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;)
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 /** Compare the in registers of two register pairs. */
100 static int compare_reg_pair(const void *a, const void *b) {
101 const reg_pair_t *pair_a = a;
102 const reg_pair_t *pair_b = b;
104 if (pair_a->in_reg->index > pair_b->in_reg->index)
110 /** returns the number register pairs marked as checked. */
111 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
112 int i, n_checked = 0;
114 for (i = 0; i < n; i++) {
115 if (pairs[i].checked)
123 * Gets the node corresponding to an IN register from an array of register pairs.
124 * NOTE: The given registers pairs and the register to look for must belong
125 * to the same register class.
127 * @param pairs The array of register pairs
128 * @param n The number of pairs
129 * @param reg The register to look for
130 * @return The corresponding node or NULL if not found
132 static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
135 for (i = 0; i < n; i++) {
136 /* in register matches */
137 if (pairs[i].in_reg->index == reg->index)
138 return pairs[i].in_node;
145 * Gets the node corresponding to an OUT register from an array of register pairs.
146 * NOTE: The given registers pairs and the register to look for must belong
147 * to the same register class.
149 * @param pairs The array of register pairs
150 * @param n The number of pairs
151 * @param reg The register to look for
152 * @return The corresponding node or NULL if not found
154 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
157 for (i = 0; i < n; i++) {
158 /* out register matches */
159 if (pairs[i].out_reg->index == reg->index)
160 return pairs[i].out_node;
167 * Gets the index in the register pair array where the in 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_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
179 for (i = 0; i < n; i++) {
180 /* in register matches */
181 if (pairs[i].in_reg->index == reg_idx)
188 * Gets the index in the register pair array where the out register
189 * corresponds to reg_idx.
191 * @param pairs The array of register pairs
192 * @param n The number of pairs
193 * @param reg The register index to look for
195 * @return The corresponding index in pairs or -1 if not found
197 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
200 for (i = 0; i < n; i++) {
201 /* out register matches */
202 if (pairs[i].out_reg->index == reg_idx)
209 * Gets an array of register pairs and tries to identify a cycle or chain
210 * starting at position start.
212 * @param cycle Variable to hold the cycle
213 * @param pairs Array of register pairs
214 * @param start Index to start
215 * @return The cycle or chain
217 static void get_perm_cycle(perm_cycle_t *const cycle,
218 reg_pair_t *const pairs,
222 int head = pairs[start].in_reg->index;
223 int cur_idx = pairs[start].out_reg->index;
224 int const n_pairs_done = get_n_checked_pairs(pairs, n);
225 perm_type_t cycle_tp = PERM_CYCLE;
228 /* We could be right in the middle of a chain, so we need to find the start */
229 while (head != cur_idx) {
230 /* goto previous register in cycle or chain */
231 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
233 if (cur_pair_idx < 0) {
234 cycle_tp = PERM_CHAIN;
237 head = pairs[cur_pair_idx].in_reg->index;
238 start = cur_pair_idx;
242 /* assume worst case: all remaining pairs build a cycle or chain */
243 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
244 cycle->n_elems = 2; /* initial number of elements is 2 */
245 cycle->elems[0] = pairs[start].in_reg;
246 cycle->elems[1] = pairs[start].out_reg;
247 cycle->type = cycle_tp;
248 cur_idx = pairs[start].out_reg->index;
251 /* check for cycle or end of a chain */
252 while (cur_idx != head) {
253 /* goto next register in cycle or chain */
254 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
256 if (cur_pair_idx < 0)
259 cur_idx = pairs[cur_pair_idx].out_reg->index;
261 /* it's not the first element: insert it */
262 if (cur_idx != head) {
263 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
266 /* we are there where we started -> CYCLE */
267 cycle->type = PERM_CYCLE;
271 /* mark all pairs having one in/out register with cycle in common as checked */
272 for (idx = 0; idx < cycle->n_elems; idx++) {
275 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
276 if (cur_pair_idx >= 0)
277 pairs[cur_pair_idx].checked = 1;
279 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
280 if (cur_pair_idx >= 0)
281 pairs[cur_pair_idx].checked = 1;
286 * Lowers a perm node. Resolves cycles and creates a bunch of
287 * copy and swap operations to permute registers.
