]>
Commit | Line | Data |
---|---|---|
4ee9c684 | 1 | /* Optimization of PHI nodes by converting them into straightline code. |
f0b5f617 | 2 | Copyright (C) 2004, 2005, 2006, 2007, 2008 Free Software Foundation, |
3 | Inc. | |
4ee9c684 | 4 | |
5 | This file is part of GCC. | |
20e5647c | 6 | |
4ee9c684 | 7 | GCC is free software; you can redistribute it and/or modify it |
8 | under the terms of the GNU General Public License as published by the | |
8c4c00c1 | 9 | Free Software Foundation; either version 3, or (at your option) any |
4ee9c684 | 10 | later version. |
20e5647c | 11 | |
4ee9c684 | 12 | GCC is distributed in the hope that it will be useful, but WITHOUT |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
20e5647c | 16 | |
4ee9c684 | 17 | You should have received a copy of the GNU General Public License |
8c4c00c1 | 18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
4ee9c684 | 20 | |
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
4ee9c684 | 25 | #include "ggc.h" |
26 | #include "tree.h" | |
27 | #include "rtl.h" | |
0beac6fc | 28 | #include "flags.h" |
4ee9c684 | 29 | #include "tm_p.h" |
30 | #include "basic-block.h" | |
31 | #include "timevar.h" | |
32 | #include "diagnostic.h" | |
33 | #include "tree-flow.h" | |
34 | #include "tree-pass.h" | |
35 | #include "tree-dump.h" | |
36 | #include "langhooks.h" | |
e6d0e152 | 37 | #include "pointer-set.h" |
38 | #include "domwalk.h" | |
4ee9c684 | 39 | |
75a70cf9 | 40 | static unsigned int tree_ssa_phiopt (void); |
e6d0e152 | 41 | static unsigned int tree_ssa_phiopt_worker (bool); |
a4844041 | 42 | static bool conditional_replacement (basic_block, basic_block, |
75a70cf9 | 43 | edge, edge, gimple, tree, tree); |
a4844041 | 44 | static bool value_replacement (basic_block, basic_block, |
75a70cf9 | 45 | edge, edge, gimple, tree, tree); |
a4844041 | 46 | static bool minmax_replacement (basic_block, basic_block, |
75a70cf9 | 47 | edge, edge, gimple, tree, tree); |
a4844041 | 48 | static bool abs_replacement (basic_block, basic_block, |
75a70cf9 | 49 | edge, edge, gimple, tree, tree); |
e6d0e152 | 50 | static bool cond_store_replacement (basic_block, basic_block, edge, edge, |
51 | struct pointer_set_t *); | |
52 | static struct pointer_set_t * get_non_trapping (void); | |
75a70cf9 | 53 | static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree); |
902929aa | 54 | |
caac37c2 | 55 | /* This pass tries to replaces an if-then-else block with an |
56 | assignment. We have four kinds of transformations. Some of these | |
57 | transformations are also performed by the ifcvt RTL optimizer. | |
58 | ||
59 | Conditional Replacement | |
60 | ----------------------- | |
61 | ||
30c5ffd2 | 62 | This transformation, implemented in conditional_replacement, |
caac37c2 | 63 | replaces |
4ee9c684 | 64 | |
65 | bb0: | |
66 | if (cond) goto bb2; else goto bb1; | |
67 | bb1: | |
68 | bb2: | |
caac37c2 | 69 | x = PHI <0 (bb1), 1 (bb0), ...>; |
4ee9c684 | 70 | |
caac37c2 | 71 | with |
20e5647c | 72 | |
2ab0a163 | 73 | bb0: |
caac37c2 | 74 | x' = cond; |
75 | goto bb2; | |
2ab0a163 | 76 | bb2: |
caac37c2 | 77 | x = PHI <x' (bb0), ...>; |
4ee9c684 | 78 | |
caac37c2 | 79 | We remove bb1 as it becomes unreachable. This occurs often due to |
80 | gimplification of conditionals. | |
20e5647c | 81 | |
caac37c2 | 82 | Value Replacement |
83 | ----------------- | |
84 | ||
85 | This transformation, implemented in value_replacement, replaces | |
0beac6fc | 86 | |
87 | bb0: | |
caac37c2 | 88 | if (a != b) goto bb2; else goto bb1; |
0beac6fc | 89 | bb1: |
90 | bb2: | |
caac37c2 | 91 | x = PHI <a (bb1), b (bb0), ...>; |
0beac6fc | 92 | |
caac37c2 | 93 | with |
0beac6fc | 94 | |
95 | bb0: | |
0beac6fc | 96 | bb2: |
caac37c2 | 97 | x = PHI <b (bb0), ...>; |
98 | ||
