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
21 * @file ir/opt/ifconv.c
22 * @brief If conversion
23 * @author Christoph Mallon
51 DEBUG_ONLY(firm_dbg_module_t *dbg);
53 /** allow every Psi to be created. */
54 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
62 static const opt_if_conv_info_t default_info = {
63 0, /* doesn't matter for Psi */
68 * Additional block info.
70 typedef struct block_info {
71 ir_node *phi; /**< head of the Phi list */
72 int has_pinned; /**< set if the block contains instructions that cannot be moved */
75 #define get_block_blockinfo(block) ((block_info *)get_irn_link(block))
78 * Returns non-zero if a Block can be emptied.
80 static int can_empty_block(ir_node *block)
82 return !get_block_blockinfo(block)->has_pinned;
86 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
91 /* No need to find the conditional block if this block cannot be emptied and
94 if (!can_empty_block(start)) return NULL;
96 arity = get_irn_arity(start);
97 for (i = 0; i < arity; ++i) {
98 ir_node* pred = get_irn_n(start, i);
99 ir_node* pred_block = get_nodes_block(pred);
101 if (pred_block == dependency) {
103 assert(get_irn_mode(pred) == mode_X);
110 assert(get_irn_mode(pred) == mode_X);
114 if (is_cdep_on(pred_block, dependency)) {
115 return walk_to_projx(pred_block, dependency);
123 * Copies the DAG starting at node to the ith predecessor block of src_block
124 * -if the node isn't in the src_block, this is a nop and the node is returned
125 * -if the node is a phi in the src_block, the ith predecessor of the phi is
127 * otherwise returns the copy of the passed node
129 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
136 if (get_nodes_block(node) != src_block) return node;
137 if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
139 copy = exact_copy(node);
140 dst_block = get_nodes_block(get_irn_n(src_block, i));
141 set_nodes_block(copy, dst_block);
143 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
144 node, dst_block, copy));
146 arity = get_irn_arity(node);
147 for (j = 0; j < arity; ++j) {
148 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
149 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
156 * Remove predecessors i and j from node and add predecessor new_pred
158 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
160 int arity = get_irn_arity(node);
165 NEW_ARR_A(ir_node *, ins, arity - 1);
168 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
169 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
170 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
172 assert(l == arity - 1);
173 set_irn_in(node, l, ins);
178 * Remove the jth predecessors from the ith predecessor of block and add it to block
180 static void split_block(ir_node* block, int i, int j)
182 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
183 int arity = get_irn_arity(block);
190 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
192 NEW_ARR_A(ir_node*, ins, arity + 1);
194 for (phi = get_block_blockinfo(block)->phi; phi != NULL; phi = get_irn_link(phi)) {
195 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
197 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
199 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
200 ins[k] = get_irn_n(phi, i);
202 set_irn_in(phi, arity + 1, ins);
205 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
206 ins[k++] = get_irn_n(pred_block, j);
207 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
208 ins[k] = get_irn_n(block, i);
210 set_irn_in(block, arity + 1, ins);
212 new_pred_arity = get_irn_arity(pred_block) - 1;
213 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
215 for (phi = get_block_blockinfo(pred_block)->phi; phi != NULL; phi = get_irn_link(phi)) {
216 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
217 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
218 assert(k == new_pred_arity);
219 if (new_pred_arity > 1) {
220 set_irn_in(phi, new_pred_arity, pred_ins);
222 exchange(phi, pred_ins[0]);
226 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