288 * Note: The caller of this function has to make sure, that irn
291 * @param irn The perm node
292 * @param block The block the perm node belongs to
293 * @param walk_env The environment
295 static void lower_perm_node(ir_node *irn, void *walk_env)
297 lower_env_t *const env = walk_env;
298 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
299 ir_graph *const irg = get_irn_irg(irn);
300 ir_node *const block = get_nodes_block(irn);
301 int const n = get_irn_arity(irn);
302 reg_pair_t *const pairs = alloca(n * sizeof(pairs[0]));
304 int do_copy = env->do_copy;
305 /* Get the schedule predecessor node to the perm.
306 * NOTE: This works with auto-magic. If we insert the new copy/exchange
307 * nodes after this node, everything should be ok. */
308 ir_node * sched_point = sched_prev(irn);
309 const ir_edge_t * edge;
312 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
313 assert(sched_point && "Perm is not scheduled or has no predecessor");
315 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
317 /* build the list of register pairs (in, out) */
319 foreach_out_edge(irn, edge) {
320 reg_pair_t *const pair = &pairs[i++];
321 ir_node *const out = get_edge_src_irn(edge);
322 long const pn = get_Proj_proj(out);
323 ir_node *const in = get_irn_n(irn, pn);
326 pair->in_reg = arch_get_irn_register(in);
327 pair->out_node = out;
328 pair->out_reg = arch_get_irn_register(out);
332 /* sort the register pairs by the indices of the in registers */
333 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
335 /* Mark all equal pairs as checked, and exchange the OUT proj with the IN
337 for (i = 0; i < n; i++) {
338 reg_pair_t *const pair = &pairs[i];
340 if (pair->in_reg->index != pair->out_reg->index)
343 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
344 irn, pair->in_node, pair->out_node, pair->out_reg->name));
346 /* reroute the edges from the proj to the argument */
347 exchange(pair->out_node, pair->in_node);
352 /* Set do_copy to 0 if it's on but we have no free register */
353 /* TODO check for free register */
358 /* check for cycles and chains */
359 while (get_n_checked_pairs(pairs, n) < n) {
363 /* go to the first not-checked pair */
364 for (i = 0; pairs[i].checked; ++i) {}
365 get_perm_cycle(&cycle, pairs, n, i);
367 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle"));
368 for (j = 0; j < cycle.n_elems; j++) {
369 DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name));
371 DB((dbg, LEVEL_1, "\n"));
373 if (n == 2 && cycle.type == PERM_CYCLE) {
374 /* We don't need to do anything if we have a Perm with two elements
375 * which represents a cycle, because those nodes already represent
379 /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and
380 * insert copy to/from it before/after the copy cascade (this
381 * reduces the cycle into a chain) */
383 /* build copy/swap nodes from back to front */
384 for (i = cycle.n_elems - 2; i >= 0; i--) {
385 ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]);
386 ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]);
388 ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]);
389 ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]);
390 /* If we have a cycle and don't copy: we need to create exchange
392 * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
393 * IN_1 = in node with register i
394 * IN_2 = in node with register i + 1
395 * OUT_1 = out node with register i + 1
396 * OUT_2 = out node with register i */
397 if (cycle.type == PERM_CYCLE && !do_copy) {
404 /* At this point we have to handle the following problem:
406 * If we have a cycle with more than two elements, then this
407 * could correspond to the following Perm node:
409 * +----+ +----+ +----+
410 * | r1 | | r2 | | r3 |
411 * +-+--+ +-+--+ +--+-+
414 * +-+--------+---------+-+
416 * +-+--------+---------+-+
419 * +-+--+ +-+--+ +--+-+
420 * |Proj| |Proj| |Proj|
421 * | r2 | | r3 | | r1 |
422 * +----+ +----+ +----+
424 * This node is about to be split up into two 2x Perm's for
425 * which we need 4 Proj's and the one additional Proj of the
426 * first Perm has to be one IN of the second. So in general
427 * we need to create one additional Proj for each "middle"
428 * Perm and set this to one in node of the successor Perm. */
430 DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
431 irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name));
432 DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
433 irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
435 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
438 /* cycle is not done yet */
439 int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index);
441 /* create intermediate proj */
442 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
444 /* set as in for next Perm */
445 pairs[pidx].in_node = res1;
448 set_Proj_pred(res2, cpyxchg);
449 set_Proj_proj(res2, 0);
450 set_Proj_pred(res1, cpyxchg);
451 set_Proj_proj(res1, 1);
453 arch_set_irn_register(res2, cycle.elems[i + 1]);