99 | This opportunity can sometimes occur as a result of other | |
100 | optimizations. | |
0beac6fc | 101 | |
caac37c2 | 102 | ABS Replacement |
103 | --------------- | |
70512b93 | 104 | |
caac37c2 | 105 | This transformation, implemented in abs_replacement, replaces |
70512b93 | 106 | |
107 | bb0: | |
caac37c2 | 108 | if (a >= 0) goto bb2; else goto bb1; |
70512b93 | 109 | bb1: |
caac37c2 | 110 | x = -a; |
70512b93 | 111 | bb2: |
caac37c2 | 112 | x = PHI <x (bb1), a (bb0), ...>; |
70512b93 | 113 | |
caac37c2 | 114 | with |
70512b93 | 115 | |
116 | bb0: | |
caac37c2 | 117 | x' = ABS_EXPR< a >; |
70512b93 | 118 | bb2: |
caac37c2 | 119 | x = PHI <x' (bb0), ...>; |
120 | ||
121 | MIN/MAX Replacement | |
122 | ------------------- | |
70512b93 | 123 | |
caac37c2 | 124 | This transformation, minmax_replacement replaces |
194899bf | 125 | |
126 | bb0: | |
caac37c2 | 127 | if (a <= b) goto bb2; else goto bb1; |
194899bf | 128 | bb1: |
194899bf | 129 | bb2: |
caac37c2 | 130 | x = PHI <b (bb1), a (bb0), ...>; |
194899bf | 131 | |
caac37c2 | 132 | with |
194899bf | 133 | |
caac37c2 | 134 | bb0: |
135 | x' = MIN_EXPR (a, b) | |
136 | bb2: | |
137 | x = PHI <x' (bb0), ...>; | |
194899bf | 138 | |
30c5ffd2 | 139 | A similar transformation is done for MAX_EXPR. */ |
70512b93 | 140 | |
2a1990e9 | 141 | static unsigned int |
4ee9c684 | 142 | tree_ssa_phiopt (void) |
e6d0e152 | 143 | { |
144 | return tree_ssa_phiopt_worker (false); | |
145 | } | |
146 | ||
147 | /* This pass tries to transform conditional stores into unconditional | |
148 | ones, enabling further simplifications with the simpler then and else | |
149 | blocks. In particular it replaces this: | |
150 | ||
151 | bb0: | |
152 | if (cond) goto bb2; else goto bb1; | |
153 | bb1: | |
154 | *p = RHS | |
155 | bb2: | |
156 | ||
157 | with | |
158 | ||
159 | bb0: | |
160 | if (cond) goto bb1; else goto bb2; | |
161 | bb1: | |
162 | condtmp' = *p; | |
163 | bb2: | |
164 | condtmp = PHI <RHS, condtmp'> | |
165 | *p = condtmp | |
166 | ||
167 | This transformation can only be done under several constraints, | |
168 | documented below. */ | |
169 | ||
170 | static unsigned int | |
171 | tree_ssa_cs_elim (void) | |
172 | { | |
173 | return tree_ssa_phiopt_worker (true); | |
174 | } | |
175 | ||
176 | /* For conditional store replacement we need a temporary to | |
177 | put the old contents of the memory in. */ | |
178 | static tree condstoretemp; | |
179 | ||
180 | /* The core routine of conditional store replacement and normal | |
181 | phi optimizations. Both share much of the infrastructure in how | |
182 | to match applicable basic block patterns. DO_STORE_ELIM is true | |
183 | when we want to do conditional store replacement, false otherwise. */ | |
184 | static unsigned int | |
185 | tree_ssa_phiopt_worker (bool do_store_elim) | |
4ee9c684 | 186 | { |
187 | basic_block bb; | |
194899bf | 188 | basic_block *bb_order; |
189 | unsigned n, i; | |
1e4b21e3 | 190 | bool cfgchanged = false; |
e6d0e152 | 191 | struct pointer_set_t *nontrap = 0; |
192 | ||
193 | if (do_store_elim) | |
194 | { | |
195 | condstoretemp = NULL_TREE; | |
196 | /* Calculate the set of non-trapping memory accesses. */ | |
197 | nontrap = get_non_trapping (); | |
198 | } | |
194899bf | 199 | |
200 | /* Search every basic block for COND_EXPR we may be able to optimize. | |
201 | ||
202 | We walk the blocks in order that guarantees that a block with | |
203 | a single predecessor is processed before the predecessor. | |
204 | This ensures that we collapse inner ifs before visiting the | |
205 | outer ones, and also that we do not try to visit a removed | |
206 | block. */ | |
207 | bb_order = blocks_in_phiopt_order (); | |
4d2e5d52 | 208 | n = n_basic_blocks - NUM_FIXED_BLOCKS; |
