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 perm_cycle_t *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;
288 * Lowers a perm node. Resolves cycles and creates a bunch of
289 * copy and swap operations to permute registers.
290 * Note: The caller of this function has to make sure, that irn
293 * @param irn The perm node
294 * @param block The block the perm node belongs to
295 * @param walk_env The environment
297 static void lower_perm_node(ir_node *irn, void *walk_env)
299 lower_env_t *const env = walk_env;
300 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
301 ir_graph *const irg = get_irn_irg(irn);
302 ir_node *const block = get_nodes_block(irn);
303 int const n = get_irn_arity(irn);
304 reg_pair_t *const pairs = alloca(n * sizeof(pairs[0]));
306 int do_copy = env->do_copy;
307 /* Get the schedule predecessor node to the perm.
308 * NOTE: This works with auto-magic. If we insert the new copy/exchange
309 * nodes after this node, everything should be ok. */
310 ir_node * sched_point = sched_prev(irn);
311 const ir_edge_t * edge;
314 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
315 assert(sched_point && "Perm is not scheduled or has no predecessor");
317 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
319 /* build the list of register pairs (in, out) */
321 foreach_out_edge(irn, edge) {
322 reg_pair_t *const pair = &pairs[i++];
323 ir_node *const out = get_edge_src_irn(edge);
324 long const pn = get_Proj_proj(out);
325 ir_node *const in = get_irn_n(irn, pn);
328 pair->in_reg = arch_get_irn_register(in);
329 pair->out_node = out;
330 pair->out_reg = arch_get_irn_register(out);
334 /* sort the register pairs by the indices of the in registers */
335 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
337 /* Mark all equal pairs as checked, and exchange the OUT proj with the IN
339 for (i = 0; i < n; i++) {
340 reg_pair_t *const pair = &pairs[i];
342 if (pair->in_reg->index != pair->out_reg->index)
345 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
346 irn, pair->in_node, pair->out_node, pair->out_reg->name));
348 /* reroute the edges from the proj to the argument */
349 exchange(pair->out_node, pair->in_node);
354 /* Set do_copy to 0 if it's on but we have no free register */
355 /* TODO check for free register */
360 /* check for cycles and chains */
361 while (get_n_checked_pairs(pairs, n) < n) {
365 /* go to the first not-checked pair */
366 for (i = 0; pairs[i].checked; ++i) {}
367 get_perm_cycle(&cycle, pairs, n, i);
369 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle.type == PERM_CHAIN ? "chain" : "cycle"));
370 for (j = 0; j < cycle.n_elems; j++) {
371 DB((dbg, LEVEL_1, " %s", cycle.elems[j]->name));
373 DB((dbg, LEVEL_1, "\n"));
375 if (n == 2 && cycle.type == PERM_CYCLE) {
376 /* We don't need to do anything if we have a Perm with two elements
377 * which represents a cycle, because those nodes already represent
381 /* TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and
382 * insert copy to/from it before/after the copy cascade (this
383 * reduces the cycle into a chain) */
385 /* build copy/swap nodes from back to front */
386 for (i = cycle.n_elems - 2; i >= 0; i--) {
387 ir_node *arg1 = get_node_for_in_register(pairs, n, cycle.elems[i]);
388 ir_node *arg2 = get_node_for_in_register(pairs, n, cycle.elems[i + 1]);
390 ir_node *res1 = get_node_for_out_register(pairs, n, cycle.elems[i]);
391 ir_node *res2 = get_node_for_out_register(pairs, n, cycle.elems[i + 1]);
392 /* If we have a cycle and don't copy: we need to create exchange
394 * NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
395 * IN_1 = in node with register i
396 * IN_2 = in node with register i + 1
397 * OUT_1 = out node with register i + 1
398 * OUT_2 = out node with register i */
399 if (cycle.type == PERM_CYCLE && !do_copy) {
406 /* At this point we have to handle the following problem:
408 * If we have a cycle with more than two elements, then this
409 * could correspond to the following Perm node:
411 * +----+ +----+ +----+
412 * | r1 | | r2 | | r3 |
413 * +-+--+ +-+--+ +--+-+
416 * +-+--------+---------+-+
418 * +-+--------+---------+-+
421 * +-+--+ +-+--+ +--+-+
422 * |Proj| |Proj| |Proj|
423 * | r2 | | r3 | | r1 |
424 * +----+ +----+ +----+
426 * This node is about to be split up into two 2x Perm's for
427 * which we need 4 Proj's and the one additional Proj of the
428 * first Perm has to be one IN of the second. So in general
429 * we need to create one additional Proj for each "middle"
430 * Perm and set this to one in node of the successor Perm. */
432 DBG((dbg, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
433 irn, arg1, cycle.elems[i]->name, arg2, cycle.elems[i + 1]->name));
434 DBG((dbg, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
435 irn, res1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
437 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
440 /* cycle is not done yet */
441 int pidx = get_pairidx_for_in_regidx(pairs, n, cycle.elems[i]->index);
443 /* create intermediate proj */
444 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
446 /* set as in for next Perm */
447 pairs[pidx].in_node = res1;
450 set_Proj_pred(res2, cpyxchg);
451 set_Proj_proj(res2, 0);
452 set_Proj_pred(res1, cpyxchg);
453 set_Proj_proj(res1, 1);
455 arch_set_irn_register(res2, cycle.elems[i + 1]);
456 arch_set_irn_register(res1, cycle.elems[i]);
458 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
459 sched_add_after(sched_point, cpyxchg);
461 DBG((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
463 /* set the new scheduling point */
468 DBG((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
469 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
471 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
472 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
474 /* exchange copy node and proj */
475 exchange(res2, cpyxchg);
477 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
478 sched_add_after(sched_point, cpyxchg);
480 /* set the new scheduling point */
481 sched_point = cpyxchg;
486 free((void*)cycle.elems);
489 /* remove the perm from schedule */
498 static int has_irn_users(const ir_node *irn) {
499 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
503 * Skip all Proj nodes.