454 arch_set_irn_register(res1, cycle.elems[i]);
456 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
457 sched_add_after(sched_point, cpyxchg);
459 DBG((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
461 /* set the new scheduling point */
466 DBG((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
467 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
469 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
470 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
472 /* exchange copy node and proj */
473 exchange(res2, cpyxchg);
475 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
476 sched_add_after(sched_point, cpyxchg);
478 /* set the new scheduling point */
479 sched_point = cpyxchg;
484 free((void*)cycle.elems);
487 /* remove the perm from schedule */
496 static int has_irn_users(const ir_node *irn) {
497 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
500 static ir_node *find_copy(ir_node *irn, ir_node *op)
504 for (cur_node = irn;;) {
505 cur_node = sched_prev(cur_node);
506 if (! be_is_Copy(cur_node))
508 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
513 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
515 ir_nodemap_t *op_set;
517 const arch_register_class_t *cls;
519 op_copy_assoc_t *entry;
521 if (arch_irn_is(other_different, ignore) ||
522 !mode_is_datab(get_irn_mode(other_different))) {
523 DBG((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
527 irg = be_get_birg_irg(env->birg);
528 op_set = &env->op_set;
529 block = get_nodes_block(irn);
530 cls = arch_get_irn_reg_class(other_different, -1);
532 /* Make a not spillable copy of the different node */
533 /* this is needed because the different irn could be */
534 /* in block far far away */
535 /* The copy is optimized later if not needed */
537 /* check if already exists such a copy in the schedule immediately before */
538 cpy = find_copy(skip_Proj(irn), other_different);
540 cpy = be_new_Copy(cls, irg, block, other_different);
541 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
542 DBG((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
544 DBG((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
547 /* Add the Keep resp. CopyKeep and reroute the users */
548 /* of the other_different irn in case of CopyKeep. */
549 if (has_irn_users(other_different)) {
550 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
551 be_node_set_reg_class(keep, 1, cls);
557 keep = be_new_Keep(cls, irg, block, 2, in);
560 DBG((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
562 /* insert copy and keep into schedule */
563 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
564 if (! sched_is_scheduled(cpy))
565 sched_add_before(skip_Proj(irn), cpy);
566 sched_add_after(irn, keep);
568 /* insert the other different and it's copies into the map */
569 entry = ir_nodemap_get(op_set, other_different);
571 entry = obstack_alloc(&env->obst, sizeof(*entry));
573 ir_nodeset_init(&entry->copies);
575 ir_nodemap_insert(op_set, other_different, entry);
579 ir_nodeset_insert(&entry->copies, cpy);
581 /* insert keep in case of CopyKeep */
582 if (be_is_CopyKeep(keep))
583 ir_nodeset_insert(&entry->copies, keep);
587 * Checks if node has a must_be_different constraint in output and adds a Keep
588 * then to assure the constraint.
590 * @param irn the node to check
591 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
592 * @param env the constraint environment
594 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
595 const arch_register_req_t *req = arch_get_register_req(irn, -1);
597 if (arch_register_req_is(req, must_be_different)) {
598 const unsigned other = req->other_different;
601 if (arch_register_req_is(req, should_be_same)) {
602 const unsigned same = req->other_same;
604 if (is_po2(other) && is_po2(same)) {
605 int idx_other = ntz(other);
606 int idx_same = ntz(same);
609 * We can safely ignore a should_be_same x must_be_different y
610 * IFF both inputs are equal!
612 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
617 for (i = 0; 1U << i <= other; ++i) {
618 if (other & (1U << i)) {
619 ir_node *different_from = get_irn_n(skipped_irn, i);
620 gen_assure_different_pattern(irn, different_from, env);
627 * Calls the functions to assure register constraints.
629 * @param block The block to be checked
630 * @param walk_env The walker environment
632 static void assure_constraints_walker(ir_node *block, void *walk_env) {
635 sched_foreach_reverse(block, irn) {
636 ir_mode *mode = get_irn_mode(irn);
638 if (mode == mode_T) {
639 const ir_edge_t *edge;
641 foreach_out_edge(irn, edge) {
642 ir_node *proj = get_edge_src_irn(edge);
644 mode = get_irn_mode(proj);
645 if (mode_is_datab(mode))
646 assure_different_constraints(proj, irn, walk_env);
648 } else if (mode_is_datab(mode)) {
649 assure_different_constraints(irn, irn, walk_env);
655 * Melt all copykeeps pointing to the same node
656 * (or Projs of the same node), copying the same operand.