4ee9c684 | 209 | |
4d2e5d52 | 210 | for (i = 0; i < n; i++) |
4ee9c684 | 211 | { |
75a70cf9 | 212 | gimple cond_stmt, phi; |
33784d89 | 213 | basic_block bb1, bb2; |
214 | edge e1, e2; | |
194899bf | 215 | tree arg0, arg1; |
216 | ||
217 | bb = bb_order[i]; | |
20e5647c | 218 | |
75a70cf9 | 219 | cond_stmt = last_stmt (bb); |
220 | /* Check to see if the last statement is a GIMPLE_COND. */ | |
221 | if (!cond_stmt | |
222 | || gimple_code (cond_stmt) != GIMPLE_COND) | |
33784d89 | 223 | continue; |
20e5647c | 224 | |
33784d89 | 225 | e1 = EDGE_SUCC (bb, 0); |
226 | bb1 = e1->dest; | |
227 | e2 = EDGE_SUCC (bb, 1); | |
228 | bb2 = e2->dest; | |
20e5647c | 229 | |
33784d89 | 230 | /* We cannot do the optimization on abnormal edges. */ |
231 | if ((e1->flags & EDGE_ABNORMAL) != 0 | |
232 | || (e2->flags & EDGE_ABNORMAL) != 0) | |
233 | continue; | |
20e5647c | 234 | |
33784d89 | 235 | /* If either bb1's succ or bb2 or bb2's succ is non NULL. */ |
ea091dfd | 236 | if (EDGE_COUNT (bb1->succs) == 0 |
33784d89 | 237 | || bb2 == NULL |
ea091dfd | 238 | || EDGE_COUNT (bb2->succs) == 0) |
33784d89 | 239 | continue; |
20e5647c | 240 | |
33784d89 | 241 | /* Find the bb which is the fall through to the other. */ |
242 | if (EDGE_SUCC (bb1, 0)->dest == bb2) | |
243 | ; | |
244 | else if (EDGE_SUCC (bb2, 0)->dest == bb1) | |
245 | { | |
246 | basic_block bb_tmp = bb1; | |
247 | edge e_tmp = e1; | |
248 | bb1 = bb2; | |
249 | bb2 = bb_tmp; | |
250 | e1 = e2; | |
251 | e2 = e_tmp; | |
252 | } | |
253 | else | |
254 | continue; | |
20e5647c | 255 | |
33784d89 | 256 | e1 = EDGE_SUCC (bb1, 0); |
20e5647c | 257 | |
33784d89 | 258 | /* Make sure that bb1 is just a fall through. */ |
db5ba14c | 259 | if (!single_succ_p (bb1) |
33784d89 | 260 | || (e1->flags & EDGE_FALLTHRU) == 0) |
261 | continue; | |
20e5647c | 262 | |
3472707f | 263 | /* Also make sure that bb1 only have one predecessor and that it |
264 | is bb. */ | |
ea091dfd | 265 | if (!single_pred_p (bb1) |
266 | || single_pred (bb1) != bb) | |
33784d89 | 267 | continue; |
20e5647c | 268 | |
e6d0e152 | 269 | if (do_store_elim) |
270 | { | |
271 | /* bb1 is the middle block, bb2 the join block, bb the split block, | |
272 | e1 the fallthrough edge from bb1 to bb2. We can't do the | |
273 | optimization if the join block has more than two predecessors. */ | |
274 | if (EDGE_COUNT (bb2->preds) > 2) | |
275 | continue; | |
276 | if (cond_store_replacement (bb1, bb2, e1, e2, nontrap)) | |
277 | cfgchanged = true; | |
278 | } | |
279 | else | |
280 | { | |
75a70cf9 | 281 | gimple_seq phis = phi_nodes (bb2); |
e6d0e152 | 282 | |
283 | /* Check to make sure that there is only one PHI node. | |
284 | TODO: we could do it with more than one iff the other PHI nodes | |
285 | have the same elements for these two edges. */ | |
75a70cf9 | 286 | if (! gimple_seq_singleton_p (phis)) |
e6d0e152 | 287 | continue; |
288 | ||
75a70cf9 | 289 | phi = gsi_stmt (gsi_start (phis)); |
290 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); | |
291 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); | |
e6d0e152 | 292 | |
293 | /* Something is wrong if we cannot find the arguments in the PHI | |
294 | node. */ | |
295 | gcc_assert (arg0 != NULL && arg1 != NULL); | |
296 | ||
297 | /* Do the replacement of conditional if it can be done. */ | |
298 | if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
299 | cfgchanged = true; | |
300 | else if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
301 | cfgchanged = true; | |
302 | else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
303 | cfgchanged = true; | |
304 | else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
305 | cfgchanged = true; | |
306 | } | |
194899bf | 307 | } |
308 | ||
309 | free (bb_order); | |
1e4b21e3 | 310 | |
