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 * Environment for if-conversion.
51 typedef struct walker_env {
52 const ir_settings_if_conv_t *params; /**< Conversion parameter. */
53 int changed; /**< Set if the graph was changed. */
56 DEBUG_ONLY(static firm_dbg_module_t *dbg);
59 * Default callback for Mux creation: don't allow any Mux nodes
61 static int default_allow_ifconv(ir_node *sel, ir_node *mux_false,
73 static const ir_settings_if_conv_t default_info = {
74 0, /* doesn't matter for Mux */
79 * Returns non-zero if a Block can be emptied.
81 * @param block the block
83 static bool can_empty_block(ir_node *block)
85 return get_Block_mark(block) == 0;
89 * Find the ProjX node leading from block dependency to block start.
91 * @param start a block that is control depended on dependency
92 * @param dependency the block that decides whether start is executed
94 * @return a ProjX node that represent the decision control flow or
95 * NULL is start is not dependent at all or a block on the way
98 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
103 /* No need to find the conditional block if this block cannot be emptied and
104 * therefore not moved */
105 if (!can_empty_block(start)) return NULL;
107 arity = get_irn_arity(start);
108 for (i = 0; i < arity; ++i) {
109 ir_node* pred = get_irn_n(start, i);
110 ir_node* pred_block = get_nodes_block(skip_Proj(pred));
112 if (pred_block == dependency) {
114 assert(get_irn_mode(pred) == mode_X);
118 /* Not a Proj? Should not happen. */
123 assert(get_irn_mode(pred) == mode_X);
124 /* another Proj but not from the control block */
128 if (is_cdep_on(pred_block, dependency)) {
129 return walk_to_projx(pred_block, dependency);
137 * Recursively copies the DAG starting at node to the i-th predecessor
139 * - if node isn't in the src_block, recursion ends and node is returned
140 * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
141 * returned and recursion ends
142 * otherwise returns a copy of the passed node created in the i-th predecessor of
145 * @param node a root of a DAG
146 * @param src_block the block of the DAG
147 * @param i the position of the predecessor the DAG
150 * @return the root of the copied DAG
152 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
158 if (get_nodes_block(node) != src_block) {
159 /* already outside src_block, do not copy */
163 /* move through the Phi to the i-th predecessor */
164 return get_irn_n(node, i);
167 /* else really need a copy */
168 copy = exact_copy(node);
169 dst_block = get_nodes_block(get_irn_n(src_block, i));
170 set_nodes_block(copy, dst_block);
172 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
173 node, dst_block, copy));
175 /* move recursively all predecessors */
176 for (j = get_irn_arity(node) - 1; j >= 0; --j) {
177 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
178 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
185 * Remove predecessors i and j (i < j) from a node and
186 * add an additional predecessor new_pred.
188 * @param node the node whose inputs are changed
189 * @param i the first index to remove
190 * @param j the second index to remove
191 * @param new_pred a node that is added as a new input to node
193 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
195 int arity = get_irn_arity(node);
200 NEW_ARR_A(ir_node *, ins, arity - 1);
203 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
204 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
205 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
207 assert(l == arity - 1);
208 set_irn_in(node, l, ins);
213 * Remove the j-th predecessors from the i-th predecessor of block and add it to block
215 static void split_block(ir_node* block, int i, int j)
217 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
218 int arity = get_irn_arity(block);
225 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
227 NEW_ARR_A(ir_node*, ins, arity + 1);
229 for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
230 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
232 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
234 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
235 ins[k] = get_irn_n(phi, i);
237 set_irn_in(phi, arity + 1, ins);
240 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
241 ins[k++] = get_irn_n(pred_block, j);
242 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
243 ins[k] = get_irn_n(block, i);
245 set_irn_in(block, arity + 1, ins);
247 new_pred_arity = get_irn_arity(pred_block) - 1;
248 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
250 for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
251 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
252 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
253 assert(k == new_pred_arity);
254 next = get_Phi_next(phi);
255 if (new_pred_arity > 1) {
256 set_irn_in(phi, new_pred_arity, pred_ins);
258 exchange(phi, pred_ins[0]);
262 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
263 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
264 assert(k == new_pred_arity);
265 if (new_pred_arity > 1) {
266 set_irn_in(pred_block, new_pred_arity, pred_ins);
268 exchange(pred_block, get_nodes_block(pred_ins[0]));
273 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
275 ir_node* pred = get_nodes_block(get_irn_n(block, i));
279 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
281 pred_arity = get_irn_arity(pred);
282 for (j = 0; j < pred_arity; ++j) {
283 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
285 if (is_cdep_on(pred_pred, dependency)) {
286 prepare_path(pred, j, dependency);
287 split_block(block, i, j);
294 * Block walker: Search for diamonds and do the if conversion.