658 static void melt_copykeeps(constraint_env_t *cenv) {
659 be_irg_t *birg = cenv->birg;
660 ir_graph *irg = be_get_birg_irg(birg);
661 ir_nodemap_iterator_t map_iter;
662 ir_nodemap_entry_t map_entry;
665 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
666 op_copy_assoc_t *entry = map_entry.data;
670 ir_nodeset_iterator_t iter;
671 ir_node **ck_arr, **melt_arr;
675 /* collect all copykeeps */
677 foreach_ir_nodeset(&entry->copies, cp, iter) {
678 if (be_is_CopyKeep(cp)) {
679 obstack_grow(&obst, &cp, sizeof(cp));
682 #ifdef KEEP_ALIVE_COPYKEEP_HACK
684 set_irn_mode(cp, mode_ANY);
687 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
690 /* compare each copykeep with all other copykeeps */
691 ck_arr = (ir_node **)obstack_finish(&obst);
692 for (idx = 0; idx < num_ck; ++idx) {
693 ir_node *ref, *ref_mode_T;
699 ir_node *sched_pt = NULL;
703 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
704 obstack_grow(&obst, &ref, sizeof(ref));
706 DBG((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
708 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
709 for (j = 0; j < num_ck; ++j) {
710 ir_node *cur_ck = ck_arr[j];
712 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
713 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
714 ir_nodeset_remove(&entry->copies, cur_ck);
715 DBG((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
718 sched_remove(cur_ck);
723 /* check, if we found some candidates for melting */
725 DBG((dbg_constr, LEVEL_1, "\tno candidate found\n"));
729 ir_nodeset_remove(&entry->copies, ref);
732 melt_arr = (ir_node **)obstack_finish(&obst);
733 /* melt all found copykeeps */
734 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
735 for (j = 0; j < n_melt; ++j) {
736 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
738 /* now, we can kill the melted keep, except the */
739 /* ref one, we still need some information */
740 if (melt_arr[j] != ref)
741 kill_node(melt_arr[j]);
744 #ifdef KEEP_ALIVE_COPYKEEP_HACK
745 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
748 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));
749 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
751 /* set register class for all kept inputs */
752 for (j = 1; j <= n_melt; ++j)
753 be_node_set_reg_class(new_ck, j, entry->cls);
755 ir_nodeset_insert(&entry->copies, new_ck);
757 /* find scheduling point */
758 sched_pt = ref_mode_T;
760 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
761 sched_pt = sched_next(sched_pt);
762 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
764 sched_add_before(sched_pt, new_ck);
765 DBG((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
767 /* finally: kill the reference copykeep */
772 obstack_free(&obst, NULL);
777 * Walks over all nodes to assure register constraints.
779 * @param birg The birg structure containing the irg
781 void assure_constraints(be_irg_t *birg) {
782 ir_graph *irg = be_get_birg_irg(birg);
783 constraint_env_t cenv;
785 ir_nodemap_iterator_t map_iter;
786 ir_nodemap_entry_t map_entry;
788 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
791 ir_nodemap_init(&cenv.op_set);
792 obstack_init(&cenv.obst);
794 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
796 /* melt copykeeps, pointing to projs of */
797 /* the same mode_T node and keeping the */
799 melt_copykeeps(&cenv);
802 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
803 op_copy_assoc_t *entry = map_entry.data;
806 ir_nodeset_iterator_t iter;
807 be_ssa_construction_env_t senv;
809 n = ir_nodeset_size(&entry->copies);
810 nodes = alloca(n * sizeof(nodes[0]));
812 /* put the node in an array */
813 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
815 /* collect all copies */
817 foreach_ir_nodeset(&entry->copies, cp, iter) {
819 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
822 DB((dbg_constr, LEVEL_1, "\n"));
824 /* introduce the copies for the operand and it's copies */
825 be_ssa_construction_init(&senv, birg);
826 be_ssa_construction_add_copy(&senv, map_entry.node);
827 be_ssa_construction_add_copies(&senv, nodes, n);
828 be_ssa_construction_fix_users(&senv, map_entry.node);
829 be_ssa_construction_destroy(&senv);
831 /* Could be that not all CopyKeeps are really needed, */
832 /* so we transform unnecessary ones into Keeps. */
833 foreach_ir_nodeset(&entry->copies, cp, iter) {
834 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
836 int n = get_irn_arity(cp);
838 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
839 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
840 sched_add_before(cp, keep);
842 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
848 ir_nodeset_destroy(&entry->copies);
851 ir_nodemap_destroy(&cenv.op_set);
852 obstack_free(&cenv.obst, NULL);
853 be_liveness_invalidate(be_get_birg_liveness(birg));
858 * Push nodes that do not need to be permed through the Perm.
