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 /** Associates an ir_node with it's copy and CopyKeep. */
54 ir_nodeset_t copies; /**< all non-spillable copies of this irn */
55 const arch_register_class_t *cls;
58 /** Environment for constraints. */
63 DEBUG_ONLY(firm_dbg_module_t *dbg;)
66 /** Lowering walker environment. */
67 typedef struct _lower_env_t {
70 DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
73 /** Holds a Perm register pair. */
74 typedef struct _reg_pair_t {
75 const arch_register_t *in_reg; /**< a perm IN register */
76 ir_node *in_node; /**< the in node to which the register belongs */
78 const arch_register_t *out_reg; /**< a perm OUT register */
79 ir_node *out_node; /**< the out node to which the register belongs */
81 int checked; /**< indicates whether the pair was check for cycle or not */
84 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 /** Compare the in registers of two register pairs. */
99 static int compare_reg_pair(const void *a, const void *b) {
100 const reg_pair_t *pair_a = a;
101 const reg_pair_t *pair_b = b;
103 if (pair_a->in_reg->index > pair_b->in_reg->index)
109 /** returns the number register pairs marked as checked. */
110 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
111 int i, n_checked = 0;
113 for (i = 0; i < n; i++) {
114 if (pairs[i].checked)
122 * Gets the node corresponding to a register from an array of register pairs.
123 * NOTE: The given registers pairs and the register to look for must belong
124 * to the same register class.
126 * @param pairs The array of register pairs
127 * @param n The number of pairs
128 * @param reg The register to look for
129 * @param in_out 0 == look for IN register, 1 == look for OUT register
130 * @return The corresponding node or NULL if not found
132 static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
136 for (i = 0; i < n; i++) {
137 /* out register matches */
138 if (pairs[i].out_reg->index == reg->index)
139 return pairs[i].out_node;
143 for (i = 0; i < n; i++) {
144 /* in register matches */
145 if (pairs[i].in_reg->index == reg->index)
146 return pairs[i].in_node;
154 * Gets the index in the register pair array where the in/out register
155 * corresponds to reg_idx.
157 * @param pairs The array of register pairs
158 * @param n The number of pairs
159 * @param reg The register index to look for
160 * @param in_out 0 == look for IN register, 1 == look for OUT register
161 * @return The corresponding index in pairs or -1 if not found
163 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
167 for (i = 0; i < n; i++) {
168 /* out register matches */
169 if ((int) pairs[i].out_reg->index == reg_idx)
174 for (i = 0; i < n; i++) {
175 /* in register matches */
176 if ((int) pairs[i].in_reg->index == reg_idx)
185 * Gets an array of register pairs and tries to identify a cycle or chain starting
188 * @param cycle Variable to hold the cycle
189 * @param pairs Array of register pairs
190 * @param start Index to start
191 * @return The cycle or chain
193 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
194 int head = pairs[start].in_reg->index;
195 int cur_idx = pairs[start].out_reg->index;
196 int cur_pair_idx = start;
197 int n_pairs_done = get_n_checked_pairs(pairs, n);
199 perm_type_t cycle_tp = PERM_CYCLE;
201 /* We could be right in the middle of a chain, so we need to find the start */
202 while (head != cur_idx) {
203 /* goto previous register in cycle or chain */
204 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
206 if (cur_pair_idx < 0) {
207 cycle_tp = PERM_CHAIN;
211 head = pairs[cur_pair_idx].in_reg->index;
212 start = cur_pair_idx;
216 /* assume worst case: all remaining pairs build a cycle or chain */
217 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
218 cycle->n_elems = 2; /* initial number of elements is 2 */
219 cycle->elems[0] = pairs[start].in_reg;
220 cycle->elems[1] = pairs[start].out_reg;
221 cycle->type = cycle_tp;
222 cur_idx = pairs[start].out_reg->index;
225 /* check for cycle or end of a chain */
226 while (cur_idx != head) {
227 /* goto next register in cycle or chain */
228 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
230 if (cur_pair_idx < 0)
233 cur_idx = pairs[cur_pair_idx].out_reg->index;
235 /* it's not the first element: insert it */
236 if (cur_idx != head) {
237 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
241 /* we are there where we started -> CYCLE */
242 cycle->type = PERM_CYCLE;
246 /* mark all pairs having one in/out register with cycle in common as checked */
247 for (idx = 0; idx < cycle->n_elems; idx++) {
248 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
250 if (cur_pair_idx >= 0)
251 pairs[cur_pair_idx].checked = 1;
253 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
255 if (cur_pair_idx >= 0)
256 pairs[cur_pair_idx].checked = 1;
263 * Lowers a perm node. Resolves cycles and creates a bunch of
264 * copy and swap operations to permute registers.
