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
37 #include "irnodeset.h"
39 #include "iredges_t.h"
46 #include "besched_t.h"
48 #include "bessaconstr.h"
49 #include "beintlive_t.h"
51 #undef KEEP_ALIVE_COPYKEEP_HACK
53 /** Associates an op with it's copy and CopyKeep. */
55 ir_node *op; /**< an irn which must be different */
56 ir_nodeset_t copies; /**< all non-spillable copies of this irn */
57 const arch_register_class_t *cls;
60 /** Environment for constraints. */
65 DEBUG_ONLY(firm_dbg_module_t *dbg;)
68 /** Lowering walker environment. */
69 typedef struct _lower_env_t {
71 const arch_env_t *arch_env;
73 DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
76 /** Holds a Perm register pair. */
77 typedef struct _reg_pair_t {
78 const arch_register_t *in_reg; /**< a perm IN register */
79 ir_node *in_node; /**< the in node to which the register belongs */
81 const arch_register_t *out_reg; /**< a perm OUT register */
82 ir_node *out_node; /**< the out node to which the register belongs */
84 int checked; /**< indicates whether the pair was check for cycle or not */
87 typedef enum _perm_type_t {
94 /** Structure to represent cycles or chains in a Perm. */
95 typedef struct _perm_cycle_t {
96 const arch_register_t **elems; /**< the registers in the cycle */
97 int n_elems; /**< number of elements in the cycle */
98 perm_type_t type; /**< type (CHAIN or CYCLE) */
101 /** Compare the two operands. */
102 static int cmp_op_copy_assoc(const void *a, const void *b) {
103 const op_copy_assoc_t *op1 = a;
104 const op_copy_assoc_t *op2 = b;
106 return op1->op != op2->op;
109 /** Compare the in registers of two register pairs. */
110 static int compare_reg_pair(const void *a, const void *b) {
111 const reg_pair_t *pair_a = a;
112 const reg_pair_t *pair_b = b;
114 if (pair_a->in_reg->index > pair_b->in_reg->index)
120 /** returns the number register pairs marked as checked. */
121 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
122 int i, n_checked = 0;
124 for (i = 0; i < n; i++) {
125 if (pairs[i].checked)
133 * Gets the node corresponding to a register from an array of register pairs.
134 * NOTE: The given registers pairs and the register to look for must belong
135 * to the same register class.
137 * @param pairs The array of register pairs
138 * @param n The number of pairs
139 * @param reg The register to look for
140 * @param in_out 0 == look for IN register, 1 == look for OUT register
141 * @return The corresponding node or NULL if not found
143 static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
147 for (i = 0; i < n; i++) {
148 /* out register matches */
149 if (pairs[i].out_reg->index == reg->index)
150 return pairs[i].out_node;
154 for (i = 0; i < n; i++) {
155 /* in register matches */
156 if (pairs[i].in_reg->index == reg->index)
157 return pairs[i].in_node;
165 * Gets the index in the register pair array where the in/out register
166 * corresponds to reg_idx.
168 * @param pairs The array of register pairs
169 * @param n The number of pairs
170 * @param reg The register index to look for
171 * @param in_out 0 == look for IN register, 1 == look for OUT register
172 * @return The corresponding index in pairs or -1 if not found
174 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
178 for (i = 0; i < n; i++) {
179 /* out register matches */
180 if ((int) pairs[i].out_reg->index == reg_idx)
185 for (i = 0; i < n; i++) {
186 /* in register matches */
187 if ((int) pairs[i].in_reg->index == reg_idx)
196 * Gets an array of register pairs and tries to identify a cycle or chain starting
199 * @param cycle Variable to hold the cycle
200 * @param pairs Array of register pairs
201 * @param start Index to start
202 * @return The cycle or chain
204 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
205 int head = pairs[start].in_reg->index;
206 int cur_idx = pairs[start].out_reg->index;
207 int cur_pair_idx = start;
208 int n_pairs_done = get_n_checked_pairs(pairs, n);
210 perm_type_t cycle_tp = PERM_CYCLE;
212 /* We could be right in the middle of a chain, so we need to find the start */
213 while (head != cur_idx) {
214 /* goto previous register in cycle or chain */
215 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
217 if (cur_pair_idx < 0) {
218 cycle_tp = PERM_CHAIN;
222 head = pairs[cur_pair_idx].in_reg->index;
223 start = cur_pair_idx;
227 /* assume worst case: all remaining pairs build a cycle or chain */
228 cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
229 cycle->n_elems = 2; /* initial number of elements is 2 */
230 cycle->elems[0] = pairs[start].in_reg;
231 cycle->elems[1] = pairs[start].out_reg;
232 cycle->type = cycle_tp;
233 cur_idx = pairs[start].out_reg->index;
236 /* check for cycle or end of a chain */
237 while (cur_idx != head) {
238 /* goto next register in cycle or chain */
239 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
241 if (cur_pair_idx < 0)
244 cur_idx = pairs[cur_pair_idx].out_reg->index;
246 /* it's not the first element: insert it */
