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 an IN 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 * @return The corresponding node or NULL if not found
131 static ir_node *get_node_for_in_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
134 for (i = 0; i < n; i++) {
135 /* in register matches */
136 if (pairs[i].in_reg->index == reg->index)
137 return pairs[i].in_node;
144 * Gets the node corresponding to an OUT register from an array of register pairs.
145 * NOTE: The given registers pairs and the register to look for must belong
146 * to the same register class.
148 * @param pairs The array of register pairs
149 * @param n The number of pairs
150 * @param reg The register to look for
151 * @return The corresponding node or NULL if not found
153 static ir_node *get_node_for_out_register(reg_pair_t *pairs, int n, const arch_register_t *reg) {
156 for (i = 0; i < n; i++) {
157 /* out register matches */
158 if (pairs[i].out_reg->index == reg->index)
159 return pairs[i].out_node;
166 * Gets the index in the register pair array where the in/out register
167 * corresponds to reg_idx.
169 * @param pairs The array of register pairs
170 * @param n The number of pairs
171 * @param reg The register index to look for
172 * @param in_out 0 == look for IN register, 1 == look for OUT register
173 * @return The corresponding index in pairs or -1 if not found
175 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
179 for (i = 0; i < n; i++) {
180 /* out register matches */
181 if ((int) pairs[i].out_reg->index == reg_idx)
186 for (i = 0; i < n; i++) {
187 /* in register matches */
188 if ((int) pairs[i].in_reg->index == reg_idx)
197 * Gets an array of register pairs and tries to identify a cycle or chain starting
200 * @param cycle Variable to hold the cycle
201 * @param pairs Array of register pairs
202 * @param start Index to start
203 * @return The cycle or chain
205 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
206 int head = pairs[start].in_reg->index;
207 int cur_idx = pairs[start].out_reg->index;
208 int cur_pair_idx = start;
209 int n_pairs_done = get_n_checked_pairs(pairs, n);
211 perm_type_t cycle_tp = PERM_CYCLE;
213 /* We could be right in the middle of a chain, so we need to find the start */
214 while (head != cur_idx) {
215 /* goto previous register in cycle or chain */
216 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
218 if (cur_pair_idx < 0) {
219 cycle_tp = PERM_CHAIN;
223 head = pairs[cur_pair_idx].in_reg->index;
224 start = cur_pair_idx;
228 /* assume worst case: all remaining pairs build a cycle or chain */
229 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
230 cycle->n_elems = 2; /* initial number of elements is 2 */
231 cycle->elems[0] = pairs[start].in_reg;
232 cycle->elems[1] = pairs[start].out_reg;
233 cycle->type = cycle_tp;
234 cur_idx = pairs[start].out_reg->index;
237 /* check for cycle or end of a chain */
238 while (cur_idx != head) {
239 /* goto next register in cycle or chain */
240 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
242 if (cur_pair_idx < 0)
245 cur_idx = pairs[cur_pair_idx].out_reg->index;
247 /* it's not the first element: insert it */
248 if (cur_idx != head) {
249 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
253 /* we are there where we started -> CYCLE */
254 cycle->type = PERM_CYCLE;
258 /* mark all pairs having one in/out register with cycle in common as checked */
259 for (idx = 0; idx < cycle->n_elems; idx++) {
260 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
262 if (cur_pair_idx >= 0)
263 pairs[cur_pair_idx].checked = 1;
265 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
267 if (cur_pair_idx >= 0)
268 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) {
285 ir_graph *irg = get_irn_irg(irn);
286 const arch_register_class_t *reg_class;
287 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;
296 DEBUG_ONLY(firm_dbg_module_t *mod;)
298 do_copy = env->do_copy;
299 DEBUG_ONLY(mod = env->dbg_module;)
300 block = get_nodes_block(irn);
303 Get the schedule predecessor node to the perm
304 NOTE: This works with auto-magic. If we insert the
305 new copy/exchange nodes after this node, everything
308 sched_point = sched_prev(irn);
309 DBG((mod, LEVEL_1, "perm: %+F\n", irn));
310 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
311 assert(sched_point && "Perm is not scheduled or has no predecessor");
313 n = get_irn_arity(irn);
314 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
316 reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
317 pairs = alloca(n * sizeof(pairs[0]));
319 /* build the list of register pairs (in, out) */
321 foreach_out_edge(irn, edge) {
322 pairs[i].out_node = get_edge_src_irn(edge);
323 pn = get_Proj_proj(pairs[i].out_node);
