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 const ir_settings_if_conv_t *params; /**< Conversion parameter. */
51 int changed; /**< Set if the graph was changed. */
54 DEBUG_ONLY(static firm_dbg_module_t *dbg);
57 * Default callback for Mux creation: allows every Mux to be created.
59 static int default_allow_ifconv(ir_node *sel, ir_node* phi_list, int i, int j)
71 static const ir_settings_if_conv_t default_info = {
72 0, /* doesn't matter for Mux */
77 * Returns non-zero if a Block can be emptied.
79 * @param block the block
81 static int can_empty_block(ir_node *block) {
82 return get_Block_mark(block) == 0;
86 * Find the ProjX node leading from block dependency to block start.
88 * @param start a block that is control depended on dependency
89 * @param dependency the block that decides whether start is executed
91 * @return a ProjX node that represent the decision control flow or
92 * NULL is start is not dependent at all or a block on the way
95 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
100 /* No need to find the conditional block if this block cannot be emptied and
101 * therefore not moved */
102 if (!can_empty_block(start)) return NULL;
104 arity = get_irn_arity(start);
105 for (i = 0; i < arity; ++i) {
106 ir_node* pred = get_irn_n(start, i);
107 ir_node* pred_block = get_nodes_block(skip_Proj(pred));
109 if (pred_block == dependency) {
111 assert(get_irn_mode(pred) == mode_X);
115 /* Not a Proj? Should not happen. */
120 assert(get_irn_mode(pred) == mode_X);
121 /* another Proj but not from the control block */
125 if (is_cdep_on(pred_block, dependency)) {
126 return walk_to_projx(pred_block, dependency);
134 * Recursively copies the DAG starting at node to the i-th predecessor
136 * - if node isn't in the src_block, recursion ends and node is returned
137 * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
138 * returned and recursion ends
139 * otherwise returns a copy of the passed node created in the i-th predecessor of
142 * @param node a root of a DAG
143 * @param src_block the block of the DAG
144 * @param i the position of the predecessor the DAG
147 * @return the root of the copied DAG
149 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
155 if (get_nodes_block(node) != src_block) {
156 /* already outside src_block, do not copy */
160 /* move through the Phi to the i-th predecessor */
161 return get_irn_n(node, i);
164 /* else really need a copy */
165 copy = exact_copy(node);
166 dst_block = get_nodes_block(get_irn_n(src_block, i));
167 set_nodes_block(copy, dst_block);
169 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
170 node, dst_block, copy));
172 /* move recursively all predecessors */
173 for (j = get_irn_arity(node) - 1; j >= 0; --j) {
174 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
175 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
182 * Remove predecessors i and j (i < j) from a node and
183 * add an additional predecessor new_pred.
185 * @param node the node whose inputs are changed
186 * @param i the first index to remove
187 * @param j the second index to remove
188 * @param new_pred a node that is added as a new input to node
190 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
192 int arity = get_irn_arity(node);
197 NEW_ARR_A(ir_node *, ins, arity - 1);
200 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
201 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
202 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
204 assert(l == arity - 1);
205 set_irn_in(node, l, ins);
210 * Remove the j-th predecessors from the i-th predecessor of block and add it to block
212 static void split_block(ir_node* block, int i, int j)
214 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
215 int arity = get_irn_arity(block);
222 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
224 NEW_ARR_A(ir_node*, ins, arity + 1);
226 for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
227 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
229 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
231 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
232 ins[k] = get_irn_n(phi, i);
234 set_irn_in(phi, arity + 1, ins);
237 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
238 ins[k++] = get_irn_n(pred_block, j);
239 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
240 ins[k] = get_irn_n(block, i);
242 set_irn_in(block, arity + 1, ins);
244 new_pred_arity = get_irn_arity(pred_block) - 1;
245 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
247 for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
248 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
249 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
250 assert(k == new_pred_arity);
251 next = get_Phi_next(phi);
252 if (new_pred_arity > 1) {
253 set_irn_in(phi, new_pred_arity, pred_ins);
255 exchange(phi, pred_ins[0]);
259 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
260 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
261 assert(k == new_pred_arity);
262 if (new_pred_arity > 1) {
263 set_irn_in(pred_block, new_pred_arity, pred_ins);
265 exchange(pred_block, get_nodes_block(pred_ins[0]));
270 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
272 ir_node* pred = get_nodes_block(get_irn_n(block, i));
276 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
278 pred_arity = get_irn_arity(pred);
279 for (j = 0; j < pred_arity; ++j) {
280 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
282 if (is_cdep_on(pred_pred, dependency)) {
283 prepare_path(pred, j, dependency);
284 split_block(block, i, j);
291 * Block walker: Search for diamonds and do the if conversion.