296 static void if_conv_walker(ir_node *block, void *ctx)
298 walker_env *env = ctx;
302 /* Bail out, if there are no Phis at all */
303 if (get_Block_phis(block) == NULL) return;
306 arity = get_irn_arity(block);
307 for (i = 0; i < arity; ++i) {
311 pred0 = get_Block_cfgpred_block(block, i);
312 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
313 const ir_node* dependency = cdep->node;
314 ir_node* projx0 = walk_to_projx(pred0, dependency);
318 if (projx0 == NULL) continue;
320 cond = get_Proj_pred(projx0);
324 /* We only handle boolean decisions, no switches */
325 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
327 for (j = i + 1; j < arity; ++j) {
338 pred1 = get_Block_cfgpred_block(block, j);
340 if (!is_cdep_on(pred1, dependency)) continue;
342 projx1 = walk_to_projx(pred1, dependency);
344 if (projx1 == NULL) continue;
346 sel = get_Cond_selector(cond);
347 phi = get_Block_phis(block);
349 negated = get_Proj_proj(projx0) == pn_Cond_false;
350 for (p = phi; p != NULL; p = get_Phi_next(p)) {
354 mux_true = get_Phi_pred(p, j);
355 mux_false = get_Phi_pred(p, i);
357 mux_true = get_Phi_pred(p, i);
358 mux_false = get_Phi_pred(p, j);
360 if (!env->params->allow_ifconv(sel, mux_false, mux_true)) {
368 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
373 prepare_path(block, i, dependency);
374 prepare_path(block, j, dependency);
375 arity = get_irn_arity(block);
377 mux_block = get_nodes_block(cond);
378 cond_dbg = get_irn_dbg_info(cond);
380 ir_node* val_i = get_irn_n(phi, i);
381 ir_node* val_j = get_irn_n(phi, j);
385 if (val_i == val_j) {
387 DB((dbg, LEVEL_2, "Generating no Mux, because both values are equal\n"));
391 /* Something is very fishy if two predecessors of a PhiM point into
392 * one block, but not at the same memory node
394 assert(get_irn_mode(phi) != mode_M);
403 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
404 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
407 next_phi = get_Phi_next(phi);
412 rewire(phi, i, j, mux);
415 } while (phi != NULL);
417 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
418 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
423 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, mux_block));
424 /* copy the block-info from the Mux-block to the block before merging */
426 mark = get_Block_mark(mux_block) | get_Block_mark(block);
427 set_Block_mark(block, mark);
428 set_Block_phis(block, get_Block_phis(mux_block));
430 set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
431 exchange_cdep(mux_block, block);
432 exchange(mux_block, block);
434 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, mux_block));
435 mark = get_Block_mark(mux_block) | get_Block_mark(block);
436 /* mark both block just to be sure, should be enough to mark mux_block */
437 set_Block_mark(mux_block, mark);
438 exchange(block, mux_block);
442 rewire(block, i, j, new_r_Jmp(mux_block));
451 * Block walker: clear block marks and Phi lists.
453 static void init_block_link(ir_node *block, void *env)
456 set_Block_mark(block, 0);
457 set_Block_phis(block, NULL);
462 * Daisy-chain all Phis in a block.
463 * If a non-movable node is encountered set the has_pinned flag in its block.
465 static void collect_phis(ir_node *node, void *env)
470 ir_node *block = get_nodes_block(node);
472 add_Block_phi(block, node);
474 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
476 * Ignore control flow nodes (except Raise), these will be removed.
478 if (!is_cfop(node) && !is_Raise(node)) {
479 ir_node *block = get_nodes_block(node);
481 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
482 set_Block_mark(block, 1);
488 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params)
492 /* get the parameters */
493 env.params = (params != NULL ? params : &default_info);
496 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
498 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
500 normalize_one_return(irg);
501 remove_critical_cf_edges(irg);
505 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
507 irg_block_walk_graph(irg, init_block_link, NULL, NULL);
508 irg_walk_graph(irg, collect_phis, NULL, NULL);
509 irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
511 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
514 local_optimize_graph(irg);
516 /* graph has changed, invalidate analysis info */
517 set_irg_outs_inconsistent(irg);
518 set_irg_extblk_inconsistent(irg);
519 set_irg_loopinfo_inconsistent(irg);
520 set_irg_doms_inconsistent(irg);
527 ir_graph_pass_t pass;
528 const ir_settings_if_conv_t *params;
532 * Wrapper for running opt_if_conv() as an ir_graph pass.
534 static int pass_wrapper(ir_graph *irg, void *context)
536 struct pass_t *pass = context;
537 opt_if_conv(irg, pass->params);
541 ir_graph_pass_t *opt_if_conv_pass(const char *name,
542 const ir_settings_if_conv_t *params)
544 struct pass_t *pass = XMALLOCZ(struct pass_t);
545 pass->params = params;
547 return def_graph_pass_constructor(
548 &pass->pass, name ? name : "ifconv", pass_wrapper);