324 pairs[i].in_node = get_irn_n(irn, pn);
326 pairs[i].in_reg = arch_get_irn_register(pairs[i].in_node);
327 pairs[i].out_reg = arch_get_irn_register(pairs[i].out_node);
329 pairs[i].checked = 0;
333 /* sort the register pairs by the indices of the in registers */
334 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
336 /* Mark all equal pairs as checked, and exchange the OUT proj with
338 for (i = 0; i < n; i++) {
339 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
340 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
341 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
343 /* reroute the edges from the proj to the argument */
344 exchange(pairs[i].out_node, pairs[i].in_node);
346 pairs[i].checked = 1;
350 /* Set do_copy to 0 if it's on but we have no free register */
355 real_size = n - get_n_checked_pairs(pairs, n);
357 /* check for cycles and chains */
358 while (get_n_checked_pairs(pairs, n) < n) {
363 /* go to the first not-checked pair */
364 while (pairs[i].checked) i++;
365 cycle = XMALLOCZ(perm_cycle_t);
366 cycle = get_perm_cycle(cycle, pairs, n, i);
368 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
369 for (j = 0; j < cycle->n_elems; j++) {
370 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
372 DB((mod, LEVEL_1, "\n"));
375 We don't need to do anything if we have a Perm with two
376 elements which represents a cycle, because those nodes
377 already represent exchange nodes
379 if (n == 2 && cycle->type == PERM_CYCLE) {
385 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
386 // the copy cascade (this reduces the cycle into a chain)
388 /* build copy/swap nodes from back to front */
389 for (i = cycle->n_elems - 2; i >= 0; i--) {
390 arg1 = get_node_for_in_register(pairs, n, cycle->elems[i]);
391 arg2 = get_node_for_in_register(pairs, n, cycle->elems[i + 1]);
393 res1 = get_node_for_out_register(pairs, n, cycle->elems[i]);
394 res2 = get_node_for_out_register(pairs, n, cycle->elems[i + 1]);
396 If we have a cycle and don't copy: we need to create exchange nodes
397 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
398 IN_1 = in node with register i
399 IN_2 = in node with register i + 1
400 OUT_1 = out node with register i + 1
401 OUT_2 = out node with register i
403 if (cycle->type == PERM_CYCLE && !do_copy) {
407 /* At this point we have to handle the following problem: */
409 /* If we have a cycle with more than two elements, then */
410 /* this could correspond to the following Perm node: */
412 /* +----+ +----+ +----+ */
413 /* | r1 | | r2 | | r3 | */
414 /* +-+--+ +-+--+ +--+-+ */
417 /* +-+--------+---------+-+ */
419 /* +-+--------+---------+-+ */
422 /* +-+--+ +-+--+ +--+-+ */
423 /* |Proj| |Proj| |Proj| */
424 /* | r2 | | r3 | | r1 | */
425 /* +----+ +----+ +----+ */
427 /* This node is about to be split up into two 2x Perm's */
428 /* for which we need 4 Proj's and the one additional Proj */
429 /* of the first Perm has to be one IN of the second. So in */
430 /* general we need to create one additional Proj for each */
431 /* "middle" Perm and set this to one in node of the successor */
434 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
435 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
436 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
437 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
439 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
443 /* cycle is not done yet */
444 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
446 /* create intermediate proj */
447 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
449 /* set as in for next Perm */
450 pairs[pidx].in_node = res1;
453 set_Proj_pred(res2, cpyxchg);
454 set_Proj_proj(res2, 0);
455 set_Proj_pred(res1, cpyxchg);
456 set_Proj_proj(res1, 1);
458 arch_set_irn_register(res2, cycle->elems[i + 1]);
459 arch_set_irn_register(res1, cycle->elems[i]);
461 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
462 sched_add_after(sched_point, cpyxchg);
464 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
466 /* set the new scheduling point */
470 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
471 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
473 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
474 arch_set_irn_register(cpyxchg, cycle->elems[i + 1]);
477 /* exchange copy node and proj */
478 exchange(res2, cpyxchg);
480 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
481 sched_add_after(sched_point, cpyxchg);
483 /* set the new scheduling point */
484 sched_point = cpyxchg;
488 free((void *) cycle->elems);
492 /* remove the perm from schedule */
501 static int has_irn_users(const ir_node *irn) {
502 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
506 * Skip all Proj nodes.
