sched_add_before(skip_Proj(irn), cpy);
554 sched_add_after(irn, keep);
556 /* insert the other different and it's copies into the map */
557 entry = ir_nodemap_get(op_set, other_different);
559 entry = obstack_alloc(&env->obst, sizeof(*entry));
561 ir_nodeset_init(&entry->copies);
563 ir_nodemap_insert(op_set, other_different, entry);
567 ir_nodeset_insert(&entry->copies, cpy);
569 /* insert keep in case of CopyKeep */
570 if (be_is_CopyKeep(keep))
571 ir_nodeset_insert(&entry->copies, keep);
575 * Checks if node has a must_be_different constraint in output and adds a Keep
576 * then to assure the constraint.
578 * @param irn the node to check
579 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
580 * @param env the constraint environment
582 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
583 const arch_register_req_t *req = arch_get_register_req(irn, -1);
585 if (arch_register_req_is(req, must_be_different)) {
586 const unsigned other = req->other_different;
589 if (arch_register_req_is(req, should_be_same)) {
590 const unsigned same = req->other_same;
592 if (is_po2(other) && is_po2(same)) {
593 int idx_other = ntz(other);
594 int idx_same = ntz(same);
597 * We can safely ignore a should_be_same x must_be_different y
598 * IFF both inputs are equal!
600 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
605 for (i = 0; 1U << i <= other; ++i) {
606 if (other & (1U << i)) {
607 ir_node *different_from = get_irn_n(skipped_irn, i);
608 gen_assure_different_pattern(irn, different_from, env);
615 * Calls the functions to assure register constraints.
617 * @param block The block to be checked
618 * @param walk_env The walker environment
620 static void assure_constraints_walker(ir_node *block, void *walk_env) {
623 sched_foreach_reverse(block, irn) {
624 ir_mode *mode = get_irn_mode(irn);
626 if (mode == mode_T) {
627 const ir_edge_t *edge;
629 foreach_out_edge(irn, edge) {
630 ir_node *proj = get_edge_src_irn(edge);
632 mode = get_irn_mode(proj);
633 if (mode_is_datab(mode))
634 assure_different_constraints(proj, irn, walk_env);
636 } else if (mode_is_datab(mode)) {
637 assure_different_constraints(irn, irn, walk_env);
643 * Melt all copykeeps pointing to the same node
644 * (or Projs of the same node), copying the same operand.
646 static void melt_copykeeps(constraint_env_t *cenv) {
647 be_irg_t *birg = cenv->birg;
648 ir_graph *irg = be_get_birg_irg(birg);
649 ir_nodemap_iterator_t map_iter;
650 ir_nodemap_entry_t map_entry;
653 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
654 op_copy_assoc_t *entry = map_entry.data;
658 ir_nodeset_iterator_t iter;
659 ir_node **ck_arr, **melt_arr;
663 /* collect all copykeeps */
665 foreach_ir_nodeset(&entry->copies, cp, iter) {
666 if (be_is_CopyKeep(cp)) {
667 obstack_grow(&obst, &cp, sizeof(cp));
670 #ifdef KEEP_ALIVE_COPYKEEP_HACK
672 set_irn_mode(cp, mode_ANY);
675 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
678 /* compare each copykeep with all other copykeeps */
679 ck_arr = (ir_node **)obstack_finish(&obst);
680 for (idx = 0; idx < num_ck; ++idx) {
681 ir_node *ref, *ref_mode_T;
687 ir_node *sched_pt = NULL;
691 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
692 obstack_grow(&obst, &ref, sizeof(ref));
694 DBG((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
696 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
697 for (j = 0; j < num_ck; ++j) {
698 ir_node *cur_ck = ck_arr[j];
700 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
701 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
702 ir_nodeset_remove(&entry->copies, cur_ck);
703 DBG((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
706 sched_remove(cur_ck);
711 /* check, if we found some candidates for melting */
713 DBG((dbg_constr, LEVEL_1, "\tno candidate found\n"));
