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 /** returns the number register pairs marked as checked. */
100 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
101 int i, n_checked = 0;
103 for (i = 0; i < n; i++) {
104 if (pairs[i].checked)
112 * Gets the node corresponding to an IN register from an array of register pairs.
113 * NOTE: The given registers pairs and the register to look for must belong
114 * to the same register class.
116 * @param pairs The array of register pairs
117 * @param n The number of pairs
118 * @param reg The register to look for
119 * @return The corresponding node or NULL if not found
121 static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
124 for (i = 0; i < n; i++) {
125 /* in register matches */
126 if (pairs[i].in_reg->index == reg->index)
127 return pairs[i].in_node;
134 * Gets the node corresponding to an OUT register from an array of register pairs.
135 * NOTE: The given registers pairs and the register to look for must belong
136 * to the same register class.
138 * @param pairs The array of register pairs
139 * @param n The number of pairs
140 * @param reg The register to look for
141 * @return The corresponding node or NULL if not found
143 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
146 for (i = 0; i < n; i++) {
147 /* out register matches */
148 if (pairs[i].out_reg->index == reg->index)
149 return pairs[i].out_node;
156 * Gets the index in the register pair array where the in register
157 * corresponds to reg_idx.
159 * @param pairs The array of register pairs
160 * @param n The number of pairs
161 * @param reg The register index to look for
163 * @return The corresponding index in pairs or -1 if not found
165 static int get_pairidx_for_in_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
168 for (i = 0; i < n; i++) {
169 /* in register matches */
170 if (pairs[i].in_reg->index == reg_idx)
177 * Gets the index in the register pair array where the out register
178 * corresponds to reg_idx.
180 * @param pairs The array of register pairs
181 * @param n The number of pairs
182 * @param reg The register index to look for
184 * @return The corresponding index in pairs or -1 if not found
186 static int get_pairidx_for_out_regidx(reg_pair_t *pairs, int n, unsigned reg_idx) {
189 for (i = 0; i < n; i++) {
190 /* out register matches */
191 if (pairs[i].out_reg->index == reg_idx)
198 * Gets an array of register pairs and tries to identify a cycle or chain
199 * starting at position start.
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 void get_perm_cycle(perm_cycle_t *const cycle,
207 reg_pair_t *const pairs,
211 int head = pairs[start].in_reg->index;
212 int cur_idx = pairs[start].out_reg->index;
213 int const n_pairs_done = get_n_checked_pairs(pairs, n);
214 perm_type_t cycle_tp = PERM_CYCLE;
217 /* We could be right in the middle of a chain, so we need to find the start */
218 while (head != cur_idx) {
219 /* goto previous register in cycle or chain */
220 int const cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, head);
222 if (cur_pair_idx < 0) {
223 cycle_tp = PERM_CHAIN;
226 head = pairs[cur_pair_idx].in_reg->index;
227 start = cur_pair_idx;
231 /* assume worst case: all remaining pairs build a cycle or chain */
232 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
233 cycle->n_elems = 2; /* initial number of elements is 2 */
234 cycle->elems[0] = pairs[start].in_reg;
235 cycle->elems[1] = pairs[start].out_reg;
236 cycle->type = cycle_tp;
237 cur_idx = pairs[start].out_reg->index;
240 /* check for cycle or end of a chain */
241 while (cur_idx != head) {
242 /* goto next register in cycle or chain */
243 int const cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cur_idx);
245 if (cur_pair_idx < 0)
248 cur_idx = pairs[cur_pair_idx].out_reg->index;
250 /* it's not the first element: insert it */
251 if (cur_idx != head) {
252 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
255 /* we are there where we started -> CYCLE */
256 cycle->type = PERM_CYCLE;
260 /* mark all pairs having one in/out register with cycle in common as checked */
261 for (idx = 0; idx < cycle->n_elems; idx++) {
264 cur_pair_idx = get_pairidx_for_in_regidx(pairs, n, cycle->elems[idx]->index);
265 if (cur_pair_idx >= 0)
266 pairs[cur_pair_idx].checked = 1;
268 cur_pair_idx = get_pairidx_for_out_regidx(pairs, n, cycle->elems[idx]->index);
269 if (cur_pair_idx >= 0)
270 pairs[cur_pair_idx].checked = 1;
275 * Lowers a perm node. Resolves cycles and creates a bunch of
276 * copy and swap operations to permute registers.
