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
31 #include "iroptimize.h"
48 * Environment for if-conversion.
50 typedef struct walker_env {
51 arch_allow_ifconv_func allow_ifconv;
52 bool changed; /**< Set if the graph was changed. */
55 DEBUG_ONLY(static firm_dbg_module_t *dbg);
58 * Default callback for Mux creation: don't allow any Mux nodes
60 static int default_allow_ifconv(ir_node *sel, ir_node *mux_false,
70 * Returns non-zero if a Block can be emptied.
72 * @param block the block
74 static bool can_empty_block(ir_node *block)
76 return get_Block_mark(block) == 0;
80 * Find the ProjX node leading from block dependency to block start.
82 * @param start a block that is control depended on dependency
83 * @param dependency the block that decides whether start is executed
85 * @return a ProjX node that represent the decision control flow or
86 * NULL is start is not dependent at all or a block on the way
89 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
94 /* No need to find the conditional block if this block cannot be emptied and
95 * therefore not moved */
96 if (!can_empty_block(start)) return NULL;
98 arity = get_irn_arity(start);
99 for (i = 0; i < arity; ++i) {
100 ir_node* pred = get_irn_n(start, i);
101 ir_node* pred_block = get_nodes_block(skip_Proj(pred));
103 if (pred_block == dependency) {
105 assert(get_irn_mode(pred) == mode_X);
109 /* Not a Proj? Should not happen. */
114 assert(get_irn_mode(pred) == mode_X);
115 /* another Proj but not from the control block */
119 if (is_cdep_on(pred_block, dependency)) {
120 return walk_to_projx(pred_block, dependency);
128 * Recursively copies the DAG starting at node to the i-th predecessor
130 * - if node isn't in the src_block, recursion ends and node is returned
131 * - if node is a Phi in the src_block, the i-th predecessor of this Phi is
132 * returned and recursion ends
133 * otherwise returns a copy of the passed node created in the i-th predecessor of
136 * @param node a root of a DAG
137 * @param src_block the block of the DAG
138 * @param i the position of the predecessor the DAG
141 * @return the root of the copied DAG
143 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
149 if (get_nodes_block(node) != src_block) {
150 /* already outside src_block, do not copy */
154 /* move through the Phi to the i-th predecessor */
155 return get_irn_n(node, i);
158 /* else really need a copy */
159 copy = exact_copy(node);
160 dst_block = get_nodes_block(get_irn_n(src_block, i));
161 set_nodes_block(copy, dst_block);
163 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
164 node, dst_block, copy));
166 /* move recursively all predecessors */
167 for (j = get_irn_arity(node) - 1; j >= 0; --j) {
168 set_irn_n(copy, j, copy_to(get_irn_n(node, j), src_block, i));
169 DB((dbg, LEVEL_2, "-- pred %d is %+F\n", j, get_irn_n(copy, j)));
176 * Remove predecessors i and j (i < j) from a node and
177 * add an additional predecessor new_pred.
179 * @param node the node whose inputs are changed
180 * @param i the first index to remove
181 * @param j the second index to remove
182 * @param new_pred a node that is added as a new input to node
184 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
186 int arity = get_irn_arity(node);
191 NEW_ARR_A(ir_node *, ins, arity - 1);
194 for (k = 0; k < i; ++k) ins[l++] = get_irn_n(node, k);
195 for (++k; k < j; ++k) ins[l++] = get_irn_n(node, k);
196 for (++k; k < arity; ++k) ins[l++] = get_irn_n(node, k);
198 assert(l == arity - 1);
199 set_irn_in(node, l, ins);
204 * Remove the j-th predecessors from the i-th predecessor of block and add it to block
206 static void split_block(ir_node* block, int i, int j)
208 ir_node* pred_block = get_nodes_block(get_irn_n(block, i));
209 int arity = get_irn_arity(block);
216 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
218 NEW_ARR_A(ir_node*, ins, arity + 1);
220 for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
221 ir_node* copy = copy_to(get_irn_n(phi, i), pred_block, j);
223 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
225 for (; k < arity; ++k) ins[k] = get_irn_n(phi, k);
226 ins[k] = get_irn_n(phi, i);
228 set_irn_in(phi, arity + 1, ins);
231 for (k = 0; k < i; ++k) ins[k] = get_irn_n(block, k);
232 ins[k++] = get_irn_n(pred_block, j);
233 for (; k < arity; ++k) ins[k] = get_irn_n(block, k);
234 ins[k] = get_irn_n(block, i);
236 set_irn_in(block, arity + 1, ins);
238 new_pred_arity = get_irn_arity(pred_block) - 1;
239 NEW_ARR_A(ir_node*, pred_ins, new_pred_arity);
241 for (phi = get_Block_phis(pred_block); phi != NULL; phi = next) {
242 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(phi, k);
243 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(phi, k + 1);
244 assert(k == new_pred_arity);
245 next = get_Phi_next(phi);
246 if (new_pred_arity > 1) {
247 set_irn_in(phi, new_pred_arity, pred_ins);
249 exchange(phi, pred_ins[0]);
253 for (k = 0; k < j; ++k) pred_ins[k] = get_irn_n(pred_block, k);
254 for (; k < new_pred_arity; ++k) pred_ins[k] = get_irn_n(pred_block, k + 1);
255 assert(k == new_pred_arity);
256 if (new_pred_arity > 1) {
257 set_irn_in(pred_block, new_pred_arity, pred_ins);
259 exchange(pred_block, get_nodes_block(pred_ins[0]));
264 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
266 ir_node* pred = get_nodes_block(get_irn_n(block, i));
270 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
272 pred_arity = get_irn_arity(pred);
273 for (j = 0; j < pred_arity; ++j) {
274 ir_node* pred_pred = get_nodes_block(get_irn_n(pred, j));
276 if (is_cdep_on(pred_pred, dependency)) {
277 prepare_path(pred, j, dependency);
278 split_block(block, i, j);
285 * Block walker: Search for diamonds and do the if conversion.