717 ir_nodeset_remove(&entry->copies, ref);
720 melt_arr = (ir_node **)obstack_finish(&obst);
721 /* melt all found copykeeps */
722 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
723 for (j = 0; j < n_melt; ++j) {
724 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
726 /* now, we can kill the melted keep, except the */
727 /* ref one, we still need some information */
728 if (melt_arr[j] != ref)
729 kill_node(melt_arr[j]);
732 #ifdef KEEP_ALIVE_COPYKEEP_HACK
733 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
736 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));
737 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
739 /* set register class for all kept inputs */
740 for (j = 1; j <= n_melt; ++j)
741 be_node_set_reg_class(new_ck, j, entry->cls);
743 ir_nodeset_insert(&entry->copies, new_ck);
745 /* find scheduling point */
746 sched_pt = ref_mode_T;
748 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
749 sched_pt = sched_next(sched_pt);
750 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
752 sched_add_before(sched_pt, new_ck);
753 DBG((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
755 /* finally: kill the reference copykeep */
760 obstack_free(&obst, NULL);
765 * Walks over all nodes to assure register constraints.
767 * @param birg The birg structure containing the irg
769 void assure_constraints(be_irg_t *birg) {
770 ir_graph *irg = be_get_birg_irg(birg);
771 constraint_env_t cenv;
773 ir_nodemap_iterator_t map_iter;
774 ir_nodemap_entry_t map_entry;
776 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
779 ir_nodemap_init(&cenv.op_set);
780 obstack_init(&cenv.obst);
782 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
784 /* melt copykeeps, pointing to projs of */
785 /* the same mode_T node and keeping the */
787 melt_copykeeps(&cenv);
790 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
791 op_copy_assoc_t *entry = map_entry.data;
794 ir_nodeset_iterator_t iter;
795 be_ssa_construction_env_t senv;
797 n = ir_nodeset_size(&entry->copies);
798 nodes = alloca(n * sizeof(nodes[0]));
800 /* put the node in an array */
801 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
803 /* collect all copies */
805 foreach_ir_nodeset(&entry->copies, cp, iter) {
807 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
810 DB((dbg_constr, LEVEL_1, "\n"));
812 /* introduce the copies for the operand and it's copies */
813 be_ssa_construction_init(&senv, birg);
814 be_ssa_construction_add_copy(&senv, map_entry.node);
815 be_ssa_construction_add_copies(&senv, nodes, n);
816 be_ssa_construction_fix_users(&senv, map_entry.node);
817 be_ssa_construction_destroy(&senv);
819 /* Could be that not all CopyKeeps are really needed, */
820 /* so we transform unnecessary ones into Keeps. */
821 foreach_ir_nodeset(&entry->copies, cp, iter) {
822 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
824 int n = get_irn_arity(cp);
826 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
827 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
828 sched_add_before(cp, keep);
830 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
836 ir_nodeset_destroy(&entry->copies);
839 ir_nodemap_destroy(&cenv.op_set);
840 obstack_free(&cenv.obst, NULL);
841 be_liveness_invalidate(be_get_birg_liveness(birg));
846 * Push nodes that do not need to be permed through the Perm.
847 * This is commonly a reload cascade at block ends.
848 * @note This routine needs interference.
849 * @note Probably, we can implement it a little more efficient.
850 * Especially searching the frontier lazily might be better.
852 * @param perm The perm
853 * @param env The lowerer environment
855 * @return 1, if there is something left to perm over.
856 * 0, if removed the complete perm.
