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/out 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
173 * @param in_out 0 == look for IN register, 1 == look for OUT register
174 * @return The corresponding index in pairs or -1 if not found
176 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
180 for (i = 0; i < n; i++) {
181 /* out register matches */
182 if ((int) pairs[i].out_reg->index == reg_idx)
187 for (i = 0; i < n; i++) {
188 /* in register matches */
189 if ((int) pairs[i].in_reg->index == reg_idx)
198 * Gets an array of register pairs and tries to identify a cycle or chain starting
201 * @param cycle Variable to hold the cycle
202 * @param pairs Array of register pairs
203 * @param start Index to start
204 * @return The cycle or chain
206 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
207 int head = pairs[start].in_reg->index;
208 int cur_idx = pairs[start].out_reg->index;
209 int cur_pair_idx = start;
210 int n_pairs_done = get_n_checked_pairs(pairs, n);
212 perm_type_t cycle_tp = PERM_CYCLE;
214 /* We could be right in the middle of a chain, so we need to find the start */
215 while (head != cur_idx) {
216 /* goto previous register in cycle or chain */
217 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
219 if (cur_pair_idx < 0) {
220 cycle_tp = PERM_CHAIN;
224 head = pairs[cur_pair_idx].in_reg->index;
225 start = cur_pair_idx;
229 /* assume worst case: all remaining pairs build a cycle or chain */
230 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
231 cycle->n_elems = 2; /* initial number of elements is 2 */
232 cycle->elems[0] = pairs[start].in_reg;
233 cycle->elems[1] = pairs[start].out_reg;
234 cycle->type = cycle_tp;
235 cur_idx = pairs[start].out_reg->index;
238 /* check for cycle or end of a chain */
239 while (cur_idx != head) {
240 /* goto next register in cycle or chain */
241 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
243 if (cur_pair_idx < 0)
246 cur_idx = pairs[cur_pair_idx].out_reg->index;
248 /* it's not the first element: insert it */
249 if (cur_idx != head) {
250 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
254 /* we are there where we started -> CYCLE */
255 cycle->type = PERM_CYCLE;
259 /* mark all pairs having one in/out register with cycle in common as checked */
260 for (idx = 0; idx < cycle->n_elems; idx++) {
261 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
263 if (cur_pair_idx >= 0)
264 pairs[cur_pair_idx].checked = 1;
266 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
268 if (cur_pair_idx >= 0)
269 pairs[cur_pair_idx].checked = 1;
276 * Lowers a perm node. Resolves cycles and creates a bunch of
277 * copy and swap operations to permute registers.
278 * Note: The caller of this function has to make sure, that irn
281 * @param irn The perm node
282 * @param block The block the perm node belongs to
283 * @param walk_env The environment
285 static void lower_perm_node(ir_node *irn, void *walk_env) {
286 ir_graph *irg = get_irn_irg(irn);
287 const arch_register_class_t *reg_class;
288 lower_env_t *env = walk_env;
290 int n, i, pn, do_copy, j, n_ops;
292 const ir_edge_t *edge;
293 ir_node *sched_point, *block, *in[2];
294 ir_node *arg1, *arg2, *res1, *res2;
295 ir_node *cpyxchg = NULL;
297 do_copy = env->do_copy;
298 block = get_nodes_block(irn);
301 Get the schedule predecessor node to the perm
302 NOTE: This works with auto-magic. If we insert the
303 new copy/exchange nodes after this node, everything
306 sched_point = sched_prev(irn);
307 DBG((dbg, LEVEL_1, "perm: %+F\n", irn));
308 DBG((dbg, LEVEL_1, "sched point is %+F\n", sched_point));
309 assert(sched_point && "Perm is not scheduled or has no predecessor");
311 n = get_irn_arity(irn);
312 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
314 reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
315 pairs = alloca(n * sizeof(pairs[0]));
317 /* build the list of register pairs (in, out) */
319 foreach_out_edge(irn, edge) {
320 pairs[i].out_node = get_edge_src_irn(edge);
321 pn = get_Proj_proj(pairs[i].out_node);
322 pairs[i].in_node = get_irn_n(irn, pn);
