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"
45 #include "besched_t.h"
47 #include "bessaconstr.h"
48 #include "beintlive_t.h"
50 #undef KEEP_ALIVE_COPYKEEP_HACK
52 /** Associates an op with it's copy and CopyKeep. */
54 ir_node *op; /**< an irn which must be different */
55 ir_nodeset_t copies; /**< all non-spillable copies of this irn */
56 const arch_register_class_t *cls;
59 /** Environment for constraints. */
64 DEBUG_ONLY(firm_dbg_module_t *dbg;)
67 /** Lowering walker environment. */
68 typedef struct _lower_env_t {
70 const arch_env_t *arch_env;
72 DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
75 /** Holds a Perm register pair. */
76 typedef struct _reg_pair_t {
77 const arch_register_t *in_reg; /**< a perm IN register */
78 ir_node *in_node; /**< the in node to which the register belongs */
80 const arch_register_t *out_reg; /**< a perm OUT register */
81 ir_node *out_node; /**< the out node to which the register belongs */
83 int checked; /**< indicates whether the pair was check for cycle or not */
86 typedef enum _perm_type_t {
93 /** Structure to represent cycles or chains in a Perm. */
94 typedef struct _perm_cycle_t {
95 const arch_register_t **elems; /**< the registers in the cycle */
96 int n_elems; /**< number of elements in the cycle */
97 perm_type_t type; /**< type (CHAIN or CYCLE) */
100 /** Compare the two operands. */
101 static int cmp_op_copy_assoc(const void *a, const void *b) {
102 const op_copy_assoc_t *op1 = a;
103 const op_copy_assoc_t *op2 = b;
105 return op1->op != op2->op;
108 /** Compare the in registers of two register pairs. */
109 static int compare_reg_pair(const void *a, const void *b) {
110 const reg_pair_t *pair_a = a;
111 const reg_pair_t *pair_b = b;
113 if (pair_a->in_reg->index > pair_b->in_reg->index)
119 /** returns the number register pairs marked as checked. */
120 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
121 int i, n_checked = 0;
123 for (i = 0; i < n; i++) {
124 if (pairs[i].checked)
132 * Gets the node corresponding to a register from an array of register pairs.
133 * NOTE: The given registers pairs and the register to look for must belong
134 * to the same register class.
136 * @param pairs The array of register pairs
137 * @param n The number of pairs
138 * @param reg The register to look for
139 * @param in_out 0 == look for IN register, 1 == look for OUT register
140 * @return The corresponding node or NULL if not found
142 static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
146 for (i = 0; i < n; i++) {
147 /* out register matches */
148 if (pairs[i].out_reg->index == reg->index)
149 return pairs[i].out_node;
153 for (i = 0; i < n; i++) {
154 /* in register matches */
155 if (pairs[i].in_reg->index == reg->index)
156 return pairs[i].in_node;
164 * Gets the index in the register pair array where the in/out register
165 * corresponds to reg_idx.
