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
33 #include "iroptimize.h"
50 DEBUG_ONLY(firm_dbg_module_t *dbg);
52 /** allow every Psi to be created. */
53 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
65 static const opt_if_conv_info_t default_info = {
66 0, /* doesn't matter for Psi */
71 * Additional block info.
73 typedef struct block_info {
74 ir_node *phi; /**< head of the Phi list */
75 int has_pinned; /**< set if the block contains instructions that cannot be moved */
79 static INLINE block_info* get_block_blockinfo(const ir_node* block)
81 return get_irn_link(block);
86 * Returns non-zero if a Block can be emptied.
88 static int can_empty_block(ir_node *block)
90 return !get_block_blockinfo(block)->has_pinned;
94 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
99 /* No need to find the conditional block if this block cannot be emptied and
100 * therefore not moved */
101 if (!can_empty_block(start)) return NULL;
103 arity = get_irn_arity(start);
104 for (i = 0; i < arity; ++i) {
105 ir_node* pred = get_irn_n(start, i);
106 ir_node* pred_block = get_nodes_block(pred);
108 if (pred_block == dependency) {
110 assert(get_irn_mode(pred) == mode_X);
117 assert(get_irn_mode(pred) == mode_X);
121 if (is_cdep_on(pred_block, dependency)) {
122 return walk_to_projx(pred_block, dependency);
130 * Copies the DAG starting at node to the ith predecessor block of src_block
131 * -if the node isn't in the src_block, this is a nop and the node is returned
132 * -if the node is a phi in the src_block, the ith predecessor of the phi is
134 * otherwise returns the copy of the passed node
136 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
143 if (get_nodes_block(node) != src_block) return node;
144 if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
146 copy = exact_copy(node);
147 dst_block = get_nodes_block(get_irn_n(src_block, i));
148 set_nodes_block(copy, dst_block);
150 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
151 node, dst_block, copy));
153 arity = get_irn_arity(node);
154 for (j = 0; j < arity; ++j) {
155 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
156 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
163 * Remove predecessors i and j from node and add predecessor new_pred
165 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
167 int arity = get_irn_arity(node);
172 NEW_ARR_A(ir_node *, ins, arity - 1);
175 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
176 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
177 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
179 assert(l == arity - 1);
180 set_irn_in(node, l, ins);
185 * Remove the jth predecessors from the ith predecessor of block and add it to block
187 static void split_block(ir_node* block, int i, int j)
189 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
190 int arity = get_irn_arity(block);
197 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
199 NEW_ARR_A(ir_node*, ins, arity + 1);
201 for (phi = get_block_blockinfo(block)->phi; phi != NULL; phi = get_irn_link(phi)) {
202 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
204 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
206 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
207 ins[k] = get_irn_n(phi, i);
209 set_irn_in(phi, arity + 1, ins);
212 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
213 ins[k++] = get_irn_n(pred_block, j);
214 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
215 ins[k] = get_irn_n(block, i);
217 set_irn_in(block, arity + 1, ins);
219 new_pred_arity = get_irn_arity(pred_block) - 1;
220 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
222 for (phi = get_block_blockinfo(pred_block)->phi; phi != NULL; phi = get_irn_link(phi)) {
223 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
224 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
225 assert(k == new_pred_arity);
226 if (new_pred_arity > 1) {
227 set_irn_in(phi, new_pred_arity, pred_ins);
229 exchange(phi, pred_ins[0]);
233 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
