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
21 * @file ir/opt/ifconv.c
22 * @brief If conversion
23 * @author Christoph Mallon
32 #include "iroptimize.h"
49 DEBUG_ONLY(static firm_dbg_module_t *dbg);
51 /** allow every Psi to be created. */
52 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
64 static const ir_settings_if_conv_t default_info = {
65 0, /* doesn't matter for Psi */
70 * Additional block info.
72 typedef struct block_info {
73 ir_node *phi; /**< head of the Phi list */
74 int has_pinned; /**< set if the block contains instructions that cannot be moved */
77 /** Returns the additional block info of a block. */
78 static INLINE block_info* get_block_blockinfo(const ir_node* block) {
79 return get_irn_link(block);
84 * Returns non-zero if a Block can be emptied.
86 static int can_empty_block(ir_node *block) {
87 return !get_block_blockinfo(block)->has_pinned;
91 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
96 /* No need to find the conditional block if this block cannot be emptied and
97 * therefore not moved */
98 if (!can_empty_block(start)) return NULL;
100 arity = get_irn_arity(start);
101 for (i = 0; i < arity; ++i) {
102 ir_node* pred = get_irn_n(start, i);
103 ir_node* pred_block = get_nodes_block(pred);
105 if (pred_block == dependency) {
107 assert(get_irn_mode(pred) == mode_X);
114 assert(get_irn_mode(pred) == mode_X);
118 if (is_cdep_on(pred_block, dependency)) {
119 return walk_to_projx(pred_block, dependency);
127 * Copies the DAG starting at node to the ith predecessor block of src_block
128 * -if the node isn't in the src_block, this is a nop and the node is returned
129 * -if the node is a phi in the src_block, the ith predecessor of the phi is
131 * otherwise returns the copy of the passed node
133 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
140 if (get_nodes_block(node) != src_block) return node;
141 if (get_irn_op(node) == op_Phi) return get_irn_n(node, i);
143 copy = exact_copy(node);
144 dst_block = get_nodes_block(get_irn_n(src_block, i));
145 set_nodes_block(copy, dst_block);
147 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
148 node, dst_block, copy));
150 arity = get_irn_arity(node);
151 for (j = 0; j < arity; ++j) {
152 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
153 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
160 * Remove predecessors i and j from node and add predecessor new_pred
162 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
164 int arity = get_irn_arity(node);
169 NEW_ARR_A(ir_node *, ins, arity - 1);
172 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
173 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
174 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
176 assert(l == arity - 1);
177 set_irn_in(node, l, ins);
182 * Remove the jth predecessors from the ith predecessor of block and add it to block
184 static void split_block(ir_node* block, int i, int j)
186 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
187 int arity = get_irn_arity(block);
194 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
196 NEW_ARR_A(ir_node*, ins, arity + 1);
198 for (phi = get_block_blockinfo(block)->phi; phi != NULL; phi = get_irn_link(phi)) {
199 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
201 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
203 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
204 ins[k] = get_irn_n(phi, i);
206 set_irn_in(phi, arity + 1, ins);
209 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
210 ins[k++] = get_irn_n(pred_block, j);
211 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
212 ins[k] = get_irn_n(block, i);
214 set_irn_in(block, arity + 1, ins);
216 new_pred_arity = get_irn_arity(pred_block) - 1;
217 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
219 for (phi = get_block_blockinfo(pred_block)->phi; phi != NULL; phi = get_irn_link(phi)) {
220 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
221 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
222 assert(k == new_pred_arity);
223 if (new_pred_arity > 1) {
224 set_irn_in(phi, new_pred_arity, pred_ins);
226 exchange(phi, pred_ins[0]);
