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 */
47 /* environment for constraints */
54 /* lowering walker environment */
55 typedef struct _lower_env_t {
56 be_chordal_env_t *chord_env;
58 DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
61 /* holds a perm register pair */
62 typedef struct _reg_pair_t {
63 const arch_register_t *in_reg; /**< a perm IN register */
64 ir_node *in_node; /**< the in node to which the register belongs */
66 const arch_register_t *out_reg; /**< a perm OUT register */
67 ir_node *out_node; /**< the out node to which the register belongs */
69 int checked; /**< indicates whether the pair was check for cycle or not */
72 typedef enum _perm_type_t {
79 /* structure to represent cycles or chains in a perm */
80 typedef struct _perm_cycle_t {
81 const arch_register_t **elems; /**< the registers in the cycle */
82 int n_elems; /**< number of elements in the cycle */
83 perm_type_t type; /**< type (CHAIN or CYCLE) */
86 /* Compare the two operands */
87 static int cmp_op_copy_assoc(const void *a, const void *b) {
88 const op_copy_assoc_t *op1 = a;
89 const op_copy_assoc_t *op2 = b;
91 return op1->op != op2->op;
94 /* Compare the in registers of two register pairs */
95 static int compare_reg_pair(const void *a, const void *b) {
96 const reg_pair_t *pair_a = a;
97 const reg_pair_t *pair_b = b;
99 if (pair_a->in_reg->index > pair_b->in_reg->index)
105 /* returns the number register pairs marked as checked */
106 static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
107 int i, n_checked = 0;
109 for (i = 0; i < n; i++) {
110 if (pairs[i].checked)
118 * Gets the node corresponding to a register from an array of register pairs.
119 * NOTE: The given registers pairs and the register to look for must belong
120 * to the same register class.
122 * @param pairs The array of register pairs
123 * @param n The number of pairs
124 * @param reg The register to look for
125 * @param in_out 0 == look for IN register, 1 == look for OUT register
126 * @return The corresponding node or NULL if not found
128 static ir_node *get_node_for_register(reg_pair_t *pairs, int n, const arch_register_t *reg, int in_out) {
132 for (i = 0; i < n; i++) {
133 /* out register matches */
134 if (pairs[i].out_reg->index == reg->index)
135 return pairs[i].out_node;
139 for (i = 0; i < n; i++) {
140 /* in register matches */
141 if (pairs[i].in_reg->index == reg->index)
142 return pairs[i].in_node;
150 * Gets the index in the register pair array where the in/out register
151 * corresponds to reg_idx.
153 * @param pairs The array of register pairs
154 * @param n The number of pairs
155 * @param reg The register index to look for
156 * @param in_out 0 == look for IN register, 1 == look for OUT register
157 * @return The corresponding index in pairs or -1 if not found
159 static int get_pairidx_for_regidx(reg_pair_t *pairs, int n, int reg_idx, int in_out) {
163 for (i = 0; i < n; i++) {
164 /* out register matches */
165 if (pairs[i].out_reg->index == reg_idx)
170 for (i = 0; i < n; i++) {
171 /* in register matches */
172 if (pairs[i].in_reg->index == reg_idx)
181 * Gets an array of register pairs and tries to identify a cycle or chain starting
184 * @param cycle Variable to hold the cycle
185 * @param pairs Array of register pairs
186 * @param start Index to start
187 * @return The cycle or chain
189 static perm_cycle_t *get_perm_cycle(perm_cycle_t *cycle, reg_pair_t *pairs, int n, int start) {
190 int head = pairs[start].in_reg->index;
191 int cur_idx = pairs[start].out_reg->index;
192 int cur_pair_idx = start;
193 int n_pairs_done = get_n_checked_pairs(pairs, n);
195 perm_type_t cycle_tp = PERM_CYCLE;
197 /* We could be right in the middle of a chain, so we need to find the start */
198 while (head != cur_idx) {
199 /* goto previous register in cycle or chain */
200 cur_pair_idx = get_pairidx_for_regidx(pairs, n, head, 1);
202 if (cur_pair_idx < 0) {
203 cycle_tp = PERM_CHAIN;
207 head = pairs[cur_pair_idx].in_reg->index;
208 start = cur_pair_idx;
212 /* assume worst case: all remaining pairs build a cycle or chain */
213 cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
214 cycle->n_elems = 2; /* initial number of elements is 2 */
215 cycle->elems[0] = pairs[start].in_reg;
216 cycle->elems[1] = pairs[start].out_reg;
217 cycle->type = cycle_tp;
218 cur_idx = pairs[start].out_reg->index;
221 /* check for cycle or end of a chain */
222 while (cur_idx != head) {
223 /* goto next register in cycle or chain */
224 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cur_idx, 0);
226 if (cur_pair_idx < 0)
229 cur_idx = pairs[cur_pair_idx].out_reg->index;
231 /* it's not the first element: insert it */
232 if (cur_idx != head) {
233 cycle->elems[idx++] = pairs[cur_pair_idx].out_reg;
237 /* we are there where we started -> CYCLE */
238 cycle->type = PERM_CYCLE;
242 /* mark all pairs having one in/out register with cycle in common as checked */
243 for (idx = 0; idx < cycle->n_elems; idx++) {
244 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 0);
246 if (cur_pair_idx >= 0)
247 pairs[cur_pair_idx].checked = 1;
249 cur_pair_idx = get_pairidx_for_regidx(pairs, n, cycle->elems[idx]->index, 1);
251 if (cur_pair_idx >= 0)
252 pairs[cur_pair_idx].checked = 1;
259 * Lowers a perm node. Resolves cycles and creates a bunch of
260 * copy and swap operations to permute registers.
