2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Performs lowering of perm nodes. Inserts copies to assure
23 * register constraints.
24 * @author Christian Wuerdig
34 #include "irnodeset.h"
35 #include "irnodehashmap.h"
37 #include "iredges_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 its 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. */
65 ir_nodehashmap_t op_set;
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 {
91 /** Structure to represent cycles or chains in a Perm. */
92 typedef struct perm_cycle_t {
93 const arch_register_t **elems; /**< the registers in the cycle */
94 int n_elems; /**< number of elements in the cycle */
95 perm_type_t type; /**< type (CHAIN or CYCLE) */
98 /** returns the number register pairs marked as checked. */
99 static int get_n_unchecked_pairs(reg_pair_t const *const pairs, int const n)
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)
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)
149 for (i = 0; i < n; i++) {
150 /* out register matches */
151 if (pairs[i].out_reg->index == reg->index)
152 return pairs[i].out_node;
159 * Gets the index in the register pair array where the in register
160 * corresponds to reg_idx.
162 * @param pairs The array of register pairs
163 * @param n The number of pairs
164 * @param reg The register index to look for
166 * @return The corresponding index in pairs or -1 if not found
168 static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx)
172 for (i = 0; i < n; i++) {
173 /* in register matches */
174 if (pairs[i].in_reg->index == reg_idx)
181 * Gets the index in the register pair array where the out register
182 * corresponds to reg_idx.
184 * @param pairs The array of register pairs
185 * @param n The number of pairs
186 * @param reg The register index to look for
188 * @return The corresponding index in pairs or -1 if not found
190 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx)
194 for (i = 0; i < n; i++) {
195 /* out register matches */
196 if (pairs[i].out_reg->index == reg_idx)
203 * Gets an array of register pairs and tries to identify a cycle or chain
204 * starting at position start.
206 * @param cycle Variable to hold the cycle
207 * @param pairs Array of register pairs
208 * @param n length of the pairs array
209 * @param start Index to start
211 * @return The cycle or chain
213 static void get_perm_cycle(perm_cycle_t *const cycle,
214 reg_pair_t *const pairs,
218 int head = pairs[start].in_reg->index;
219 int cur_idx = pairs[start].out_reg->index;
220 int const n_pairs_todo = get_n_unchecked_pairs(pairs, n);
221 perm_type_t cycle_tp = PERM_CYCLE;
224 /* We could be right in the middle of a chain, so we need to find the start */
225 while (head != cur_idx) {
226 /* goto previous register in cycle or chain */
227 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
229 if (cur_pair_idx < 0) {
230 cycle_tp = PERM_CHAIN;
233 head = pairs[cur_pair_idx].in_reg->index;
234 start = cur_pair_idx;
238 /* assume worst case: all remaining pairs build a cycle or chain */
239 cycle->elems = XMALLOCNZ(const arch_register_t*, n_pairs_todo * 2);
240 cycle->n_elems = 2; /* initial number of elements is 2 */
241 cycle->elems[0] = pairs[start].in_reg;
242 cycle->elems[1] = pairs[start].out_reg;
243 cycle->type = cycle_tp;
244 cur_idx = pairs[start].out_reg->index;
247 /* check for cycle or end of a chain */
248 while (cur_idx != head) {
249 /* goto next register in cycle or chain */
250 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
252 if (cur_pair_idx < 0)
255 cur_idx = pairs[cur_pair_idx].out_reg->index;
257 /* it's not the first element: insert it */
258 if (cur_idx != head) {
259 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
262 /* we are there where we started -> CYCLE */
263 cycle->type = PERM_CYCLE;
267 /* mark all pairs having one in/out register with cycle in common as checked */
268 for (idx = 0; idx < cycle->n_elems; idx++) {
271 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
272 if (cur_pair_idx >= 0)
273 pairs[cur_pair_idx].checked = 1;
275 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
276 if (cur_pair_idx >= 0)
277 pairs[cur_pair_idx].checked = 1;
282 * Lowers a perm node. Resolves cycles and creates a bunch of
283 * copy and swap operations to permute registers.
284 * Note: The caller of this function has to make sure, that irn
287 * @param irn The perm node
288 * @param block The block the perm node belongs to
289 * @param env The lowerer environment
291 static void lower_perm_node(ir_node *irn, lower_env_t *env)
293 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
294 ir_node *const block = get_nodes_block(irn);
295 int const arity = get_irn_arity(irn);
296 reg_pair_t *const pairs = ALLOCAN(reg_pair_t, arity);
298 int do_copy = env->do_copy;
299 /* Get the schedule predecessor node to the perm.