230 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
231 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
232 assert(k == new_pred_arity);
233 if (new_pred_arity > 1) {
234 set_irn_in(pred_block, new_pred_arity, pred_ins);
236 exchange(pred_block, get_nodes_block(pred_ins[0]));
241 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
243 ir_node* pred = get_nodes_block(get_irn_n(block, i));
247 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
249 pred_arity = get_irn_arity(pred);
250 for (j = 0; j < pred_arity; ++j) {
251 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
253 if (is_cdep_on(pred_pred, dependency)) {
254 prepare_path(pred, j, dependency);
255 split_block(block, i, j);
262 static void if_conv_walker(ir_node* block, void* env)
264 ir_settings_if_conv_t* opt_info = env;
268 /* Bail out, if there are no Phis at all */
269 if (get_block_blockinfo(block)->phi == NULL) return;
272 arity = get_irn_arity(block);
273 for (i = 0; i < arity; ++i) {
277 pred0 = get_nodes_block(get_irn_n(block, i));
278 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
279 const ir_node* dependency = cdep->node;
280 ir_node* projx0 = walk_to_projx(pred0, dependency);
284 if (projx0 == NULL) continue;
286 cond = get_Proj_pred(projx0);
287 if (get_irn_op(cond) != op_Cond) continue;
289 /* We only handle boolean decisions, no switches */
290 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
292 for (j = i + 1; j < arity; ++j) {
300 pred1 = get_nodes_block(get_irn_n(block, j));
302 if (!is_cdep_on(pred1, dependency)) continue;
304 projx1 = walk_to_projx(pred1, dependency);
306 if (projx1 == NULL) continue;
308 phi = get_block_blockinfo(block)->phi;
309 if (!opt_info->allow_ifconv(get_Cond_selector(cond), phi, i, j)) continue;
311 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
315 prepare_path(block, i, dependency);
316 prepare_path(block, j, dependency);
317 arity = get_irn_arity(block);
319 conds[0] = get_Cond_selector(cond);
321 psi_block = get_nodes_block(cond);
322 cond_dbg = get_irn_dbg_info(cond);
324 ir_node* val_i = get_irn_n(phi, i);
325 ir_node* val_j = get_irn_n(phi, j);
329 if (val_i == val_j) {
331 DB((dbg, LEVEL_2, "Generating no psi, because both values are equal\n"));
335 /* Something is very fishy if two predecessors of a PhiM point into
336 * one block, but not at the same memory node
338 assert(get_irn_mode(phi) != mode_M);
339 if (get_Proj_proj(projx0) == pn_Cond_true) {
347 psi = new_rd_Psi(cond_dbg, current_ir_graph, psi_block, 1, conds, vals, get_irn_mode(phi));
348 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", psi, phi));
351 next_phi = get_irn_link(phi);
356 rewire(phi, i, j, psi);
360 } while (phi != NULL);
362 exchange(get_nodes_block(get_irn_n(block, i)), psi_block);
363 exchange(get_nodes_block(get_irn_n(block, j)), psi_block);
367 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, psi_block));
368 /* copy the block-info from the Psi-block to the block before merging */
369 get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
370 set_irn_link(block, get_irn_link(psi_block));
372 set_irn_in(block, get_irn_arity(psi_block), get_irn_in(psi_block) + 1);
373 exchange_cdep(psi_block, block);
374 exchange(psi_block, block);
376 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, psi_block));
377 get_block_blockinfo(psi_block)->has_pinned |= get_block_blockinfo(block)->has_pinned;
378 exchange(block, psi_block);
382 rewire(block, i, j, new_r_Jmp(current_ir_graph, psi_block));
391 * Block walker: add additional data
393 static void init_block_link(ir_node *block, void *env)
395 struct obstack *obst = env;
396 block_info *bi = obstack_alloc(obst, sizeof(*bi));
400 set_irn_link(block, bi);
405 * Daisy-chain all phis in a block
406 * If a non-movable node is encountered set the has_pinned flag in its block.
408 static void collect_phis(ir_node *node, void *env) {
412 ir_node *block = get_nodes_block(node);
413 block_info *bi = get_block_blockinfo(block);
415 set_irn_link(node, bi->phi);
418 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
420 * Ignore control flow nodes, these will be removed.
421 * This ignores Raise. That is surely bad. FIXME.