261 * Note: The caller of this function has to make sure, that irn
264 * @param irn The perm node
265 * @param block The block the perm node belongs to
266 * @param walk_env The environment
268 static void lower_perm_node(ir_node *irn, void *walk_env) {
269 const arch_register_class_t *reg_class;
270 const arch_env_t *arch_env;
271 lower_env_t *env = walk_env;
273 int n, i, pn, do_copy, j, n_ops;
275 const ir_edge_t *edge;
277 ir_node *sched_point, *block, *in[2];
278 ir_node *arg1, *arg2, *res1, *res2;
279 ir_node *cpyxchg = NULL;
280 DEBUG_ONLY(firm_dbg_module_t *mod;)
282 arch_env = env->chord_env->birg->main_env->arch_env;
283 do_copy = env->do_copy;
284 DEBUG_ONLY(mod = env->dbg_module;)
285 block = get_nodes_block(irn);
288 Get the schedule predecessor node to the perm
289 NOTE: This works with auto-magic. If we insert the
290 new copy/exchange nodes after this node, everything
293 sched_point = sched_prev(irn);
294 DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
295 assert(sched_point && "Perm is not scheduled or has no predecessor");
297 n = get_irn_arity(irn);
298 assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
300 reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
301 pairs = alloca(n * sizeof(pairs[0]));
303 /* build the list of register pairs (in, out) */
305 foreach_out_edge(irn, edge) {
306 pairs[i].out_node = get_edge_src_irn(edge);
307 pn = get_Proj_proj(pairs[i].out_node);
308 pairs[i].in_node = get_irn_n(irn, pn);
310 pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node);
311 pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
313 pairs[i].checked = 0;
317 /* sort the register pairs by the indices of the in registers */
318 qsort(pairs, n, sizeof(pairs[0]), compare_reg_pair);
320 /* Mark all equal pairs as checked, and exchange the OUT proj with
322 for (i = 0; i < n; i++) {
323 if (pairs[i].in_reg->index == pairs[i].out_reg->index) {
324 DBG((mod, LEVEL_1, "%+F removing equal perm register pair (%+F, %+F, %s)\n",
325 irn, pairs[i].in_node, pairs[i].out_node, pairs[i].out_reg->name));
327 /* We have to check for a special case:
328 The in-node could be a Proj from a Perm. In this case,
329 we need to correct the projnum */
330 if (is_Perm(arch_env, pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
331 set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
334 /* remove the proj from the schedule */
335 sched_remove(pairs[i].out_node);
337 /* reroute the edges from the proj to the argument */
338 edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg);
340 pairs[i].checked = 1;
344 /* Set do_copy to 0 if it's on but we have no free register */
349 real_size = n - get_n_checked_pairs(pairs, n);
351 be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
353 /* check for cycles and chains */
354 while (get_n_checked_pairs(pairs, n) < n) {
357 /* go to the first not-checked pair */
358 while (pairs[i].checked) i++;
359 cycle = xcalloc(1, sizeof(*cycle));
360 cycle = get_perm_cycle(cycle, pairs, n, i);
362 DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
363 for (j = 0; j < cycle->n_elems; j++) {
364 DB((mod, LEVEL_1, " %s", cycle->elems[j]->name));
366 DB((mod, LEVEL_1, "\n"));
368 /* We don't need to do anything if we have a Perm with two
369 elements which represents a cycle, because those nodes
370 already represent exchange nodes */
371 if (n == 2 && cycle->type == PERM_CYCLE) {
376 //TODO: - iff PERM_CYCLE && do_copy -> determine free temp reg and insert copy to/from it before/after
377 // the copy cascade (this reduces the cycle into a chain)
379 /* build copy/swap nodes from back to front */
380 for (i = cycle->n_elems - 2; i >= 0; i--) {
381 arg1 = get_node_for_register(pairs, n, cycle->elems[i], 0);
382 arg2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 0);
384 res1 = get_node_for_register(pairs, n, cycle->elems[i], 1);
385 res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1);
387 If we have a cycle and don't copy: we need to create exchange nodes
388 NOTE: An exchange node is a perm node with 2 INs and 2 OUTs
389 IN_1 = in node with register i
390 IN_2 = in node with register i + 1
391 OUT_1 = out node with register i + 1
392 OUT_2 = out node with register i
394 if (cycle->type == PERM_CYCLE && !do_copy) {
398 /* At this point we have to handle the following problem: */
400 /* If we have a cycle with more than two elements, then */
401 /* this could correspond to the following Perm node: */
403 /* +----+ +----+ +----+ */
404 /* | r1 | | r2 | | r3 | */
405 /* +-+--+ +-+--+ +--+-+ */
408 /* +-+--------+---------+-+ */
410 /* +-+--------+---------+-+ */
413 /* +-+--+ +-+--+ +--+-+ */
414 /* |Proj| |Proj| |Proj| */
