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"
47 * Environment for if-conversion.
49 typedef struct walker_env {
50 arch_allow_ifconv_func allow_ifconv;
51 bool changed; /**< Set if the graph was changed. */
54 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
57 * Returns non-zero if a Block can be emptied.
59 * @param block the block
61 static bool can_empty_block(ir_node *block)
63 return get_Block_mark(block) == 0;
67 * Find the ProjX node leading from block dependency to block start.
69 * @param start a block that is control depended on dependency
70 * @param dependency the block that decides whether start is executed
72 * @return a ProjX node that represent the decision control flow or
73 * NULL is start is not dependent at all or a block on the way
76 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
81 /* No need to find the conditional block if this block cannot be emptied and
82 * therefore not moved */
83 if (!can_empty_block(start)) return NULL;
85 arity = get_irn_arity(start);
86 for (i = 0; i < arity; ++i) {
87 ir_node* pred = get_irn_n(start, i);
88 ir_node* pred_block = get_nodes_block(skip_Proj(pred));
90 if (pred_block == dependency) {
92 assert(get_irn_mode(pred) == mode_X);
96 /* Not a Proj? Should not happen. */
101 assert(get_irn_mode(pred) == mode_X);
102 /* another Proj but not from the control block */
106 if (is_cdep_on(pred_block, dependency)) {
107 return walk_to_projx(pred_block, dependency);
115 * Recursively copies the DAG starting at node to the i-th predecessor
117 * - if node isn't in the src_block, recursion ends and node is returned
118 * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
119 * returned and recursion ends
120 * otherwise returns a copy of the passed node created in the i-th predecessor of
123 * @param node a root of a DAG
124 * @param src_block the block of the DAG
125 * @param i the position of the predecessor the DAG
128 * @return the root of the copied DAG
130 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
136 if (get_nodes_block(node) != src_block) {
137 /* already outside src_block, do not copy */
141 /* move through the Phi to the i-th predecessor */
142 return get_irn_n(node, i);
145 /* else really need a copy */
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 /* move recursively all predecessors */
154 for (j = get_irn_arity(node) - 1; j >= 0; --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 (i < j) from a node and
164 * add an additional predecessor new_pred.
166 * @param node the node whose inputs are changed
167 * @param i the first index to remove
168 * @param j the second index to remove
169 * @param new_pred a node that is added as a new input to node
171 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
173 int arity = get_irn_arity(node);
178 NEW_ARR_A(ir_node *, ins, arity - 1);
181 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
182 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
183 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
185 assert(l == arity - 1);
186 set_irn_in(node, l, ins);
191 * Remove the j-th predecessor from the i-th predecessor of block and add it to block
193 static void split_block(ir_node* block, int i, int j)
195 ir_node *pred_block = get_nodes_block(get_irn_n(block, i));
196 int arity = get_irn_arity(block);
197 ir_node **ins = ALLOCAN(ir_node*, arity + 1);
204 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
206 for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
207 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
209 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
211 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
212 ins[k++] = get_irn_n(phi, i);
213 set_irn_in(phi, k, ins);
216 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
217 ins[k++] = get_irn_n(pred_block, j);
218 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
219 ins[k++] = get_irn_n(block, i);
220 set_irn_in(block, k, ins);
222 new_pred_arity = get_irn_arity(pred_block) - 1;
223 pred_ins = ALLOCAN(ir_node*, new_pred_arity);
225 for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
226 next = get_Phi_next(phi);
227 for (k = 0; k != j; ++k) pred_ins[k] = get_irn_n(phi, k);
228 for (; k != new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
230 exchange(phi, pred_ins[0]);
232 set_irn_in(phi, k, pred_ins);
236 for (k = 0; k != j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
237 for (; k != new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
239 exchange(pred_block, get_nodes_block(pred_ins[0]));
241 set_irn_in(pred_block, k, pred_ins);
246 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
248 ir_node* pred = get_nodes_block(get_irn_n(block, i));
252 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
254 pred_arity = get_irn_arity(pred);
255 for (j = 0; j < pred_arity; ++j) {
256 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
258 if (pred_pred != dependency && is_cdep_on(pred_pred, dependency)) {
259 prepare_path(pred, j, dependency);
260 split_block(block, i, j);
267 * Block walker: Search for diamonds and do the if conversion.