247 if (cur_idx != head) {
248 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
252 /* we are there where we started -> CYCLE */
253 cycle->type = PERM_CYCLE;
257 /* mark all pairs having one in/out register with cycle in common as checked */
258 for (idx = 0; idx < cycle->n_elems; idx++) {
259 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
261 if (cur_pair_idx >= 0)
262 pairs[cur_pair_idx].checked = 1;
264 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
266 if (cur_pair_idx >= 0)
267 pairs[cur_pair_idx].checked = 1;
274 * Lowers a perm node. Resolves cycles and creates a bunch of
275 * copy and swap operations to permute registers.
276 * Note: The caller of this function has to make sure, that irn
279 * @param irn The perm node
280 * @param block The block the perm node belongs to
281 * @param walk_env The environment
283 static void lower_perm_node(ir_node *irn, void *walk_env) {
284 ir_graph *irg = get_irn_irg(irn);
285 const arch_register_class_t *reg_class;
286 lower_env_t *env = walk_env;
289 int n, i, pn, do_copy, j, n_ops;
291 const ir_edge_t *edge;
292 ir_node *sched_point, *block, *in[2];
293 ir_node *arg1, *arg2, *res1, *res2;
294 ir_node *cpyxchg = NULL;
295 DEBUG_ONLY(firm_dbg_module_t *mod;)
297 do_copy = env->do_copy;
298 DEBUG_ONLY(mod = env->dbg_module;)
299 block = get_nodes_block(irn);
302 Get the schedule predecessor node to the perm
303 NOTE: This works with auto-magic. If we insert the
304 new copy/exchange nodes after this node, everything
307 sched_point = sched_prev(irn);
308 DBG((mod, LEVEL_1, "perm: %+F\n", irn));
309 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
310 assert(sched_point && "Perm is not scheduled or has no predecessor");
312 n = get_irn_arity(irn);
313 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
315 reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
316 pairs = alloca(n * sizeof(pairs[0]));
318 /* build the list of register pairs (in, out) */
320 foreach_out_edge(irn, edge) {
321 pairs[i].out_node = get_edge_src_irn(edge);
322 pn = get_Proj_proj(pairs[i].out_node);
323 pairs[i].in_node = get_irn_n(irn, pn);
325 pairs[i].in_reg = arch_get_irn_register(pairs[i].in_node);
326 pairs[i].out_reg = arch_get_irn_register(pairs[i].out_node);
328 pairs[i].checked = 0;
332 /* sort the register pairs by the indices of the in registers */
333 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
335 /* Mark all equal pairs as checked, and exchange the OUT proj with
337 for (i = 0; i < n; i++) {
338 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
339 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
340 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
342 /* We have to check for a special case:
343 The in-node could be a Proj from a Perm. In this case,
344 we need to correct the projnum */
345 if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
346 set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
349 /* reroute the edges from the proj to the argument */
350 exchange(pairs[i].out_node, pairs[i].in_node);
351 //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
352 //set_irn_n(pairs[i].out_node, 0, new_Bad());
354 pairs[i].checked = 1;
358 /* Set do_copy to 0 if it's on but we have no free register */
363 real_size = n - get_n_checked_pairs(pairs, n);
365 /* check for cycles and chains */
366 while (get_n_checked_pairs(pairs, n) < n) {
371 /* go to the first not-checked pair */
372 while (pairs[i].checked) i++;
373 cycle = XMALLOCZ(perm_cycle_t);
374 cycle = get_perm_cycle(cycle, pairs, n, i);
376 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
377 for (j = 0; j < cycle->n_elems; j++) {
378 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
380 DB((mod, LEVEL_1, "\n"));
383 We don't need to do anything if we have a Perm with two
384 elements which represents a cycle, because those nodes
385 already represent exchange nodes
387 if (n == 2 && cycle->type == PERM_CYCLE) {
393 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
394 // the copy cascade (this reduces the cycle into a chain)
396 /* build copy/swap nodes from back to front */
397 for (i = cycle->n_elems - 2; i >= 0; i--) {
398 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
399 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
401 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
402 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
404 If we have a cycle and don't copy: we need to create exchange nodes
405 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
406 IN_1 = in node with register i
407 IN_2 = in node with register i + 1
408 OUT_1 = out node with register i + 1
409 OUT_2 = out node with register i
411 if (cycle->type == PERM_CYCLE && !do_copy) {
415 /* At this point we have to handle the following problem: */
417 /* If we have a cycle with more than two elements, then */
418 /* this could correspond to the following Perm node: */
420 /* +----+ +----+ +----+ */
421 /* | r1 | | r2 | | r3 | */
422 /* +-+--+ +-+--+ +--+-+ */