415 /* | r2 | | r3 | | r1 | */
416 /* +----+ +----+ +----+ */
418 /* This node is about to be split up into two 2x Perm's */
419 /* for which we need 4 Proj's and the one additional Proj */
420 /* of the first Perm has to be one IN of the second. So in */
421 /* general we need to create one additional Proj for each */
422 /* "middle" Perm and set this to one in node of the successor */
425 DBG((mod, LEVEL_1, "%+F creating exchange node (%+F, %s) and (%+F, %s) with\n",
426 irn, arg1, cycle->elems[i]->name, arg2, cycle->elems[i + 1]->name));
427 DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
428 irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
430 cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
434 /* cycle is not done yet */
435 int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
437 /* create intermediate proj */
438 res1 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0);
440 /* set as in for next Perm */
441 pairs[pidx].in_node = res1;
449 set_Proj_pred(res2, cpyxchg);
450 set_Proj_proj(res2, 0);
451 set_Proj_pred(res1, cpyxchg);
452 set_Proj_proj(res1, 1);
454 sched_add_after(sched_point, res1);
455 sched_add_after(sched_point, res2);
457 arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
458 arch_set_irn_register(arch_env, res1, cycle->elems[i]);
460 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
461 sched_add_after(sched_point, cpyxchg);
463 DBG((mod, LEVEL_1, "replacing %+F with %+F, placed new node after %+F\n", irn, cpyxchg, sched_point));
465 /* set the new scheduling point */
469 DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
470 irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
472 cpyxchg = be_new_Copy(reg_class, env->chord_env->irg, block, arg1);
473 arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
476 /* remove the proj from the schedule */
479 /* exchange copy node and proj */
480 exchange(res2, cpyxchg);
482 /* insert the copy/exchange node in schedule after the magic schedule node (see above) */
483 sched_add_after(sched_point, cpyxchg);
485 /* set the new scheduling point */
486 sched_point = cpyxchg;
490 be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
492 free((void *) cycle->elems);
498 /* remove the perm from schedule */
504 static int get_n_out_edges(const ir_node *irn) {
505 const ir_edge_t *edge;
508 foreach_out_edge(irn, edge) {
515 static ir_node *belower_skip_proj(ir_node *irn) {
517 irn = get_Proj_pred(irn);
521 static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
522 be_irg_t *birg = env->birg;
523 pset *op_set = env->op_set;
524 const arch_env_t *arch_env = birg->main_env->arch_env;
525 ir_node *block = get_nodes_block(irn);
526 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
527 ir_node *in[2], *keep, *cpy;
528 op_copy_assoc_t key, *entry;
529 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower");
531 if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
532 DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
536 /* Make a not spillable copy of the different node */
537 /* this is needed because the different irn could be */
538 /* in block far far away */
539 /* The copy is optimized later if not needed */
541 cpy = be_new_Copy(cls, birg->irg, block, other_different);
542 be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
547 /* Add the Keep resp. CopyKeep and reroute the users */
548 /* of the other_different irn in case of CopyKeep. */
549 if (get_n_out_edges(other_different) == 0) {
550 keep = be_new_Keep(cls, birg->irg, block, 2, in);
553 keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different));
554 be_node_set_reg_class(keep, 1, cls);
557 /* let the copy point to the other_different irn */
558 set_irn_n(cpy, 0, other_different);
560 /* insert copy and keep into schedule */
561 assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
562 sched_add_before(belower_skip_proj(irn), cpy);
563 sched_add_after(irn, keep);
565 /* insert the other different and it's copies into the set */
566 key.op = other_different;
568 entry = pset_find(op_set, &key, HASH_PTR(other_different));
571 entry = obstack_alloc(&env->obst, sizeof(*entry));
572 entry->copies = pset_new_ptr_default();
573 entry->op = other_different;
577 pset_insert_ptr(entry->copies, cpy);
579 /* insert keep in case of CopyKeep */
580 if (be_is_CopyKeep(keep))
581 pset_insert_ptr(entry->copies, keep);
583 pset_insert(op_set, entry, HASH_PTR(other_different));
585 DBG((mod, LEVEL_1, "created %+F for %+F to assure should_be_different\n", keep, irn));
589 * Checks if node has a should_be_different constraint in output
590 * and adds a Keep then to assure the constraint.
