2 * Copyright (C) 1995-2007 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"
45 #include "besched_t.h"
47 #include "bessaconstr.h"
48 #include "benodesets.h"
49 #include "beintlive_t.h"
51 #undef KEEP_ALIVE_COPYKEEP_HACK
53 /* associates 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) */
102 /* Compare the two operands */
103 static int cmp_op_copy_assoc(const void *a, const void *b) {
104 const op_copy_assoc_t *op1 = a;
105 const op_copy_assoc_t *op2 = b;
107 return op1->op != op2->op;
110 /* Compare the in registers of two register pairs */
111 static int compare_reg_pair(const void *a, const void *b) {
112 const reg_pair_t *pair_a = a;
113 const reg_pair_t *pair_b = b;
115 if (pair_a->in_reg->index > pair_b->in_reg->index)
121 /* returns the number register pairs marked as checked */
122 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
123 int i, n_checked = 0;
125 for (i = 0; i < n; i++) {
126 if (pairs[i].checked)
134 * Gets the node corresponding to a 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 * @param in_out 0 == look for IN register, 1 == look for OUT register
142 * @return The corresponding node or NULL if not found
144 static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
148 for (i = 0; i < n; i++) {
149 /* out register matches */
150 if (pairs[i].out_reg->index == reg->index)
151 return pairs[i].out_node;
155 for (i = 0; i < n; i++) {
156 /* in register matches */
157 if (pairs[i].in_reg->index == reg->index)
158 return pairs[i].in_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 (pairs[i].out_reg->index == reg_idx)
186 for (i = 0; i < n; i++) {
187 /* in register matches */
188 if (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 = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
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 const arch_env_t *arch_env;
288 lower_env_t *env = walk_env;
291 int n, i, pn, do_copy, j, n_ops;
293 const ir_edge_t *edge;
295 ir_node *sched_point, *block, *in[2];
296 ir_node *arg1, *arg2, *res1, *res2;
297 ir_node *cpyxchg = NULL;
298 DEBUG_ONLY(firm_dbg_module_t *mod;)
300 arch_env = env->arch_env;
301 do_copy = env->do_copy;
302 DEBUG_ONLY(mod = env->dbg_module;)
303 block = get_nodes_block(irn);
306 Get the schedule predecessor node to the perm
307 NOTE: This works with auto-magic. If we insert the
308 new copy/exchange nodes after this node, everything
311 sched_point = sched_prev(irn);
312 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
313 assert(sched_point && "Perm is not scheduled or has no predecessor");
315 n = get_irn_arity(irn);
316 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
318 reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
319 pairs = alloca(n * sizeof(pairs[0]));
321 /* build the list of register pairs (in, out) */
323 foreach_out_edge(irn, edge) {
324 pairs[i].out_node = get_edge_src_irn(edge);
325 pn = get_Proj_proj(pairs[i].out_node);
326 pairs[i].in_node = get_irn_n(irn, pn);
328 pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node);
329 pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
331 pairs[i].checked = 0;
335 /* sort the register pairs by the indices of the in registers */
336 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
338 /* Mark all equal pairs as checked, and exchange the OUT proj with
340 for (i = 0; i < n; i++) {
341 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
342 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
343 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
345 /* We have to check for a special case:
346 The in-node could be a Proj from a Perm. In this case,
347 we need to correct the projnum */
348 if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
349 set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
352 /* remove the proj from the schedule */
353 sched_remove(pairs[i].out_node);
355 /* reroute the edges from the proj to the argument */
356 exchange(pairs[i].out_node, pairs[i].in_node);
357 //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
358 //set_irn_n(pairs[i].out_node, 0, new_Bad());
360 pairs[i].checked = 1;
364 /* Set do_copy to 0 if it's on but we have no free register */
369 real_size = n - get_n_checked_pairs(pairs, n);
371 be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
373 /* check for cycles and chains */
374 while (get_n_checked_pairs(pairs, n) < n) {
377 /* go to the first not-checked pair */
378 while (pairs[i].checked) i++;
379 cycle = xcalloc(1, sizeof(*cycle));
380 cycle = get_perm_cycle(cycle, pairs, n, i);
382 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
383 for (j = 0; j < cycle->n_elems; j++) {
384 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
386 DB((mod, LEVEL_1, "\n"));
389 We don't need to do anything if we have a Perm with two
390 elements which represents a cycle, because those nodes
391 already represent exchange nodes
393 if (n == 2 && cycle->type == PERM_CYCLE) {
399 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
400 // the copy cascade (this reduces the cycle into a chain)