265 * Note: The caller of this function has to make sure, that irn
268 * @param irn The perm node
269 * @param block The block the perm node belongs to
270 * @param walk_env The environment
272 static void lower_perm_node(ir_node *irn, void *walk_env) {
273 ir_graph *irg = get_irn_irg(irn);
274 const arch_register_class_t *reg_class;
275 lower_env_t *env = walk_env;
278 int n, i, pn, do_copy, j, n_ops;
280 const ir_edge_t *edge;
281 ir_node *sched_point, *block, *in[2];
282 ir_node *arg1, *arg2, *res1, *res2;
283 ir_node *cpyxchg = NULL;
284 DEBUG_ONLY(firm_dbg_module_t *mod;)
286 do_copy = env->do_copy;
287 DEBUG_ONLY(mod = env->dbg_module;)
288 block = get_nodes_block(irn);
291 Get the schedule predecessor node to the perm
292 NOTE: This works with auto-magic. If we insert the
293 new copy/exchange nodes after this node, everything
296 sched_point = sched_prev(irn);
297 DBG((mod, LEVEL_1, "perm: %+F\n", irn));
298 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
299 assert(sched_point && "Perm is not scheduled or has no predecessor");
301 n = get_irn_arity(irn);
302 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
304 reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
305 pairs = alloca(n * sizeof(pairs[0]));
307 /* build the list of register pairs (in, out) */
309 foreach_out_edge(irn, edge) {
310 pairs[i].out_node = get_edge_src_irn(edge);
311 pn = get_Proj_proj(pairs[i].out_node);
312 pairs[i].in_node = get_irn_n(irn, pn);
314 pairs[i].in_reg = arch_get_irn_register(pairs[i].in_node);
315 pairs[i].out_reg = arch_get_irn_register(pairs[i].out_node);
317 pairs[i].checked = 0;
321 /* sort the register pairs by the indices of the in registers */
322 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
324 /* Mark all equal pairs as checked, and exchange the OUT proj with
326 for (i = 0; i < n; i++) {
327 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
328 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
329 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
331 /* reroute the edges from the proj to the argument */
332 exchange(pairs[i].out_node, pairs[i].in_node);
334 pairs[i].checked = 1;
338 /* Set do_copy to 0 if it's on but we have no free register */
343 real_size = n - get_n_checked_pairs(pairs, n);
345 /* check for cycles and chains */
346 while (get_n_checked_pairs(pairs, n) < n) {
351 /* go to the first not-checked pair */
352 while (pairs[i].checked) i++;
353 cycle = XMALLOCZ(perm_cycle_t);
354 cycle = get_perm_cycle(cycle, pairs, n, i);
356 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
357 for (j = 0; j < cycle->n_elems; j++) {
358 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
360 DB((mod, LEVEL_1, "\n"));
363 We don't need to do anything if we have a Perm with two
364 elements which represents a cycle, because those nodes
365 already represent exchange nodes
367 if (n == 2 && cycle->type == PERM_CYCLE) {
373 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
374 // the copy cascade (this reduces the cycle into a chain)
376 /* build copy/swap nodes from back to front */
377 for (i = cycle->n_elems - 2; i >= 0; i--) {
378 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
379 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
381 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
382 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
384 If we have a cycle and don't copy: we need to create exchange nodes
385 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
386 IN_1 = in node with register i
387 IN_2 = in node with register i + 1
388 OUT_1 = out node with register i + 1
389 OUT_2 = out node with register i
391 if (cycle->type == PERM_CYCLE && !do_copy) {
395 /* At this point we have to handle the following problem: */
397 /* If we have a cycle with more than two elements, then */
398 /* this could correspond to the following Perm node: */
400 /* +----+ +----+ +----+ */
401 /* | r1 | | r2 | | r3 | */
402 /* +-+--+ +-+--+ +--+-+ */
405 /* +-+--------+---------+-+ */
407 /* +-+--------+---------+-+ */
410 /* +-+--+ +-+--+ +--+-+ */
411 /* |Proj| |Proj| |Proj| */
412 /* | r2 | | r3 | | r1 | */
413 /* +----+ +----+ +----+ */
415 /* This node is about to be split up into two 2x Perm's */
416 /* for which we need 4 Proj's and the one additional Proj */
417 /* of the first Perm has to be one IN of the second. So in */
418 /* general we need to create one additional Proj for each */
419 /* "middle" Perm and set this to one in node of the successor */
422 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
423 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
424 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
425 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
427 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
431 /* cycle is not done yet */
432 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
434 /* create intermediate proj */
435 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
437 /* set as in for next Perm */
438 pairs[pidx].in_node = res1;
441 set_Proj_pred(res2, cpyxchg);
442 set_Proj_proj(res2, 0);
443 set_Proj_pred(res1, cpyxchg);
444 set_Proj_proj(res1, 1);
446 arch_set_irn_register(res2, cycle->elems[i + 1]);
447 arch_set_irn_register(res1, cycle->elems[i]);
449 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
450 sched_add_after(sched_point, cpyxchg);
452 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
454 /* set the new scheduling point */
458 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
459 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
461 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
462 arch_set_irn_register(cpyxchg, cycle->elems[i + 1]);
465 /* exchange copy node and proj */
466 exchange(res2, cpyxchg);
468 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
469 sched_add_after(sched_point, cpyxchg);
471 /* set the new scheduling point */
472 sched_point = cpyxchg;
476 free((void *) cycle->elems);
480 /* remove the perm from schedule */
489 static int has_irn_users(const ir_node *irn) {
490 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
494 * Skip all Proj nodes.