425 /* +-+--------+---------+-+ */
427 /* +-+--------+---------+-+ */
430 /* +-+--+ +-+--+ +--+-+ */
431 /* |Proj| |Proj| |Proj| */
432 /* | r2 | | r3 | | r1 | */
433 /* +----+ +----+ +----+ */
435 /* This node is about to be split up into two 2x Perm's */
436 /* for which we need 4 Proj's and the one additional Proj */
437 /* of the first Perm has to be one IN of the second. So in */
438 /* general we need to create one additional Proj for each */
439 /* "middle" Perm and set this to one in node of the successor */
442 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
443 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
444 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
445 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
447 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
451 /* cycle is not done yet */
452 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
454 /* create intermediate proj */
455 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
457 /* set as in for next Perm */
458 pairs[pidx].in_node = res1;
461 set_Proj_pred(res2, cpyxchg);
462 set_Proj_proj(res2, 0);
463 set_Proj_pred(res1, cpyxchg);
464 set_Proj_proj(res1, 1);
466 arch_set_irn_register(res2, cycle->elems[i + 1]);
467 arch_set_irn_register(res1, cycle->elems[i]);
469 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
470 sched_add_after(sched_point, cpyxchg);
472 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
474 /* set the new scheduling point */
478 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
479 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
481 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
482 arch_set_irn_register(cpyxchg, cycle->elems[i + 1]);
485 /* exchange copy node and proj */
486 exchange(res2, cpyxchg);
488 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
489 sched_add_after(sched_point, cpyxchg);
491 /* set the new scheduling point */
492 sched_point = cpyxchg;
496 free((void *) cycle->elems);
500 /* remove the perm from schedule */
509 static int has_irn_users(const ir_node *irn) {
510 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
514 * Skip all Proj nodes.
516 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
518 irn = get_Proj_pred(irn);
522 static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) {
523 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
524 ir_node *block = get_nodes_block(irn);
527 for (cur_node = sched_prev(irn);
528 ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
529 cur_node = sched_prev(cur_node))
531 if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill))
538 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
539 be_irg_t *birg = env->birg;
540 ir_graph *irg = be_get_birg_irg(birg);
541 pset *op_set = env->op_set;
542 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
543 ir_node *block = get_nodes_block(irn);
544 const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1);
545 ir_node *in[2], *keep, *cpy;
546 op_copy_assoc_t key, *entry;
547 DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
549 if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
550 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
554 /* Make a not spillable copy of the different node */
555 /* this is needed because the different irn could be */
556 /* in block far far away */
557 /* The copy is optimized later if not needed */
559 /* check if already exists such a copy in the schedule immediately before */
560 cpy = find_copy(env, belower_skip_proj(irn), other_different);
562 cpy = be_new_Copy(cls, irg, block, other_different);
563 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
564 DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
567 DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
573 /* Add the Keep resp. CopyKeep and reroute the users */
574 /* of the other_different irn in case of CopyKeep. */
575 if (has_irn_users(other_different)) {
576 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
577 be_node_set_reg_class(keep, 1, cls);
580 keep = be_new_Keep(cls, irg, block, 2, in);
583 DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
585 /* insert copy and keep into schedule */
586 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
587 if (! sched_is_scheduled(cpy))
588 sched_add_before(belower_skip_proj(irn), cpy);
589 sched_add_after(irn, keep);
591 /* insert the other different and it's copies into the set */
592 key.op = other_different;
593 entry = pset_find(op_set, &key, hash_irn(other_different));
596 entry = obstack_alloc(&env->obst, sizeof(*entry));
597 ir_nodeset_init(&entry->copies);
598 entry->op = other_different;
603 ir_nodeset_insert(&entry->copies, cpy);
605 /* insert keep in case of CopyKeep */
606 if (be_is_CopyKeep(keep)) {
607 ir_nodeset_insert(&entry->copies, keep);
610 pset_insert(op_set, entry, hash_irn(other_different));
614 * Checks if node has a must_be_different constraint in output and adds a Keep
615 * then to assure the constraint.