269 static void if_conv_walker(ir_node *block, void *ctx)
271 walker_env *env = (walker_env*)ctx;
275 /* Bail out, if there are no Phis at all */
276 if (get_Block_phis(block) == NULL) return;
279 arity = get_irn_arity(block);
280 for (i = 0; i < arity; ++i) {
284 pred0 = get_Block_cfgpred_block(block, i);
285 if (pred0 == block) continue;
287 for (cdep = find_cdep(pred0); cdep != NULL; cdep = get_cdep_next(cdep)) {
288 const ir_node* dependency = get_cdep_node(cdep);
289 ir_node* projx0 = walk_to_projx(pred0, dependency);
293 if (projx0 == NULL) continue;
295 cond = get_Proj_pred(projx0);
299 /* We only handle boolean decisions, no switches */
300 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
302 for (j = i + 1; j < arity; ++j) {
313 pred1 = get_Block_cfgpred_block(block, j);
314 if (pred1 == block) continue;
316 if (!is_cdep_on(pred1, dependency)) continue;
318 projx1 = walk_to_projx(pred1, dependency);
320 if (projx1 == NULL) continue;
322 sel = get_Cond_selector(cond);
323 phi = get_Block_phis(block);
325 negated = get_Proj_proj(projx0) == pn_Cond_false;
326 for (p = phi; p != NULL; p = get_Phi_next(p)) {
330 mux_true = get_Phi_pred(p, j);
331 mux_false = get_Phi_pred(p, i);
333 mux_true = get_Phi_pred(p, i);
334 mux_false = get_Phi_pred(p, j);
336 if (mux_true == mux_false)
338 ir_mode *mode = get_irn_mode(mux_true);
340 || !env->allow_ifconv(sel, mux_false, mux_true)) {
348 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
352 /* remove critical edges */
354 prepare_path(block, i, dependency);
355 prepare_path(block, j, dependency);
356 arity = get_irn_arity(block);
358 mux_block = get_nodes_block(cond);
359 cond_dbg = get_irn_dbg_info(cond);
360 do { /* generate Mux nodes in mux_block for Phis in block */
361 ir_node* val_i = get_irn_n(phi, i);
362 ir_node* val_j = get_irn_n(phi, j);
366 if (val_i == val_j) {
368 DB((dbg, LEVEL_2, "Generating no Mux, because both values are equal\n"));
372 /* Something is very fishy if two predecessors of a PhiM point into
373 * one block, but not at the same memory node
375 assert(get_irn_mode(phi) != mode_M);
384 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
385 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
388 next_phi = get_Phi_next(phi);
393 rewire(phi, i, j, mux);
396 } while (phi != NULL);
398 /* move mux operands into mux_block */
399 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
400 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
405 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, mux_block));
406 /* copy the block-info from the Mux-block to the block before merging */
408 mark = get_Block_mark(mux_block) | get_Block_mark(block);
409 set_Block_mark(block, mark);
410 set_Block_phis(block, get_Block_phis(mux_block));
412 set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
413 exchange_cdep(mux_block, block);
414 exchange(mux_block, block);
416 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, mux_block));
417 mark = get_Block_mark(mux_block) | get_Block_mark(block);
418 /* mark both block just to be sure, should be enough to mark mux_block */
419 set_Block_mark(mux_block, mark);
420 exchange(block, mux_block);
424 rewire(block, i, j, new_r_Jmp(mux_block));
433 * Block walker: clear block marks and Phi lists.
435 static void init_block_link(ir_node *block, void *env)
438 set_Block_mark(block, 0);
439 set_Block_phis(block, NULL);
444 * Daisy-chain all Phis in a block.
445 * If a non-movable node is encountered set the has_pinned flag in its block.
447 static void collect_phis(ir_node *node, void *env)
452 ir_node *block = get_nodes_block(node);
454 add_Block_phi(block, node);
456 if (!is_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
458 * Ignore control flow nodes (except Raise), these will be removed.
460 if (!is_cfop(node) && !is_Raise(node)) {
461 ir_node *block = get_nodes_block(node);
463 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
464 set_Block_mark(block, 1);
470 void opt_if_conv(ir_graph *irg)
473 const backend_params *be_params = be_get_backend_param();
475 assure_irg_properties(irg,
476 IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES
477 | IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
478 | IR_GRAPH_PROPERTY_NO_BADS
479 | IR_GRAPH_PROPERTY_ONE_RETURN);
481 /* get the parameters */
482 env.allow_ifconv = be_params->allow_ifconv;
485 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
487 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
491 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
493 irg_block_walk_graph(irg, init_block_link, NULL, NULL);
494 irg_walk_graph(irg, collect_phis, NULL, NULL);
495 irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
497 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
500 local_optimize_graph(irg);
505 confirm_irg_properties(irg,
506 IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES
507 | IR_GRAPH_PROPERTY_ONE_RETURN);
510 ir_graph_pass_t *opt_if_conv_pass(const char *name)
512 return def_graph_pass(name ? name : "ifconv", opt_if_conv);