859 * This is commonly a reload cascade at block ends.
860 * @note This routine needs interference.
861 * @note Probably, we can implement it a little more efficient.
862 * Especially searching the frontier lazily might be better.
863 * @param perm The perm.
864 * @param data The walker data (lower_env_t).
865 * @return 1, if there is something left to perm over.
866 * 0, if removed the complete perm.
868 static int push_through_perm(ir_node *perm, void *data)
870 lower_env_t *env = data;
872 ir_graph *irg = get_irn_irg(perm);
873 ir_node *bl = get_nodes_block(perm);
875 int arity = get_irn_arity(perm);
878 bitset_t *moved = bitset_alloca(arity);
881 ir_node *frontier = bl;
882 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
885 const ir_edge_t *edge;
886 ir_node *one_proj = NULL, *irn;
887 const arch_register_class_t *cls = NULL;
889 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
891 /* get some Proj and find out the register class of that Proj. */
892 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
893 one_proj = get_edge_src_irn(edge);
894 assert(is_Proj(one_proj));
895 cls = arch_get_irn_reg_class(one_proj, -1);
897 /* Find the point in the schedule after which the
898 * potentially movable nodes must be defined.
899 * A Perm will only be pushed up to first instruction
900 * which lets an operand of itself die.
901 * If we would allow to move the Perm above this instruction,
902 * the former dead operand would be live now at the point of
903 * the Perm, increasing the register pressure by one.
905 sched_foreach_reverse_from (sched_prev(perm), irn) {
906 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
907 ir_node *op = get_irn_n(irn, i);
908 if (arch_irn_consider_in_reg_alloc(cls, op) &&
909 !values_interfere(env->birg, op, one_proj)) {
917 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
919 node = sched_prev(perm);
921 while(!sched_is_begin(node)) {
922 const arch_register_req_t *req;
926 /* search if node is a INPUT of Perm */
927 foreach_out_edge(perm, edge) {
928 ir_node *out = get_edge_src_irn(edge);
929 int pn = get_Proj_proj(out);
930 ir_node *in = get_irn_n(perm, pn);
937 /* it wasn't an input to the perm, we can't do anything more */
940 if(!sched_comes_after(frontier, node))
942 if (arch_irn_is(node, modify_flags))
945 req = arch_get_register_req(get_Proj_pred(node),
946 -1 - get_Proj_proj(node));
948 req = arch_get_register_req(node, -1);
950 if(req->type != arch_register_req_type_normal)
952 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
953 ir_node *opop = get_irn_n(node, i);
954 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
961 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
963 /* move the movable node in front of the Perm */
965 sched_add_after(perm, node);
967 /* give it the proj's register */
968 arch_set_irn_register(node, arch_get_irn_register(proj));
970 /* reroute all users of the proj to the moved node. */
971 edges_reroute(proj, node, irg);
974 set_Proj_pred(proj, new_Bad());
977 bitset_set(moved, input);
980 node = sched_prev(node);
983 /* well, we could not push anything through the perm */
987 new_size = arity - n_moved;
992 map = alloca(new_size * sizeof(map[0]));
993 proj_map = alloca(arity * sizeof(proj_map[0]));
994 memset(proj_map, -1, sizeof(proj_map[0]));
996 for(i = 0; i < arity; ++i) {
997 if(bitset_is_set(moved, i))
1003 assert(n == new_size);
1004 foreach_out_edge(perm, edge) {
1005 ir_node *proj = get_edge_src_irn(edge);
1006 int pn = get_Proj_proj(proj);
1009 set_Proj_proj(proj, pn);
1012 be_Perm_reduce(perm, new_size, map);
1017 * Calls the corresponding lowering function for the node.
1019 * @param irn The node to be checked for lowering
1020 * @param walk_env The walker environment
1022 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
1026 if (!be_is_Perm(irn))
1029 perm_stayed = push_through_perm(irn, walk_env);
1033 lower_perm_node(irn, walk_env);
1037 * Walks over all blocks in an irg and performs lowering need to be
1038 * done after register allocation (e.g. perm lowering).
1040 * @param birg The birg object
1041 * @param do_copy 1 == resolve cycles with a free reg if available
1043 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1047 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1050 env.do_copy = do_copy;
1052 /* we will need interference */
1053 be_liveness_assure_chk(be_get_birg_liveness(birg));
1055 irg = be_get_birg_irg(birg);
1056 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);