402 /* build copy/swap nodes from back to front */
403 for (i = cycle->n_elems - 2; i >= 0; i--) {
404 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
405 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
407 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
408 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
410 If we have a cycle and don't copy: we need to create exchange nodes
411 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
412 IN_1 = in node with register i
413 IN_2 = in node with register i + 1
414 OUT_1 = out node with register i + 1
415 OUT_2 = out node with register i
417 if (cycle->type == PERM_CYCLE && !do_copy) {
421 /* At this point we have to handle the following problem: */
423 /* If we have a cycle with more than two elements, then */
424 /* this could correspond to the following Perm node: */
426 /* +----+ +----+ +----+ */
427 /* | r1 | | r2 | | r3 | */
428 /* +-+--+ +-+--+ +--+-+ */
431 /* +-+--------+---------+-+ */
433 /* +-+--------+---------+-+ */
436 /* +-+--+ +-+--+ +--+-+ */
437 /* |Proj| |Proj| |Proj| */
438 /* | r2 | | r3 | | r1 | */
439 /* +----+ +----+ +----+ */
441 /* This node is about to be split up into two 2x Perm's */
442 /* for which we need 4 Proj's and the one additional Proj */
443 /* of the first Perm has to be one IN of the second. So in */
444 /* general we need to create one additional Proj for each */
445 /* "middle" Perm and set this to one in node of the successor */
448 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
449 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
450 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
451 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
453 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
457 /* cycle is not done yet */
458 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
460 /* create intermediate proj */
461 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
463 /* set as in for next Perm */
464 pairs[pidx].in_node = res1;
472 set_Proj_pred(res2, cpyxchg);
473 set_Proj_proj(res2, 0);
474 set_Proj_pred(res1, cpyxchg);
475 set_Proj_proj(res1, 1);
477 sched_add_after(sched_point, res1);
478 sched_add_after(sched_point, res2);
480 arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
481 arch_set_irn_register(arch_env, res1, cycle->elems[i]);
483 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
484 sched_add_after(sched_point, cpyxchg);
486 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
488 /* set the new scheduling point */
492 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
493 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
495 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
496 arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
499 /* remove the proj from the schedule */
502 /* exchange copy node and proj */
503 exchange(res2, cpyxchg);
505 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
506 sched_add_after(sched_point, cpyxchg);
508 /* set the new scheduling point */
509 sched_point = cpyxchg;
513 be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
515 free((void *) cycle->elems);
519 /* remove the perm from schedule */
528 static int get_n_out_edges(const ir_node *irn) {
529 const ir_edge_t *edge;
532 foreach_out_edge(irn, edge) {
539 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
541 irn = get_Proj_pred(irn);
545 static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) {
546 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
547 ir_node *block = get_nodes_block(irn);
550 for (cur_node = sched_prev(irn);
551 ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
552 cur_node = sched_prev(cur_node))
554 if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill))
561 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
562 be_irg_t *birg = env->birg;
563 ir_graph *irg = be_get_birg_irg(birg);
564 pset *op_set = env->op_set;
565 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
566 ir_node *block = get_nodes_block(irn);
567 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
568 ir_node *in[2], *keep, *cpy;
569 op_copy_assoc_t key, *entry;
570 DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
572 if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
573 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
577 /* Make a not spillable copy of the different node */
578 /* this is needed because the different irn could be */
579 /* in block far far away */
580 /* The copy is optimized later if not needed */
582 /* check if already exists such a copy in the schedule immediatly before */
583 cpy = find_copy(env, belower_skip_proj(irn), other_different);
585 cpy = be_new_Copy(cls, irg, block, other_different);
586 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
587 DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
590 DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
596 /* Add the Keep resp. CopyKeep and reroute the users */
597 /* of the other_different irn in case of CopyKeep. */