300 * NOTE: This works with auto-magic. If we insert the new copy/exchange
301 * nodes after this node, everything should be ok. */
302 ir_node * sched_point = sched_prev(irn);
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) {
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 = in_reg;
331 pair->out_node = out;
332 pair->out_reg = out_reg;
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 */
384 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, 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(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 DB((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
444 DB((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
445 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
447 cpyxchg = be_new_Copy(block, arg1);
448 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
450 /* exchange copy node and proj */
451 exchange(res2, cpyxchg);
454 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
455 sched_add_after(sched_point, cpyxchg);
457 /* set the new scheduling point */
458 sched_point = cpyxchg;
462 free((void*)cycle.elems);
465 /* remove the perm from schedule */
474 static int has_irn_users(const ir_node *irn)
476 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
479 static ir_node *find_copy(ir_node *irn, ir_node *op)
483 for (cur_node = irn;;) {
484 cur_node = sched_prev(cur_node);
485 if (! be_is_Copy(cur_node))
487 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
492 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env)
494 ir_nodehashmap_t *op_set;
496 const arch_register_class_t *cls;
498 op_copy_assoc_t *entry;
500 arch_register_req_t const *const req = arch_get_irn_register_req(other_different);
501 if (arch_register_req_is(req, ignore) ||
502 !mode_is_datab(get_irn_mode(other_different))) {
503 DB((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
507 op_set = &env->op_set;
508 block = get_nodes_block(irn);
511 /* Make a not spillable copy of the different node */
512 /* this is needed because the different irn could be */
513 /* in block far far away */
514 /* The copy is optimized later if not needed */
516 /* check if already exists such a copy in the schedule immediately before */
517 cpy = find_copy(skip_Proj(irn), other_different);
519 cpy = be_new_Copy(block, other_different);
520 arch_set_irn_flags(cpy, arch_irn_flags_dont_spill);
521 DB((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
523 DB((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
526 /* Add the Keep resp. CopyKeep and reroute the users */
527 /* of the other_different irn in case of CopyKeep. */
528 if (has_irn_users(other_different)) {
529 keep = be_new_CopyKeep_single(block, cpy, irn);
530 be_node_set_reg_class_in(keep, 1, cls);
536 keep = be_new_Keep(block, 2, in);
539 DB((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
541 /* insert copy and keep into schedule */
542 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
543 if (! sched_is_scheduled(cpy))
544 sched_add_before(skip_Proj(irn), cpy);
545 sched_add_after(skip_Proj(irn), keep);
547 /* insert the other different and its copies into the map */
548 entry = ir_nodehashmap_get(op_copy_assoc_t, op_set, other_different);
550 entry = OALLOC(&env->obst, op_copy_assoc_t);
552 ir_nodeset_init(&entry->copies);
554 ir_nodehashmap_insert(op_set, other_different, entry);
558 ir_nodeset_insert(&entry->copies, cpy);
560 /* insert keep in case of CopyKeep */
561 if (be_is_CopyKeep(keep))
562 ir_nodeset_insert(&entry->copies, keep);
566 * Checks if node has a must_be_different constraint in output and adds a Keep
567 * then to assure the constraint.
569 * @param irn the node to check
570 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
571 * @param env the constraint environment
573 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env)
575 const arch_register_req_t *req = arch_get_irn_register_req(irn);
577 if (arch_register_req_is(req, must_be_different)) {
578 const unsigned other = req->other_different;
581 if (arch_register_req_is(req, should_be_same)) {
582 const unsigned same = req->other_same;
584 if (is_po2(other) && is_po2(same)) {
585 int idx_other = ntz(other);
586 int idx_same = ntz(same);
589 * We can safely ignore a should_be_same x must_be_different y
590 * IFF both inputs are equal!
592 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
597 for (i = 0; 1U << i <= other; ++i) {
598 if (other & (1U << i)) {
599 ir_node *different_from = get_irn_n(skipped_irn, i);
600 gen_assure_different_pattern(irn, different_from, env);
607 * Calls the functions to assure register constraints.
609 * @param block The block to be checked
610 * @param walk_env The walker environment
612 static void assure_constraints_walker(ir_node *block, void *walk_env)
614 constraint_env_t *env = (constraint_env_t*)walk_env;
616 sched_foreach_reverse(block, irn) {
617 be_foreach_value(irn, value,
618 if (mode_is_datab(get_irn_mode(value)))
619 assure_different_constraints(value, irn, env);
625 * Melt all copykeeps pointing to the same node
626 * (or Projs of the same node), copying the same operand.