167 * @param pairs The array of register pairs
168 * @param n The number of pairs
169 * @param reg The register index to look for
170 * @param in_out 0 == look for IN register, 1 == look for OUT register
171 * @return The corresponding index in pairs or -1 if not found
173 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
177 for (i = 0; i < n; i++) {
178 /* out register matches */
179 if ((int) pairs[i].out_reg->index == reg_idx)
184 for (i = 0; i < n; i++) {
185 /* in register matches */
186 if ((int) pairs[i].in_reg->index == reg_idx)
195 * Gets an array of register pairs and tries to identify a cycle or chain starting
198 * @param cycle Variable to hold the cycle
199 * @param pairs Array of register pairs
200 * @param start Index to start
201 * @return The cycle or chain
203 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
204 int head = pairs[start].in_reg->index;
205 int cur_idx = pairs[start].out_reg->index;
206 int cur_pair_idx = start;
207 int n_pairs_done = get_n_checked_pairs(pairs, n);
209 perm_type_t cycle_tp = PERM_CYCLE;
211 /* We could be right in the middle of a chain, so we need to find the start */
212 while (head != cur_idx) {
213 /* goto previous register in cycle or chain */
214 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
216 if (cur_pair_idx < 0) {
217 cycle_tp = PERM_CHAIN;
221 head = pairs[cur_pair_idx].in_reg->index;
222 start = cur_pair_idx;
226 /* assume worst case: all remaining pairs build a cycle or chain */
227 cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
228 cycle->n_elems = 2; /* initial number of elements is 2 */
229 cycle->elems[0] = pairs[start].in_reg;
230 cycle->elems[1] = pairs[start].out_reg;
231 cycle->type = cycle_tp;
232 cur_idx = pairs[start].out_reg->index;
235 /* check for cycle or end of a chain */
236 while (cur_idx != head) {
237 /* goto next register in cycle or chain */
238 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
240 if (cur_pair_idx < 0)
243 cur_idx = pairs[cur_pair_idx].out_reg->index;
245 /* it's not the first element: insert it */
246 if (cur_idx != head) {
247 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
251 /* we are there where we started -> CYCLE */
252 cycle->type = PERM_CYCLE;
256 /* mark all pairs having one in/out register with cycle in common as checked */
257 for (idx = 0; idx < cycle->n_elems; idx++) {
258 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
260 if (cur_pair_idx >= 0)
261 pairs[cur_pair_idx].checked = 1;
263 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
265 if (cur_pair_idx >= 0)
266 pairs[cur_pair_idx].checked = 1;
273 * Lowers a perm node. Resolves cycles and creates a bunch of
274 * copy and swap operations to permute registers.
275 * Note: The caller of this function has to make sure, that irn
278 * @param irn The perm node
279 * @param block The block the perm node belongs to
280 * @param walk_env The environment
282 static void lower_perm_node(ir_node *irn, void *walk_env) {
283 ir_graph *irg = get_irn_irg(irn);
284 const arch_register_class_t *reg_class;
285 const arch_env_t *arch_env;
286 lower_env_t *env = walk_env;
289 int n, i, pn, do_copy, j, n_ops;
291 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 arch_env = env->arch_env;
299 do_copy = env->do_copy;
300 DEBUG_ONLY(mod = env->dbg_module;)
301 block = get_nodes_block(irn);
304 Get the schedule predecessor node to the perm
305 NOTE: This works with auto-magic. If we insert the
306 new copy/exchange nodes after this node, everything
309 sched_point = sched_prev(irn);
310 DBG((mod, LEVEL_1, "perm: %+F\n", irn));
311 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
312 assert(sched_point && "Perm is not scheduled or has no predecessor");
314 n = get_irn_arity(irn);
315 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
317 reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
318 pairs = alloca(n * sizeof(pairs[0]));
320 /* build the list of register pairs (in, out) */
322 foreach_out_edge(irn, edge) {
323 pairs[i].out_node = get_edge_src_irn(edge);
324 pn = get_Proj_proj(pairs[i].out_node);
325 pairs[i].in_node = get_irn_n(irn, pn);
327 pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node);
328 pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
330 pairs[i].checked = 0;
334 /* sort the register pairs by the indices of the in registers */
335 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
337 /* Mark all equal pairs as checked, and exchange the OUT proj with
339 for (i = 0; i < n; i++) {
340 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
341 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
342 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
344 /* We have to check for a special case:
345 The in-node could be a Proj from a Perm. In this case,
346 we need to correct the projnum */
347 if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
348 set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
351 /* reroute the edges from the proj to the argument */
352 exchange(pairs[i].out_node, pairs[i].in_node);
353 //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
354 //set_irn_n(pairs[i].out_node, 0, new_Bad());
356 pairs[i].checked = 1;
360 /* Set do_copy to 0 if it's on but we have no free register */
365 real_size = n - get_n_checked_pairs(pairs, n);
367 be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
369 /* check for cycles and chains */
370 while (get_n_checked_pairs(pairs, n) < n) {
373 /* go to the first not-checked pair */
374 while (pairs[i].checked) i++;
375 cycle = xcalloc(1, sizeof(*cycle));
376 cycle = get_perm_cycle(cycle, pairs, n, i);
378 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
379 for (j = 0; j < cycle->n_elems; j++) {
380 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
382 DB((mod, LEVEL_1, "\n"));
385 We don't need to do anything if we have a Perm with two
386 elements which represents a cycle, because those nodes
387 already represent exchange nodes
389 if (n == 2 && cycle->type == PERM_CYCLE) {
395 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
