2 * Author: Christian Wuerdig
4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
8 * Performs lowering of perm nodes and spill/reload optimization.
23 #include "bechordal_t.h"
24 #include "besched_t.h"
28 #include "iredges_t.h"
39 #define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
41 /* associates op with it's copy and CopyKeep */
43 ir_node *op; /* an irn which must be different */
44 pset *copies; /* all non-spillable copies of this irn */
45 pset *copy_keeps; /* the CopyKeep's of this irn */
48 /* environment for constraints */
55 /* lowering walker environment */
56 typedef struct _lower_env_t {
57 be_chordal_env_t *chord_env;
59 DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
62 /* holds a perm register pair */
63 typedef struct _reg_pair_t {
64 const arch_register_t *in_reg; /**< a perm IN register */
65 ir_node *in_node; /**< the in node to which the register belongs */
67 const arch_register_t *out_reg; /**< a perm OUT register */
68 ir_node *out_node; /**< the out node to which the register belongs */
70 int checked; /**< indicates whether the pair was check for cycle or not */
73 typedef enum _perm_type_t {
80 /* structure to represent cycles or chains in a perm */
81 typedef struct _perm_cycle_t {
82 const arch_register_t **elems; /**< the registers in the cycle */
83 int n_elems; /**< number of elements in the cycle */
84 perm_type_t type; /**< type (CHAIN or CYCLE) */
87 /* Compare the two operands */
88 static int cmp_op_copy_assoc(const void *a, const void *b) {
89 const op_copy_assoc_t *op1 = a;
90 const op_copy_assoc_t *op2 = b;
92 return op1->op != op2->op;
95 /* Compare the in registers of two register pairs */
96 static int compare_reg_pair(const void *a, const void *b) {
97 const reg_pair_t *pair_a = a;
98 const reg_pair_t *pair_b = b;
100 if (pair_a->in_reg->index > pair_b->in_reg->index)
106 /* returns the number register pairs marked as checked */
107 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
108 int i, n_checked = 0;
110 for (i = 0; i < n; i++) {
111 if (pairs[i].checked)
119 * Gets the node corresponding to a register from an array of register pairs.
120 * NOTE: The given registers pairs and the register to look for must belong
121 * to the same register class.
123 * @param pairs The array of register pairs
124 * @param n The number of pairs
125 * @param reg The register to look for
126 * @param in_out 0 == look for IN register, 1 == look for OUT register
127 * @return The corresponding node or NULL if not found
129 static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
133 for (i = 0; i < n; i++) {
134 /* out register matches */
135 if (pairs[i].out_reg->index == reg->index)
136 return pairs[i].out_node;
140 for (i = 0; i < n; i++) {
141 /* in register matches */
142 if (pairs[i].in_reg->index == reg->index)
143 return pairs[i].in_node;
151 * Gets the index in the register pair array where the in/out register
152 * corresponds to reg_idx.
154 * @param pairs The array of register pairs
155 * @param n The number of pairs
156 * @param reg The register index to look for
157 * @param in_out 0 == look for IN register, 1 == look for OUT register
158 * @return The corresponding index in pairs or -1 if not found
160 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
164 for (i = 0; i < n; i++) {
165 /* out register matches */
166 if (pairs[i].out_reg->index == reg_idx)
171 for (i = 0; i < n; i++) {
172 /* in register matches */
173 if (pairs[i].in_reg->index == reg_idx)
182 * Gets an array of register pairs and tries to identify a cycle or chain starting
185 * @param cycle Variable to hold the cycle
186 * @param pairs Array of register pairs
187 * @param start Index to start
188 * @return The cycle or chain
190 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
191 int head = pairs[start].in_reg->index;
192 int cur_idx = pairs[start].out_reg->index;
193 int cur_pair_idx = start;
194 int n_pairs_done = get_n_checked_pairs(pairs, n);
196 perm_type_t cycle_tp = PERM_CYCLE;
198 /* We could be right in the middle of a chain, so we need to find the start */
199 while (head != cur_idx) {
200 /* goto previous register in cycle or chain */
201 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
203 if (cur_pair_idx < 0) {
204 cycle_tp = PERM_CHAIN;
208 head = pairs[cur_pair_idx].in_reg->index;
209 start = cur_pair_idx;
213 /* assume worst case: all remaining pairs build a cycle or chain */
214 cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
215 cycle->n_elems = 2; /* initial number of elements is 2 */
216 cycle->elems[0] = pairs[start].in_reg;
217 cycle->elems[1] = pairs[start].out_reg;
218 cycle->type = cycle_tp;
219 cur_idx = pairs[start].out_reg->index;
222 /* check for cycle or end of a chain */
223 while (cur_idx != head) {
224 /* goto next register in cycle or chain */
225 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
227 if (cur_pair_idx < 0)
230 cur_idx = pairs[cur_pair_idx].out_reg->index;
232 /* it's not the first element: insert it */
233 if (cur_idx != head) {
234 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
238 /* we are there where we started -> CYCLE */
239 cycle->type = PERM_CYCLE;
243 /* mark all pairs having one in/out register with cycle in common as checked */
244 for (idx = 0; idx < cycle->n_elems; idx++) {
245 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
247 if (cur_pair_idx >= 0)
248 pairs[cur_pair_idx].checked = 1;
250 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
252 if (cur_pair_idx >= 0)
253 pairs[cur_pair_idx].checked = 1;
260 * Lowers a perm node. Resolves cycles and creates a bunch of
261 * copy and swap operations to permute registers.
