1 /*
2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
3 *
4 * This file is part of libFirm.
5 *
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.
10 *
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.
14 *
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
17 * PURPOSE.
18 */
19
20 /*
21 * @file ir/opt/ifconv.c
22 * @brief If conversion
23 * @author Christoph Mallon
24 */
25 #include "config.h"
26
27 #include <assert.h>
28 #include <stdbool.h>
29
30 #include "iroptimize.h"
31 #include "obst.h"
32 #include "irnode_t.h"
33 #include "cdep_t.h"
34 #include "ircons.h"
35 #include "irgmod.h"
36 #include "irgopt.h"
37 #include "irgwalk.h"
38 #include "irtools.h"
39 #include "array_t.h"
40 #include "irpass_t.h"
41 #include "be.h"
42
43 #include "irdump.h"
44 #include "debug.h"
45
46 /**
47 * Environment for if-conversion.
48 */
49 typedef struct walker_env {
50 arch_allow_ifconv_func allow_ifconv;
51 bool changed; /**< Set if the graph was changed. */
52 } walker_env;
53
DEBUG_ONLY(static firm_dbg_module_t * dbg;)54 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
55
56 /**
57 * Returns non-zero if a Block can be emptied.
58 *
59 * @param block the block
60 */
61 static bool can_empty_block(ir_node *block)
62 {
63 return get_Block_mark(block) == 0;
64 }
65
66 /**
67 * Find the ProjX node leading from block dependency to block start.
68 *
69 * @param start a block that is control depended on dependency
70 * @param dependency the block that decides whether start is executed
71 *
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
74 * cannot be emptied
75 */
walk_to_projx(ir_node * start,const ir_node * dependency)76 static ir_node* walk_to_projx(ir_node* start, const ir_node* dependency)
77 {
78 int arity;
79 int i;
80
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;
84
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));
89
90 if (pred_block == dependency) {
91 if (is_Proj(pred)) {
92 assert(get_irn_mode(pred) == mode_X);
93 /* we found it */
94 return pred;
95 }
96 /* Not a Proj? Should not happen. */
97 return NULL;
98 }
99
100 if (is_Proj(pred)) {
101 assert(get_irn_mode(pred) == mode_X);
102 /* another Proj but not from the control block */
103 return NULL;
104 }
105
106 if (is_cdep_on(pred_block, dependency)) {
107 return walk_to_projx(pred_block, dependency);
108 }
109 }
110 return NULL;
111 }
112
113
114 /**
115 * Recursively copies the DAG starting at node to the i-th predecessor
116 * block of src_block
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
121 * src_block.
122 *
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
126 * is moved to
127 *
128 * @return the root of the copied DAG
129 */
copy_to(ir_node * node,ir_node * src_block,int i)130 static ir_node* copy_to(ir_node* node, ir_node* src_block, int i)
131 {
132 ir_node* dst_block;
133 ir_node* copy;
134 int j;
135
136 if (get_nodes_block(node) != src_block) {
137 /* already outside src_block, do not copy */
138 return node;
139 }
140 if (is_Phi(node)) {
141 /* move through the Phi to the i-th predecessor */
142 return get_irn_n(node, i);
143 }
144
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);
149
150 DB((dbg, LEVEL_1, "Copying node %+F to block %+F, copy is %+F\n",
151 node, dst_block, copy));
152
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)));
157 }
158 return copy;
159 }
160
161
162 /**
163 * Remove predecessors i and j (i < j) from a node and
164 * add an additional predecessor new_pred.
