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 ((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 = 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, "perm: %+F\n", irn));
313 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
314 assert(sched_point && "Perm is not scheduled or has no predecessor");
316 n = get_irn_arity(irn);
317 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
319 reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
320 pairs = alloca(n * sizeof(pairs[0]));
322 /* build the list of register pairs (in, out) */
324 foreach_out_edge(irn, edge) {
325 pairs[i].out_node = get_edge_src_irn(edge);
326 pn = get_Proj_proj(pairs[i].out_node);
327 pairs[i].in_node = get_irn_n(irn, pn);
329 pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node);
330 pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
332 pairs[i].checked = 0;
336 /* sort the register pairs by the indices of the in registers */
337 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
339 /* Mark all equal pairs as checked, and exchange the OUT proj with
341 for (i = 0; i < n; i++) {
342 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
343 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
344 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
346 /* We have to check for a special case:
347 The in-node could be a Proj from a Perm. In this case,
348 we need to correct the projnum */
349 if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
350 set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
353 /* reroute the edges from the proj to the argument */
354 exchange(pairs[i].out_node, pairs[i].in_node);
355 //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
356 //set_irn_n(pairs[i].out_node, 0, new_Bad());
358 pairs[i].checked = 1;
362 /* Set do_copy to 0 if it's on but we have no free register */
367 real_size = n - get_n_checked_pairs(pairs, n);
369 be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
371 /* check for cycles and chains */
372 while (get_n_checked_pairs(pairs, n) < n) {
375 /* go to the first not-checked pair */
376 while (pairs[i].checked) i++;
377 cycle = xcalloc(1, sizeof(*cycle));
378 cycle = get_perm_cycle(cycle, pairs, n, i);
380 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
381 for (j = 0; j < cycle->n_elems; j++) {
382 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
384 DB((mod, LEVEL_1, "\n"));
387 We don't need to do anything if we have a Perm with two
388 elements which represents a cycle, because those nodes
389 already represent exchange nodes
391 if (n == 2 && cycle->type == PERM_CYCLE) {
397 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
398 // the copy cascade (this reduces the cycle into a chain)
400 /* build copy/swap nodes from back to front */
401 for (i = cycle->n_elems - 2; i >= 0; i--) {
402 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
403 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
405 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
406 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
408 If we have a cycle and don't copy: we need to create exchange nodes
409 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
410 IN_1 = in node with register i
411 IN_2 = in node with register i + 1
412 OUT_1 = out node with register i + 1
413 OUT_2 = out node with register i
415 if (cycle->type == PERM_CYCLE && !do_copy) {
419 /* At this point we have to handle the following problem: */
421 /* If we have a cycle with more than two elements, then */
422 /* this could correspond to the following Perm node: */
424 /* +----+ +----+ +----+ */
425 /* | r1 | | r2 | | r3 | */
426 /* +-+--+ +-+--+ +--+-+ */
429 /* +-+--------+---------+-+ */
431 /* +-+--------+---------+-+ */
434 /* +-+--+ +-+--+ +--+-+ */
435 /* |Proj| |Proj| |Proj| */
436 /* | r2 | | r3 | | r1 | */
437 /* +----+ +----+ +----+ */
439 /* This node is about to be split up into two 2x Perm's */
440 /* for which we need 4 Proj's and the one additional Proj */
441 /* of the first Perm has to be one IN of the second. So in */
442 /* general we need to create one additional Proj for each */
443 /* "middle" Perm and set this to one in node of the successor */
446 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
447 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
448 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
449 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
451 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
455 /* cycle is not done yet */
456 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
458 /* create intermediate proj */
459 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
461 /* set as in for next Perm */
462 pairs[pidx].in_node = res1;
465 set_Proj_pred(res2, cpyxchg);
466 set_Proj_proj(res2, 0);
467 set_Proj_pred(res1, cpyxchg);
468 set_Proj_proj(res1, 1);
470 arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
