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
30 #include "iroptimize.h"
46 * Environment for if-conversion.
48 typedef struct walker_env {
49 const ir_settings_if_conv_t *params; /**< Conversion parameter. */
50 int changed; /**< Set if the graph was changed. */
53 DEBUG_ONLY(static firm_dbg_module_t *dbg);
56 * Default callback for Mux creation: allows every Mux to be created.
58 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
70 static const ir_settings_if_conv_t default_info = {
71 0, /* doesn't matter for Mux */
76 * Returns non-zero if a Block can be emptied.
78 * @param block the block
80 static int can_empty_block(ir_node *block) {
81 return get_Block_mark(block) == 0;
85 * Find the ProjX node leading from block dependency to block start.
87 * @param start a block that is control depended on dependency
88 * @param dependency the block that decides whether start is executed
90 * @return a ProjX node that represent the decision control flow or
91 * NULL is start is not dependent at all or a block on the way
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(skip_Proj(pred));
108 if (pred_block == dependency) {
110 assert(get_irn_mode(pred) == mode_X);
114 /* Not a Proj? Should not happen. */
119 assert(get_irn_mode(pred) == mode_X);
120 /* another Proj but not from the control block */
124 if (is_cdep_on(pred_block, dependency)) {
125 return walk_to_projx(pred_block, dependency);
133 * Recursively copies the DAG starting at node to the i-th predecessor
135 * - if node isn't in the src_block, recursion ends and node is returned
136 * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
137 * returned and recursion ends
138 * otherwise returns a copy of the passed node created in the i-th predecessor of
141 * @param node a root of a DAG
142 * @param src_block the block of the DAG
143 * @param i the position of the predecessor the DAG
146 * @return the root of the copied DAG
148 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
154 if (get_nodes_block(node) != src_block) {
155 /* already outside src_block, do not copy */
159 /* move through the Phi to the i-th predecessor */
160 return get_irn_n(node, i);
163 /* else really need a copy */
164 copy = exact_copy(node);
165 dst_block = get_nodes_block(get_irn_n(src_block, i));
166 set_nodes_block(copy, dst_block);
168 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
169 node, dst_block, copy));
171 /* move recursively all predecessors */
172 for (j = get_irn_arity(node) - 1; j >= 0; --j) {
173 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
174 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
181 * Remove predecessors i and j (i < j) from a node and
182 * add an additional predecessor new_pred.
184 * @param node the node whose inputs are changed
185 * @param i the first index to remove
186 * @param j the second index to remove
187 * @param new_pred a node that is added as a new input to node
189 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
191 int arity = get_irn_arity(node);
196 NEW_ARR_A(ir_node *, ins, arity - 1);
199 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
200 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
201 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
203 assert(l == arity - 1);
204 set_irn_in(node, l, ins);
209 * Remove the j-th predecessors from the i-th predecessor of block and add it to block
211 static void split_block(ir_node* block, int i, int j)
213 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
214 int arity = get_irn_arity(block);
221 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
223 NEW_ARR_A(ir_node*, ins, arity + 1);
225 for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
226 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
228 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
230 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
231 ins[k] = get_irn_n(phi, i);
233 set_irn_in(phi, arity + 1, ins);
236 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
237 ins[k++] = get_irn_n(pred_block, j);
238 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
239 ins[k] = get_irn_n(block, i);
241 set_irn_in(block, arity + 1, ins);
243 new_pred_arity = get_irn_arity(pred_block) - 1;
244 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
246 for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
247 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
248 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
249 assert(k == new_pred_arity);
250 next = get_Phi_next(phi);
251 if (new_pred_arity > 1) {
252 set_irn_in(phi, new_pred_arity, pred_ins);
254 exchange(phi, pred_ins[0]);
258 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
259 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
260 assert(k == new_pred_arity);
261 if (new_pred_arity > 1) {
262 set_irn_in(pred_block, new_pred_arity, pred_ins);
264 exchange(pred_block, get_nodes_block(pred_ins[0]));
269 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
271 ir_node* pred = get_nodes_block(get_irn_n(block, i));
275 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
277 pred_arity = get_irn_arity(pred);
278 for (j = 0; j < pred_arity; ++j) {
279 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
281 if (is_cdep_on(pred_pred, dependency)) {
282 prepare_path(pred, j, dependency);
283 split_block(block, i, j);
290 * Block walker: Search for diamonds and do the if conversion.