234 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
235 assert(k == new_pred_arity);
236 if (new_pred_arity > 1) {
237 set_irn_in(pred_block, new_pred_arity, pred_ins);
239 exchange(pred_block, get_nodes_block(pred_ins[0]));
244 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
246 ir_node* pred = get_nodes_block(get_irn_n(block, i));
250 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
252 pred_arity = get_irn_arity(pred);
253 for (j = 0; j < pred_arity; ++j) {
254 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
256 if (is_cdep_on(pred_pred, dependency)) {
257 prepare_path(pred, j, dependency);
258 split_block(block, i, j);
265 static void if_conv_walker(ir_node* block, void* env)
267 opt_if_conv_info_t* opt_info = env;
271 /* Bail out, if there are no Phis at all */
272 if (get_block_blockinfo(block)->phi == NULL) return;
275 arity = get_irn_arity(block);
276 for (i = 0; i < arity; ++i) {
280 pred0 = get_nodes_block(get_irn_n(block, i));
281 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
282 const ir_node* dependency = cdep->node;
283 ir_node* projx0 = walk_to_projx(pred0, dependency);
287 if (projx0 == NULL) continue;
289 cond = get_Proj_pred(projx0);
290 if (get_irn_op(cond) != op_Cond) continue;
292 /* We only handle boolean decisions, no switches */
293 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
295 for (j = i + 1; j < arity; ++j) {
303 pred1 = get_nodes_block(get_irn_n(block, j));
305 if (!is_cdep_on(pred1, dependency)) continue;
307 projx1 = walk_to_projx(pred1, dependency);
309 if (projx1 == NULL) continue;
311 phi = get_block_blockinfo(block)->phi;
312 if (!opt_info->allow_ifconv(get_Cond_selector(cond), phi, i, j)) continue;
314 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
318 prepare_path(block, i, dependency);
319 prepare_path(block, j, dependency);
320 arity = get_irn_arity(block);
322 conds[0] = get_Cond_selector(cond);
324 psi_block = get_nodes_block(cond);
325 cond_dbg = get_irn_dbg_info(cond);
327 ir_node* val_i = get_irn_n(phi, i);
328 ir_node* val_j = get_irn_n(phi, j);
332 if (val_i == val_j) {
334 DB((dbg, LEVEL_2, "Generating no psi, because both values are equal\n"));
338 /* Something is very fishy if two predecessors of a PhiM point into
339 * one block, but not at the same memory node
341 assert(get_irn_mode(phi) != mode_M);
342 if (get_Proj_proj(projx0) == pn_Cond_true) {
350 psi = new_rd_Psi(cond_dbg, current_ir_graph, psi_block, 1, conds, vals, get_irn_mode(phi));
351 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", psi, phi));
354 next_phi = get_irn_link(phi);
359 rewire(phi, i, j, psi);
363 } while (phi != NULL);
365 exchange(get_nodes_block(get_irn_n(block, i)), psi_block);
366 exchange(get_nodes_block(get_irn_n(block, j)), psi_block);
370 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, psi_block));
371 /* copy the block-info from the Psi-block to the block before merging */
372 get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
373 set_irn_link(block, get_irn_link(psi_block));
375 set_irn_in(block, get_irn_arity(psi_block), get_irn_in(psi_block) + 1);
376 exchange_cdep(psi_block, block);
377 exchange(psi_block, block);
379 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, psi_block));
380 get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
381 exchange(block, psi_block);
385 rewire(block, i, j, new_r_Jmp(current_ir_graph, psi_block));
394 * Block walker: add additional data
396 static void init_block_link(ir_node *block, void *env)
398 struct obstack *obst = env;
399 block_info *bi = obstack_alloc(obst, sizeof(*bi));
403 set_irn_link(block, bi);
408 * Daisy-chain all phis in a block
409 * If a non-movable node is encountered set the has_pinned flag
411 static void collect_phis(ir_node *node, void *env)
416 ir_node *block = get_nodes_block(node);
417 block_info *bi = get_block_blockinfo(block);
419 set_irn_link(node, bi->phi);
422 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
424 * Ignore control flow nodes, these will be removed.
425 * This ignores Raise. That is surely bad. FIXME.