277 * Note: The caller of this function has to make sure, that irn
280 * @param irn The perm node
281 * @param block The block the perm node belongs to
282 * @param walk_env The environment
284 static void lower_perm_node(ir_node *irn, void *walk_env)
286 lower_env_t *const env = walk_env;
287 const arch_register_class_t *const reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
288 ir_graph *const irg = get_irn_irg(irn);
289 ir_node *const block = get_nodes_block(irn);
290 int const n = get_irn_arity(irn);
291 reg_pair_t *const pairs = alloca(n * sizeof(pairs[0]));
293 int do_copy = env->do_copy;
294 /* Get the schedule predecessor node to the perm.
295 * NOTE: This works with auto-magic. If we insert the new copy/exchange
296 * nodes after this node, everything should be ok. */
297 ir_node * sched_point = sched_prev(irn);
298 const ir_edge_t * edge;
301 DBG((dbg, LEVEL_1, "perm: %+F, sched point is %+F\n", irn, sched_point));
302 assert(sched_point && "Perm is not scheduled or has no predecessor");
304 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
306 /* build the list of register pairs (in, out) */
308 foreach_out_edge(irn, edge) {
309 reg_pair_t *const pair = &pairs[i++];
310 ir_node *const out = get_edge_src_irn(edge);
311 long const pn = get_Proj_proj(out);
312 ir_node *const in = get_irn_n(irn, pn);
315 pair->in_reg = arch_get_irn_register(in);
316 pair->out_node = out;
317 pair->out_reg = arch_get_irn_register(out);
321 /* Mark all equal pairs as checked, and exchange the OUT proj with the IN
323 for (i = 0; i < n; i++) {
324 reg_pair_t *const pair = &pairs[i];
326 if (pair->in_reg->index != pair->out_reg->index)
329 DBG((dbg, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
330 irn, pair->in_node, pair->out_node, pair->out_reg->name));
332 /* reroute the edges from the proj to the argument */
333 exchange(pair->out_node, pair->in_node);
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_checked_pairs(pairs, n) < n) {
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 (n == 2 && cycle.type == PERM_CYCLE) {
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 */
383 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, irg, 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(irg, block, 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 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
443 sched_add_after(sched_point, cpyxchg);
445 DBG((dbg, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
447 /* set the new scheduling point */
452 DBG((dbg, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
453 irn, arg1, cycle.elems[i]->name, res2, cycle.elems[i + 1]->name));
455 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
456 arch_set_irn_register(cpyxchg, cycle.elems[i + 1]);
458 /* exchange copy node and proj */
459 exchange(res2, cpyxchg);
461 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
462 sched_add_after(sched_point, cpyxchg);
464 /* set the new scheduling point */
465 sched_point = cpyxchg;
470 free((void*)cycle.elems);
473 /* remove the perm from schedule */
482 static int has_irn_users(const ir_node *irn) {
483 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
486 static ir_node *find_copy(ir_node *irn, ir_node *op)
490 for (cur_node = irn;;) {
491 cur_node = sched_prev(cur_node);
492 if (! be_is_Copy(cur_node))
494 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
499 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
501 ir_nodemap_t *op_set;
503 const arch_register_class_t *cls;
505 op_copy_assoc_t *entry;
507 if (arch_irn_is(other_different, ignore) ||
508 !mode_is_datab(get_irn_mode(other_different))) {
509 DBG((dbg_constr, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
513 irg = be_get_birg_irg(env->birg);
514 op_set = &env->op_set;
515 block = get_nodes_block(irn);
516 cls = arch_get_irn_reg_class(other_different, -1);
518 /* Make a not spillable copy of the different node */
519 /* this is needed because the different irn could be */
520 /* in block far far away */
521 /* The copy is optimized later if not needed */
523 /* check if already exists such a copy in the schedule immediately before */
524 cpy = find_copy(skip_Proj(irn), other_different);
526 cpy = be_new_Copy(cls, irg, block, other_different);
527 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
528 DBG((dbg_constr, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
530 DBG((dbg_constr, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
533 /* Add the Keep resp. CopyKeep and reroute the users */
534 /* of the other_different irn in case of CopyKeep. */
535 if (has_irn_users(other_different)) {
536 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
537 be_node_set_reg_class(keep, 1, cls);
543 keep = be_new_Keep(cls, irg, block, 2, in);
546 DBG((dbg_constr, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
548 /* insert copy and keep into schedule */
549 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
550 if (! sched_is_scheduled(cpy))
551 sched_add_before(skip_Proj(irn), cpy);
552 sched_add_after(irn, keep);
554 /* insert the other different and it's copies into the map */
555 entry = ir_nodemap_get(op_set, other_different);
557 entry = obstack_alloc(&env->obst, sizeof(*entry));
559 ir_nodeset_init(&entry->copies);
561 ir_nodemap_insert(op_set, other_different, entry);
565 ir_nodeset_insert(&entry->copies, cpy);
567 /* insert keep in case of CopyKeep */
568 if (be_is_CopyKeep(keep))
569 ir_nodeset_insert(&entry->copies, keep);
573 * Checks if node has a must_be_different constraint in output and adds a Keep
574 * then to assure the constraint.