293 static void if_conv_walker(ir_node *block, void *ctx)
295 walker_env *env = ctx;
299 /* Bail out, if there are no Phis at all */
300 if (get_Block_phis(block) == NULL) return;
303 arity = get_irn_arity(block);
304 for (i = 0; i < arity; ++i) {
308 pred0 = get_Block_cfgpred_block(block, i);
309 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
310 const ir_node* dependency = cdep->node;
311 ir_node* projx0 = walk_to_projx(pred0, dependency);
315 if (projx0 == NULL) continue;
317 cond = get_Proj_pred(projx0);
321 /* We only handle boolean decisions, no switches */
322 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
324 for (j = i + 1; j < arity; ++j) {
332 pred1 = get_Block_cfgpred_block(block, j);
334 if (!is_cdep_on(pred1, dependency)) continue;
336 projx1 = walk_to_projx(pred1, dependency);
338 if (projx1 == NULL) continue;
340 phi = get_Block_phis(block);
341 if (!env->params->allow_ifconv(get_Cond_selector(cond), phi, i, j))
344 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
349 prepare_path(block, i, dependency);
350 prepare_path(block, j, dependency);
351 arity = get_irn_arity(block);
353 sel = get_Cond_selector(cond);
355 mux_block = get_nodes_block(cond);
356 cond_dbg = get_irn_dbg_info(cond);
358 ir_node* val_i = get_irn_n(phi, i);
359 ir_node* val_j = get_irn_n(phi, j);
363 if (val_i == val_j) {
365 DB((dbg, LEVEL_2, "Generating no Mux, because both values are equal\n"));
369 /* Something is very fishy if two predecessors of a PhiM point into
370 * one block, but not at the same memory node
372 assert(get_irn_mode(phi) != mode_M);
373 if (get_Proj_proj(projx0) == pn_Cond_true) {
381 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
382 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
385 next_phi = get_Phi_next(phi);
390 rewire(phi, i, j, mux);
393 } while (phi != NULL);
395 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
396 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
401 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, mux_block));
402 /* copy the block-info from the Mux-block to the block before merging */
404 mark = get_Block_mark(mux_block) | get_Block_mark(block);
405 set_Block_mark(block, mark);
406 set_Block_phis(block, get_Block_phis(mux_block));
408 set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
409 exchange_cdep(mux_block, block);
410 exchange(mux_block, block);
412 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, mux_block));
413 mark = get_Block_mark(mux_block) | get_Block_mark(block);
414 /* mark both block just to be sure, should be enough to mark mux_block */
415 set_Block_mark(mux_block, mark);
416 exchange(block, mux_block);
420 rewire(block, i, j, new_r_Jmp(mux_block));
429 * Block walker: clear block marks and Phi lists.
431 static void init_block_link(ir_node *block, void *env)
434 set_Block_mark(block, 0);
435 set_Block_phis(block, NULL);
440 * Daisy-chain all Phis in a block.
441 * If a non-movable node is encountered set the has_pinned flag in its block.
443 static void collect_phis(ir_node *node, void *env) {
447 ir_node *block = get_nodes_block(node);
449 add_Block_phi(block, node);
451 if (is_no_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
453 * Ignore control flow nodes (except Raise), these will be removed.
455 if (!is_cfop(node) && !is_Raise(node)) {
456 ir_node *block = get_nodes_block(node);
458 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
459 set_Block_mark(block, 1);
465 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params)
469 /* get the parameters */
470 env.params = (params != NULL ? params : &default_info);
473 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
475 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
477 normalize_one_return(irg);
478 remove_critical_cf_edges(irg);
482 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
484 irg_block_walk_graph(irg, init_block_link, NULL, NULL);
485 irg_walk_graph(irg, collect_phis, NULL, NULL);
486 irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
488 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
491 local_optimize_graph(irg);
493 /* graph has changed, invalidate analysis info */
494 set_irg_outs_inconsistent(irg);
495 set_irg_extblk_inconsistent(irg);
496 set_irg_loopinfo_inconsistent(irg);
497 set_irg_doms_inconsistent(irg);
504 ir_graph_pass_t pass;
505 const ir_settings_if_conv_t *params;
509 * Wrapper for running opt_if_conv() as an ir_graph pass.
511 static int pass_wrapper(ir_graph *irg, void *context) {
512 struct pass_t *pass = context;
513 opt_if_conv(irg, pass->params);
517 ir_graph_pass_t *opt_if_conv_pass(
518 const char *name, const ir_settings_if_conv_t *params)
520 struct pass_t *pass = xmalloc(sizeof(*pass));
522 pass->pass.kind = k_ir_prog_pass;
523 pass->pass.run_on_irg = pass_wrapper;
524 pass->pass.context = pass;
525 pass->pass.name = name ? name : "if_conv";
527 pass->params = params;
529 INIT_LIST_HEAD(&pass->pass.list);