427 if (!is_cfop(node)) {
428 ir_node *block = get_nodes_block(node);
429 block_info *bi = get_block_blockinfo(block);
431 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
440 * Transform multiple cascaded Psis into one Psi
442 static ir_node* fold_psi(ir_node* psi)
444 int arity = get_Psi_n_conds(psi);
453 for (i = 0; i < arity; ++i) {
454 n = get_Psi_val(psi, i);
455 if (get_irn_op(n) == op_Psi) {
456 new_arity += get_Psi_n_conds(n) + 1;
461 n = get_Psi_default(psi);
462 if (get_irn_op(n) == op_Psi) {
463 new_arity += get_Psi_n_conds(n);
466 if (arity == new_arity) return psi; // no attached Psis found
467 DB((dbg, LEVEL_1, "Folding %+F from %d to %d conds\n", psi, arity, new_arity));
469 NEW_ARR_A(ir_node *, conds, new_arity);
470 NEW_ARR_A(ir_node *, vals, new_arity + 1);
472 for (i = 0; i < arity; ++i) {
473 ir_node* c = get_Psi_cond(psi, i);
475 n = get_Psi_val(psi, i);
476 if (get_irn_op(n) == op_Psi) {
477 int a = get_Psi_n_conds(n);
480 for (k = 0; k < a; ++k) {
481 conds[j] = new_r_And(
482 current_ir_graph, get_nodes_block(psi),
483 c, get_Psi_cond(n, k), mode_b
485 vals[j] = get_Psi_val(n, k);
489 vals[j] = get_Psi_default(n);
496 n = get_Psi_default(psi);
497 if (get_irn_op(n) == op_Psi) {
498 int a = get_Psi_n_conds(n);
501 for (k = 0; k < a; ++k) {
502 conds[j] = get_Psi_cond(n, k);
503 vals[j] = get_Psi_val(n, k);
506 vals[j] = get_Psi_default(n);
510 assert(j == new_arity);
512 current_ir_graph, get_nodes_block(psi),
513 new_arity, conds, vals, get_irn_mode(psi)
515 DB((dbg, LEVEL_1, "Folded %+F into new %+F\n", psi, new_psi));
516 exchange(psi, new_psi);
522 * Merge consecutive psi inputs if the data inputs are the same
524 static ir_node* meld_psi(ir_node* psi)
526 int arity = get_Psi_n_conds(psi);
537 val = get_Psi_val(psi, 0);
538 DB((dbg, LEVEL_1, "Pred 0 of %+F is %+F\n", psi, val));
539 for (i = 1; i < arity; ++i) {
540 ir_node* v = get_Psi_val(psi, i);
541 DB((dbg, LEVEL_1, "Pred %2d of %+F is %+F\n", i, psi, v));
547 DB((dbg, LEVEL_1, "Default of %+F is %+F\n", psi, get_Psi_default(psi)));
548 if (val == get_Psi_default(psi)) --new_arity;
550 DB((dbg, LEVEL_1, "Melding Psi %+F from %d conds to %d\n", psi, arity, new_arity));
552 if (new_arity == arity) return psi;
554 /* If all data inputs of the Psi are equal, exchange the Psi with that value */
555 if (new_arity == 0) {
560 NEW_ARR_A(ir_node *, conds, new_arity);
561 NEW_ARR_A(ir_node *, vals, new_arity + 1);
562 cond = get_Psi_cond(psi, 0);
563 val = get_Psi_val(psi, 0);
565 for (i = 1; i < arity; ++i) {
566 ir_node* v = get_Psi_val(psi, i);
570 current_ir_graph, get_nodes_block(psi),
571 cond, get_Psi_cond(psi, i), mode_b
577 cond = get_Psi_cond(psi, i)
581 if (val != get_Psi_default(psi)) {
586 vals[j] = get_Psi_default(psi);
587 assert(j == new_arity);
589 current_ir_graph, get_nodes_block(psi),
590 new_arity, conds, vals, get_irn_mode(psi)
592 DB((dbg, LEVEL_1, "Molded %+F into %+F\n", psi, new_psi));
593 exchange(psi, new_psi);
599 * Split a Psi with multiple conditions into multiple Psis with one condtition
602 static ir_node* split_psi(ir_node* psi)
604 int arity = get_Psi_n_conds(psi);
610 if (arity == 1) return psi;
612 mode = get_irn_mode(psi);
613 block = get_nodes_block(psi);
614 rval = get_Psi_default(psi);
615 for (i = arity - 1; i >= 0; --i) {
619 conds[0] = get_Psi_cond(psi, i);
620 vals[0] = get_Psi_val(psi, i);
623 current_ir_graph, block, 1, conds, vals, mode
631 static void optimise_psis(ir_node* node, void* env)
635 if (get_irn_op(node) != op_Psi) return;
637 node = fold_psi(node);
640 node = meld_psi(node);
643 node = split_psi(node);
648 void opt_if_conv(ir_graph *irg, const opt_if_conv_info_t *params)
651 opt_if_conv_info_t p;
653 if (! get_opt_if_conversion())
656 /* get the parameters */
657 p = (params != NULL ? *params : default_info);
659 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
661 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
663 normalize_one_return(irg);
664 remove_critical_cf_edges(irg);
670 irg_block_walk_graph(irg, init_block_link, NULL, &obst);
671 irg_walk_graph(irg, collect_phis, NULL, NULL);
672 irg_block_walk_graph(irg, NULL, if_conv_walker, &p);
674 local_optimize_graph(irg);
676 irg_walk_graph(irg, NULL, optimise_psis, NULL);
678 obstack_free(&obst, NULL);