576 * @param irn the node to check
577 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
578 * @param env the constraint environment
580 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
581 const arch_register_req_t *req = arch_get_register_req(irn, -1);
583 if (arch_register_req_is(req, must_be_different)) {
584 const unsigned other = req->other_different;
587 if (arch_register_req_is(req, should_be_same)) {
588 const unsigned same = req->other_same;
590 if (is_po2(other) && is_po2(same)) {
591 int idx_other = ntz(other);
592 int idx_same = ntz(same);
595 * We can safely ignore a should_be_same x must_be_different y
596 * IFF both inputs are equal!
598 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
603 for (i = 0; 1U << i <= other; ++i) {
604 if (other & (1U << i)) {
605 ir_node *different_from = get_irn_n(skipped_irn, i);
606 gen_assure_different_pattern(irn, different_from, env);
613 * Calls the functions to assure register constraints.
615 * @param block The block to be checked
616 * @param walk_env The walker environment
618 static void assure_constraints_walker(ir_node *block, void *walk_env) {
621 sched_foreach_reverse(block, irn) {
622 ir_mode *mode = get_irn_mode(irn);
624 if (mode == mode_T) {
625 const ir_edge_t *edge;
627 foreach_out_edge(irn, edge) {
628 ir_node *proj = get_edge_src_irn(edge);
630 mode = get_irn_mode(proj);
631 if (mode_is_datab(mode))
632 assure_different_constraints(proj, irn, walk_env);
634 } else if (mode_is_datab(mode)) {
635 assure_different_constraints(irn, irn, walk_env);
641 * Melt all copykeeps pointing to the same node
642 * (or Projs of the same node), copying the same operand.
644 static void melt_copykeeps(constraint_env_t *cenv) {
645 be_irg_t *birg = cenv->birg;
646 ir_graph *irg = be_get_birg_irg(birg);
647 ir_nodemap_iterator_t map_iter;
648 ir_nodemap_entry_t map_entry;
651 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
652 op_copy_assoc_t *entry = map_entry.data;
656 ir_nodeset_iterator_t iter;
657 ir_node **ck_arr, **melt_arr;
661 /* collect all copykeeps */
663 foreach_ir_nodeset(&entry->copies, cp, iter) {
664 if (be_is_CopyKeep(cp)) {
665 obstack_grow(&obst, &cp, sizeof(cp));
668 #ifdef KEEP_ALIVE_COPYKEEP_HACK
670 set_irn_mode(cp, mode_ANY);
673 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
676 /* compare each copykeep with all other copykeeps */
677 ck_arr = (ir_node **)obstack_finish(&obst);
678 for (idx = 0; idx < num_ck; ++idx) {
679 ir_node *ref, *ref_mode_T;
685 ir_node *sched_pt = NULL;
689 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
690 obstack_grow(&obst, &ref, sizeof(ref));
692 DBG((dbg_constr, LEVEL_1, "Trying to melt %+F:\n", ref));
694 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
695 for (j = 0; j < num_ck; ++j) {
696 ir_node *cur_ck = ck_arr[j];
698 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
699 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
700 ir_nodeset_remove(&entry->copies, cur_ck);
701 DBG((dbg_constr, LEVEL_1, "\t%+F\n", cur_ck));
704 sched_remove(cur_ck);
709 /* check, if we found some candidates for melting */
711 DBG((dbg_constr, LEVEL_1, "\tno candidate found\n"));
715 ir_nodeset_remove(&entry->copies, ref);
718 melt_arr = (ir_node **)obstack_finish(&obst);
719 /* melt all found copykeeps */
720 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
721 for (j = 0; j < n_melt; ++j) {
722 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
724 /* now, we can kill the melted keep, except the */