396 // the copy cascade (this reduces the cycle into a chain)
398 /* build copy/swap nodes from back to front */
399 for (i = cycle->n_elems - 2; i >= 0; i--) {
400 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
401 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
403 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
404 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
406 If we have a cycle and don't copy: we need to create exchange nodes
407 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
408 IN_1 = in node with register i
409 IN_2 = in node with register i + 1
410 OUT_1 = out node with register i + 1
411 OUT_2 = out node with register i
413 if (cycle->type == PERM_CYCLE && !do_copy) {
417 /* At this point we have to handle the following problem: */
419 /* If we have a cycle with more than two elements, then */
420 /* this could correspond to the following Perm node: */
422 /* +----+ +----+ +----+ */
423 /* | r1 | | r2 | | r3 | */
424 /* +-+--+ +-+--+ +--+-+ */
427 /* +-+--------+---------+-+ */
429 /* +-+--------+---------+-+ */
432 /* +-+--+ +-+--+ +--+-+ */
433 /* |Proj| |Proj| |Proj| */
434 /* | r2 | | r3 | | r1 | */
435 /* +----+ +----+ +----+ */
437 /* This node is about to be split up into two 2x Perm's */
438 /* for which we need 4 Proj's and the one additional Proj */
439 /* of the first Perm has to be one IN of the second. So in */
440 /* general we need to create one additional Proj for each */
441 /* "middle" Perm and set this to one in node of the successor */
444 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
445 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
446 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
447 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
449 cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
453 /* cycle is not done yet */
454 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
456 /* create intermediate proj */
457 res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
459 /* set as in for next Perm */
460 pairs[pidx].in_node = res1;
463 set_Proj_pred(res2, cpyxchg);
464 set_Proj_proj(res2, 0);
465 set_Proj_pred(res1, cpyxchg);
466 set_Proj_proj(res1, 1);
468 arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
469 arch_set_irn_register(arch_env, res1, cycle->elems[i]);
471 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
472 sched_add_after(sched_point, cpyxchg);
474 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
476 /* set the new scheduling point */
480 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
481 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
483 cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
484 arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
487 /* exchange copy node and proj */
488 exchange(res2, cpyxchg);
490 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
491 sched_add_after(sched_point, cpyxchg);
493 /* set the new scheduling point */
494 sched_point = cpyxchg;
498 be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
500 free((void *) cycle->elems);
504 /* remove the perm from schedule */
513 static int has_irn_users(const ir_node *irn) {
514 return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
518 * Skip all Proj nodes.
520 static INLINE ir_node *belower_skip_proj(ir_node *irn) {
522 irn = get_Proj_pred(irn);
526 static ir_node *find_copy(constraint_env_t *env, ir_node *irn, ir_node *op) {
527 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
528 ir_node *block = get_nodes_block(irn);
531 for (cur_node = sched_prev(irn);
532 ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
533 cur_node = sched_prev(cur_node))
535 if (be_get_Copy_op(cur_node) == op && arch_irn_is(arch_env, cur_node, dont_spill))
542 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
543 be_irg_t *birg = env->birg;
544 ir_graph *irg = be_get_birg_irg(birg);
545 pset *op_set = env->op_set;
546 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
547 ir_node *block = get_nodes_block(irn);
548 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
549 ir_node *in[2], *keep, *cpy;
550 op_copy_assoc_t key, *entry;
551 DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
553 if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
554 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
558 /* Make a not spillable copy of the different node */
559 /* this is needed because the different irn could be */
560 /* in block far far away */
561 /* The copy is optimized later if not needed */
563 /* check if already exists such a copy in the schedule immediately before */
564 cpy = find_copy(env, belower_skip_proj(irn), other_different);
566 cpy = be_new_Copy(cls, irg, block, other_different);
567 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
568 DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
571 DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
577 /* Add the Keep resp. CopyKeep and reroute the users */
578 /* of the other_different irn in case of CopyKeep. */
579 if (has_irn_users(other_different)) {
580 keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
581 be_node_set_reg_class(keep, 1, cls);
584 keep = be_new_Keep(cls, irg, block, 2, in);
587 DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
589 /* insert copy and keep into schedule */
590 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
591 if (! sched_is_scheduled(cpy))
592 sched_add_before(belower_skip_proj(irn), cpy);
593 sched_add_after(irn, keep);
595 /* insert the other different and it's copies into the set */
596 key.op = other_different;
597 entry = pset_find(op_set, &key, hash_irn(other_different));
600 entry = obstack_alloc(&env->obst, sizeof(*entry));
601 ir_nodeset_init(&entry->copies);
602 entry->op = other_different;
607 ir_nodeset_insert(&entry->copies, cpy);
609 /* insert keep in case of CopyKeep */
610 if (be_is_CopyKeep(keep)) {
611 ir_nodeset_insert(&entry->copies, keep);
614 pset_insert(op_set, entry, hash_irn(other_different));
618 * Checks if node has a must_be_different constraint in output and adds a Keep
619 * then to assure the constraint.