471 arch_set_irn_register(arch_env, res1, cycle->elems[i]);
473 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
474 sched_add_after(sched_point, cpyxchg);
476 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
478 /* set the new scheduling point */
482 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
483 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
485 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
486 arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
489 /* exchange copy node and proj */
490 exchange(res2, cpyxchg);
492 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
493 sched_add_after(sched_point, cpyxchg);
495 /* set the new scheduling point */
496 sched_point = cpyxchg;
500 be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
502 free((void *) cycle->elems);
506 /* remove the perm from schedule */
515 static int get_n_out_edges(const ir_node *irn) {
516 const ir_edge_t *edge;
519 foreach_out_edge(irn, edge) {
526 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
528 irn = get_Proj_pred(irn);
532 static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) {
533 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
534 ir_node *block = get_nodes_block(irn);
537 for (cur_node = sched_prev(irn);
538 ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
539 cur_node = sched_prev(cur_node))
541 if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill))
548 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
549 be_irg_t *birg = env->birg;
550 ir_graph *irg = be_get_birg_irg(birg);
551 pset *op_set = env->op_set;
552 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
553 ir_node *block = get_nodes_block(irn);
554 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
555 ir_node *in[2], *keep, *cpy;
556 op_copy_assoc_t key, *entry;
557 DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
559 if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
560 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
564 /* Make a not spillable copy of the different node */
565 /* this is needed because the different irn could be */
566 /* in block far far away */
567 /* The copy is optimized later if not needed */
569 /* check if already exists such a copy in the schedule immediatly before */
570 cpy = find_copy(env, belower_skip_proj(irn), other_different);
572 cpy = be_new_Copy(cls, irg, block, other_different);
573 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
574 DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
577 DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
583 /* Add the Keep resp. CopyKeep and reroute the users */
584 /* of the other_different irn in case of CopyKeep. */
585 if (get_n_out_edges(other_different) == 0) {
586 keep = be_new_Keep(cls, irg, block, 2, in);
589 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
590 be_node_set_reg_class(keep, 1, cls);
593 DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
595 /* insert copy and keep into schedule */
596 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
597 if (! sched_is_scheduled(cpy))
598 sched_add_before(belower_skip_proj(irn), cpy);
599 sched_add_after(irn, keep);
601 /* insert the other different and it's copies into the set */
602 key.op = other_different;
603 entry = pset_find(op_set, &key, nodeset_hash(other_different));
606 entry = obstack_alloc(&env->obst, sizeof(*entry));
607 ir_nodeset_init(&entry->copies);
608 entry->op = other_different;
613 ir_nodeset_insert(&entry->copies, cpy);
615 /* insert keep in case of CopyKeep */
616 if (be_is_CopyKeep(keep)) {
617 ir_nodeset_insert(&entry->copies, keep);
620 pset_insert(op_set, entry, nodeset_hash(other_different));
624 * Checks if node has a should_be_different constraint in output
625 * and adds a Keep then to assure the constraint.
627 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
628 const arch_register_req_t *req;
629 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
631 req = arch_get_register_req(arch_env, irn, -1);
633 if (arch_register_req_is(req, should_be_different)) {
634 ir_node *different_from = get_irn_n(belower_skip_proj(irn), req->other_different);
635 gen_assure_different_pattern(irn, different_from, env);
636 } else if (arch_register_req_is(req, should_be_different_from_all)) {
637 int i, n = get_irn_arity(belower_skip_proj(irn));
638 for (i = 0; i < n; i++) {
639 gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env);
647 * Calls the functions to assure register constraints.
649 * @param irn The node to be checked for lowering
650 * @param walk_env The walker environment
652 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
656 if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
657 assure_different_constraints(irn, walk_env);
663 * Melt all copykeeps pointing to the same node
664 * (or Projs of the same node), copying the same operand.