e6d0e152 | 311 | if (do_store_elim) |
312 | pointer_set_destroy (nontrap); | |
313 | /* If the CFG has changed, we should cleanup the CFG. */ | |
314 | if (cfgchanged && do_store_elim) | |
315 | { | |
316 | /* In cond-store replacement we have added some loads on edges | |
317 | and new VOPS (as we moved the store, and created a load). */ | |
75a70cf9 | 318 | gsi_commit_edge_inserts (); |
e6d0e152 | 319 | return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals; |
320 | } | |
321 | else if (cfgchanged) | |
322 | return TODO_cleanup_cfg; | |
323 | return 0; | |
194899bf | 324 | } |
325 | ||
326 | /* Returns the list of basic blocks in the function in an order that guarantees | |
327 | that if a block X has just a single predecessor Y, then Y is after X in the | |
328 | ordering. */ | |
329 | ||
8530c7be | 330 | basic_block * |
194899bf | 331 | blocks_in_phiopt_order (void) |
332 | { | |
333 | basic_block x, y; | |
945865c5 | 334 | basic_block *order = XNEWVEC (basic_block, n_basic_blocks); |
4d2e5d52 | 335 | unsigned n = n_basic_blocks - NUM_FIXED_BLOCKS; |
336 | unsigned np, i; | |
337 | sbitmap visited = sbitmap_alloc (last_basic_block); | |
194899bf | 338 | |
4d2e5d52 | 339 | #define MARK_VISITED(BB) (SET_BIT (visited, (BB)->index)) |
340 | #define VISITED_P(BB) (TEST_BIT (visited, (BB)->index)) | |
194899bf | 341 | |
342 | sbitmap_zero (visited); | |
343 | ||
344 | MARK_VISITED (ENTRY_BLOCK_PTR); | |
345 | FOR_EACH_BB (x) | |
346 | { | |
347 | if (VISITED_P (x)) | |
348 | continue; | |
349 | ||
350 | /* Walk the predecessors of x as long as they have precisely one | |
351 | predecessor and add them to the list, so that they get stored | |
352 | after x. */ | |
353 | for (y = x, np = 1; | |
354 | single_pred_p (y) && !VISITED_P (single_pred (y)); | |
355 | y = single_pred (y)) | |
356 | np++; | |
357 | for (y = x, i = n - np; | |
358 | single_pred_p (y) && !VISITED_P (single_pred (y)); | |
359 | y = single_pred (y), i++) | |
360 | { | |
361 | order[i] = y; | |
362 | MARK_VISITED (y); | |
2ab0a163 | 363 | } |
194899bf | 364 | order[i] = y; |
365 | MARK_VISITED (y); | |
366 | ||
367 | gcc_assert (i == n - 1); | |
368 | n -= np; | |
4ee9c684 | 369 | } |
194899bf | 370 | |
371 | sbitmap_free (visited); | |
372 | gcc_assert (n == 0); | |
373 | return order; | |
374 | ||
375 | #undef MARK_VISITED | |
376 | #undef VISITED_P | |
4ee9c684 | 377 | } |
378 | ||
47aaf6e6 | 379 | |
70512b93 | 380 | /* Return TRUE if block BB has no executable statements, otherwise return |
381 | FALSE. */ | |
47aaf6e6 | 382 | |
c91e8223 | 383 | bool |
47aaf6e6 | 384 | empty_block_p (basic_block bb) |
70512b93 | 385 | { |
70512b93 | 386 | /* BB must have no executable statements. */ |
75a70cf9 | 387 | return gsi_end_p (gsi_after_labels (bb)); |
70512b93 | 388 | } |
389 | ||
fccee353 | 390 | /* Replace PHI node element whose edge is E in block BB with variable NEW. |
33784d89 | 391 | Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK |
902929aa | 392 | is known to have two edges, one of which must reach BB). */ |
393 | ||
394 | static void | |
a4844041 | 395 | replace_phi_edge_with_variable (basic_block cond_block, |
75a70cf9 | 396 | edge e, gimple phi, tree new_tree) |
902929aa | 397 | { |
75a70cf9 | 398 | basic_block bb = gimple_bb (phi); |
0e1a77e1 | 399 | basic_block block_to_remove; |
75a70cf9 | 400 | gimple_stmt_iterator gsi; |
33784d89 | 401 | |
20e5647c | 402 | /* Change the PHI argument to new. */ |
f0d6e81c | 403 | SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree); |
0e1a77e1 | 404 | |
0e1a77e1 | 405 | /* Remove the empty basic block. */ |
cd665a06 | 406 | if (EDGE_SUCC (cond_block, 0)->dest == bb) |
902929aa | 407 | { |
cd665a06 | 408 | EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU; |