725 /* ref one, we still need some information */
726 if (melt_arr[j] != ref)
727 kill_node(melt_arr[j]);
730 #ifdef KEEP_ALIVE_COPYKEEP_HACK
731 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
734 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));
735 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
737 /* set register class for all kept inputs */
738 for (j = 1; j <= n_melt; ++j)
739 be_node_set_reg_class(new_ck, j, entry->cls);
741 ir_nodeset_insert(&entry->copies, new_ck);
743 /* find scheduling point */
744 sched_pt = ref_mode_T;
746 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
747 sched_pt = sched_next(sched_pt);
748 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
750 sched_add_before(sched_pt, new_ck);
751 DBG((dbg_constr, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
753 /* finally: kill the reference copykeep */
758 obstack_free(&obst, NULL);
763 * Walks over all nodes to assure register constraints.
765 * @param birg The birg structure containing the irg
767 void assure_constraints(be_irg_t *birg) {
768 ir_graph *irg = be_get_birg_irg(birg);
769 constraint_env_t cenv;
771 ir_nodemap_iterator_t map_iter;
772 ir_nodemap_entry_t map_entry;
774 FIRM_DBG_REGISTER(dbg_constr, "firm.be.lower.constr");
777 ir_nodemap_init(&cenv.op_set);
778 obstack_init(&cenv.obst);
780 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
782 /* melt copykeeps, pointing to projs of */
783 /* the same mode_T node and keeping the */
785 melt_copykeeps(&cenv);
788 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
789 op_copy_assoc_t *entry = map_entry.data;
792 ir_nodeset_iterator_t iter;
793 be_ssa_construction_env_t senv;
795 n = ir_nodeset_size(&entry->copies);
796 nodes = alloca(n * sizeof(nodes[0]));
798 /* put the node in an array */
799 DBG((dbg_constr, LEVEL_1, "introduce copies for %+F ", map_entry.node));
801 /* collect all copies */
803 foreach_ir_nodeset(&entry->copies, cp, iter) {
805 DB((dbg_constr, LEVEL_1, ", %+F ", cp));
808 DB((dbg_constr, LEVEL_1, "\n"));
810 /* introduce the copies for the operand and it's copies */
811 be_ssa_construction_init(&senv, birg);
812 be_ssa_construction_add_copy(&senv, map_entry.node);
813 be_ssa_construction_add_copies(&senv, nodes, n);
814 be_ssa_construction_fix_users(&senv, map_entry.node);
815 be_ssa_construction_destroy(&senv);
817 /* Could be that not all CopyKeeps are really needed, */
818 /* so we transform unnecessary ones into Keeps. */
819 foreach_ir_nodeset(&entry->copies, cp, iter) {
820 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
822 int n = get_irn_arity(cp);
824 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
825 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
826 sched_add_before(cp, keep);
828 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
834 ir_nodeset_destroy(&entry->copies);
837 ir_nodemap_destroy(&cenv.op_set);
838 obstack_free(&cenv.obst, NULL);
839 be_liveness_invalidate(be_get_birg_liveness(birg));
844 * Push nodes that do not need to be permed through the Perm.
845 * This is commonly a reload cascade at block ends.
846 * @note This routine needs interference.
847 * @note Probably, we can implement it a little more efficient.
848 * Especially searching the frontier lazily might be better.
849 * @param perm The perm.
850 * @param data The walker data (lower_env_t).
851 * @return 1, if there is something left to perm over.
852 * 0, if removed the complete perm.