324 pairs[i].in_reg = arch_get_irn_register(pairs[i].in_node);
325 pairs[i].out_reg = arch_get_irn_register(pairs[i].out_node);
327 pairs[i].checked = 0;
331 /* sort the register pairs by the indices of the in registers */
332 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
334 /* Mark all equal pairs as checked, and exchange the OUT proj with
336 for (i = 0; i < n; i++) {
337 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
338 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
339 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
341 /* reroute the edges from the proj to the argument */
342 exchange(pairs[i].out_node, pairs[i].in_node);
344 pairs[i].checked = 1;
348 /* Set do_copy to 0 if it's on but we have no free register */
353 /* check for cycles and chains */
354 while (get_n_checked_pairs(pairs, n) < n) {
359 /* go to the first not-checked pair */
360 while (pairs[i].checked) i++;
361 cycle = XMALLOCZ(perm_cycle_t);
362 cycle = get_perm_cycle(cycle, pairs, n, i);
364 DB((dbg, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
365 for (j = 0; j < cycle->n_elems; j++) {
366 DB((dbg, LEVEL_1, " %s", cycle->elems[j]->name));
368 DB((dbg, LEVEL_1, "\n"));
371 We don't need to do anything if we have a Perm with two
372 elements which represents a cycle, because those nodes
373 already represent exchange nodes
375 if (n == 2 && cycle->type == PERM_CYCLE) {
381 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
382 // the copy cascade (this reduces the cycle into a chain)
384 /* build copy/swap nodes from back to front */
385 for (i = cycle->n_elems - 2; i >= 0; i--) {
386 arg1 = get_node_for_in_register(pairs, n, cycle->elems[i]);
387 arg2 = get_node_for_in_register(pairs, n, cycle->elems[i + 1]);
389 res1 = get_node_for_out_register(pairs, n, cycle->elems[i]);
390 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 nodes
393 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
394 IN_1 = in node with register i
395 IN_2 = in node with register i + 1
396 OUT_1 = out node with register i + 1
397 OUT_2 = out node with register i
399 if (cycle->type == PERM_CYCLE && !do_copy) {
403 /* At this point we have to handle the following problem: */
405 /* If we have a cycle with more than two elements, then */
406 /* this could correspond to the following Perm node: */
408 /* +----+ +----+ +----+ */
409 /* | r1 | | r2 | | r3 | */
410 /* +-+--+ +-+--+ +--+-+ */
413 /* +-+--------+---------+-+ */
415 /* +-+--------+---------+-+ */
418 /* +-+--+ +-+--+ +--+-+ */
419 /* |Proj| |Proj| |Proj| */
420 /* | r2 | | r3 | | r1 | */
421 /* +----+ +----+ +----+ */
423 /* This node is about to be split up into two 2x Perm's */
424 /* for which we need 4 Proj's and the one additional Proj */
425 /* of the first Perm has to be one IN of the second. So in */
426 /* general we need to create one additional Proj for each */
427 /* "middle" Perm and set this to one in node of the successor */
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);
439 /* cycle is not done yet */
440 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
442 /* create intermediate proj */
443 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
445 /* set as in for next Perm */
446 pairs[pidx].in_node = res1;
449 set_Proj_pred(res2, cpyxchg);
450 set_Proj_proj(res2, 0);
451 set_Proj_pred(res1, cpyxchg);
452 set_Proj_proj(res1, 1);
454 arch_set_irn_register(res2, cycle->elems[i + 1]);
455 arch_set_irn_register(res1, cycle->elems[i]);
457 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
458 sched_add_after(sched_point, cpyxchg);
460 DBG((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
462 /* 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]);
473 /* exchange copy node and proj */
474 exchange(res2, cpyxchg);
476 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
477 sched_add_after(sched_point, cpyxchg);
479 /* set the new scheduling point */
480 sched_point = cpyxchg;
484 free((void *) cycle->elems);
488 /* remove the perm from schedule */
497 static int has_irn_users(const ir_node *irn) {
498 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
502 * Skip all Proj nodes.