409 | EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); | |
81c5be57 | 410 | EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE; |
411 | EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count; | |
0e1a77e1 | 412 | |
cd665a06 | 413 | block_to_remove = EDGE_SUCC (cond_block, 1)->dest; |
902929aa | 414 | } |
415 | else | |
416 | { | |
cd665a06 | 417 | EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU; |
418 | EDGE_SUCC (cond_block, 1)->flags | |
902929aa | 419 | &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |
81c5be57 | 420 | EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE; |
421 | EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count; | |
0e1a77e1 | 422 | |
cd665a06 | 423 | block_to_remove = EDGE_SUCC (cond_block, 0)->dest; |
902929aa | 424 | } |
0e1a77e1 | 425 | delete_basic_block (block_to_remove); |
20e5647c | 426 | |
902929aa | 427 | /* Eliminate the COND_EXPR at the end of COND_BLOCK. */ |
75a70cf9 | 428 | gsi = gsi_last_bb (cond_block); |
429 | gsi_remove (&gsi, true); | |
20e5647c | 430 | |
902929aa | 431 | if (dump_file && (dump_flags & TDF_DETAILS)) |
432 | fprintf (dump_file, | |
433 | "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n", | |
434 | cond_block->index, | |
435 | bb->index); | |
436 | } | |
437 | ||
438 | /* The function conditional_replacement does the main work of doing the | |
439 | conditional replacement. Return true if the replacement is done. | |
440 | Otherwise return false. | |
441 | BB is the basic block where the replacement is going to be done on. ARG0 | |
dac49aa5 | 442 | is argument 0 from PHI. Likewise for ARG1. */ |
902929aa | 443 | |
444 | static bool | |
33784d89 | 445 | conditional_replacement (basic_block cond_bb, basic_block middle_bb, |
75a70cf9 | 446 | edge e0, edge e1, gimple phi, |
33784d89 | 447 | tree arg0, tree arg1) |
902929aa | 448 | { |
449 | tree result; | |
75a70cf9 | 450 | gimple stmt, new_stmt; |
451 | tree cond; | |
452 | gimple_stmt_iterator gsi; | |
902929aa | 453 | edge true_edge, false_edge; |
75a70cf9 | 454 | tree new_var, new_var2; |
902929aa | 455 | |
435e1a75 | 456 | /* FIXME: Gimplification of complex type is too hard for now. */ |
457 | if (TREE_CODE (TREE_TYPE (arg0)) == COMPLEX_TYPE | |
458 | || TREE_CODE (TREE_TYPE (arg1)) == COMPLEX_TYPE) | |
459 | return false; | |
460 | ||
902929aa | 461 | /* The PHI arguments have the constants 0 and 1, then convert |
462 | it to the conditional. */ | |
463 | if ((integer_zerop (arg0) && integer_onep (arg1)) | |
464 | || (integer_zerop (arg1) && integer_onep (arg0))) | |
465 | ; | |
466 | else | |
467 | return false; | |
20e5647c | 468 | |
33784d89 | 469 | if (!empty_block_p (middle_bb)) |
902929aa | 470 | return false; |
20e5647c | 471 | |
75a70cf9 | 472 | /* At this point we know we have a GIMPLE_COND with two successors. |
2ab0a163 | 473 | One successor is BB, the other successor is an empty block which |
474 | falls through into BB. | |
20e5647c | 475 | |
2ab0a163 | 476 | There is a single PHI node at the join point (BB) and its arguments |
477 | are constants (0, 1). | |
20e5647c | 478 | |
2ab0a163 | 479 | So, given the condition COND, and the two PHI arguments, we can |
20e5647c | 480 | rewrite this PHI into non-branching code: |
481 | ||
2ab0a163 | 482 | dest = (COND) or dest = COND' |
20e5647c | 483 | |
2ab0a163 | 484 | We use the condition as-is if the argument associated with the |
485 | true edge has the value one or the argument associated with the | |
486 | false edge as the value zero. Note that those conditions are not | |
75a70cf9 | 487 | the same since only one of the outgoing edges from the GIMPLE_COND |
2ab0a163 | 488 | will directly reach BB and thus be associated with an argument. */ |
ae5a4794 | 489 | |