505 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
507 irn = get_Proj_pred(irn);
511 static ir_node *find_copy(ir_node *irn, ir_node *op)
515 for (cur_node = irn;;) {
516 cur_node = sched_prev(cur_node);
517 if (! be_is_Copy(cur_node))
519 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
524 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
525 be_irg_t *birg = env->birg;
526 ir_graph *irg = be_get_birg_irg(birg);
527 ir_nodemap_t *op_set = &env->op_set;
528 ir_node *block = get_nodes_block(irn);
529 const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1);
530 ir_node *in[2], *keep, *cpy;
531 op_copy_assoc_t *entry;
533 if (arch_irn_is(other_different, ignore) ||
534 !mode_is_datab(get_irn_mode(other_different))) {
535 DBG((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
539 /* Make a not spillable copy of the different node */
540 /* this is needed because the different irn could be */
541 /* in block far far away */
542 /* The copy is optimized later if not needed */
544 /* check if already exists such a copy in the schedule immediately before */
545 cpy = find_copy(belower_skip_proj(irn), other_different);
547 cpy = be_new_Copy(cls, irg, block, other_different);
548 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
549 DBG((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
552 DBG((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
558 /* Add the Keep resp. CopyKeep and reroute the users */
559 /* of the other_different irn in case of CopyKeep. */
560 if (has_irn_users(other_different)) {
561 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
562 be_node_set_reg_class(keep, 1, cls);
565 keep = be_new_Keep(cls, irg, block, 2, in);
568 DBG((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
570 /* insert copy and keep into schedule */
571 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
572 if (! sched_is_scheduled(cpy))
573 sched_add_before(belower_skip_proj(irn), cpy);
574 sched_add_after(irn, keep);
576 /* insert the other different and it's copies into the map */
577 entry = ir_nodemap_get(op_set, other_different);
579 entry = obstack_alloc(&env->obst, sizeof(*entry));
581 ir_nodeset_init(&entry->copies);
583 ir_nodemap_insert(op_set, other_different, entry);
587 ir_nodeset_insert(&entry->copies, cpy);
589 /* insert keep in case of CopyKeep */
590 if (be_is_CopyKeep(keep)) {
591 ir_nodeset_insert(&entry->copies, keep);
596 * Checks if node has a must_be_different constraint in output and adds a Keep
597 * then to assure the constraint.
599 * @param irn the node to check
600 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
601 * @param env the constraint environment
603 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
604 const arch_register_req_t *req = arch_get_register_req(irn, -1);
606 if (arch_register_req_is(req, must_be_different)) {
607 const unsigned other = req->other_different;
610 if (arch_register_req_is(req, should_be_same)) {
611 const unsigned same = req->other_same;
613 if (is_po2(other) && is_po2(same)) {
614 int idx_other = ntz(other);
615 int idx_same = ntz(same);
618 * We can safely ignore a should_be_same x must_be_different y
619 * IFF both inputs are equal!
621 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
626 for (i = 0; 1U << i <= other; ++i) {
627 if (other & (1U << i)) {
628 ir_node *different_from = get_irn_n(skipped_irn, i);
629 gen_assure_different_pattern(irn, different_from, env);
636 * Calls the functions to assure register constraints.