496 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
498 irn = get_Proj_pred(irn);
502 static ir_node *find_copy(ir_node *irn, ir_node *op)
504 ir_node *block = get_nodes_block(irn);
507 for (cur_node = sched_prev(irn);
508 ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
509 cur_node = sched_prev(cur_node))
511 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
518 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
519 be_irg_t *birg = env->birg;
520 ir_graph *irg = be_get_birg_irg(birg);
521 ir_nodemap_t *op_set = &env->op_set;
522 ir_node *block = get_nodes_block(irn);
523 const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1);
524 ir_node *in[2], *keep, *cpy;
525 op_copy_assoc_t *entry;
526 DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
528 if (arch_irn_is(other_different, ignore) ||
529 !mode_is_datab(get_irn_mode(other_different))) {
530 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
534 /* Make a not spillable copy of the different node */
535 /* this is needed because the different irn could be */
536 /* in block far far away */
537 /* The copy is optimized later if not needed */
539 /* check if already exists such a copy in the schedule immediately before */
540 cpy = find_copy(belower_skip_proj(irn), other_different);
542 cpy = be_new_Copy(cls, irg, block, other_different);
543 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
544 DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
547 DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
553 /* Add the Keep resp. CopyKeep and reroute the users */
554 /* of the other_different irn in case of CopyKeep. */
555 if (has_irn_users(other_different)) {
556 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
557 be_node_set_reg_class(keep, 1, cls);
560 keep = be_new_Keep(cls, irg, block, 2, in);
563 DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
565 /* insert copy and keep into schedule */
566 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
567 if (! sched_is_scheduled(cpy))
568 sched_add_before(belower_skip_proj(irn), cpy);
569 sched_add_after(irn, keep);
571 /* insert the other different and it's copies into the map */
572 entry = ir_nodemap_get(op_set, other_different);
574 entry = obstack_alloc(&env->obst, sizeof(*entry));
576 ir_nodeset_init(&entry->copies);
578 ir_nodemap_insert(op_set, other_different, entry);
582 ir_nodeset_insert(&entry->copies, cpy);
584 /* insert keep in case of CopyKeep */
585 if (be_is_CopyKeep(keep)) {
586 ir_nodeset_insert(&entry->copies, keep);
591 * Checks if node has a must_be_different constraint in output and adds a Keep
592 * then to assure the constraint.
594 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
595 ir_node *skipped_irn = belower_skip_proj(irn);
596 const arch_register_req_t *req = arch_get_register_req(irn, -1);
598 if (arch_register_req_is(req, must_be_different)) {
599 const unsigned other = req->other_different;
602 if (arch_register_req_is(req, should_be_same)) {
603 const unsigned same = req->other_same;
605 if (is_po2(other) && is_po2(same)) {
606 int idx_other = ntz(other);
607 int idx_same = ntz(same);
610 * We can safely ignore a should_be_same x must_be_different y
611 * IFF both inputs are equal!
613 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
618 for (i = 0; 1U << i <= other; ++i) {
619 if (other & (1U << i)) {
620 ir_node *different_from = get_irn_n(skipped_irn, i);
621 gen_assure_different_pattern(irn, different_from, env);
628 * Calls the functions to assure register constraints.
630 * @param irn The node to be checked for lowering
631 * @param walk_env The walker environment
633 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
637 if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
638 assure_different_constraints(irn, walk_env);
642 * Melt all copykeeps pointing to the same node
643 * (or Projs of the same node), copying the same operand.