165 *
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
170 */
rewire(ir_node * node,int i,int j,ir_node * new_pred)171 static void rewire(ir_node* node, int i, int j, ir_node* new_pred)
172 {
173 int arity = get_irn_arity(node);
174 ir_node **ins;
175 int k;
176 int l;
177
178 NEW_ARR_A(ir_node *, ins, arity - 1);
179
180 l = 0;
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);
184 ins[l++] = new_pred;
185 assert(l == arity - 1);
186 set_irn_in(node, l, ins);
187 }
188
189
190 /**
191 * Remove the j-th predecessor from the i-th predecessor of block and add it to block
192 */
split_block(ir_node * block,int i,int j)193 static void split_block(ir_node* block, int i, int j)
194 {
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);
198 int new_pred_arity;
199 ir_node *phi;
200 ir_node *next;
201 ir_node **pred_ins;
202 int k;
203
204 DB((dbg, LEVEL_1, "Splitting predecessor %d of predecessor %d of %+F\n", j, i, block));
205
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);
208
209 for (k = 0; k < i; ++k) ins[k] = get_irn_n(phi, k);
210 ins[k++] = copy;
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);
214 }
215
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);
221
222 new_pred_arity = get_irn_arity(pred_block) - 1;
223 pred_ins = ALLOCAN(ir_node*, new_pred_arity);
224
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);
229 if (k == 1) {
230 exchange(phi, pred_ins[0]);
231 } else {
232 set_irn_in(phi, k, pred_ins);
233 }
234 }
235
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);
238 if (k == 1) {
239 exchange(pred_block, get_nodes_block(pred_ins[0]));
240 } else {
241 set_irn_in(pred_block, k, pred_ins);
242 }
243 }
244
245
prepare_path(ir_node * block,int i,const ir_node * dependency)246 static void prepare_path(ir_node* block, int i, const ir_node* dependency)
247 {
248 ir_node* pred = get_nodes_block(get_irn_n(block, i));
249 int pred_arity;
250 int j;
251
252 DB((dbg, LEVEL_1, "Preparing predecessor %d of %+F\n", i, block));
253
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));
257
258 if (pred_pred != dependency && is_cdep_on(pred_pred, dependency)) {
259 prepare_path(pred, j, dependency);
260 split_block(block, i, j);
261 break;
262 }
263 }
264 }
265
266 /**
267 * Block walker: Search for diamonds and do the if conversion.
268 */
if_conv_walker(ir_node * block,void * ctx)269 static void if_conv_walker(ir_node *block, void *ctx)
270 {
271 walker_env *env = (walker_env*)ctx;
272 int arity;
273 int i;
274
275 /* Bail out, if there are no Phis at all */
276 if (get_Block_phis(block) == NULL) return;
277
278 restart:
279 arity = get_irn_arity(block);
280 for (i = 0; i < arity; ++i) {
281 ir_node* pred0;
282 ir_cdep* cdep;
283
284 pred0 = get_Block_cfgpred_block(block, i);
285 if (pred0 == block) continue;
286
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);
290 ir_node* cond;
291 int j;
292
293 if (projx0 == NULL) continue;
294
295 cond = get_Proj_pred(projx0);
296 if (! is_Cond(cond))
297 continue;
298
299 /* We only handle boolean decisions, no switches */
300 if (get_irn_mode(get_Cond_selector(cond)) != mode_b) continue;
301
302 for (j = i + 1; j < arity; ++j) {
303 ir_node* projx1;
304 ir_node* sel;
305 ir_node* mux_block;
306 ir_node* phi;
307 ir_node* p;
308 ir_node* pred1;
309 bool supported;
310 bool negated;
311 dbg_info* cond_dbg;
312
313 pred1 = get_Block_cfgpred_block(block, j);
314 if (pred1 == block) continue;
315
316 if (!is_cdep_on(pred1, dependency)) continue;
317
318 projx1 = walk_to_projx(pred1, dependency);
319
320 if (projx1 == NULL) continue;
321
322 sel = get_Cond_selector(cond);
323 phi = get_Block_phis(block);
324 supported = true;
325 negated = get_Proj_proj(projx0) == pn_Cond_false;
326 for (p = phi; p != NULL; p = get_Phi_next(p)) {
327 ir_node *mux_false;
328 ir_node *mux_true;
329 if (negated) {
330 mux_true = get_Phi_pred(p, j);
331 mux_false = get_Phi_pred(p, i);
332 } else {