638 * @param block The block to be checked
639 * @param walk_env The walker environment
641 static void assure_constraints_walker(ir_node *block, void *walk_env) {
644 sched_foreach_reverse(block, irn) {
645 ir_mode *mode = get_irn_mode(irn);
647 if (mode == mode_T) {
648 const ir_edge_t *edge;
650 foreach_out_edge(irn, edge) {
651 ir_node *proj = get_edge_src_irn(edge);
653 mode = get_irn_mode(proj);
654 if (mode_is_datab(mode))
655 assure_different_constraints(proj, irn, walk_env);
657 } else if (mode_is_datab(mode)) {
658 assure_different_constraints(irn, irn, walk_env);
664 * Melt all copykeeps pointing to the same node
665 * (or Projs of the same node), copying the same operand.
667 static void melt_copykeeps(constraint_env_t *cenv) {
668 be_irg_t *birg = cenv->birg;
669 ir_graph *irg = be_get_birg_irg(birg);
670 ir_nodemap_iterator_t map_iter;
671 ir_nodemap_entry_t map_entry;
674 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
675 op_copy_assoc_t *entry = map_entry.data;
679 ir_nodeset_iterator_t iter;
680 ir_node **ck_arr, **melt_arr;
684 /* collect all copykeeps */
686 foreach_ir_nodeset(&entry->copies, cp, iter) {
687 if (be_is_CopyKeep(cp)) {
688 obstack_grow(&obst, &cp, sizeof(cp));
691 #ifdef KEEP_ALIVE_COPYKEEP_HACK
693 set_irn_mode(cp, mode_ANY);
696 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
699 /* compare each copykeep with all other copykeeps */
700 ck_arr = (ir_node **)obstack_finish(&obst);
701 for (idx = 0; idx < num_ck; ++idx) {
702 ir_node *ref, *ref_mode_T;
708 ir_node *sched_pt = NULL;
712 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
713 obstack_grow(&obst, &ref, sizeof(ref));
715 DBG((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
717 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
718 for (j = 0; j < num_ck; ++j) {
719 ir_node *cur_ck = ck_arr[j];
721 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
722 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
723 ir_nodeset_remove(&entry->copies, cur_ck);
724 DBG((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
727 sched_remove(cur_ck);
732 /* check, if we found some candidates for melting */
734 DBG((dbg_constr, LEVEL_1, "\tno candidate found\n"));
738 ir_nodeset_remove(&entry->copies, ref);
741 melt_arr = (ir_node **)obstack_finish(&obst);
742 /* melt all found copykeeps */
743 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
744 for (j = 0; j < n_melt; ++j) {
745 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
747 /* now, we can kill the melted keep, except the */
748 /* ref one, we still need some information */
749 if (melt_arr[j] != ref)
750 kill_node(melt_arr[j]);
753 #ifdef KEEP_ALIVE_COPYKEEP_HACK
754 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
757 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));
758 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
760 /* set register class for all kept inputs */
761 for (j = 1; j <= n_melt; ++j)
762 be_node_set_reg_class(new_ck, j, entry->cls);
764 ir_nodeset_insert(&entry->copies, new_ck);
766 /* find scheduling point */
767 sched_pt = ref_mode_T;
769 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
770 sched_pt = sched_next(sched_pt);
771 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
773 sched_add_before(sched_pt, new_ck);
774 DBG((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
776 /* finally: kill the reference copykeep */
781 obstack_free(&obst, NULL);
786 * Walks over all nodes to assure register constraints.
788 * @param birg The birg structure containing the irg
790 void assure_constraints(be_irg_t *birg) {
791 ir_graph *irg = be_get_birg_irg(birg);
792 constraint_env_t cenv;
794 ir_nodemap_iterator_t map_iter;
795 ir_nodemap_entry_t map_entry;
797 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
800 ir_nodemap_init(&cenv.op_set);
801 obstack_init(&cenv.obst);
803 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
805 /* melt copykeeps, pointing to projs of */
806 /* the same mode_T node and keeping the */
808 melt_copykeeps(&cenv);
811 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
812 op_copy_assoc_t *entry = map_entry.data;
815 ir_nodeset_iterator_t iter;
816 be_ssa_construction_env_t senv;
818 n = ir_nodeset_size(&entry->copies);
819 nodes = alloca(n * sizeof(nodes[0]));
821 /* put the node in an array */
822 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
824 /* collect all copies */
826 foreach_ir_nodeset(&entry->copies, cp, iter) {
828 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
831 DB((dbg_constr, LEVEL_1, "\n"));
833 /* introduce the copies for the operand and it's copies */
834 be_ssa_construction_init(&senv, birg);
835 be_ssa_construction_add_copy(&senv, map_entry.node);
836 be_ssa_construction_add_copies(&senv, nodes, n);
837 be_ssa_construction_fix_users(&senv, map_entry.node);
838 be_ssa_construction_destroy(&senv);
840 /* Could be that not all CopyKeeps are really needed, */
841 /* so we transform unnecessary ones into Keeps. */
842 foreach_ir_nodeset(&entry->copies, cp, iter) {
843 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
845 int n = get_irn_arity(cp);
847 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
848 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
849 sched_add_before(cp, keep);
851 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
857 ir_nodeset_destroy(&entry->copies);
860 ir_nodemap_destroy(&cenv.op_set);
861 obstack_free(&cenv.obst, NULL);
862 be_liveness_invalidate(be_get_birg_liveness(birg));
867 * Push nodes that do not need to be permed through the Perm.