854 static int push_through_perm(ir_node *perm, void *data)
856 lower_env_t *env = data;
858 ir_graph *irg = get_irn_irg(perm);
859 ir_node *bl = get_nodes_block(perm);
861 int arity = get_irn_arity(perm);
864 bitset_t *moved = bitset_alloca(arity);
867 ir_node *frontier = bl;
868 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
871 const ir_edge_t *edge;
872 ir_node *one_proj = NULL, *irn;
873 const arch_register_class_t *cls = NULL;
875 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
877 /* get some Proj and find out the register class of that Proj. */
878 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
879 one_proj = get_edge_src_irn(edge);
880 assert(is_Proj(one_proj));
881 cls = arch_get_irn_reg_class(one_proj, -1);
883 /* Find the point in the schedule after which the
884 * potentially movable nodes must be defined.
885 * A Perm will only be pushed up to first instruction
886 * which lets an operand of itself die.
887 * If we would allow to move the Perm above this instruction,
888 * the former dead operand would be live now at the point of
889 * the Perm, increasing the register pressure by one.
891 sched_foreach_reverse_from (sched_prev(perm), irn) {
892 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
893 ir_node *op = get_irn_n(irn, i);
894 if (arch_irn_consider_in_reg_alloc(cls, op) &&
895 !values_interfere(env->birg, op, one_proj)) {
903 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
905 node = sched_prev(perm);
907 while(!sched_is_begin(node)) {
908 const arch_register_req_t *req;
912 /* search if node is a INPUT of Perm */
913 foreach_out_edge(perm, edge) {
914 ir_node *out = get_edge_src_irn(edge);
915 int pn = get_Proj_proj(out);
916 ir_node *in = get_irn_n(perm, pn);
923 /* it wasn't an input to the perm, we can't do anything more */
926 if(!sched_comes_after(frontier, node))
928 if (arch_irn_is(node, modify_flags))
931 req = arch_get_register_req(get_Proj_pred(node),
932 -1 - get_Proj_proj(node));
934 req = arch_get_register_req(node, -1);
936 if(req->type != arch_register_req_type_normal)
938 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
939 ir_node *opop = get_irn_n(node, i);
940 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
947 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
949 /* move the movable node in front of the Perm */
951 sched_add_after(perm, node);
953 /* give it the proj's register */
954 arch_set_irn_register(node, arch_get_irn_register(proj));
956 /* reroute all users of the proj to the moved node. */
957 edges_reroute(proj, node, irg);
960 set_Proj_pred(proj, new_Bad());
963 bitset_set(moved, input);
966 node = sched_prev(node);
969 /* well, we could not push anything through the perm */
973 new_size = arity - n_moved;
978 map = alloca(new_size * sizeof(map[0]));
979 proj_map = alloca(arity * sizeof(proj_map[0]));
980 memset(proj_map, -1, sizeof(proj_map[0]));
982 for(i = 0; i < arity; ++i) {
983 if(bitset_is_set(moved, i))
989 assert(n == new_size);
990 foreach_out_edge(perm, edge) {
991 ir_node *proj = get_edge_src_irn(edge);
992 int pn = get_Proj_proj(proj);
995 set_Proj_proj(proj, pn);
998 be_Perm_reduce(perm, new_size, map);
1003 * Calls the corresponding lowering function for the node.
1005 * @param irn The node to be checked for lowering
1006 * @param walk_env The walker environment
1008 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env)
1012 if (!be_is_Perm(irn))
1015 perm_stayed = push_through_perm(irn, walk_env);
1019 lower_perm_node(irn, walk_env);
1023 * Walks over all blocks in an irg and performs lowering need to be
1024 * done after register allocation (e.g. perm lowering).
1026 * @param birg The birg object
1027 * @param do_copy 1 == resolve cycles with a free reg if available
1029 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1033 FIRM_DBG_REGISTER(dbg, "firm.be.lower");
1036 env.do_copy = do_copy;
1038 /* we will need interference */
1039 be_liveness_assure_chk(be_get_birg_liveness(birg));
1041 irg = be_get_birg_irg(birg);
1042 irg_walk_graph(irg, NULL, lower_nodes_after_ra_walker, &env);