617 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
618 ir_node *skipped_irn = belower_skip_proj(irn);
619 const arch_register_req_t *req = arch_get_register_req(irn, -1);
621 if (arch_register_req_is(req, must_be_different)) {
622 const unsigned other = req->other_different;
625 if (arch_register_req_is(req, should_be_same)) {
626 const unsigned same = req->other_same;
628 if (is_po2(other) && is_po2(same)) {
629 int idx_other = ntz(other);
630 int idx_same = ntz(same);
633 * We can safely ignore a should_be_same x must_be_different y
634 * IFF both inputs are equal!
636 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
641 for (i = 0; 1U << i <= other; ++i) {
642 if (other & (1U << i)) {
643 ir_node *different_from = get_irn_n(skipped_irn, i);
644 gen_assure_different_pattern(irn, different_from, env);
651 * Calls the functions to assure register constraints.
653 * @param irn The node to be checked for lowering
654 * @param walk_env The walker environment
656 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
660 if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
661 assure_different_constraints(irn, walk_env);
665 * Melt all copykeeps pointing to the same node
666 * (or Projs of the same node), copying the same operand.
668 static void melt_copykeeps(constraint_env_t *cenv) {
669 be_irg_t *birg = cenv->birg;
670 ir_graph *irg = be_get_birg_irg(birg);
671 op_copy_assoc_t *entry;
674 foreach_pset(cenv->op_set, entry) {
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((cenv->dbg, 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((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
726 sched_remove(cur_ck);
731 /* check, if we found some candidates for melting */
733 DBG((cenv->dbg, 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((cenv->dbg, 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;
792 op_copy_assoc_t *entry;
794 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
796 DEBUG_ONLY(cenv.dbg = mod;)
798 cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
799 obstack_init(&cenv.obst);
801 irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
803 /* melt copykeeps, pointing to projs of */
804 /* the same mode_T node and keeping the */
806 melt_copykeeps(&cenv);
809 foreach_pset(cenv.op_set, entry) {
812 ir_nodeset_iterator_t iter;
813 be_ssa_construction_env_t senv;
815 n = ir_nodeset_size(&entry->copies);
816 nodes = alloca(n * sizeof(nodes[0]));
818 /* put the node in an array */
819 DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
821 /* collect all copies */
823 foreach_ir_nodeset(&entry->copies, cp, iter) {
825 DB((mod, LEVEL_1, ", %+F ", cp));
828 DB((mod, LEVEL_1, "\n"));
830 /* introduce the copies for the operand and it's copies */
831 be_ssa_construction_init(&senv, birg);
832 be_ssa_construction_add_copy(&senv, entry->op);
833 be_ssa_construction_add_copies(&senv, nodes, n);
834 be_ssa_construction_fix_users(&senv, entry->op);
835 be_ssa_construction_destroy(&senv);
837 /* Could be that not all CopyKeeps are really needed, */
838 /* so we transform unnecessary ones into Keeps. */
839 foreach_ir_nodeset(&entry->copies, cp, iter) {
840 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
842 int n = get_irn_arity(cp);
844 keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
845 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
846 sched_add_before(cp, keep);
848 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
854 ir_nodeset_destroy(&entry->copies);
857 del_pset(cenv.op_set);
858 obstack_free(&cenv.obst, NULL);
859 be_liveness_invalidate(be_get_birg_liveness(birg));
864 * Push nodes that do not need to be permed through the Perm.