592 static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
593 const arch_register_req_t *req;
594 arch_register_req_t req_temp;
596 req = arch_get_register_req(env->birg->main_env->arch_env, &req_temp, irn, -1);
599 if (arch_register_req_is(req, should_be_different)) {
600 gen_assure_different_pattern(irn, req->other_different, env);
602 else if (arch_register_req_is(req, should_be_different_from_all)) {
603 int i, n = get_irn_arity(belower_skip_proj(irn));
604 for (i = 0; i < n; i++) {
605 gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env);
614 * Calls the functions to assure register constraints.
616 * @param irn The node to be checked for lowering
617 * @param walk_env The walker environment
619 static void assure_constraints_walker(ir_node *irn, void *walk_env) {
623 if (mode_is_datab(get_irn_mode(irn)))
624 assure_different_constraints(irn, walk_env);
632 * Walks over all nodes to assure register constraints.
634 * @param birg The birg structure containing the irg
636 void assure_constraints(be_irg_t *birg) {
637 constraint_env_t cenv;
638 op_copy_assoc_t *entry;
639 dom_front_info_t *dom;
641 FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower");
644 cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
645 obstack_init(&cenv.obst);
647 irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, &cenv);
649 /* introduce copies needs dominance information */
650 dom = be_compute_dominance_frontiers(birg->irg);
653 foreach_pset(cenv.op_set, entry) {
657 n = pset_count(entry->copies);
658 nodes = alloca((n + 1) * sizeof(nodes[0]));
660 /* put the node in an array */
662 nodes[n++] = entry->op;
663 DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
665 /* collect all copies */
666 foreach_pset(entry->copies, cp) {
668 DB((mod, LEVEL_1, ", %+F ", cp));
671 DB((mod, LEVEL_1, "\n"));
673 /* introduce the copies for the operand and it's copies */
674 be_ssa_constr(dom, NULL, n, nodes);
677 /* Could be that not all CopyKeeps are really needed, */
678 /* so we transform unnecessary ones into Keeps. */
679 foreach_pset(entry->copies, cp) {
680 if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
682 int n = get_irn_arity(cp);
684 keep = be_new_Keep(arch_get_irn_reg_class(birg->main_env->arch_env, cp, -1),
685 birg->irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]);
686 sched_add_before(cp, keep);
689 /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
691 set_irn_n(cp, n, get_irg_bad(birg->irg));
695 del_pset(entry->copies);
698 be_free_dominance_frontiers(dom);
700 del_pset(cenv.op_set);
701 obstack_free(&cenv.obst, NULL);
707 * Calls the corresponding lowering function for the node.
709 * @param irn The node to be checked for lowering
710 * @param walk_env The walker environment
712 static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
713 lower_env_t *env = walk_env;
714 const arch_env_t *arch_env = env->chord_env->birg->main_env->arch_env;
716 if (!is_Block(irn) && !is_Proj(irn)) {
717 if (is_Perm(arch_env, irn)) {
718 lower_perm_node(irn, walk_env);
726 * Walks over all blocks in an irg and performs lowering need to be
727 * done after register allocation (e.g. perm lowering).
729 * @param chord_env The chordal environment containing the irg
730 * @param do_copy 1 == resolve cycles with a free reg if available
732 void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy) {
735 env.chord_env = chord_env;
736 env.do_copy = do_copy;
737 FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
739 irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);