508 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
510 irn = get_Proj_pred(irn);
514 static ir_node *find_copy(ir_node *irn, ir_node *op)
516 ir_node *block = get_nodes_block(irn);
519 for (cur_node = irn;;) {
520 cur_node = sched_prev(cur_node);
521 if (! be_is_Copy(cur_node))
523 if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
528 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
529 be_irg_t *birg = env->birg;
530 ir_graph *irg = be_get_birg_irg(birg);
531 ir_nodemap_t *op_set = &env->op_set;
532 ir_node *block = get_nodes_block(irn);
533 const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1);
534 ir_node *in[2], *keep, *cpy;
535 op_copy_assoc_t *entry;
536 DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
538 if (arch_irn_is(other_different, ignore) ||
539 !mode_is_datab(get_irn_mode(other_different))) {
540 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
544 /* Make a not spillable copy of the different node */
545 /* this is needed because the different irn could be */
546 /* in block far far away */
547 /* The copy is optimized later if not needed */
549 /* check if already exists such a copy in the schedule immediately before */
550 cpy = find_copy(belower_skip_proj(irn), other_different);
552 cpy = be_new_Copy(cls, irg, block, other_different);
553 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
554 DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
557 DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
563 /* Add the Keep resp. CopyKeep and reroute the users */
564 /* of the other_different irn in case of CopyKeep. */
565 if (has_irn_users(other_different)) {
566 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
567 be_node_set_reg_class(keep, 1, cls);
570 keep = be_new_Keep(cls, irg, block, 2, in);
573 DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
575 /* insert copy and keep into schedule */
576 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
577 if (! sched_is_scheduled(cpy))
578 sched_add_before(belower_skip_proj(irn), cpy);
579 sched_add_after(irn, keep);
581 /* insert the other different and it's copies into the map */
582 entry = ir_nodemap_get(op_set, other_different);
584 entry = obstack_alloc(&env->obst, sizeof(*entry));
586 ir_nodeset_init(&entry->copies);
588 ir_nodemap_insert(op_set, other_different, entry);
592 ir_nodeset_insert(&entry->copies, cpy);
594 /* insert keep in case of CopyKeep */
595 if (be_is_CopyKeep(keep)) {
596 ir_nodeset_insert(&entry->copies, keep);
601 * Checks if node has a must_be_different constraint in output and adds a Keep
602 * then to assure the constraint.
604 * @param irn the node to check
605 * @param skipped_irn if irn is a Proj node, its predecessor, else irn
606 * @param env the constraint environment
608 static void assure_different_constraints(ir_node *irn, ir_node *skipped_irn, constraint_env_t *env) {
609 const arch_register_req_t *req = arch_get_register_req(irn, -1);
611 if (arch_register_req_is(req, must_be_different)) {
612 const unsigned other = req->other_different;
615 if (arch_register_req_is(req, should_be_same)) {
616 const unsigned same = req->other_same;
618 if (is_po2(other) && is_po2(same)) {
619 int idx_other = ntz(other);
620 int idx_same = ntz(same);
623 * We can safely ignore a should_be_same x must_be_different y
624 * IFF both inputs are equal!
626 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
631 for (i = 0; 1U << i <= other; ++i) {
632 if (other & (1U << i)) {
633 ir_node *different_from = get_irn_n(skipped_irn, i);
634 gen_assure_different_pattern(irn, different_from, env);
641 * Calls the functions to assure register constraints.
643 * @param block The block to be checked
644 * @param walk_env The walker environment
646 static void assure_constraints_walker(ir_node *block, void *walk_env) {
649 sched_foreach_reverse(block, irn) {
650 ir_mode *mode = get_irn_mode(irn);
652 if (mode == mode_T) {
653 const ir_edge_t *edge;
655 foreach_out_edge(irn, edge) {
656 ir_node *proj = get_edge_src_irn(edge);
658 mode = get_irn_mode(proj);
659 if (mode_is_datab(mode))
660 assure_different_constraints(proj, irn, walk_env);
662 } else if (mode_is_datab(mode)) {
663 assure_different_constraints(irn, irn, walk_env);
669 * Melt all copykeeps pointing to the same node
670 * (or Projs of the same node), copying the same operand.