621 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
622 const arch_register_req_t *req;
623 const arch_env_t *arch_env = be_get_birg_arch_env(env->birg);
624 ir_node *skipped_irn = belower_skip_proj(irn);
626 req = arch_get_register_req(arch_env, irn, -1);
628 if (arch_register_req_is(req, must_be_different)) {
629 const unsigned other = req->other_different;
632 if (arch_register_req_is(req, should_be_same)) {
633 const unsigned same = req->other_same;
635 if (is_po2(other) && is_po2(same)) {
636 int idx_other = ntz(other);
637 int idx_same = ntz(same);
640 * We can safely ignore a should_be_same x must_be_different y
641 * IFF both inputs are equal!
643 if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
648 for (i = 0; 1U << i <= other; ++i) {
649 if (other & (1U << i)) {
650 ir_node *different_from = get_irn_n(skipped_irn, i);
651 gen_assure_different_pattern(irn, different_from, env);
658 * Calls the functions to assure register constraints.
660 * @param irn The node to be checked for lowering
661 * @param walk_env The walker environment
663 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
667 if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
668 assure_different_constraints(irn, walk_env);
672 * Melt all copykeeps pointing to the same node
673 * (or Projs of the same node), copying the same operand.
675 static void melt_copykeeps(constraint_env_t *cenv) {
676 be_irg_t *birg = cenv->birg;
677 ir_graph *irg = be_get_birg_irg(birg);
678 op_copy_assoc_t *entry;
681 foreach_pset(cenv->op_set, entry) {
685 ir_nodeset_iterator_t iter;
686 ir_node **ck_arr, **melt_arr;
690 /* collect all copykeeps */
692 foreach_ir_nodeset(&entry->copies, cp, iter) {
693 if (be_is_CopyKeep(cp)) {
694 obstack_grow(&obst, &cp, sizeof(cp));
697 #ifdef KEEP_ALIVE_COPYKEEP_HACK
699 set_irn_mode(cp, mode_ANY);
702 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
705 /* compare each copykeep with all other copykeeps */
706 ck_arr = (ir_node **)obstack_finish(&obst);
707 for (idx = 0; idx < num_ck; ++idx) {
708 ir_node *ref, *ref_mode_T;
714 ir_node *sched_pt = NULL;
718 ref_mode_T = skip_Proj(get_irn_n(ref, 1));
719 obstack_grow(&obst, &ref, sizeof(ref));
721 DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
723 /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
724 for (j = 0; j < num_ck; ++j) {
725 ir_node *cur_ck = ck_arr[j];
727 if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
728 obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
729 ir_nodeset_remove(&entry->copies, cur_ck);
730 DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
733 sched_remove(cur_ck);
738 /* check, if we found some candidates for melting */
740 DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
744 ir_nodeset_remove(&entry->copies, ref);
747 melt_arr = (ir_node **)obstack_finish(&obst);
748 /* melt all found copykeeps */
749 NEW_ARR_A(ir_node *, new_ck_in, n_melt);
750 for (j = 0; j < n_melt; ++j) {
751 new_ck_in[j] = get_irn_n(melt_arr[j], 1);
753 /* now, we can kill the melted keep, except the */
754 /* ref one, we still need some information */
755 if (melt_arr[j] != ref)
756 kill_node(melt_arr[j]);
759 #ifdef KEEP_ALIVE_COPYKEEP_HACK
760 new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
763 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));
764 #endif /* KEEP_ALIVE_COPYKEEP_HACK */
766 /* set register class for all kept inputs */
767 for (j = 1; j <= n_melt; ++j)
768 be_node_set_reg_class(new_ck, j, entry->cls);
770 ir_nodeset_insert(&entry->copies, new_ck);
772 /* find scheduling point */
773 sched_pt = ref_mode_T;
775 /* just walk along the schedule until a non-Keep/CopyKeep node is found */
776 sched_pt = sched_next(sched_pt);
777 } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
779 sched_add_before(sched_pt, new_ck);
780 DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
782 /* finally: kill the reference copykeep */
787 obstack_free(&obst, NULL);
792 * Walks over all nodes to assure register constraints.