227 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
228 assert(k == new_pred_arity);
229 if (new_pred_arity > 1) {
230 set_irn_in(pred_block, new_pred_arity, pred_ins);
232 exchange(pred_block, get_nodes_block(pred_ins[0]));
237 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
239 ir_node* pred = get_nodes_block(get_irn_n(block, i));
243 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
245 pred_arity = get_irn_arity(pred);
246 for (j = 0; j < pred_arity; ++j) {
247 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
249 if (is_cdep_on(pred_pred, dependency)) {
250 prepare_path(pred, j, dependency);
251 split_block(block, i, j);
258 static void if_conv_walker(ir_node* block, void* env)
262 opt_if_conv_info_t *opt_info = env;
264 /* Bail out, if there are no Phis at all */
265 if (get_block_blockinfo(block)->phi == NULL) return;
268 arity = get_irn_arity(block);
269 for (i = 0; i < arity; ++i) {
273 pred = get_nodes_block(get_irn_n(block, i));
274 for (cdep = find_cdep(pred); cdep != NULL; cdep = cdep->next) {
275 const ir_node* dependency = cdep->node;
276 ir_node* projx0 = walk_to_projx(pred, dependency);
280 if (projx0 == NULL) continue;
282 cond = get_Proj_pred(projx0);
283 if (get_irn_op(cond) != op_Cond) continue;
285 /* We only handle boolean decisions, no switches */
286 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
288 for (j = i + 1; j < arity; ++j) {
296 pred = get_nodes_block(get_irn_n(block, j));
298 if (!is_cdep_on(pred, dependency)) continue;
300 projx1 = walk_to_projx(pred, dependency);
302 if (projx1 == NULL) continue;
304 phi = get_block_blockinfo(block)->phi;
305 if (!opt_info->allow_ifconv(get_Cond_selector(cond), phi, i, j)) continue;
307 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
311 prepare_path(block, i, dependency);
312 prepare_path(block, j, dependency);
313 arity = get_irn_arity(block);
315 conds[0] = get_Cond_selector(cond);
317 psi_block = get_nodes_block(cond);
319 ir_node* val_i = get_irn_n(phi, i);
320 ir_node* val_j = get_irn_n(phi, j);
322 if (val_i == val_j) {
324 DB((dbg, LEVEL_2, "Generating no psi, because both values are equal\n"));
326 /* Something is very fishy if two predecessors of a PhiM point into
327 * one block, but not at the same memory node
329 assert(get_irn_mode(phi) != mode_M);
330 if (get_Proj_proj(projx0) == pn_Cond_true) {
339 current_ir_graph, psi_block, 1, conds, vals, get_irn_mode(phi)
341 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", psi, phi));
347 rewire(phi, i, j, psi);
350 phi = get_irn_link(phi);
351 } while (phi != NULL);
353 exchange(get_nodes_block(get_irn_n(block, i)), psi_block);
354 exchange(get_nodes_block(get_irn_n(block, j)), psi_block);
358 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, psi_block));
359 /* copy the block-info from the Psi-block to the block before merging */
360 get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
361 set_irn_link(block, get_irn_link(psi_block));
363 set_irn_in(block, get_irn_arity(psi_block), get_irn_in(psi_block) + 1);
364 exchange_cdep(psi_block, block);
365 exchange(psi_block, block);
367 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, psi_block));
368 get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
369 exchange(block, psi_block);
373 rewire(block, i, j, new_r_Jmp(current_ir_graph, psi_block));
382 * Block walker: add additional data
384 static void init_block_link(ir_node *block, void *env)
386 struct obstack *obst = env;
387 block_info *bi = obstack_alloc(obst, sizeof(*bi));
391 set_irn_link(block, bi);
396 * Daisy-chain all phis in a block
397 * If a non-movable node is encountered set the has_pinned flag
399 static void collect_phis(ir_node *node, void *env)
402 ir_node *block = get_nodes_block(node);
403 block_info *bi = get_block_blockinfo(block);
405 set_irn_link(node, bi->phi);
409 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
411 * Ignore control flow nodes, these will be removed.
412 * This ignores Raise. That is surely bad. FIXME.