262 * Note: The caller of this function has to make sure, that irn
265 * @param irn The perm node
266 * @param block The block the perm node belongs to
267 * @param walk_env The environment
269 static void lower_perm_node(ir_node *irn, void *walk_env) {
270 const arch_register_class_t *reg_class;
271 const arch_env_t *arch_env;
272 lower_env_t *env = walk_env;
274 int n, i, pn, do_copy, j, n_ops;
276 const ir_edge_t *edge;
278 ir_node *sched_point, *block, *in[2];
279 ir_node *arg1, *arg2, *res1, *res2;
280 ir_node *cpyxchg = NULL;
281 DEBUG_ONLY(firm_dbg_module_t *mod;)
283 arch_env = env->chord_env->birg->main_env->arch_env;
284 do_copy = env->do_copy;
285 DEBUG_ONLY(mod = env->dbg_module;)
286 block = get_nodes_block(irn);
289 Get the schedule predecessor node to the perm
290 NOTE: This works with auto-magic. If we insert the
291 new copy/exchange nodes after this node, everything
294 sched_point = sched_prev(irn);
295 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
296 assert(sched_point && "Perm is not scheduled or has no predecessor");
298 n = get_irn_arity(irn);
299 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
301 reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
302 pairs = alloca(n * sizeof(pairs[0]));
304 /* build the list of register pairs (in, out) */
306 foreach_out_edge(irn, edge) {
307 pairs[i].out_node = get_edge_src_irn(edge);
308 pn = get_Proj_proj(pairs[i].out_node);
309 pairs[i].in_node = get_irn_n(irn, pn);
311 pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node);
312 pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
314 pairs[i].checked = 0;
318 /* sort the register pairs by the indices of the in registers */
319 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
321 /* Mark all equal pairs as checked, and exchange the OUT proj with
323 for (i = 0; i < n; i++) {
324 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
325 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
326 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
328 /* We have to check for a special case:
329 The in-node could be a Proj from a Perm. In this case,
330 we need to correct the projnum */
331 if (is_Perm(arch_env, pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
332 set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
335 /* remove the proj from the schedule */
336 sched_remove(pairs[i].out_node);
338 /* reroute the edges from the proj to the argument */
339 edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg);
341 pairs[i].checked = 1;
345 /* Set do_copy to 0 if it's on but we have no free register */
350 real_size = n - get_n_checked_pairs(pairs, n);
352 be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
354 /* check for cycles and chains */
355 while (get_n_checked_pairs(pairs, n) < n) {
358 /* go to the first not-checked pair */
359 while (pairs[i].checked) i++;
360 cycle = xcalloc(1, sizeof(*cycle));
361 cycle = get_perm_cycle(cycle, pairs, n, i);
363 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
364 for (j = 0; j < cycle->n_elems; j++) {
365 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
367 DB((mod, LEVEL_1, "\n"));
369 /* We don't need to do anything if we have a Perm with two
370 elements which represents a cycle, because those nodes
371 already represent exchange nodes */
372 if (n == 2 && cycle->type == PERM_CYCLE) {
377 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
378 // the copy cascade (this reduces the cycle into a chain)
380 /* build copy/swap nodes from back to front */
381 for (i = cycle->n_elems - 2; i >= 0; i--) {
382 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
383 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
385 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
386 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
388 If we have a cycle and don't copy: we need to create exchange nodes
389 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
390 IN_1 = in node with register i
391 IN_2 = in node with register i + 1
392 OUT_1 = out node with register i + 1
393 OUT_2 = out node with register i
395 if (cycle->type == PERM_CYCLE && !do_copy) {
399 /* At this point we have to handle the following problem: */
401 /* If we have a cycle with more than two elements, then */
402 /* this could correspond to the following Perm node: */
404 /* +----+ +----+ +----+ */
405 /* | r1 | | r2 | | r3 | */
406 /* +-+--+ +-+--+ +--+-+ */
409 /* +-+--------+---------+-+ */
411 /* +-+--------+---------+-+ */
414 /* +-+--+ +-+--+ +--+-+ */
415 /* |Proj| |Proj| |Proj| */
416 /* | r2 | | r3 | | r1 | */
417 /* +----+ +----+ +----+ */
419 /* This node is about to be split up into two 2x Perm's */