504 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
506 irn = get_Proj_pred(irn);
510 static ir_node *find_copy(ir_node *irn, ir_node *op)
514 for (cur_node = irn;;) {
515 cur_node = sched_prev(cur_node);
516 if (! be_is_Copy(cur_node))
518 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
523 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
524 be_irg_t *birg = env->birg;
525 ir_graph *irg = be_get_birg_irg(birg);
526 ir_nodemap_t *op_set = &env->op_set;
527 ir_node *block = get_nodes_block(irn);
528 const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1);
529 ir_node *in[2], *keep, *cpy;
530 op_copy_assoc_t *entry;
532 if (arch_irn_is(other_different, ignore) ||
533 !mode_is_datab(get_irn_mode(other_different))) {
534 DBG((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
538 /* Make a not spillable copy of the different node */
539 /* this is needed because the different irn could be */
540 /* in block far far away */
541 /* The copy is optimized later if not needed */
543 /* check if already exists such a copy in the schedule immediately before */
544 cpy = find_copy(belower_skip_proj(irn), other_different);
546 cpy = be_new_Copy(cls, irg, block, other_different);
547 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
548 DBG((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
551 DBG((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
557 /* Add the Keep resp. CopyKeep and reroute the users */
558 /* of the other_different irn in case of CopyKeep. */
559 if (has_irn_users(other_different)) {
560 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
561 be_node_set_reg_class(keep, 1, cls);
564 keep = be_new_Keep(cls, irg, block, 2, in);
567 DBG((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
569 /* insert copy and keep into schedule */
570 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
571 if (! sched_is_scheduled(cpy))
572 sched_add_before(belower_skip_proj(irn), cpy);
573 sched_add_after(irn, keep);
575 /* insert the other different and it's copies into the map */
576 entry = ir_nodemap_get(op_set, other_different);
578 entry = obstack_alloc(&env->obst, sizeof(*entry));
580 ir_nodeset_init(&entry->copies);
582 ir_nodemap_insert(op_set, other_different, entry);
586 ir_nodeset_insert(&entry->copies, cpy);
588 /* insert keep in case of CopyKeep */
589 if (be_is_CopyKeep(keep)) {
590 ir_nodeset_insert(&entry->copies, keep);
595 * Checks if node has a must_be_different constraint in output and adds a Keep
596 * then to assure the constraint.
598 * @param irn the node to check
599 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
600 * @param env the constraint environment
602 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
603 const arch_register_req_t *req = arch_get_register_req(irn, -1);
605 if (arch_register_req_is(req, must_be_different)) {
606 const unsigned other = req->other_different;
609 if (arch_register_req_is(req, should_be_same)) {
610 const unsigned same = req->other_same;
612 if (is_po2(other) && is_po2(same)) {
613 int idx_other = ntz(other);
614 int idx_same = ntz(same);
617 * We can safely ignore a should_be_same x must_be_different y
618 * IFF both inputs are equal!
620 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
625 for (i = 0; 1U << i <= other; ++i) {
626 if (other & (1U << i)) {
627 ir_node *different_from = get_irn_n(skipped_irn, i);
628 gen_assure_different_pattern(irn, different_from, env);
635 * Calls the functions to assure register constraints.
637 * @param block The block to be checked
638 * @param walk_env The walker environment
640 static void assure_constraints_walker(ir_node *block, void *walk_env) {
643 sched_foreach_reverse(block, irn) {
644 ir_mode *mode = get_irn_mode(irn);
646 if (mode == mode_T) {
647 const ir_edge_t *edge;
649 foreach_out_edge(irn, edge) {
650 ir_node *proj = get_edge_src_irn(edge);
652 mode = get_irn_mode(proj);
653 if (mode_is_datab(mode))
654 assure_different_constraints(proj, irn, walk_env);
656 } else if (mode_is_datab(mode)) {
657 assure_different_constraints(irn, irn, walk_env);
663 * Melt all copykeeps pointing to the same node
664 * (or Projs of the same node), copying the same operand.