423 if (!is_cfop(node)) {
424 ir_node *block = get_nodes_block(node);
425 block_info *bi = get_block_blockinfo(block);
427 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
434 static void optimise_psis_0(ir_node* psi, void* env)
441 if (!is_Psi(psi)) return;
443 t = get_Psi_val(psi, 0);
444 f = get_Psi_default(psi);
446 DB((dbg, LEVEL_3, "Simplify %+F T=%+F F=%+F\n", psi, t, f));
449 DB((dbg, LEVEL_3, "Replace Psi with unknown operand by %+F\n", f));
454 DB((dbg, LEVEL_3, "Replace Psi with unknown operand by %+F\n", t));
460 ir_graph* irg = current_ir_graph;
461 ir_node* block = get_nodes_block(psi);
462 ir_mode* mode = get_irn_mode(psi);
463 ir_node* c0 = get_Psi_cond(psi, 0);
464 ir_node* c1 = get_Psi_cond(t, 0);
465 ir_node* t1 = get_Psi_val(t, 0);
466 ir_node* f1 = get_Psi_default(t);
468 /* Psi(c0, Psi(c1, x, y), y) -> typical if (c0 && c1) x else y */
469 ir_node* and_ = new_r_And(irg, block, c0, c1, mode_b);
470 ir_node* vals[2] = { t1, f1 };
471 ir_node* new_psi = new_r_Psi(irg, block, 1, &and_, vals, mode);
472 exchange(psi, new_psi);
473 } else if (f == t1) {
474 /* Psi(c0, Psi(c1, x, y), x) */
475 ir_node* not_c1 = new_r_Not(irg, block, c1, mode_b);
476 ir_node* and_ = new_r_And(irg, block, c0, not_c1, mode_b);
477 ir_node* vals[2] = { f1, t1 };
478 ir_node* new_psi = new_r_Psi(irg, block, 1, &and_, vals, mode);
479 exchange(psi, new_psi);
481 } else if (is_Psi(f)) {
482 ir_graph* irg = current_ir_graph;
483 ir_node* block = get_nodes_block(psi);
484 ir_mode* mode = get_irn_mode(psi);
485 ir_node* c0 = get_Psi_cond(psi, 0);
486 ir_node* c1 = get_Psi_cond(f, 0);
487 ir_node* t1 = get_Psi_val(f, 0);
488 ir_node* f1 = get_Psi_default(f);
490 /* Psi(c0, x, Psi(c1, x, y)) -> typical if (c0 || c1) x else y */
491 ir_node* or_ = new_r_Or(irg, block, c0, c1, mode_b);
492 ir_node* vals[2] = { t1, f1 };
493 ir_node* new_psi = new_r_Psi(irg, block, 1, &or_, vals, mode);
494 exchange(psi, new_psi);
495 } else if (t == f1) {
496 /* Psi(c0, x, Psi(c1, y, x)) */
497 ir_node* not_c1 = new_r_Not(irg, block, c1, mode_b);
498 ir_node* or_ = new_r_Or(irg, block, c0, not_c1, mode_b);
499 ir_node* vals[2] = { f1, t1 };
500 ir_node* new_psi = new_r_Psi(irg, block, 1, &or_, vals, mode);
501 exchange(psi, new_psi);
507 static void optimise_psis_1(ir_node* psi, void* env)
515 if (!is_Psi(psi)) return;
517 t = get_Psi_val(psi, 0);
518 f = get_Psi_default(psi);
520 DB((dbg, LEVEL_3, "Simplify %+F T=%+F F=%+F\n", psi, t, f));
522 mode = get_irn_mode(psi);
524 if (is_Const(t) && is_Const(f) && (mode_is_int(mode))) {
525 ir_node* block = get_nodes_block(psi);
526 ir_node* c = get_Psi_cond(psi, 0);
527 tarval* tv_t = get_Const_tarval(t);
528 tarval* tv_f = get_Const_tarval(f);
529 if (tarval_is_one(tv_t) && tarval_is_null(tv_f)) {
530 ir_node* conv = new_r_Conv(current_ir_graph, block, c, mode);
532 } else if (tarval_is_null(tv_t) && tarval_is_one(tv_f)) {
533 ir_node* not_ = new_r_Not(current_ir_graph, block, c, mode_b);
534 ir_node* conv = new_r_Conv(current_ir_graph, block, not_, mode);
541 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params)
544 ir_settings_if_conv_t p;
546 /* get the parameters */
547 p = (params != NULL ? *params : default_info);
549 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
551 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
553 normalize_one_return(irg);
554 remove_critical_cf_edges(irg);
559 set_using_irn_link(irg);
562 irg_block_walk_graph(irg, init_block_link, NULL, &obst);
563 irg_walk_graph(irg, collect_phis, NULL, NULL);
564 irg_block_walk_graph(irg, NULL, if_conv_walker, &p);
566 clear_using_irn_link(irg);
568 local_optimize_graph(irg);
570 irg_walk_graph(irg, NULL, optimise_psis_0, NULL);
572 irg_walk_graph(irg, NULL, optimise_psis_1, NULL);
575 obstack_free(&obst, NULL);
577 /* TODO: graph might be changed, handle more gracefull */
578 set_irg_outs_inconsistent(irg);
579 set_irg_extblk_inconsistent(irg);
580 set_irg_loopinfo_inconsistent(irg);