420 /* for which we need 4 Proj's and the one additional Proj */
421 /* of the first Perm has to be one IN of the second. So in */
422 /* general we need to create one additional Proj for each */
423 /* "middle" Perm and set this to one in node of the successor */
426 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
427 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
428 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
429 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
431 cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
435 /* cycle is not done yet */
436 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
438 /* create intermediate proj */
439 res1 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0);
441 /* set as in for next Perm */
442 pairs[pidx].in_node = res1;
450 set_Proj_pred(res2, cpyxchg);
451 set_Proj_proj(res2, 0);
452 set_Proj_pred(res1, cpyxchg);
453 set_Proj_proj(res1, 1);
455 sched_add_after(sched_point, res1);
456 sched_add_after(sched_point, res2);
458 arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
459 arch_set_irn_register(arch_env, res1, cycle->elems[i]);
461 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
462 sched_add_after(sched_point, cpyxchg);
464 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
466 /* set the new scheduling point */
470 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
471 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
473 cpyxchg = be_new_Copy(reg_class, env->chord_env->irg, block, arg1);
474 arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
477 /* remove the proj from the schedule */
480 /* exchange copy node and proj */
481 exchange(res2, cpyxchg);
483 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
484 sched_add_after(sched_point, cpyxchg);
486 /* set the new scheduling point */
487 sched_point = cpyxchg;
491 be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
493 free((void *) cycle->elems);
499 /* remove the perm from schedule */
505 static int get_n_out_edges(const ir_node *irn) {
506 const ir_edge_t *edge;
509 foreach_out_edge(irn, edge) {
516 static ir_node *belower_skip_proj(ir_node *irn) {
518 irn = get_Proj_pred(irn);
522 static void fix_in(ir_node *irn, ir_node *old, ir_node *nw) {
525 irn = belower_skip_proj(irn);
526 n = get_irn_arity(irn);
528 for (i = 0; i < n; i++) {
529 if (get_irn_n(irn, i) == old) {
530 set_irn_n(irn, i, nw);
536 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
537 be_irg_t *birg = env->birg;
538 pset *op_set = env->op_set;
539 const arch_env_t *arch_env = birg->main_env->arch_env;
540 ir_node *block = get_nodes_block(irn);
541 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
542 ir_node *in[2], *keep, *cpy;
543 op_copy_assoc_t key, *entry;
544 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower");
546 if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
547 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
551 /* Make a not spillable copy of the different node */
552 /* this is needed because the different irn could be */
553 /* in block far far away */
554 /* The copy is optimized later if not needed */
556 cpy = be_new_Copy(cls, birg->irg, block, other_different);
557 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
562 /* Add the Keep resp. CopyKeep and reroute the users */
563 /* of the other_different irn in case of CopyKeep. */
564 if (get_n_out_edges(other_different) == 0) {
565 keep = be_new_Keep(cls, birg->irg, block, 2, in);
568 keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different));
569 be_node_set_reg_class(keep, 1, cls);
572 /* let the copy point to the other_different irn */
573 set_irn_n(cpy, 0, other_different);
575 /* insert copy and keep into schedule */
576 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
577 sched_add_before(belower_skip_proj(irn), cpy);
578 sched_add_after(irn, keep);
580 /* insert the other different and it's copies into the set */
581 key.op = other_different;
583 key.copy_keeps = NULL;
584 entry = pset_find(op_set, &key, HASH_PTR(other_different));
587 entry = obstack_alloc(&env->obst, sizeof(*entry));
588 entry->copies = pset_new_ptr_default();
589 entry->copy_keeps = pset_new_ptr_default();
590 entry->op = other_different;
594 pset_insert_ptr(entry->copies, cpy);
596 /* insert keep in case of CopyKeep */
597 if (be_is_CopyKeep(keep))
598 pset_insert_ptr(entry->copy_keeps, keep);
600 pset_insert(op_set, entry, HASH_PTR(other_different));
602 DBG((mod, LEVEL_1, "created %+F for %+F to assure should_be_different\n", keep, irn));
606 * Checks if node has a should_be_different constraint in output
607 * and adds a Keep then to assure the constraint.