628 static void melt_copykeeps(constraint_env_t *cenv)
630 ir_nodehashmap_iterator_t map_iter;
631 ir_nodehashmap_entry_t map_entry;
634 foreach_ir_nodehashmap(&cenv->op_set, map_entry, map_iter) {
635 op_copy_assoc_t *entry = (op_copy_assoc_t*)map_entry.data;
638 ir_node **ck_arr, **melt_arr;
642 /* collect all copykeeps */
644 foreach_ir_nodeset(&entry->copies, cp, iter) {
645 if (be_is_CopyKeep(cp)) {
646 obstack_grow(&obst, &cp, sizeof(cp));
649 #ifdef KEEP_ALIVE_COPYKEEP_HACK
651 set_irn_mode(cp, mode_ANY);
654 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
657 /* compare each copykeep with all other copykeeps */
658 ck_arr = (ir_node **)obstack_finish(&obst);
659 for (idx = 0; idx < num_ck; ++idx) {
660 ir_node *ref, *ref_mode_T;
665 ir_node *sched_pt = NULL;
669 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
670 obstack_grow(&obst, &ref, sizeof(ref));
672 DB((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
674 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
675 for (j = 0; j < num_ck; ++j) {
676 ir_node *cur_ck = ck_arr[j];
678 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
679 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
680 ir_nodeset_remove(&entry->copies, cur_ck);
681 DB((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
684 sched_remove(cur_ck);
689 /* check, if we found some candidates for melting */
691 DB((dbg_constr, LEVEL_1, "\tno candidate found\n"));
695 ir_nodeset_remove(&entry->copies, ref);
698 melt_arr = (ir_node **)obstack_finish(&obst);
699 /* melt all found copykeeps */
700 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
701 for (j = 0; j < n_melt; ++j) {
702 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
704 /* now, we can kill the melted keep, except the */
705 /* ref one, we still need some information */
706 if (melt_arr[j] != ref)
707 kill_node(melt_arr[j]);
710 ir_node *const new_ck = be_new_CopyKeep(get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in);
711 #ifdef KEEP_ALIVE_COPYKEEP_HACK
713 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
715 /* set register class for all kept inputs */
716 for (j = 1; j <= n_melt; ++j)
717 be_node_set_reg_class_in(new_ck, j, entry->cls);
719 ir_nodeset_insert(&entry->copies, new_ck);
721 /* find scheduling point */
722 sched_pt = ref_mode_T;
724 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
725 sched_pt = sched_next(sched_pt);
726 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
728 sched_add_before(sched_pt, new_ck);
729 DB((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
731 /* finally: kill the reference copykeep */
736 obstack_free(&obst, NULL);
740 void assure_constraints(ir_graph *irg)
742 constraint_env_t cenv;
743 ir_nodehashmap_iterator_t map_iter;
744 ir_nodehashmap_entry_t map_entry;
746 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
749 ir_nodehashmap_init(&cenv.op_set);
750 obstack_init(&cenv.obst);
752 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
754 /* melt copykeeps, pointing to projs of */
755 /* the same mode_T node and keeping the */
757 melt_copykeeps(&cenv);
760 foreach_ir_nodehashmap(&cenv.op_set, map_entry, map_iter) {
761 op_copy_assoc_t *entry = (op_copy_assoc_t*)map_entry.data;
762 size_t n = ir_nodeset_size(&entry->copies);
763 ir_node **nodes = ALLOCAN(ir_node*, n);
764 be_ssa_construction_env_t senv;
766 /* put the node in an array */
767 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
769 /* collect all copies */
771 foreach_ir_nodeset(&entry->copies, cp, iter) {
773 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
776 DB((dbg_constr, LEVEL_1, "\n"));
778 /* introduce the copies for the operand and its copies */
779 be_ssa_construction_init(&senv, irg);
780 be_ssa_construction_add_copy(&senv, map_entry.node);
781 be_ssa_construction_add_copies(&senv, nodes, n);
782 be_ssa_construction_fix_users(&senv, map_entry.node);
783 be_ssa_construction_destroy(&senv);
785 /* Could be that not all CopyKeeps are really needed, */
786 /* so we transform unnecessary ones into Keeps. */
787 foreach_ir_nodeset(&entry->copies, cp, iter) {
788 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
789 int n = get_irn_arity(cp);
792 keep = be_new_Keep(get_nodes_block(cp), n, get_irn_in(cp) + 1);
793 sched_add_before(cp, keep);
795 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
801 ir_nodeset_destroy(&entry->copies);
804 ir_nodehashmap_destroy(&cenv.op_set);
805 obstack_free(&cenv.obst, NULL);
806 be_invalidate_live_sets(irg);
810 * Push nodes that do not need to be permed through the Perm.