598 if (get_n_out_edges(other_different) == 0) {
599 keep = be_new_Keep(cls, irg, block, 2, in);
602 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
603 be_node_set_reg_class(keep, 1, cls);
606 DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
608 /* insert copy and keep into schedule */
609 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
610 if (! sched_is_scheduled(cpy))
611 sched_add_before(belower_skip_proj(irn), cpy);
612 sched_add_after(irn, keep);
614 /* insert the other different and it's copies into the set */
615 key.op = other_different;
616 entry = pset_find(op_set, &key, nodeset_hash(other_different));
619 entry = obstack_alloc(&env->obst, sizeof(*entry));
620 ir_nodeset_init(&entry->copies);
621 entry->op = other_different;
626 ir_nodeset_insert(&entry->copies, cpy);
628 /* insert keep in case of CopyKeep */
629 if (be_is_CopyKeep(keep)) {
630 ir_nodeset_insert(&entry->copies, keep);
633 pset_insert(op_set, entry, nodeset_hash(other_different));
637 * Checks if node has a should_be_different constraint in output
638 * and adds a Keep then to assure the constraint.
640 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
641 const arch_register_req_t *req;
642 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
644 req = arch_get_register_req(arch_env, irn, -1);
646 if (arch_register_req_is(req, should_be_different)) {
647 ir_node *different_from = get_irn_n(belower_skip_proj(irn), req->other_different);
648 gen_assure_different_pattern(irn, different_from, env);
649 } else if (arch_register_req_is(req, should_be_different_from_all)) {
650 int i, n = get_irn_arity(belower_skip_proj(irn));
651 for (i = 0; i < n; i++) {
652 gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env);
660 * Calls the functions to assure register constraints.
662 * @param irn The node to be checked for lowering
663 * @param walk_env The walker environment
665 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
669 if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
670 assure_different_constraints(irn, walk_env);
676 * Melt all copykeeps pointing to the same node
677 * (or Projs of the same node), copying the same operand.
679 static void melt_copykeeps(constraint_env_t *cenv) {
680 be_irg_t *birg = cenv->birg;
681 ir_graph *irg = be_get_birg_irg(birg);
682 op_copy_assoc_t *entry;
685 foreach_pset(cenv->op_set, entry) {
689 ir_nodeset_iterator_t iter;
690 ir_node **ck_arr, **melt_arr;
694 /* collect all copykeeps */
696 foreach_ir_nodeset(&entry->copies, cp, iter) {
697 if (be_is_CopyKeep(cp)) {
698 obstack_grow(&obst, &cp, sizeof(cp));
701 #ifdef KEEP_ALIVE_COPYKEEP_HACK
703 set_irn_mode(cp, mode_ANY);
706 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
709 /* compare each copykeep with all other copykeeps */
710 ck_arr = (ir_node **)obstack_finish(&obst);
711 for (idx = 0; idx < num_ck; ++idx) {
712 ir_node *ref, *ref_mode_T;
718 ir_node *sched_pt = NULL;
722 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
723 obstack_grow(&obst, &ref, sizeof(ref));
725 DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
727 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
728 for (j = 0; j < num_ck; ++j) {
729 ir_node *cur_ck = ck_arr[j];
731 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
732 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
733 ir_nodeset_remove(&entry->copies, cur_ck);
734 DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
737 sched_remove(cur_ck);
742 /* check, if we found some candidates for melting */
744 DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
748 ir_nodeset_remove(&entry->copies, ref);
751 melt_arr = (ir_node **)obstack_finish(&obst);
752 /* melt all found copykeeps */
753 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
754 for (j = 0; j < n_melt; ++j) {
755 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
757 /* now, we can kill the melted keep, except the */
758 /* ref one, we still need some information */
759 if (melt_arr[j] != ref)
760 be_kill_node(melt_arr[j]);
763 #ifdef KEEP_ALIVE_COPYKEEP_HACK
764 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
767 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));
768 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
770 /* set register class for all keeped inputs */
771 for (j = 1; j <= n_melt; ++j)
772 be_node_set_reg_class(new_ck, j, entry->cls);
774 ir_nodeset_insert(&entry->copies, new_ck);
776 /* find scheduling point */
777 if (get_irn_mode(ref_mode_T) == mode_T) {
778 /* walk along the Projs */
779 for (sched_pt = sched_next(ref_mode_T); is_Proj(sched_pt) || be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt); sched_pt = sched_next(sched_pt))
780 /* just walk along the schedule until a non-Proj/Keep/CopyKeep node is found*/ ;
783 sched_pt = ref_mode_T;
786 sched_add_before(sched_pt, new_ck);
787 DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
789 /* finally: kill the reference copykeep */
794 obstack_free(&obst, NULL);
799 * Walks over all nodes to assure register constraints.