75a70cf9 | 490 | stmt = last_stmt (cond_bb); |
491 | result = PHI_RESULT (phi); | |
b2a02a0e | 492 | |
75a70cf9 | 493 | /* To handle special cases like floating point comparison, it is easier and |
494 | less error-prone to build a tree and gimplify it on the fly though it is | |
495 | less efficient. */ | |
496 | cond = fold_build2 (gimple_cond_code (stmt), boolean_type_node, | |
497 | gimple_cond_lhs (stmt), gimple_cond_rhs (stmt)); | |
4ee9c684 | 498 | |
75a70cf9 | 499 | /* We need to know which is the true edge and which is the false |
500 | edge so that we know when to invert the condition below. */ | |
501 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); | |
502 | if ((e0 == true_edge && integer_zerop (arg0)) | |
503 | || (e0 == false_edge && integer_onep (arg0)) | |
504 | || (e1 == true_edge && integer_zerop (arg1)) | |
505 | || (e1 == false_edge && integer_onep (arg1))) | |
506 | cond = fold_build1 (TRUTH_NOT_EXPR, TREE_TYPE (cond), cond); | |
507 | ||
508 | /* Insert our new statements at the end of conditional block before the | |
509 | COND_STMT. */ | |
510 | gsi = gsi_for_stmt (stmt); | |
511 | new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true, | |
512 | GSI_SAME_STMT); | |
513 | ||
514 | if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var))) | |
515 | { | |
516 | new_var2 = create_tmp_var (TREE_TYPE (result), NULL); | |
517 | add_referenced_var (new_var2); | |
518 | new_stmt = gimple_build_assign_with_ops (CONVERT_EXPR, new_var2, | |
519 | new_var, NULL); | |
520 | new_var2 = make_ssa_name (new_var2, new_stmt); | |
521 | gimple_assign_set_lhs (new_stmt, new_var2); | |
522 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); | |
523 | new_var = new_var2; | |
4ee9c684 | 524 | } |
20e5647c | 525 | |
75a70cf9 | 526 | replace_phi_edge_with_variable (cond_bb, e1, phi, new_var); |
902929aa | 527 | |
4ee9c684 | 528 | /* Note that we optimized this PHI. */ |
529 | return true; | |
530 | } | |
531 | ||
0beac6fc | 532 | /* The function value_replacement does the main work of doing the value |
533 | replacement. Return true if the replacement is done. Otherwise return | |
534 | false. | |
535 | BB is the basic block where the replacement is going to be done on. ARG0 | |
dac49aa5 | 536 | is argument 0 from the PHI. Likewise for ARG1. */ |
0beac6fc | 537 | |
538 | static bool | |
33784d89 | 539 | value_replacement (basic_block cond_bb, basic_block middle_bb, |
75a70cf9 | 540 | edge e0, edge e1, gimple phi, |
33784d89 | 541 | tree arg0, tree arg1) |
0beac6fc | 542 | { |
75a70cf9 | 543 | gimple cond; |
0beac6fc | 544 | edge true_edge, false_edge; |
75a70cf9 | 545 | enum tree_code code; |
0beac6fc | 546 | |
547 | /* If the type says honor signed zeros we cannot do this | |
dac49aa5 | 548 | optimization. */ |
0beac6fc | 549 | if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1)))) |
550 | return false; | |
551 | ||
33784d89 | 552 | if (!empty_block_p (middle_bb)) |
0beac6fc | 553 | return false; |
554 | ||
75a70cf9 | 555 | cond = last_stmt (cond_bb); |
556 | code = gimple_cond_code (cond); | |
0beac6fc | 557 | |
558 | /* This transformation is only valid for equality comparisons. */ | |
75a70cf9 | 559 | if (code != NE_EXPR && code != EQ_EXPR) |
0beac6fc | 560 | return false; |
561 | ||
562 | /* We need to know which is the true edge and which is the false | |
563 | edge so that we know if have abs or negative abs. */ | |
33784d89 | 564 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |
0beac6fc | 565 | |
566 | /* At this point we know we have a COND_EXPR with two successors. | |
567 | One successor is BB, the other successor is an empty block which | |
568 | falls through into BB. | |
569 | ||
570 | The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR. | |
571 | ||