414 if (! is_cfop(node)) {
415 ir_node *block = get_nodes_block(node);
416 block_info *bi = get_block_blockinfo(block);
418 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
427 * Transform multiple cascaded Psis into one Psi
429 static ir_node* fold_psi(ir_node* psi)
431 int arity = get_Psi_n_conds(psi);
442 for (i = 0; i < arity; ++i) {
443 n = get_Psi_val(psi, i);
444 if (get_irn_op(n) == op_Psi) {
445 new_arity += get_Psi_n_conds(n) + 1;
450 n = get_Psi_default(psi);
451 if (get_irn_op(n) == op_Psi) {
452 new_arity += get_Psi_n_conds(n);
455 if (arity == new_arity) return psi; // no attached Psis found
456 DB((dbg, LEVEL_1, "Folding %+F from %d to %d conds\n", psi, arity, new_arity));
458 NEW_ARR_A(ir_node *, conds, new_arity);
459 NEW_ARR_A(ir_node *, vals, new_arity + 1);
461 for (i = 0; i < arity; ++i) {
462 ir_node* c = get_Psi_cond(psi, i);
464 n = get_Psi_val(psi, i);
465 if (get_irn_op(n) == op_Psi) {
466 a = get_Psi_n_conds(n);
467 for (k = 0; k < a; ++k) {
468 conds[j] = new_r_And(
469 current_ir_graph, get_nodes_block(psi),
470 c, get_Psi_cond(n, k), mode_b
472 vals[j] = get_Psi_val(n, k);
476 vals[j] = get_Psi_default(n);
483 n = get_Psi_default(psi);
484 if (get_irn_op(n) == op_Psi) {
485 a = get_Psi_n_conds(n);
486 for (k = 0; k < a; ++k) {
487 conds[j] = get_Psi_cond(n, k);
488 vals[j] = get_Psi_val(n, k);
491 vals[j] = get_Psi_default(n);
495 assert(j == new_arity);
497 current_ir_graph, get_nodes_block(psi),
498 new_arity, conds, vals, get_irn_mode(psi)
500 DB((dbg, LEVEL_1, "Folded %+F into new %+F\n", psi, new_psi));
501 exchange(psi, new_psi);
507 * Merge consecutive psi inputs if the data inputs are the same
509 static ir_node* meld_psi(ir_node* psi)
511 int arity = get_Psi_n_conds(psi);
522 val = get_Psi_val(psi, 0);
523 DB((dbg, LEVEL_1, "Pred 0 of %+F is %+F\n", psi, val));
524 for (i = 1; i < arity; ++i) {
525 ir_node* v = get_Psi_val(psi, i);
526 DB((dbg, LEVEL_1, "Pred %2d of %+F is %+F\n", i, psi, v));
532 DB((dbg, LEVEL_1, "Default of %+F is %+F\n", psi, get_Psi_default(psi)));
533 if (val == get_Psi_default(psi)) --new_arity;
535 DB((dbg, LEVEL_1, "Melding Psi %+F from %d conds to %d\n", psi, arity, new_arity));
537 if (new_arity == arity) return psi;
539 /* If all data inputs of the Psi are equal, exchange the Psi with that value */
540 if (new_arity == 0) {
545 NEW_ARR_A(ir_node *, conds, new_arity);
546 NEW_ARR_A(ir_node *, vals, new_arity + 1);
547 cond = get_Psi_cond(psi, 0);
548 val = get_Psi_val(psi, 0);
550 for (i = 1; i < arity; ++i) {
551 ir_node* v = get_Psi_val(psi, i);
555 current_ir_graph, get_nodes_block(psi),
556 cond, get_Psi_cond(psi, i), mode_b
565 if (val != get_Psi_default(psi)) {
570 vals[j] = get_Psi_default(psi);
571 assert(j == new_arity);
573 current_ir_graph, get_nodes_block(psi),
574 new_arity, conds, vals, get_irn_mode(psi)
576 DB((dbg, LEVEL_1, "Molded %+F into %+F\n", psi, new_psi));
577 exchange(psi, new_psi);
583 * Split a Psi with multiple conditions into multiple Psis with one condtition
586 static ir_node* split_psi(ir_node* psi)
588 int arity = get_Psi_n_conds(psi);
594 if (arity == 1) return psi;
596 mode = get_irn_mode(psi);
597 block = get_nodes_block(psi);
598 rval = get_Psi_default(psi);
599 for (i = arity - 1; i >= 0; --i) {
603 conds[0] = get_Psi_cond(psi, i);
604 vals[0] = get_Psi_val(psi, i);
607 current_ir_graph, block, 1, conds, vals, mode
615 static void optimise_psis(ir_node* node, void* env)
617 if (get_irn_op(node) != op_Psi) return;
619 node = fold_psi(node);
622 node = meld_psi(node);
625 node = split_psi(node);
630 void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params)
633 opt_if_conv_info_t p;
635 if (! get_opt_if_conversion())
638 /* get the parameters */
639 p = (params != NULL ? *params : default_info);
641 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
643 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
645 normalize_one_return(irg);
646 remove_critical_cf_edges(irg);
652 irg_block_walk_graph(irg, init_block_link, NULL, &obst);
653 irg_walk_graph(irg, collect_phis, NULL, NULL);
654 irg_block_walk_graph(irg, NULL, if_conv_walker, &p);
656 local_optimize_graph(irg);
658 irg_walk_graph(irg, NULL, optimise_psis, NULL);
660 obstack_free(&obst, NULL);