287 static void if_conv_walker(ir_node *block, void *ctx)
289 walker_env *env = ctx;
293 /* Bail out, if there are no Phis at all */
294 if (get_Block_phis(block) == NULL) return;
297 arity = get_irn_arity(block);
298 for (i = 0; i < arity; ++i) {
302 pred0 = get_Block_cfgpred_block(block, i);
303 for (cdep = find_cdep(pred0); cdep != NULL; cdep = cdep->next) {
304 const ir_node* dependency = cdep->node;
305 ir_node* projx0 = walk_to_projx(pred0, dependency);
309 if (projx0 == NULL) continue;
311 cond = get_Proj_pred(projx0);
315 /* We only handle boolean decisions, no switches */
316 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
318 for (j = i + 1; j < arity; ++j) {
329 pred1 = get_Block_cfgpred_block(block, j);
331 if (!is_cdep_on(pred1, dependency)) continue;
333 projx1 = walk_to_projx(pred1, dependency);
335 if (projx1 == NULL) continue;
337 sel = get_Cond_selector(cond);
338 phi = get_Block_phis(block);
340 negated = get_Proj_proj(projx0) == pn_Cond_false;
341 for (p = phi; p != NULL; p = get_Phi_next(p)) {
345 mux_true = get_Phi_pred(p, j);
346 mux_false = get_Phi_pred(p, i);
348 mux_true = get_Phi_pred(p, i);
349 mux_false = get_Phi_pred(p, j);
351 if (!env->allow_ifconv(sel, mux_false, mux_true)) {
359 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
364 prepare_path(block, i, dependency);
365 prepare_path(block, j, dependency);
366 arity = get_irn_arity(block);
368 mux_block = get_nodes_block(cond);
369 cond_dbg = get_irn_dbg_info(cond);
371 ir_node* val_i = get_irn_n(phi, i);
372 ir_node* val_j = get_irn_n(phi, j);
376 if (val_i == val_j) {
378 DB((dbg, LEVEL_2, "Generating no Mux, because both values are equal\n"));
382 /* Something is very fishy if two predecessors of a PhiM point into
383 * one block, but not at the same memory node
385 assert(get_irn_mode(phi) != mode_M);
394 mux = new_rd_Mux(cond_dbg, mux_block, sel, f, t, get_irn_mode(phi));
395 DB((dbg, LEVEL_2, "Generating %+F for %+F\n", mux, phi));
398 next_phi = get_Phi_next(phi);
403 rewire(phi, i, j, mux);
406 } while (phi != NULL);
408 exchange(get_nodes_block(get_irn_n(block, i)), mux_block);
409 exchange(get_nodes_block(get_irn_n(block, j)), mux_block);
414 DB((dbg, LEVEL_1, "Welding block %+F and %+F\n", block, mux_block));
415 /* copy the block-info from the Mux-block to the block before merging */
417 mark = get_Block_mark(mux_block) | get_Block_mark(block);
418 set_Block_mark(block, mark);
419 set_Block_phis(block, get_Block_phis(mux_block));
421 set_irn_in(block, get_irn_arity(mux_block), get_irn_in(mux_block) + 1);
422 exchange_cdep(mux_block, block);
423 exchange(mux_block, block);
425 DB((dbg, LEVEL_1, "Welding block %+F to %+F\n", block, mux_block));
426 mark = get_Block_mark(mux_block) | get_Block_mark(block);
427 /* mark both block just to be sure, should be enough to mark mux_block */
428 set_Block_mark(mux_block, mark);
429 exchange(block, mux_block);
433 rewire(block, i, j, new_r_Jmp(mux_block));
442 * Block walker: clear block marks and Phi lists.
444 static void init_block_link(ir_node *block, void *env)
447 set_Block_mark(block, 0);
448 set_Block_phis(block, NULL);
453 * Daisy-chain all Phis in a block.
454 * If a non-movable node is encountered set the has_pinned flag in its block.
456 static void collect_phis(ir_node *node, void *env)
461 ir_node *block = get_nodes_block(node);
463 add_Block_phi(block, node);
465 if (!is_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
467 * Ignore control flow nodes (except Raise), these will be removed.
469 if (!is_cfop(node) && !is_Raise(node)) {
470 ir_node *block = get_nodes_block(node);
472 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
473 set_Block_mark(block, 1);
479 void opt_if_conv(ir_graph *irg)
482 const backend_params *be_params = be_get_backend_param();
484 /* get the parameters */
485 if (be_params->allow_ifconv != NULL)
486 env.allow_ifconv = be_params->allow_ifconv;
488 env.allow_ifconv = default_allow_ifconv;
491 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
493 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
495 normalize_one_return(irg);
496 remove_critical_cf_edges(irg);
500 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
502 irg_block_walk_graph(irg, init_block_link, NULL, NULL);
503 irg_walk_graph(irg, collect_phis, NULL, NULL);
504 irg_block_walk_graph(irg, NULL, if_conv_walker, &env);
506 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
509 local_optimize_graph(irg);
511 /* graph has changed, invalidate analysis info */
512 set_irg_outs_inconsistent(irg);
513 set_irg_extblk_inconsistent(irg);
514 set_irg_loopinfo_inconsistent(irg);
515 set_irg_doms_inconsistent(irg);
521 ir_graph_pass_t *opt_if_conv_pass(const char *name)
523 return def_graph_pass(name ? name : "ifconv", opt_if_conv);