572 | There is a single PHI node at the join point (BB) with two arguments. | |
573 | ||
574 | We now need to verify that the two arguments in the PHI node match | |
575 | the two arguments to the equality comparison. */ | |
20e5647c | 576 | |
75a70cf9 | 577 | if ((operand_equal_for_phi_arg_p (arg0, gimple_cond_lhs (cond)) |
578 | && operand_equal_for_phi_arg_p (arg1, gimple_cond_rhs (cond))) | |
579 | || (operand_equal_for_phi_arg_p (arg1, gimple_cond_lhs (cond)) | |
580 | && operand_equal_for_phi_arg_p (arg0, gimple_cond_rhs (cond)))) | |
0beac6fc | 581 | { |
582 | edge e; | |
583 | tree arg; | |
584 | ||
50737d20 | 585 | /* For NE_EXPR, we want to build an assignment result = arg where |
586 | arg is the PHI argument associated with the true edge. For | |
587 | EQ_EXPR we want the PHI argument associated with the false edge. */ | |
75a70cf9 | 588 | e = (code == NE_EXPR ? true_edge : false_edge); |
50737d20 | 589 | |
590 | /* Unfortunately, E may not reach BB (it may instead have gone to | |
591 | OTHER_BLOCK). If that is the case, then we want the single outgoing | |
592 | edge from OTHER_BLOCK which reaches BB and represents the desired | |
593 | path from COND_BLOCK. */ | |
33784d89 | 594 | if (e->dest == middle_bb) |
ea091dfd | 595 | e = single_succ_edge (e->dest); |
50737d20 | 596 | |
597 | /* Now we know the incoming edge to BB that has the argument for the | |
598 | RHS of our new assignment statement. */ | |
33784d89 | 599 | if (e0 == e) |
0beac6fc | 600 | arg = arg0; |
601 | else | |
602 | arg = arg1; | |
603 | ||
a4844041 | 604 | replace_phi_edge_with_variable (cond_bb, e1, phi, arg); |
0beac6fc | 605 | |
606 | /* Note that we optimized this PHI. */ | |
607 | return true; | |
608 | } | |
609 | return false; | |
610 | } | |
611 | ||
194899bf | 612 | /* The function minmax_replacement does the main work of doing the minmax |
613 | replacement. Return true if the replacement is done. Otherwise return | |
614 | false. | |
615 | BB is the basic block where the replacement is going to be done on. ARG0 | |
616 | is argument 0 from the PHI. Likewise for ARG1. */ | |
617 | ||
618 | static bool | |
619 | minmax_replacement (basic_block cond_bb, basic_block middle_bb, | |
75a70cf9 | 620 | edge e0, edge e1, gimple phi, |
194899bf | 621 | tree arg0, tree arg1) |
622 | { | |
623 | tree result, type; | |
75a70cf9 | 624 | gimple cond, new_stmt; |
194899bf | 625 | edge true_edge, false_edge; |
626 | enum tree_code cmp, minmax, ass_code; | |
627 | tree smaller, larger, arg_true, arg_false; | |
75a70cf9 | 628 | gimple_stmt_iterator gsi, gsi_from; |
194899bf | 629 | |
630 | type = TREE_TYPE (PHI_RESULT (phi)); | |
631 | ||
632 | /* The optimization may be unsafe due to NaNs. */ | |
633 | if (HONOR_NANS (TYPE_MODE (type))) | |
634 | return false; | |
635 | ||
75a70cf9 | 636 | cond = last_stmt (cond_bb); |
637 | cmp = gimple_cond_code (cond); | |
194899bf | 638 | result = PHI_RESULT (phi); |
639 | ||
640 | /* This transformation is only valid for order comparisons. Record which | |
641 | operand is smaller/larger if the result of the comparison is true. */ | |
642 | if (cmp == LT_EXPR || cmp == LE_EXPR) | |
643 | { | |
75a70cf9 | 644 | smaller = gimple_cond_lhs (cond); |
645 | larger = gimple_cond_rhs (cond); | |
194899bf | 646 | } |
647 | else if (cmp == GT_EXPR || cmp == GE_EXPR) | |
648 | { | |
75a70cf9 | 649 | smaller = gimple_cond_rhs (cond); |
650 | larger = gimple_cond_lhs (cond); | |
194899bf | 651 | } |
652 | else | |
653 | return false; | |
654 | ||
655 | /* We need to know which is the true edge and which is the false | |
656 | edge so that we know if have abs or negative abs. */ | |
657 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); | |
658 | ||
659 | /* Forward the edges over the middle basic block. */ | |