801 * @param birg The birg structure containing the irg
803 void assure_constraints(be_irg_t *birg) {
804 ir_graph *irg = be_get_birg_irg(birg);
805 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
806 constraint_env_t cenv;
807 op_copy_assoc_t *entry;
809 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
811 be_assure_dom_front(birg);
813 DEBUG_ONLY(cenv.dbg = mod;)
815 cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
816 obstack_init(&cenv.obst);
818 irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
820 /* melt copykeeps, pointing to projs of */
821 /* the same mode_T node and keeping the */
823 melt_copykeeps(&cenv);
826 foreach_pset(cenv.op_set, entry) {
829 ir_nodeset_iterator_t iter;
830 be_ssa_construction_env_t senv;
832 n = ir_nodeset_size(&entry->copies);
833 nodes = alloca(n * sizeof(nodes[0]));
835 /* put the node in an array */
836 DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
838 /* collect all copies */
840 foreach_ir_nodeset(&entry->copies, cp, iter) {
842 DB((mod, LEVEL_1, ", %+F ", cp));
845 DB((mod, LEVEL_1, "\n"));
847 /* introduce the copies for the operand and it's copies */
848 be_ssa_construction_init(&senv, birg);
849 be_ssa_construction_add_copy(&senv, entry->op);
850 be_ssa_construction_add_copies(&senv, nodes, n);
851 be_ssa_construction_fix_users(&senv, entry->op);
852 be_ssa_construction_destroy(&senv);
854 /* Could be that not all CopyKeeps are really needed, */
855 /* so we transform unnecessary ones into Keeps. */
856 foreach_ir_nodeset(&entry->copies, cp, iter) {
857 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
859 int n = get_irn_arity(cp);
861 keep = be_new_Keep(arch_get_irn_reg_class(arch_env, cp, -1),
862 irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
863 sched_add_before(cp, keep);
865 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
871 ir_nodeset_destroy(&entry->copies);
874 del_pset(cenv.op_set);
875 obstack_free(&cenv.obst, NULL);
876 be_liveness_invalidate(be_get_birg_liveness(birg));
881 * Push nodes that do not need to be permed through the Perm.
882 * This is commonly a reload cascade at block ends.
883 * @note This routine needs interference.
884 * @note Probably, we can implement it a little more efficient.
885 * Especially searching the frontier lazily might be better.
886 * @param perm The perm.
887 * @param data The walker data (lower_env_t).
888 * @return 1, if there is something left to perm over.
889 * 0, if removed the complete perm.
891 static int push_through_perm(ir_node *perm, void *data)
893 lower_env_t *env = data;
894 const arch_env_t *aenv = env->arch_env;
896 ir_graph *irg = get_irn_irg(perm);
897 ir_node *bl = get_nodes_block(perm);
898 int n = get_irn_arity(perm);
899 int *map = alloca(n * sizeof(map[0]));
900 ir_node **projs = alloca(n * sizeof(projs[0]));
901 bitset_t *keep = bitset_alloca(n);
902 ir_node *frontier = sched_first(bl);
903 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
905 int i, new_size, n_keep;
906 const ir_edge_t *edge;
907 ir_node *last_proj, *irn;
908 const arch_register_class_t *cls;
910 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
912 /* get some proj and find out the register class of the proj. */
913 foreach_out_edge (perm, edge) {
914 last_proj = get_edge_src_irn(edge);
915 cls = arch_get_irn_reg_class(aenv, last_proj, -1);
916 assert(is_Proj(last_proj));
920 /* find the point in the schedule after which the
921 * potentially movable nodes must be defined.