868 * This is commonly a reload cascade at block ends.
869 * @note This routine needs interference.
870 * @note Probably, we can implement it a little more efficient.
871 * Especially searching the frontier lazily might be better.
872 * @param perm The perm.
873 * @param data The walker data (lower_env_t).
874 * @return 1, if there is something left to perm over.
875 * 0, if removed the complete perm.
877 static int push_through_perm(ir_node *perm, void *data)
879 lower_env_t *env = data;
881 ir_graph *irg = get_irn_irg(perm);
882 ir_node *bl = get_nodes_block(perm);
884 int arity = get_irn_arity(perm);
887 bitset_t *moved = bitset_alloca(arity);
890 ir_node *frontier = bl;
891 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
894 const ir_edge_t *edge;
895 ir_node *one_proj = NULL, *irn;
896 const arch_register_class_t *cls = NULL;
898 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
900 /* get some Proj and find out the register class of that Proj. */
901 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
902 one_proj = get_edge_src_irn(edge);
903 assert(is_Proj(one_proj));
904 cls = arch_get_irn_reg_class(one_proj, -1);
906 /* Find the point in the schedule after which the
907 * potentially movable nodes must be defined.
908 * A Perm will only be pushed up to first instruction
909 * which lets an operand of itself die.
910 * If we would allow to move the Perm above this instruction,
911 * the former dead operand would be live now at the point of
912 * the Perm, increasing the register pressure by one.
914 sched_foreach_reverse_from (sched_prev(perm), irn) {
915 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
916 ir_node *op = get_irn_n(irn, i);
917 if (arch_irn_consider_in_reg_alloc(cls, op) &&
918 !values_interfere(env->birg, op, one_proj)) {
926 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
928 node = sched_prev(perm);
930 while(!sched_is_begin(node)) {
931 const arch_register_req_t *req;
935 /* search if node is a INPUT of Perm */
936 foreach_out_edge(perm, edge) {
937 ir_node *out = get_edge_src_irn(edge);
938 int pn = get_Proj_proj(out);
939 ir_node *in = get_irn_n(perm, pn);
946 /* it wasn't an input to the perm, we can't do anything more */
949 if(!sched_comes_after(frontier, node))
951 if (arch_irn_is(node, modify_flags))
954 req = arch_get_register_req(get_Proj_pred(node),
955 -1 - get_Proj_proj(node));
957 req = arch_get_register_req(node, -1);
959 if(req->type != arch_register_req_type_normal)
961 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
962 ir_node *opop = get_irn_n(node, i);
963 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
970 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
972 /* move the movable node in front of the Perm */
974 sched_add_after(perm, node);
976 /* give it the proj's register */
977 arch_set_irn_register(node, arch_get_irn_register(proj));
979 /* reroute all users of the proj to the moved node. */
980 edges_reroute(proj, node, irg);
983 set_Proj_pred(proj, new_Bad());
986 bitset_set(moved, input);
989 node = sched_prev(node);
992 /* well, we could not push anything through the perm */
996 new_size = arity - n_moved;
1001 map = alloca(new_size * sizeof(map[0]));
1002 proj_map = alloca(arity * sizeof(proj_map[0]));
1003 memset(proj_map, -1, sizeof(proj_map[0]));
1005 for(i = 0; i < arity; ++i) {
1006 if(bitset_is_set(moved, i))
1012 assert(n == new_size);
1013 foreach_out_edge(perm, edge) {
1014 ir_node *proj = get_edge_src_irn(edge);
1015 int pn = get_Proj_proj(proj);
1018 set_Proj_proj(proj, pn);
1021 be_Perm_reduce(perm, new_size, map);
1026 * Calls the corresponding lowering function for the node.
1028 * @param irn The node to be checked for lowering
1029 * @param walk_env The walker environment
1031 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
1035 if (!be_is_Perm(irn))
1038 perm_stayed = push_through_perm(irn, walk_env);
1042 lower_perm_node(irn, walk_env);
1046 * Walks over all blocks in an irg and performs lowering need to be
1047 * done after register allocation (e.g. perm lowering).
1049 * @param birg The birg object
1050 * @param do_copy 1 == resolve cycles with a free reg if available
1052 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1056 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1059 env.do_copy = do_copy;
1061 /* we will need interference */
1062 be_liveness_assure_chk(be_get_birg_liveness(birg));
1064 irg = be_get_birg_irg(birg);
1065 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);