645 static void melt_copykeeps(constraint_env_t *cenv) {
646 be_irg_t *birg = cenv->birg;
647 ir_graph *irg = be_get_birg_irg(birg);
648 ir_nodemap_iterator_t map_iter;
649 ir_nodemap_entry_t map_entry;
652 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
653 op_copy_assoc_t *entry = map_entry.data;
657 ir_nodeset_iterator_t iter;
658 ir_node **ck_arr, **melt_arr;
662 /* collect all copykeeps */
664 foreach_ir_nodeset(&entry->copies, cp, iter) {
665 if (be_is_CopyKeep(cp)) {
666 obstack_grow(&obst, &cp, sizeof(cp));
669 #ifdef KEEP_ALIVE_COPYKEEP_HACK
671 set_irn_mode(cp, mode_ANY);
674 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
677 /* compare each copykeep with all other copykeeps */
678 ck_arr = (ir_node **)obstack_finish(&obst);
679 for (idx = 0; idx < num_ck; ++idx) {
680 ir_node *ref, *ref_mode_T;
686 ir_node *sched_pt = NULL;
690 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
691 obstack_grow(&obst, &ref, sizeof(ref));
693 DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
695 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
696 for (j = 0; j < num_ck; ++j) {
697 ir_node *cur_ck = ck_arr[j];
699 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
700 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
701 ir_nodeset_remove(&entry->copies, cur_ck);
702 DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
705 sched_remove(cur_ck);
710 /* check, if we found some candidates for melting */
712 DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
716 ir_nodeset_remove(&entry->copies, ref);
719 melt_arr = (ir_node **)obstack_finish(&obst);
720 /* melt all found copykeeps */
721 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
722 for (j = 0; j < n_melt; ++j) {
723 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
725 /* now, we can kill the melted keep, except the */
726 /* ref one, we still need some information */
727 if (melt_arr[j] != ref)
728 kill_node(melt_arr[j]);
731 #ifdef KEEP_ALIVE_COPYKEEP_HACK
732 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
735 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));
736 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
738 /* set register class for all kept inputs */
739 for (j = 1; j <= n_melt; ++j)
740 be_node_set_reg_class(new_ck, j, entry->cls);
742 ir_nodeset_insert(&entry->copies, new_ck);
744 /* find scheduling point */
745 sched_pt = ref_mode_T;
747 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
748 sched_pt = sched_next(sched_pt);
749 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
751 sched_add_before(sched_pt, new_ck);
752 DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
754 /* finally: kill the reference copykeep */
759 obstack_free(&obst, NULL);
764 * Walks over all nodes to assure register constraints.
766 * @param birg The birg structure containing the irg
768 void assure_constraints(be_irg_t *birg) {
769 ir_graph *irg = be_get_birg_irg(birg);
770 constraint_env_t cenv;
772 ir_nodemap_iterator_t map_iter;
773 ir_nodemap_entry_t map_entry;
774 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
776 DEBUG_ONLY(cenv.dbg = mod;)
778 ir_nodemap_init(&cenv.op_set);
779 obstack_init(&cenv.obst);
781 irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
783 /* melt copykeeps, pointing to projs of */
784 /* the same mode_T node and keeping the */
786 melt_copykeeps(&cenv);
789 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
790 op_copy_assoc_t *entry = map_entry.data;
793 ir_nodeset_iterator_t iter;
794 be_ssa_construction_env_t senv;
796 n = ir_nodeset_size(&entry->copies);
797 nodes = alloca(n * sizeof(nodes[0]));
799 /* put the node in an array */
800 DBG((mod, LEVEL_1, "introduce copies for %+F ", map_entry.node));
802 /* collect all copies */
804 foreach_ir_nodeset(&entry->copies, cp, iter) {
806 DB((mod, LEVEL_1, ", %+F ", cp));
809 DB((mod, LEVEL_1, "\n"));
811 /* introduce the copies for the operand and it's copies */
812 be_ssa_construction_init(&senv, birg);
813 be_ssa_construction_add_copy(&senv, map_entry.node);
814 be_ssa_construction_add_copies(&senv, nodes, n);
815 be_ssa_construction_fix_users(&senv, map_entry.node);
816 be_ssa_construction_destroy(&senv);
818 /* Could be that not all CopyKeeps are really needed, */
819 /* so we transform unnecessary ones into Keeps. */
820 foreach_ir_nodeset(&entry->copies, cp, iter) {
821 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
823 int n = get_irn_arity(cp);
825 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
826 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
827 sched_add_before(cp, keep);
829 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
835 ir_nodeset_destroy(&entry->copies);
838 ir_nodemap_destroy(&cenv.op_set);
839 obstack_free(&cenv.obst, NULL);
840 be_liveness_invalidate(be_get_birg_liveness(birg));
845 * Push nodes that do not need to be permed through the Perm.