672 static void melt_copykeeps(constraint_env_t *cenv) {
673 be_irg_t *birg = cenv->birg;
674 ir_graph *irg = be_get_birg_irg(birg);
675 ir_nodemap_iterator_t map_iter;
676 ir_nodemap_entry_t map_entry;
679 foreach_ir_nodemap(&cenv->op_set, map_entry, map_iter) {
680 op_copy_assoc_t *entry = map_entry.data;
684 ir_nodeset_iterator_t iter;
685 ir_node **ck_arr, **melt_arr;
689 /* collect all copykeeps */
691 foreach_ir_nodeset(&entry->copies, cp, iter) {
692 if (be_is_CopyKeep(cp)) {
693 obstack_grow(&obst, &cp, sizeof(cp));
696 #ifdef KEEP_ALIVE_COPYKEEP_HACK
698 set_irn_mode(cp, mode_ANY);
701 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
704 /* compare each copykeep with all other copykeeps */
705 ck_arr = (ir_node **)obstack_finish(&obst);
706 for (idx = 0; idx < num_ck; ++idx) {
707 ir_node *ref, *ref_mode_T;
713 ir_node *sched_pt = NULL;
717 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
718 obstack_grow(&obst, &ref, sizeof(ref));
720 DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
722 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
723 for (j = 0; j < num_ck; ++j) {
724 ir_node *cur_ck = ck_arr[j];
726 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
727 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
728 ir_nodeset_remove(&entry->copies, cur_ck);
729 DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
732 sched_remove(cur_ck);
737 /* check, if we found some candidates for melting */
739 DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
743 ir_nodeset_remove(&entry->copies, ref);
746 melt_arr = (ir_node **)obstack_finish(&obst);
747 /* melt all found copykeeps */
748 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
749 for (j = 0; j < n_melt; ++j) {
750 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
752 /* now, we can kill the melted keep, except the */
753 /* ref one, we still need some information */
754 if (melt_arr[j] != ref)
755 kill_node(melt_arr[j]);
758 #ifdef KEEP_ALIVE_COPYKEEP_HACK
759 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
762 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));
763 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
765 /* set register class for all kept inputs */
766 for (j = 1; j <= n_melt; ++j)
767 be_node_set_reg_class(new_ck, j, entry->cls);
769 ir_nodeset_insert(&entry->copies, new_ck);
771 /* find scheduling point */
772 sched_pt = ref_mode_T;
774 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
775 sched_pt = sched_next(sched_pt);
776 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
778 sched_add_before(sched_pt, new_ck);
779 DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
781 /* finally: kill the reference copykeep */
786 obstack_free(&obst, NULL);
791 * Walks over all nodes to assure register constraints.
793 * @param birg The birg structure containing the irg
795 void assure_constraints(be_irg_t *birg) {
796 ir_graph *irg = be_get_birg_irg(birg);
797 constraint_env_t cenv;
799 ir_nodemap_iterator_t map_iter;
800 ir_nodemap_entry_t map_entry;
801 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
803 DEBUG_ONLY(cenv.dbg = mod;)
805 ir_nodemap_init(&cenv.op_set);
806 obstack_init(&cenv.obst);
808 irg_block_walk_graph(irg, NULL, assure_constraints_walker, &cenv);
810 /* melt copykeeps, pointing to projs of */
811 /* the same mode_T node and keeping the */
813 melt_copykeeps(&cenv);
816 foreach_ir_nodemap(&cenv.op_set, map_entry, map_iter) {
817 op_copy_assoc_t *entry = map_entry.data;
820 ir_nodeset_iterator_t iter;
821 be_ssa_construction_env_t senv;
823 n = ir_nodeset_size(&entry->copies);
824 nodes = alloca(n * sizeof(nodes[0]));
826 /* put the node in an array */
827 DBG((mod, LEVEL_1, "introduce copies for %+F ", map_entry.node));
829 /* collect all copies */
831 foreach_ir_nodeset(&entry->copies, cp, iter) {
833 DB((mod, LEVEL_1, ", %+F ", cp));
836 DB((mod, LEVEL_1, "\n"));
838 /* introduce the copies for the operand and it's copies */
839 be_ssa_construction_init(&senv, birg);
840 be_ssa_construction_add_copy(&senv, map_entry.node);
841 be_ssa_construction_add_copies(&senv, nodes, n);
842 be_ssa_construction_fix_users(&senv, map_entry.node);
843 be_ssa_construction_destroy(&senv);
845 /* Could be that not all CopyKeeps are really needed, */
846 /* so we transform unnecessary ones into Keeps. */
847 foreach_ir_nodeset(&entry->copies, cp, iter) {
848 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
850 int n = get_irn_arity(cp);
852 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
853 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
854 sched_add_before(cp, keep);
856 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
862 ir_nodeset_destroy(&entry->copies);
865 ir_nodemap_destroy(&cenv.op_set);
866 obstack_free(&cenv.obst, NULL);
867 be_liveness_invalidate(be_get_birg_liveness(birg));
872 * Push nodes that do not need to be permed through the Perm.