666 static void melt_copykeeps(constraint_env_t *cenv) {
667 be_irg_t *birg = cenv->birg;
668 ir_graph *irg = be_get_birg_irg(birg);
669 op_copy_assoc_t *entry;
672 foreach_pset(cenv->op_set, entry) {
676 ir_nodeset_iterator_t iter;
677 ir_node **ck_arr, **melt_arr;
681 /* collect all copykeeps */
683 foreach_ir_nodeset(&entry->copies, cp, iter) {
684 if (be_is_CopyKeep(cp)) {
685 obstack_grow(&obst, &cp, sizeof(cp));
688 #ifdef KEEP_ALIVE_COPYKEEP_HACK
690 set_irn_mode(cp, mode_ANY);
693 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
696 /* compare each copykeep with all other copykeeps */
697 ck_arr = (ir_node **)obstack_finish(&obst);
698 for (idx = 0; idx < num_ck; ++idx) {
699 ir_node *ref, *ref_mode_T;
705 ir_node *sched_pt = NULL;
709 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
710 obstack_grow(&obst, &ref, sizeof(ref));
712 DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
714 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
715 for (j = 0; j < num_ck; ++j) {
716 ir_node *cur_ck = ck_arr[j];
718 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
719 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
720 ir_nodeset_remove(&entry->copies, cur_ck);
721 DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
724 sched_remove(cur_ck);
729 /* check, if we found some candidates for melting */
731 DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
735 ir_nodeset_remove(&entry->copies, ref);
738 melt_arr = (ir_node **)obstack_finish(&obst);
739 /* melt all found copykeeps */
740 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
741 for (j = 0; j < n_melt; ++j) {
742 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
744 /* now, we can kill the melted keep, except the */
745 /* ref one, we still need some information */
746 if (melt_arr[j] != ref)
747 be_kill_node(melt_arr[j]);
750 #ifdef KEEP_ALIVE_COPYKEEP_HACK
751 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
754 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));
755 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
757 /* set register class for all keeped inputs */
758 for (j = 1; j <= n_melt; ++j)
759 be_node_set_reg_class(new_ck, j, entry->cls);
761 ir_nodeset_insert(&entry->copies, new_ck);
763 /* find scheduling point */
764 if (get_irn_mode(ref_mode_T) == mode_T) {
765 /* walk along the Projs */
766 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))
767 /* just walk along the schedule until a non-Proj/Keep/CopyKeep node is found*/ ;
770 sched_pt = ref_mode_T;
773 sched_add_before(sched_pt, new_ck);
774 DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
776 /* finally: kill the reference copykeep */
781 obstack_free(&obst, NULL);
786 * Walks over all nodes to assure register constraints.
788 * @param birg The birg structure containing the irg
790 void assure_constraints(be_irg_t *birg) {
791 ir_graph *irg = be_get_birg_irg(birg);
792 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
793 constraint_env_t cenv;
794 op_copy_assoc_t *entry;
796 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
798 be_assure_dom_front(birg);
800 DEBUG_ONLY(cenv.dbg = mod;)
802 cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
803 obstack_init(&cenv.obst);
805 irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
807 /* melt copykeeps, pointing to projs of */
808 /* the same mode_T node and keeping the */
810 melt_copykeeps(&cenv);
813 foreach_pset(cenv.op_set, entry) {
816 ir_nodeset_iterator_t iter;
817 be_ssa_construction_env_t senv;
819 n = ir_nodeset_size(&entry->copies);
820 nodes = alloca(n * sizeof(nodes[0]));
822 /* put the node in an array */
823 DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
825 /* collect all copies */
827 foreach_ir_nodeset(&entry->copies, cp, iter) {
829 DB((mod, LEVEL_1, ", %+F ", cp));
832 DB((mod, LEVEL_1, "\n"));
834 /* introduce the copies for the operand and it's copies */
835 be_ssa_construction_init(&senv, birg);
836 be_ssa_construction_add_copy(&senv, entry->op);
837 be_ssa_construction_add_copies(&senv, nodes, n);
838 be_ssa_construction_fix_users(&senv, entry->op);
839 be_ssa_construction_destroy(&senv);
841 /* Could be that not all CopyKeeps are really needed, */
842 /* so we transform unnecessary ones into Keeps. */
843 foreach_ir_nodeset(&entry->copies, cp, iter) {
844 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
846 int n = get_irn_arity(cp);
848 keep = be_new_Keep(arch_get_irn_reg_class(arch_env, cp, -1),
849 irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
850 sched_add_before(cp, keep);
852 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
858 ir_nodeset_destroy(&entry->copies);
861 del_pset(cenv.op_set);
862 obstack_free(&cenv.obst, NULL);
863 be_liveness_invalidate(be_get_birg_liveness(birg));
868 * Push nodes that do not need to be permed through the Perm.
869 * This is commonly a reload cascade at block ends.
870 * @note This routine needs interference.
871 * @note Probably, we can implement it a little more efficient.
872 * Especially searching the frontier lazily might be better.
873 * @param perm The perm.
874 * @param data The walker data (lower_env_t).
875 * @return 1, if there is something left to perm over.
876 * 0, if removed the complete perm.