333 mux_true = get_Phi_pred(p, i);
334 mux_false = get_Phi_pred(p, j);
335 }
336 if (mux_true == mux_false)
337 continue;
338 ir_mode *mode = get_irn_mode(mux_true);
339 if (mode == mode_M
340 || !env->allow_ifconv(sel, mux_false, mux_true)) {
341 supported = false;
342 break;
343 }
344 }
345 if (!supported)
346 continue;
347
348 DB((dbg, LEVEL_1, "Found Cond %+F with proj %+F and %+F\n",
349 cond, projx0, projx1
350 ));
351
352 /* remove critical edges */
353 env->changed = true;
354 prepare_path(block, i, dependency);
355 prepare_path(block, j, dependency);
356 arity = get_irn_arity(block);
357
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);
363 ir_node* mux;
364 ir_node* next_phi;
365
366 if (val_i == val_j) {
367 mux = val_i;
368 DB((dbg, LEVEL_2, "Generating no Mux, because both values are equal\n"));
369 } else {
370 ir_node *t, *f;
371
372 /* Something is very fishy if two predecessors of a PhiM point into
373 * one block, but not at the same memory node
374 */
375 assert(get_irn_mode(phi) != mode_M);
376 if (negated) {
377 t = val_j;
378 f = val_i;
379 } else {
380 t = val_i;
381 f = val_j;
382 }
383
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));
386 }
387
388 next_phi = get_Phi_next(phi);
389
390 if (arity == 2) {
391 exchange(phi, mux);
392 } else {
393 rewire(phi, i, j, mux);
394 }
395 phi = next_phi;
396 } while (phi != NULL);
397
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);
401
402 if (arity == 2) {
403 unsigned mark;
404 #if 0
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 */
407
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));
411
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);
415 #else
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);
421 #endif
422 return;
423 } else {
424 rewire(block, i, j, new_r_Jmp(mux_block));
425 goto restart;
426 }
427 }
428 }
429 }
430 }
431
432 /**
433 * Block walker: clear block marks and Phi lists.
434 */
init_block_link(ir_node * block,void * env)435 static void init_block_link(ir_node *block, void *env)
436 {
437 (void)env;
438 set_Block_mark(block, 0);
439 set_Block_phis(block, NULL);
440 }
441
442
443 /**
444 * Daisy-chain all Phis in a block.
445 * If a non-movable node is encountered set the has_pinned flag in its block.
446 */
collect_phis(ir_node * node,void * env)447 static void collect_phis(ir_node *node, void *env)
448 {
449 (void) env;
450
451 if (is_Phi(node)) {
452 ir_node *block = get_nodes_block(node);
453
454 add_Block_phi(block, node);
455 } else {
456 if (!is_Block(node) && get_irn_pinned(node) == op_pin_state_pinned) {
457 /*
458 * Ignore control flow nodes (except Raise), these will be removed.
459 */
460 if (!is_cfop(node) && !is_Raise(node)) {
461 ir_node *block = get_nodes_block(node);
462
463 DB((dbg, LEVEL_2, "Node %+F in block %+F is unmovable\n", node, block));
464 set_Block_mark(block, 1);
465 }
466 }
467 }
468 }
469
opt_if_conv(ir_graph * irg)470 void opt_if_conv(ir_graph *irg)
471 {
472 walker_env env;
473 const backend_params *be_params = be_get_backend_param();
474
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);
480
481 /* get the parameters */
482 env.allow_ifconv = be_params->allow_ifconv;
483 env.changed = false;
484
485 FIRM_DBG_REGISTER(dbg, "firm.opt.ifconv");
486
487 DB((dbg, LEVEL_1, "Running if-conversion on %+F\n", irg));
488
489 compute_cdep(irg);
490
491 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
492
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);
496
497 ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
498
499 if (env.changed) {
500 local_optimize_graph(irg);
501 }
502
503 free_cdep(irg);
504
505 confirm_irg_properties(irg,
506 IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES
507 | IR_GRAPH_PROPERTY_ONE_RETURN);
508 }
509
opt_if_conv_pass(const char * name)510 ir_graph_pass_t *opt_if_conv_pass(const char *name)
511 {
512 return def_graph_pass(name ? name : "ifconv", opt_if_conv);
513 }
514