873 * This is commonly a reload cascade at block ends.
874 * @note This routine needs interference.
875 * @note Probably, we can implement it a little more efficient.
876 * Especially searching the frontier lazily might be better.
877 * @param perm The perm.
878 * @param data The walker data (lower_env_t).
879 * @return 1, if there is something left to perm over.
880 * 0, if removed the complete perm.
882 static int push_through_perm(ir_node *perm, void *data)
884 lower_env_t *env = data;
886 ir_graph *irg = get_irn_irg(perm);
887 ir_node *bl = get_nodes_block(perm);
889 int arity = get_irn_arity(perm);
892 bitset_t *moved = bitset_alloca(arity);
895 ir_node *frontier = bl;
896 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
899 const ir_edge_t *edge;
900 ir_node *one_proj = NULL, *irn;
901 const arch_register_class_t *cls = NULL;
903 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
905 /* get some Proj and find out the register class of that Proj. */
906 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
907 one_proj = get_edge_src_irn(edge);
908 assert(is_Proj(one_proj));
909 cls = arch_get_irn_reg_class(one_proj, -1);
911 /* Find the point in the schedule after which the
912 * potentially movable nodes must be defined.
913 * A Perm will only be pushed up to first instruction
914 * which lets an operand of itself die.
915 * If we would allow to move the Perm above this instruction,
916 * the former dead operand would be live now at the point of
917 * the Perm, increasing the register pressure by one.
919 sched_foreach_reverse_from (sched_prev(perm), irn) {
920 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
921 ir_node *op = get_irn_n(irn, i);
922 if (arch_irn_consider_in_reg_alloc(cls, op) &&
923 !values_interfere(env->birg, op, one_proj)) {
931 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
933 node = sched_prev(perm);
935 while(!sched_is_begin(node)) {
936 const arch_register_req_t *req;
940 /* search if node is a INPUT of Perm */
941 foreach_out_edge(perm, edge) {
942 ir_node *out = get_edge_src_irn(edge);
943 int pn = get_Proj_proj(out);
944 ir_node *in = get_irn_n(perm, pn);
951 /* it wasn't an input to the perm, we can't do anything more */
954 if(!sched_comes_after(frontier, node))
956 if (arch_irn_is(node, modify_flags))
959 req = arch_get_register_req(get_Proj_pred(node),
960 -1 - get_Proj_proj(node));
962 req = arch_get_register_req(node, -1);
964 if(req->type != arch_register_req_type_normal)
966 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
967 ir_node *opop = get_irn_n(node, i);
968 if (arch_irn_consider_in_reg_alloc(cls, opop)) {
975 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
977 /* move the movable node in front of the Perm */
979 sched_add_after(perm, node);
981 /* give it the proj's register */
982 arch_set_irn_register(node, arch_get_irn_register(proj));
984 /* reroute all users of the proj to the moved node. */
985 edges_reroute(proj, node, irg);
988 set_Proj_pred(proj, new_Bad());
991 bitset_set(moved, input);
994 node = sched_prev(node);
997 /* well, we could not push anything through the perm */
1001 new_size = arity - n_moved;
1006 map = alloca(new_size * sizeof(map[0]));
1007 proj_map = alloca(arity * sizeof(proj_map[0]));
1008 memset(proj_map, -1, sizeof(proj_map[0]));
1010 for(i = 0; i < arity; ++i) {
1011 if(bitset_is_set(moved, i))
1017 assert(n == new_size);
1018 foreach_out_edge(perm, edge) {
1019 ir_node *proj = get_edge_src_irn(edge);
1020 int pn = get_Proj_proj(proj);
1023 set_Proj_proj(proj, pn);
1026 be_Perm_reduce(perm, new_size, map);
1031 * Calls the corresponding lowering function for the node.
1033 * @param irn The node to be checked for lowering
1034 * @param walk_env The walker environment
1036 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
1039 if (is_Block(irn) || is_Proj(irn))
1041 if (!be_is_Perm(irn))
1044 perm_stayed = push_through_perm(irn, walk_env);
1048 lower_perm_node(irn, walk_env);
1052 * Walks over all blocks in an irg and performs lowering need to be
1053 * done after register allocation (e.g. perm lowering).
1055 * @param birg The birg object
1056 * @param do_copy 1 == resolve cycles with a free reg if available
1058 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1060 ir_graph *irg = be_get_birg_irg(birg);
1063 env.do_copy = do_copy;
1064 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
1066 /* we will need interference */
1067 be_liveness_assure_chk(be_get_birg_liveness(birg));
1069 irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);