846 * This is commonly a reload cascade at block ends.
847 * @note This routine needs interference.
848 * @note Probably, we can implement it a little more efficient.
849 * Especially searching the frontier lazily might be better.
850 * @param perm The perm.
851 * @param data The walker data (lower_env_t).
852 * @return 1, if there is something left to perm over.
853 * 0, if removed the complete perm.
855 static int push_through_perm(ir_node *perm, void *data)
857 lower_env_t *env = data;
859 ir_graph *irg = get_irn_irg(perm);
860 ir_node *bl = get_nodes_block(perm);
862 int arity = get_irn_arity(perm);
865 bitset_t *moved = bitset_alloca(arity);
868 ir_node *frontier = bl;
869 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
872 const ir_edge_t *edge;
873 ir_node *one_proj = NULL, *irn;
874 const arch_register_class_t *cls = NULL;
876 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
878 /* get some Proj and find out the register class of that Proj. */
879 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
880 one_proj = get_edge_src_irn(edge);
881 assert(is_Proj(one_proj));
882 cls = arch_get_irn_reg_class(one_proj, -1);
884 /* Find the point in the schedule after which the
885 * potentially movable nodes must be defined.
886 * A Perm will only be pushed up to first instruction
887 * which lets an operand of itself die.
888 * If we would allow to move the Perm above this instruction,
889 * the former dead operand would be live now at the point of
890 * the Perm, increasing the register pressure by one.
892 sched_foreach_reverse_from (sched_prev(perm), irn) {
893 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
894 ir_node *op = get_irn_n(irn, i);
895 if (arch_irn_consider_in_reg_alloc(cls, op) &&
896 !values_interfere(env->birg, op, one_proj)) {
904 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
906 node = sched_prev(perm);
908 while(!sched_is_begin(node)) {
909 const arch_register_req_t *req;
913 /* search if node is a INPUT of Perm */
914 foreach_out_edge(perm, edge) {
915 ir_node *out = get_edge_src_irn(edge);
916 int pn = get_Proj_proj(out);
917 ir_node *in = get_irn_n(perm, pn);
924 /* it wasn't an input to the perm, we can't do anything more */
927 if(!sched_comes_after(frontier, node))
929 if (arch_irn_is(node, modify_flags))
932 req = arch_get_register_req(get_Proj_pred(node),
933 -1 - get_Proj_proj(node));
935 req = arch_get_register_req(node, -1);
937 if(req->type != arch_register_req_type_normal)
939 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
940 ir_node *opop = get_irn_n(node, i);
941 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
948 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
950 /* move the movable node in front of the Perm */
952 sched_add_after(perm, node);
954 /* give it the proj's register */
955 arch_set_irn_register(node, arch_get_irn_register(proj));
957 /* reroute all users of the proj to the moved node. */
958 edges_reroute(proj, node, irg);
961 set_Proj_pred(proj, new_Bad());
964 bitset_set(moved, input);
967 node = sched_prev(node);
970 /* well, we could not push anything through the perm */
974 new_size = arity - n_moved;
979 map = alloca(new_size * sizeof(map[0]));
980 proj_map = alloca(arity * sizeof(proj_map[0]));
981 memset(proj_map, -1, sizeof(proj_map[0]));
983 for(i = 0; i < arity; ++i) {
984 if(bitset_is_set(moved, i))
990 assert(n == new_size);
991 foreach_out_edge(perm, edge) {
992 ir_node *proj = get_edge_src_irn(edge);
993 int pn = get_Proj_proj(proj);
996 set_Proj_proj(proj, pn);
999 be_Perm_reduce(perm, new_size, map);
1004 * Calls the corresponding lowering function for the node.
1006 * @param irn The node to be checked for lowering
1007 * @param walk_env The walker environment
1009 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
1012 if (is_Block(irn) || is_Proj(irn))
1014 if (!be_is_Perm(irn))
1017 perm_stayed = push_through_perm(irn, walk_env);
1021 lower_perm_node(irn, walk_env);
1025 * Walks over all blocks in an irg and performs lowering need to be
1026 * done after register allocation (e.g. perm lowering).
1028 * @param birg The birg object
1029 * @param do_copy 1 == resolve cycles with a free reg if available
1031 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1033 ir_graph *irg = be_get_birg_irg(birg);
1036 env.do_copy = do_copy;
1037 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
1039 /* we will need interference */
1040 be_liveness_assure_chk(be_get_birg_liveness(birg));
1042 irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);