865 * This is commonly a reload cascade at block ends.
866 * @note This routine needs interference.
867 * @note Probably, we can implement it a little more efficient.
868 * Especially searching the frontier lazily might be better.
869 * @param perm The perm.
870 * @param data The walker data (lower_env_t).
871 * @return 1, if there is something left to perm over.
872 * 0, if removed the complete perm.
874 static int push_through_perm(ir_node *perm, void *data)
876 lower_env_t *env = data;
877 const arch_env_t *aenv = env->arch_env;
879 ir_graph *irg = get_irn_irg(perm);
880 ir_node *bl = get_nodes_block(perm);
882 int arity = get_irn_arity(perm);
885 bitset_t *moved = bitset_alloca(arity);
888 ir_node *frontier = bl;
889 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
892 const ir_edge_t *edge;
893 ir_node *one_proj = NULL, *irn;
894 const arch_register_class_t *cls = NULL;
896 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
898 /* get some Proj and find out the register class of that Proj. */
899 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
900 one_proj = get_edge_src_irn(edge);
901 assert(is_Proj(one_proj));
902 cls = arch_get_irn_reg_class(one_proj, -1);
904 /* Find the point in the schedule after which the
905 * potentially movable nodes must be defined.
906 * A Perm will only be pushed up to first instruction
907 * which lets an operand of itself die.
908 * If we would allow to move the Perm above this instruction,
909 * the former dead operand would be live now at the point of
910 * the Perm, increasing the register pressure by one.
912 sched_foreach_reverse_from (sched_prev(perm), irn) {
913 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
914 ir_node *op = get_irn_n(irn, i);
915 if (arch_irn_consider_in_reg_alloc(aenv, cls, op) &&
916 !values_interfere(env->birg, op, one_proj)) {
924 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
926 node = sched_prev(perm);
928 while(!sched_is_begin(node)) {
929 const arch_register_req_t *req;
933 /* search if node is a INPUT of Perm */
934 foreach_out_edge(perm, edge) {
935 ir_node *out = get_edge_src_irn(edge);
936 int pn = get_Proj_proj(out);
937 ir_node *in = get_irn_n(perm, pn);
944 /* it wasn't an input to the perm, we can't do anything more */
947 if(!sched_comes_after(frontier, node))
949 if(arch_irn_is(aenv, node, modify_flags))
952 req = arch_get_register_req(get_Proj_pred(node),
953 -1 - get_Proj_proj(node));
955 req = arch_get_register_req(node, -1);
957 if(req->type != arch_register_req_type_normal)
959 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
960 ir_node *opop = get_irn_n(node, i);
961 if (arch_irn_consider_in_reg_alloc(aenv, cls, opop)) {
968 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
970 /* move the movable node in front of the Perm */
972 sched_add_after(perm, node);
974 /* give it the proj's register */
975 arch_set_irn_register(node, arch_get_irn_register(proj));
977 /* reroute all users of the proj to the moved node. */
978 edges_reroute(proj, node, irg);
981 set_Proj_pred(proj, new_Bad());
984 bitset_set(moved, input);
987 node = sched_prev(node);
990 /* well, we could not push anything through the perm */
994 new_size = arity - n_moved;
999 map = alloca(new_size * sizeof(map[0]));
1000 proj_map = alloca(arity * sizeof(proj_map[0]));
1001 memset(proj_map, -1, sizeof(proj_map[0]));
1003 for(i = 0; i < arity; ++i) {
1004 if(bitset_is_set(moved, i))
1010 assert(n == new_size);
1011 foreach_out_edge(perm, edge) {
1012 ir_node *proj = get_edge_src_irn(edge);
1013 int pn = get_Proj_proj(proj);
1016 set_Proj_proj(proj, pn);
1019 be_Perm_reduce(perm, new_size, map);
1024 * Calls the corresponding lowering function for the node.
1026 * @param irn The node to be checked for lowering
1027 * @param walk_env The walker environment
1029 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
1032 if (is_Block(irn) || is_Proj(irn))
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) {
1053 ir_graph *irg = be_get_birg_irg(birg);
1056 env.arch_env = be_get_birg_arch_env(birg);
1057 env.do_copy = do_copy;
1058 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
1060 /* we will need interference */
1061 be_liveness_assure_chk(be_get_birg_liveness(birg));
1063 irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);