858 static int push_through_perm(ir_node *perm, lower_env_t *env)
860 ir_graph *irg = get_irn_irg(perm);
861 ir_node *bl = get_nodes_block(perm);
863 int arity = get_irn_arity(perm);
866 bitset_t *moved = bitset_alloca(arity);
869 ir_node *frontier = bl;
872 const ir_edge_t *edge;
873 ir_node *one_proj = NULL, *irn;
874 const arch_register_class_t *cls = NULL;
876 DBG((dbg_permmove, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
878 /* get some Proj and find out the register class of that Proj. */
879 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
880 one_proj = get_edge_src_irn(edge);
881 assert(is_Proj(one_proj));
882 cls = arch_get_irn_reg_class(one_proj, -1);
884 /* Find the point in the schedule after which the
885 * potentially movable nodes must be defined.
886 * A Perm will only be pushed up to first instruction
887 * which lets an operand of itself die.
888 * If we would allow to move the Perm above this instruction,
889 * the former dead operand would be live now at the point of
890 * the Perm, increasing the register pressure by one.
892 sched_foreach_reverse_from(sched_prev(perm), irn) {
893 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
894 ir_node *op = get_irn_n(irn, i);
895 if (arch_irn_consider_in_reg_alloc(cls, op) &&
896 !values_interfere(env->birg, op, one_proj)) {
904 DBG((dbg_permmove, LEVEL_2, "\tfrontier: %+F\n", frontier));
906 node = sched_prev(perm);
908 while (!sched_is_begin(node)) {
909 const arch_register_req_t *req;
913 /* search if node is a INPUT of Perm */
914 foreach_out_edge(perm, edge) {
915 ir_node *out = get_edge_src_irn(edge);
916 int pn = get_Proj_proj(out);
917 ir_node *in = get_irn_n(perm, pn);
924 /* it wasn't an input to the perm, we can't do anything more */
927 if (!sched_comes_after(frontier, node))
929 if (arch_irn_is(node, modify_flags))
932 req = arch_get_register_req(get_Proj_pred(node),
933 -1 - get_Proj_proj(node));
935 req = arch_get_register_req(node, -1);
937 if (req->type != arch_register_req_type_normal)
939 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
940 ir_node *opop = get_irn_n(node, i);
941 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
948 DBG((dbg_permmove, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
950 /* move the movable node in front of the Perm */
952 sched_add_after(perm, node);
954 /* give it the proj's register */
955 arch_set_irn_register(node, arch_get_irn_register(proj));
957 /* reroute all users of the proj to the moved node. */
958 edges_reroute(proj, node, irg);
961 set_Proj_pred(proj, new_Bad());
964 bitset_set(moved, input);
967 node = sched_prev(node);
970 /* well, we could not push anything through the perm */
974 new_size = arity - n_moved;
979 map = alloca(new_size * sizeof(map[0]));
980 proj_map = alloca(arity * sizeof(proj_map[0]));
981 memset(proj_map, -1, sizeof(proj_map[0]));
983 for (i = 0; i < arity; ++i) {
984 if (bitset_is_set(moved, i))
990 assert(n == new_size);
991 foreach_out_edge(perm, edge) {
992 ir_node *proj = get_edge_src_irn(edge);
993 int pn = get_Proj_proj(proj);
996 set_Proj_proj(proj, pn);
999 be_Perm_reduce(perm, new_size, map);
1004 * Calls the corresponding lowering function for the node.
1006 * @param irn The node to be checked for lowering
1007 * @param walk_env The walker environment
1009 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
1013 if (!be_is_Perm(irn))
1016 perm_stayed = push_through_perm(irn, walk_env);
1018 lower_perm_node(irn, walk_env);
1022 * Walks over all blocks in an irg and performs lowering need to be
1023 * done after register allocation (e.g. perm lowering).
1025 * @param birg The birg object
1026 * @param do_copy 1 == resolve cycles with a free reg if available
1028 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1032 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1033 FIRM_DBG_REGISTER(dbg_permmove, "firm.be.lower.permmove");
1036 env.do_copy = do_copy;
1038 /* we will need interference */
1039 be_liveness_assure_chk(be_get_birg_liveness(birg));
1041 irg = be_get_birg_irg(birg);
1042 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);