666 static void melt_copykeeps(constraint_env_t *cenv) {
667 be_irg_t *birg = cenv->birg;
668 ir_graph *irg = be_get_birg_irg(birg);
669 ir_nodemap_iterator_t map_iter;
670 ir_nodemap_entry_t map_entry;
673 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
674 op_copy_assoc_t *entry = map_entry.data;
678 ir_nodeset_iterator_t iter;
679 ir_node **ck_arr, **melt_arr;
683 /* collect all copykeeps */
685 foreach_ir_nodeset(&entry->copies, cp, iter) {
686 if (be_is_CopyKeep(cp)) {
687 obstack_grow(&obst, &cp, sizeof(cp));
690 #ifdef KEEP_ALIVE_COPYKEEP_HACK
692 set_irn_mode(cp, mode_ANY);
695 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
698 /* compare each copykeep with all other copykeeps */
699 ck_arr = (ir_node **)obstack_finish(&obst);
700 for (idx = 0; idx < num_ck; ++idx) {
701 ir_node *ref, *ref_mode_T;
707 ir_node *sched_pt = NULL;
711 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
712 obstack_grow(&obst, &ref, sizeof(ref));
714 DBG((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
716 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
717 for (j = 0; j < num_ck; ++j) {
718 ir_node *cur_ck = ck_arr[j];
720 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
721 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
722 ir_nodeset_remove(&entry->copies, cur_ck);
723 DBG((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
726 sched_remove(cur_ck);
731 /* check, if we found some candidates for melting */
733 DBG((dbg_constr, LEVEL_1, "\tno candidate found\n"));
737 ir_nodeset_remove(&entry->copies, ref);
740 melt_arr = (ir_node **)obstack_finish(&obst);
741 /* melt all found copykeeps */
742 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
743 for (j = 0; j < n_melt; ++j) {
744 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
746 /* now, we can kill the melted keep, except the */
747 /* ref one, we still need some information */
748 if (melt_arr[j] != ref)
749 kill_node(melt_arr[j]);
752 #ifdef KEEP_ALIVE_COPYKEEP_HACK
753 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
756 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));
757 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
759 /* set register class for all kept inputs */
760 for (j = 1; j <= n_melt; ++j)
761 be_node_set_reg_class(new_ck, j, entry->cls);
763 ir_nodeset_insert(&entry->copies, new_ck);
765 /* find scheduling point */
766 sched_pt = ref_mode_T;
768 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
769 sched_pt = sched_next(sched_pt);
770 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
772 sched_add_before(sched_pt, new_ck);
773 DBG((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
775 /* finally: kill the reference copykeep */
780 obstack_free(&obst, NULL);
785 * Walks over all nodes to assure register constraints.
787 * @param birg The birg structure containing the irg
789 void assure_constraints(be_irg_t *birg) {
790 ir_graph *irg = be_get_birg_irg(birg);
791 constraint_env_t cenv;
793 ir_nodemap_iterator_t map_iter;
794 ir_nodemap_entry_t map_entry;
796 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
799 ir_nodemap_init(&cenv.op_set);
800 obstack_init(&cenv.obst);
802 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
804 /* melt copykeeps, pointing to projs of */
805 /* the same mode_T node and keeping the */
807 melt_copykeeps(&cenv);
810 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
811 op_copy_assoc_t *entry = map_entry.data;
814 ir_nodeset_iterator_t iter;
815 be_ssa_construction_env_t senv;
817 n = ir_nodeset_size(&entry->copies);
818 nodes = alloca(n * sizeof(nodes[0]));
820 /* put the node in an array */
821 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
823 /* collect all copies */
825 foreach_ir_nodeset(&entry->copies, cp, iter) {
827 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
830 DB((dbg_constr, LEVEL_1, "\n"));
832 /* introduce the copies for the operand and it's copies */
833 be_ssa_construction_init(&senv, birg);
834 be_ssa_construction_add_copy(&senv, map_entry.node);
835 be_ssa_construction_add_copies(&senv, nodes, n);
836 be_ssa_construction_fix_users(&senv, map_entry.node);
837 be_ssa_construction_destroy(&senv);
839 /* Could be that not all CopyKeeps are really needed, */
840 /* so we transform unnecessary ones into Keeps. */
841 foreach_ir_nodeset(&entry->copies, cp, iter) {
842 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
844 int n = get_irn_arity(cp);
846 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
847 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
848 sched_add_before(cp, keep);
850 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
856 ir_nodeset_destroy(&entry->copies);
859 ir_nodemap_destroy(&cenv.op_set);
860 obstack_free(&cenv.obst, NULL);
861 be_liveness_invalidate(be_get_birg_liveness(birg));
866 * Push nodes that do not need to be permed through the Perm.
867 * This is commonly a reload cascade at block ends.
868 * @note This routine needs interference.
869 * @note Probably, we can implement it a little more efficient.