922 * A perm will only be pushed up to first instruction
923 * which lets an operand of itself die. */
925 sched_foreach_reverse_from (sched_prev(perm), irn) {
926 for(i = get_irn_arity(irn) - 1; i >= 0; --i) {
927 ir_node *op = get_irn_n(irn, i);
928 if(arch_irn_consider_in_reg_alloc(aenv, cls, op)
929 && !values_interfere(env->birg, op, last_proj)) {
930 frontier = sched_next(irn);
937 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
939 foreach_out_edge (perm, edge) {
940 ir_node *proj = get_edge_src_irn(edge);
941 int nr = get_Proj_proj(proj);
942 ir_node *op = get_irn_n(perm, nr);
946 /* we will need the last Proj as an insertion point
947 * for the instruction(s) pushed through the Perm */
948 if (sched_comes_after(last_proj, proj))
953 bitset_set(keep, nr);
954 if (!is_Proj(op) && get_nodes_block(op) == bl
955 && (op == frontier || sched_comes_after(frontier, op))) {
956 for (i = get_irn_arity(op) - 1; i >= 0; --i) {
957 ir_node *opop = get_irn_n(op, i);
958 if (!arch_irn_consider_in_reg_alloc(aenv, cls, opop)) {
959 bitset_clear(keep, nr);
966 n_keep = bitset_popcnt(keep);
968 /* well, we could not push enything through the perm */
972 assert(is_Proj(last_proj));
974 DBG((mod, LEVEL_2, "\tkeep: %d, total: %d, mask: %b\n", n_keep, n, keep));
975 last_proj = sched_next(last_proj);
976 for (new_size = 0, i = 0; i < n; ++i) {
977 ir_node *proj = projs[i];
979 if (bitset_is_set(keep, i)) {
981 set_Proj_proj(proj, map[i]);
982 DBG((mod, LEVEL_1, "\targ %d remap to %d\n", i, map[i]));
986 ir_node *move = get_irn_n(perm, i);
988 DBG((mod, LEVEL_2, "\tmoving %+F before %+F, killing %+F\n", move, last_proj, proj));
990 /* move the movable node in front of the Perm */
992 sched_add_before(last_proj, move);
994 /* give it the proj's register */
995 arch_set_irn_register(aenv, move, arch_get_irn_register(aenv, proj));
997 /* reroute all users of the proj to the moved node. */
998 edges_reroute(proj, move, irg);
1000 /* remove the proj from the schedule. */
1003 /* and like it to bad so it is no more in the use array of the perm */
1004 set_Proj_pred(proj, get_irg_bad(irg));
1012 be_Perm_reduce(perm, new_size, map);
1018 * Calls the corresponding lowering function for the node.
1020 * @param irn The node to be checked for lowering
1021 * @param walk_env The walker environment
1023 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
1024 if (! is_Block(irn) && ! is_Proj(irn)) {
1025 if (be_is_Perm(irn)) {
1026 int perm_stayed = push_through_perm(irn, walk_env);
1028 lower_perm_node(irn, walk_env);
1036 * Walks over all blocks in an irg and performs lowering need to be
1037 * done after register allocation (e.g. perm lowering).
1039 * @param birg The birg object
1040 * @param do_copy 1 == resolve cycles with a free reg if available
1042 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1044 ir_graph *irg = be_get_birg_irg(birg);
1047 env.arch_env = be_get_birg_arch_env(birg);
1048 env.do_copy = do_copy;
1049 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
1051 /* we will need interference */
1052 be_liveness_assure_chk(be_get_birg_liveness(birg));
1054 irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);