609 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
610 const arch_register_req_t *req;
611 arch_register_req_t req_temp;
613 req = arch_get_register_req(env->birg->main_env->arch_env, &req_temp, irn, -1);
616 if (arch_register_req_is(req, should_be_different)) {
617 gen_assure_different_pattern(irn, req->other_different, env);
619 else if (arch_register_req_is(req, should_be_different_from_all)) {
620 int i, n = get_irn_arity(belower_skip_proj(irn));
621 for (i = 0; i < n; i++) {
622 gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env);
631 * Calls the functions to assure register constraints.
633 * @param irn The node to be checked for lowering
634 * @param walk_env The walker environment
636 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
640 if (mode_is_datab(get_irn_mode(irn)))
641 assure_different_constraints(irn, walk_env);
649 * Walks over all nodes to assure register constraints.
651 * @param birg The birg structure containing the irg
653 void assure_constraints(be_irg_t *birg) {
654 constraint_env_t cenv;
655 op_copy_assoc_t *entry;
656 dom_front_info_t *dom;
658 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower");
661 cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
662 obstack_init(&cenv.obst);
664 irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, &cenv);
666 /* introduce copies needs dominance information */
667 dom = be_compute_dominance_frontiers(birg->irg);
670 foreach_pset(cenv.op_set, entry) {
674 n = pset_count(entry->copies);
675 n += pset_count(entry->copy_keeps);
677 nodes = alloca((n + 1) * sizeof(nodes[0]));
679 /* put the node in an array */
681 nodes[n++] = entry->op;
682 DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
684 /* collect all copies */
685 foreach_pset(entry->copies, cp) {
687 DB((mod, LEVEL_1, ", %+F ", cp));
690 /* collect all CopyKeeps */
691 foreach_pset(entry->copy_keeps, cp) {
693 DB((mod, LEVEL_1, ", %+F ", cp));
696 DB((mod, LEVEL_1, "\n"));
698 /* introduce the copies for the operand and it's copies */
699 be_ssa_constr(dom, n, nodes);
702 /* Could be that not all CopyKeeps are really needed, */
703 /* so we transform unnecessary ones into Keeps. */
704 foreach_pset(entry->copy_keeps, cp) {
705 if (get_irn_n_edges(cp) < 1) {
707 int n = get_irn_arity(cp);
709 keep = be_new_Keep(arch_get_irn_reg_class(birg->main_env->arch_env, cp, -1),
710 birg->irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
711 sched_add_before(cp, keep);
714 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
716 set_irn_n(cp, n, get_irg_bad(birg->irg));
720 del_pset(entry->copies);
721 del_pset(entry->copy_keeps);
724 be_free_dominance_frontiers(dom);
726 del_pset(cenv.op_set);
727 obstack_free(&cenv.obst, NULL);
733 * Calls the corresponding lowering function for the node.
735 * @param irn The node to be checked for lowering
736 * @param walk_env The walker environment
738 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
739 lower_env_t *env = walk_env;
740 const arch_env_t *arch_env = env->chord_env->birg->main_env->arch_env;
742 if (!is_Block(irn) && !is_Proj(irn)) {
743 if (is_Perm(arch_env, irn)) {
744 lower_perm_node(irn, walk_env);
752 * Walks over all blocks in an irg and performs lowering need to be
753 * done after register allocation (e.g. perm lowering).
755 * @param chord_env The chordal environment containing the irg
756 * @param do_copy 1 == resolve cycles with a free reg if available
758 void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy) {
761 env.chord_env = chord_env;
762 env.do_copy = do_copy;
763 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
765 irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);