811 * This is commonly a reload cascade at block ends.
812 * @note This routine needs interference.
813 * @note Probably, we can implement it a little more efficient.
814 * Especially searching the frontier lazily might be better.
816 * @param perm The perm
817 * @param env The lowerer environment
819 * @return 1, if there is something left to perm over.
820 * 0, if removed the complete perm.
822 static int push_through_perm(ir_node *perm)
824 ir_graph *irg = get_irn_irg(perm);
825 ir_node *bl = get_nodes_block(perm);
826 int arity = get_irn_arity(perm);
829 bitset_t *moved = bitset_alloca(arity);
832 ir_node *frontier = bl;
834 be_lv_t *lv = be_get_irg_liveness(irg);
836 /* get some Proj and find out the register class of that Proj. */
837 ir_node *one_proj = get_edge_src_irn(get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL));
838 const arch_register_class_t *cls = arch_get_irn_reg_class(one_proj);
839 assert(is_Proj(one_proj));
841 DB((dbg_permmove, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
843 /* Find the point in the schedule after which the
844 * potentially movable nodes must be defined.
845 * A Perm will only be pushed up to first instruction
846 * which lets an operand of itself die.
847 * If we would allow to move the Perm above this instruction,
848 * the former dead operand would be live now at the point of
849 * the Perm, increasing the register pressure by one.
851 sched_foreach_reverse_from(sched_prev(perm), irn) {
852 be_foreach_use(irn, cls, in_req_, op, op_req_,
853 if (!be_values_interfere(lv, op, one_proj)) {
861 DB((dbg_permmove, LEVEL_2, "\tfrontier: %+F\n", frontier));
865 ir_node *const node = sched_prev(perm);
866 if (node == frontier)
869 const arch_register_req_t *req;
871 ir_node *proj = NULL;
873 /* search if node is a INPUT of Perm */
874 foreach_out_edge(perm, edge) {
875 ir_node *out = get_edge_src_irn(edge);
876 int pn = get_Proj_proj(out);
877 ir_node *in = get_irn_n(perm, pn);
884 /* it wasn't an input to the perm, we can't do anything more */
887 if (arch_irn_is(node, modify_flags))
889 req = arch_get_irn_register_req(node);
890 if (req->type != arch_register_req_type_normal)
892 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
893 ir_node *opop = get_irn_n(node, i);
894 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
901 DBG((dbg_permmove, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
903 /* move the movable node in front of the Perm */
905 sched_add_after(perm, node);
907 /* give it the proj's register */
908 arch_set_irn_register(node, arch_get_irn_register(proj));
910 /* reroute all users of the proj to the moved node. */
911 exchange(proj, node);
913 bitset_set(moved, input);
917 /* well, we could not push anything through the perm */
921 new_size = arity - n_moved;
928 map = ALLOCAN(int, new_size);
929 proj_map = ALLOCAN(int, arity);
930 memset(proj_map, -1, sizeof(proj_map[0]));
932 for (i = 0; i < arity; ++i) {
933 if (bitset_is_set(moved, i))
939 assert(n == new_size);
940 foreach_out_edge(perm, edge) {
941 ir_node *proj = get_edge_src_irn(edge);
942 int pn = get_Proj_proj(proj);
945 set_Proj_proj(proj, pn);
948 be_Perm_reduce(perm, new_size, map);
953 * Calls the corresponding lowering function for the node.
955 * @param irn The node to be checked for lowering
956 * @param walk_env The walker environment
958 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
962 if (!be_is_Perm(irn))
965 perm_stayed = push_through_perm(irn);
967 lower_perm_node(irn, (lower_env_t*)walk_env);
970 void lower_nodes_after_ra(ir_graph *irg, int do_copy)
974 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
975 FIRM_DBG_REGISTER(dbg_permmove, "firm.be.lower.permmove");
978 env.do_copy = do_copy;
980 /* we will need interference */
981 be_assure_live_chk(irg);
983 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);