878 static int push_through_perm(ir_node *perm, void *data)
880 lower_env_t *env = data;
881 const arch_env_t *aenv = env->arch_env;
883 ir_graph *irg = get_irn_irg(perm);
884 ir_node *bl = get_nodes_block(perm);
886 int arity = get_irn_arity(perm);
889 bitset_t *moved = bitset_alloca(arity);
892 ir_node *frontier = sched_first(bl);
893 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
896 const ir_edge_t *edge;
897 ir_node *last_proj, *irn;
898 const arch_register_class_t *cls = NULL;
900 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
902 /* get some proj and find out the register class of the proj. */
903 foreach_out_edge (perm, edge) {
904 last_proj = get_edge_src_irn(edge);
905 cls = arch_get_irn_reg_class(aenv, last_proj, -1);
906 assert(is_Proj(last_proj));
910 /* find the point in the schedule after which the
911 * potentially movable nodes must be defined.
912 * A perm will only be pushed up to first instruction
913 * which lets an operand of itself die. */
915 sched_foreach_reverse_from (sched_prev(perm), irn) {
916 for(i = get_irn_arity(irn) - 1; i >= 0; --i) {
917 ir_node *op = get_irn_n(irn, i);
918 if(arch_irn_consider_in_reg_alloc(aenv, cls, op)
919 && !values_interfere(env->birg, op, last_proj)) {
920 frontier = sched_next(irn);
927 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
929 node = sched_prev(perm);
931 while(!sched_is_begin(node)) {
932 const arch_register_req_t *req;
936 foreach_out_edge(perm, edge) {
937 ir_node *out = get_edge_src_irn(edge);
938 int pn = get_Proj_proj(out);
939 ir_node *in = get_irn_n(perm, pn);
946 /* it wasn't an input to the perm, we can't do anything more */
949 if(!sched_comes_after(frontier, node))
951 if(arch_irn_is(aenv, node, modify_flags))
954 req = arch_get_register_req(aenv, get_Proj_pred(node),
955 -1 - get_Proj_proj(node));
957 req = arch_get_register_req(aenv, node, -1);
959 if(req->type != arch_register_req_type_normal)
961 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
962 ir_node *opop = get_irn_n(node, i);
963 if (arch_irn_consider_in_reg_alloc(aenv, cls, opop)) {
970 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
972 /* move the movable node in front of the Perm */
974 sched_add_after(perm, node);
976 /* give it the proj's register */
977 arch_set_irn_register(aenv, node, arch_get_irn_register(aenv, proj));
979 /* reroute all users of the proj to the moved node. */
980 edges_reroute(proj, node, irg);
983 set_Proj_pred(proj, new_Bad());
986 bitset_set(moved, input);
989 node = sched_prev(node);
992 /* well, we could not push anything through the perm */
996 new_size = arity - n_moved;
1001 map = alloca(new_size * sizeof(map[0]));
1002 proj_map = alloca(arity * sizeof(proj_map[0]));
1003 memset(proj_map, -1, sizeof(proj_map[0]));
1005 for(i = 0; i < arity; ++i) {
1006 if(bitset_is_set(moved, i))
1012 assert(n == new_size);
1013 foreach_out_edge(perm, edge) {
1014 ir_node *proj = get_edge_src_irn(edge);
1015 int pn = get_Proj_proj(proj);
1018 set_Proj_proj(proj, pn);
1021 be_Perm_reduce(perm, new_size, map);
1026 * Calls the corresponding lowering function for the node.
1028 * @param irn The node to be checked for lowering
1029 * @param walk_env The walker environment
1031 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
1034 if (is_Block(irn) || is_Proj(irn))
1036 if (!be_is_Perm(irn))
1039 perm_stayed = push_through_perm(irn, walk_env);
1043 lower_perm_node(irn, walk_env);
1047 * Walks over all blocks in an irg and performs lowering need to be
1048 * done after register allocation (e.g. perm lowering).
1050 * @param birg The birg object
1051 * @param do_copy 1 == resolve cycles with a free reg if available
1053 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1055 ir_graph *irg = be_get_birg_irg(birg);
1058 env.arch_env = be_get_birg_arch_env(birg);
1059 env.do_copy = do_copy;
1060 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
1062 /* we will need interference */
1063 be_liveness_assure_chk(be_get_birg_liveness(birg));
1065 irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);