870 * Especially searching the frontier lazily might be better.
871 * @param perm The perm.
872 * @param data The walker data (lower_env_t).
873 * @return 1, if there is something left to perm over.
874 * 0, if removed the complete perm.
876 static int push_through_perm(ir_node *perm, void *data)
878 lower_env_t *env = data;
880 ir_graph *irg = get_irn_irg(perm);
881 ir_node *bl = get_nodes_block(perm);
883 int arity = get_irn_arity(perm);
886 bitset_t *moved = bitset_alloca(arity);
889 ir_node *frontier = bl;
890 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
893 const ir_edge_t *edge;
894 ir_node *one_proj = NULL, *irn;
895 const arch_register_class_t *cls = NULL;
897 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
899 /* get some Proj and find out the register class of that Proj. */
900 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
901 one_proj = get_edge_src_irn(edge);
902 assert(is_Proj(one_proj));
903 cls = arch_get_irn_reg_class(one_proj, -1);
905 /* Find the point in the schedule after which the
906 * potentially movable nodes must be defined.
907 * A Perm will only be pushed up to first instruction
908 * which lets an operand of itself die.
909 * If we would allow to move the Perm above this instruction,
910 * the former dead operand would be live now at the point of
911 * the Perm, increasing the register pressure by one.
913 sched_foreach_reverse_from (sched_prev(perm), irn) {
914 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
915 ir_node *op = get_irn_n(irn, i);
916 if (arch_irn_consider_in_reg_alloc(cls, op) &&
917 !values_interfere(env->birg, op, one_proj)) {
925 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
927 node = sched_prev(perm);
929 while(!sched_is_begin(node)) {
930 const arch_register_req_t *req;
934 /* search if node is a INPUT of Perm */
935 foreach_out_edge(perm, edge) {
936 ir_node *out = get_edge_src_irn(edge);
937 int pn = get_Proj_proj(out);
938 ir_node *in = get_irn_n(perm, pn);
945 /* it wasn't an input to the perm, we can't do anything more */
948 if(!sched_comes_after(frontier, node))
950 if (arch_irn_is(node, modify_flags))
953 req = arch_get_register_req(get_Proj_pred(node),
954 -1 - get_Proj_proj(node));
956 req = arch_get_register_req(node, -1);
958 if(req->type != arch_register_req_type_normal)
960 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
961 ir_node *opop = get_irn_n(node, i);
962 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
969 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
971 /* move the movable node in front of the Perm */
973 sched_add_after(perm, node);
975 /* give it the proj's register */
976 arch_set_irn_register(node, arch_get_irn_register(proj));
978 /* reroute all users of the proj to the moved node. */
979 edges_reroute(proj, node, irg);
982 set_Proj_pred(proj, new_Bad());
985 bitset_set(moved, input);
988 node = sched_prev(node);
991 /* well, we could not push anything through the perm */
995 new_size = arity - n_moved;
1000 map = alloca(new_size * sizeof(map[0]));
1001 proj_map = alloca(arity * sizeof(proj_map[0]));
1002 memset(proj_map, -1, sizeof(proj_map[0]));
1004 for(i = 0; i < arity; ++i) {
1005 if(bitset_is_set(moved, i))
1011 assert(n == new_size);
1012 foreach_out_edge(perm, edge) {
1013 ir_node *proj = get_edge_src_irn(edge);
1014 int pn = get_Proj_proj(proj);
1017 set_Proj_proj(proj, pn);
1020 be_Perm_reduce(perm, new_size, map);
1025 * Calls the corresponding lowering function for the node.
1027 * @param irn The node to be checked for lowering
1028 * @param walk_env The walker environment
1030 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
1034 if (!be_is_Perm(irn))
1037 perm_stayed = push_through_perm(irn, walk_env);
1041 lower_perm_node(irn, walk_env);
1045 * Walks over all blocks in an irg and performs lowering need to be
1046 * done after register allocation (e.g. perm lowering).
1048 * @param birg The birg object
1049 * @param do_copy 1 == resolve cycles with a free reg if available
1051 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1055 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1058 env.do_copy = do_copy;
1060 /* we will need interference */
1061 be_liveness_assure_chk(be_get_birg_liveness(birg));
1063 irg = be_get_birg_irg(birg);
1064 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);