794 * @param birg The birg structure containing the irg
796 void assure_constraints(be_irg_t *birg) {
797 ir_graph *irg = be_get_birg_irg(birg);
798 const arch_env_t *arch_env = be_get_birg_arch_env(birg);
799 constraint_env_t cenv;
800 op_copy_assoc_t *entry;
802 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
804 be_assure_dom_front(birg);
806 DEBUG_ONLY(cenv.dbg = mod;)
808 cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
809 obstack_init(&cenv.obst);
811 irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
813 /* melt copykeeps, pointing to projs of */
814 /* the same mode_T node and keeping the */
816 melt_copykeeps(&cenv);
819 foreach_pset(cenv.op_set, entry) {
822 ir_nodeset_iterator_t iter;
823 be_ssa_construction_env_t senv;
825 n = ir_nodeset_size(&entry->copies);
826 nodes = alloca(n * sizeof(nodes[0]));
828 /* put the node in an array */
829 DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
831 /* collect all copies */
833 foreach_ir_nodeset(&entry->copies, cp, iter) {
835 DB((mod, LEVEL_1, ", %+F ", cp));
838 DB((mod, LEVEL_1, "\n"));
840 /* introduce the copies for the operand and it's copies */
841 be_ssa_construction_init(&senv, birg);
842 be_ssa_construction_add_copy(&senv, entry->op);
843 be_ssa_construction_add_copies(&senv, nodes, n);
844 be_ssa_construction_fix_users(&senv, entry->op);
845 be_ssa_construction_destroy(&senv);
847 /* Could be that not all CopyKeeps are really needed, */
848 /* so we transform unnecessary ones into Keeps. */
849 foreach_ir_nodeset(&entry->copies, cp, iter) {
850 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
852 int n = get_irn_arity(cp);
854 keep = be_new_Keep(arch_get_irn_reg_class(arch_env, cp, -1),
855 irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
856 sched_add_before(cp, keep);
858 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
864 ir_nodeset_destroy(&entry->copies);
867 del_pset(cenv.op_set);
868 obstack_free(&cenv.obst, NULL);
869 be_liveness_invalidate(be_get_birg_liveness(birg));
874 * Push nodes that do not need to be permed through the Perm.
875 * This is commonly a reload cascade at block ends.
876 * @note This routine needs interference.
877 * @note Probably, we can implement it a little more efficient.
878 * Especially searching the frontier lazily might be better.
879 * @param perm The perm.
880 * @param data The walker data (lower_env_t).
881 * @return 1, if there is something left to perm over.
882 * 0, if removed the complete perm.