292 static void if_conv_walker(ir_node *block, void *ctx)
294 walker_env *env = ctx;
298 /* Bail out, if there are no Phis at all */
299 if (get_Block_phis(block) == NULL) return;
302 arity = get_irn_arity(block);
303 for (i = 0; i < arity; ++i) {
307 pred0 = get_Block_cfgpred_block(block, i);
308 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
309 const ir_node* dependency = cdep->node;
310 ir_node* projx0 = walk_to_projx(pred0, dependency);
314 if (projx0 == NULL) continue;
316 cond = get_Proj_pred(projx0);
320 /* We only handle boolean decisions, no switches */
321 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
323 for (j = i + 1; j < arity; ++j) {
331 pred1 = get_Block_cfgpred_block(block, j);
333 if (!is_cdep_on(pred1, dependency)) continue;
335 projx1 = walk_to_projx(pred1, dependency);
337 if (projx1 == NULL) continue;
339 phi = get_Block_phis(block);
340 if (!env->params->allow_ifconv(get_Cond_selector(cond), phi, i, j))
343 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
348 prepare_path(block, i, dependency);
349 prepare_path(block, j, dependency);
350 arity = get_irn_arity(block);
352 sel = get_Cond_selector(cond);
354 mux_block = get_nodes_block(cond);
355 cond_dbg = get_irn_dbg_info(cond);
357 ir_node* val_i = get_irn_n(phi, i);
358 ir_node* val_j = get_irn_n(phi, j);
362 if (val_i == val_j) {
364 DB((dbg, LEVEL_2, "Generating no Mux, because both values are equal\n"));
368 /* Something is very fishy if two predecessors of a PhiM point into
369 * one block, but not at the same memory node
371 assert(get_irn_mode(phi) != mode_M);
372 if (get_Proj_proj(projx0) == pn_Cond_true) {
380 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
381 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
384 next_phi = get_Phi_next(phi);
389 rewire(phi, i, j, mux);
392 } while (phi != NULL);
394 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
395 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
400 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, mux_block));
401 /* copy the block-info from the Mux-block to the block before merging */
403 mark = get_Block_mark(mux_block) | get_Block_mark(block);
404 set_Block_mark(block, mark);
405 set_Block_phis(block, get_Block_phis(mux_block));
407 set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
408 exchange_cdep(mux_block, block);
409 exchange(mux_block, block);
411 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, mux_block));
412 mark = get_Block_mark(mux_block) | get_Block_mark(block);
413 /* mark both block just to be sure, should be enough to mark mux_block */
414 set_Block_mark(mux_block, mark);
415 exchange(block, mux_block);
419 rewire(block, i, j, new_r_Jmp(mux_block));
428 * Block walker: clear block marks and Phi lists.
430 static void init_block_link(ir_node *block, void *env)
433 set_Block_mark(block, 0);
434 set_Block_phis(block, NULL);
439 * Daisy-chain all Phis in a block.
440 * If a non-movable node is encountered set the has_pinned flag in its block.
442 static void collect_phis(ir_node *node, void *env) {
446 ir_node *block = get_nodes_block(node);
448 add_Block_phi(block, node);
450 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
452 * Ignore control flow nodes (except Raise), these will be removed.
454 if (!is_cfop(node) && !is_Raise(node)) {
455 ir_node *block = get_nodes_block(node);
457 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
458 set_Block_mark(block, 1);
464 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params)
468 /* get the parameters */
469 env.params = (params != NULL ? params : &default_info);
472 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
474 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
476 normalize_one_return(irg);
477 remove_critical_cf_edges(irg);
481 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
483 irg_block_walk_graph(irg, init_block_link, NULL, NULL);
484 irg_walk_graph(irg, collect_phis, NULL, NULL);
485 irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
487 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
490 local_optimize_graph(irg);
492 /* graph has changed, invalidate analysis info */
493 set_irg_outs_inconsistent(irg);
494 set_irg_extblk_inconsistent(irg);
495 set_irg_loopinfo_inconsistent(irg);
496 set_irg_doms_inconsistent(irg);