884 static int push_through_perm(ir_node *perm, void *data)
886 lower_env_t *env = data;
887 const arch_env_t *aenv = env->arch_env;
889 ir_graph *irg = get_irn_irg(perm);
890 ir_node *bl = get_nodes_block(perm);
892 int arity = get_irn_arity(perm);
895 bitset_t *moved = bitset_alloca(arity);
898 ir_node *frontier = bl;
899 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
902 const ir_edge_t *edge;
903 ir_node *one_proj = NULL, *irn;
904 const arch_register_class_t *cls = NULL;
906 DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
908 /* get some Proj and find out the register class of that Proj. */
909 edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
910 one_proj = get_edge_src_irn(edge);
911 assert(is_Proj(one_proj));
912 cls = arch_get_irn_reg_class(aenv, one_proj, -1);
914 /* Find the point in the schedule after which the
915 * potentially movable nodes must be defined.
916 * A Perm will only be pushed up to first instruction
917 * which lets an operand of itself die.
918 * If we would allow to move the Perm above this instruction,
919 * the former dead operand would be live now at the point of
920 * the Perm, increasing the register pressure by one.
922 sched_foreach_reverse_from (sched_prev(perm), irn) {
923 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
924 ir_node *op = get_irn_n(irn, i);
925 if (arch_irn_consider_in_reg_alloc(aenv, cls, op) &&
926 !values_interfere(env->birg, op, one_proj)) {
934 DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
936 node = sched_prev(perm);
938 while(!sched_is_begin(node)) {
939 const arch_register_req_t *req;
943 /* search if node is a INPUT of Perm */
944 foreach_out_edge(perm, edge) {
945 ir_node *out = get_edge_src_irn(edge);
946 int pn = get_Proj_proj(out);
947 ir_node *in = get_irn_n(perm, pn);
954 /* it wasn't an input to the perm, we can't do anything more */
957 if(!sched_comes_after(frontier, node))
959 if(arch_irn_is(aenv, node, modify_flags))
962 req = arch_get_register_req(aenv, get_Proj_pred(node),
963 -1 - get_Proj_proj(node));
965 req = arch_get_register_req(aenv, node, -1);
967 if(req->type != arch_register_req_type_normal)
969 for(i = get_irn_arity(node) - 1; i >= 0; --i) {
970 ir_node *opop = get_irn_n(node, i);
971 if (arch_irn_consider_in_reg_alloc(aenv, cls, opop)) {
978 DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
980 /* move the movable node in front of the Perm */
982 sched_add_after(perm, node);
984 /* give it the proj's register */
985 arch_set_irn_register(aenv, node, arch_get_irn_register(aenv, proj));
987 /* reroute all users of the proj to the moved node. */
988 edges_reroute(proj, node, irg);
991 set_Proj_pred(proj, new_Bad());
994 bitset_set(moved, input);
997 node = sched_prev(node);
1000 /* well, we could not push anything through the perm */
1004 new_size = arity - n_moved;
1009 map = alloca(new_size * sizeof(map[0]));
1010 proj_map = alloca(arity * sizeof(proj_map[0]));
1011 memset(proj_map, -1, sizeof(proj_map[0]));
1013 for(i = 0; i < arity; ++i) {
1014 if(bitset_is_set(moved, i))
1020 assert(n == new_size);
1021 foreach_out_edge(perm, edge) {
1022 ir_node *proj = get_edge_src_irn(edge);
1023 int pn = get_Proj_proj(proj);
1026 set_Proj_proj(proj, pn);
1029 be_Perm_reduce(perm, new_size, map);
1034 * Calls the corresponding lowering function for the node.
1036 * @param irn The node to be checked for lowering
1037 * @param walk_env The walker environment
1039 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
1042 if (is_Block(irn) || is_Proj(irn))
1044 if (!be_is_Perm(irn))
1047 perm_stayed = push_through_perm(irn, walk_env);
1051 lower_perm_node(irn, walk_env);
1055 * Walks over all blocks in an irg and performs lowering need to be
1056 * done after register allocation (e.g. perm lowering).
1058 * @param birg The birg object
1059 * @param do_copy 1 == resolve cycles with a free reg if available
1061 void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
1063 ir_graph *irg = be_get_birg_irg(birg);
1066 env.arch_env = be_get_birg_arch_env(birg);
1067 env.do_copy = do_copy;
1068 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
1070 /* we will need interference */
1071 be_liveness_assure_chk(be_get_birg_liveness(birg));
1073 irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);