660 | if (true_edge->dest == middle_bb) | |
661 | true_edge = EDGE_SUCC (true_edge->dest, 0); | |
662 | if (false_edge->dest == middle_bb) | |
663 | false_edge = EDGE_SUCC (false_edge->dest, 0); | |
664 | ||
665 | if (true_edge == e0) | |
666 | { | |
667 | gcc_assert (false_edge == e1); | |
668 | arg_true = arg0; | |
669 | arg_false = arg1; | |
670 | } | |
671 | else | |
672 | { | |
673 | gcc_assert (false_edge == e0); | |
674 | gcc_assert (true_edge == e1); | |
675 | arg_true = arg1; | |
676 | arg_false = arg0; | |
677 | } | |
678 | ||
679 | if (empty_block_p (middle_bb)) | |
680 | { | |
681 | if (operand_equal_for_phi_arg_p (arg_true, smaller) | |
682 | && operand_equal_for_phi_arg_p (arg_false, larger)) | |
683 | { | |
684 | /* Case | |
685 | ||
686 | if (smaller < larger) | |
687 | rslt = smaller; | |
688 | else | |
689 | rslt = larger; */ | |
690 | minmax = MIN_EXPR; | |
691 | } | |
692 | else if (operand_equal_for_phi_arg_p (arg_false, smaller) | |
693 | && operand_equal_for_phi_arg_p (arg_true, larger)) | |
694 | minmax = MAX_EXPR; | |
695 | else | |
696 | return false; | |
697 | } | |
698 | else | |
699 | { | |
700 | /* Recognize the following case, assuming d <= u: | |
701 | ||
702 | if (a <= u) | |
703 | b = MAX (a, d); | |
704 | x = PHI <b, u> | |
705 | ||
706 | This is equivalent to | |
707 | ||
708 | b = MAX (a, d); | |
709 | x = MIN (b, u); */ | |
710 | ||
75a70cf9 | 711 | gimple assign = last_and_only_stmt (middle_bb); |
712 | tree lhs, op0, op1, bound; | |
194899bf | 713 | |
714 | if (!assign | |
75a70cf9 | 715 | || gimple_code (assign) != GIMPLE_ASSIGN) |
194899bf | 716 | return false; |
717 | ||
75a70cf9 | 718 | lhs = gimple_assign_lhs (assign); |
719 | ass_code = gimple_assign_rhs_code (assign); | |
194899bf | 720 | if (ass_code != MAX_EXPR && ass_code != MIN_EXPR) |
721 | return false; | |
75a70cf9 | 722 | op0 = gimple_assign_rhs1 (assign); |
723 | op1 = gimple_assign_rhs2 (assign); | |
194899bf | 724 | |
725 | if (true_edge->src == middle_bb) | |
726 | { | |
727 | /* We got here if the condition is true, i.e., SMALLER < LARGER. */ | |
728 | if (!operand_equal_for_phi_arg_p (lhs, arg_true)) | |
729 | return false; | |
730 | ||
731 | if (operand_equal_for_phi_arg_p (arg_false, larger)) | |
732 | { | |
733 | /* Case | |
734 | ||
735 | if (smaller < larger) | |
736 | { | |
737 | r' = MAX_EXPR (smaller, bound) | |
738 | } | |
739 | r = PHI <r', larger> --> to be turned to MIN_EXPR. */ | |
740 | if (ass_code != MAX_EXPR) | |
741 | return false; | |
742 | ||
743 | minmax = MIN_EXPR; | |
744 | if (operand_equal_for_phi_arg_p (op0, smaller)) | |
745 | bound = op1; | |
746 | else if (operand_equal_for_phi_arg_p (op1, smaller)) | |
747 | bound = op0; | |
748 | else | |
749 | return false; | |
750 | ||
751 | /* We need BOUND <= LARGER. */ | |
49d00087 | 752 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |
753 | bound, larger))) | |
194899bf | 754 | return false; |
755 | } | |
756 | else if (operand_equal_for_phi_arg_p (arg_false, smaller)) | |
757 | { | |
758 | /* Case | |
759 | ||
760 | if (smaller < larger) | |
761 | { | |
762 | r' = MIN_EXPR (larger, bound) | |
763 | } | |
764 | r = PHI <r', smaller> --> to be turned to MAX_EXPR. */ | |
765 | if (ass_code != MIN_EXPR) | |
766 | return false; | |
767 | ||
768 | minmax = MAX_EXPR; | |
769 | if (operand_equal_for_phi_arg_p (op0, larger)) | |
770 | bound = op1; | |
771 | else if (operand_equal_for_phi_arg_p (op1, larger)) | |
772 | bound = op0; | |
773 | else | |
774 | return false; | |
775 | ||
776 | /* We need BOUND >= SMALLER. */ | |
49d00087 | 777 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |
778 | bound, smaller))) | |
194899bf | 779 | return false; |
780 | } | |
781 | else | |
782 | return false; | |
783 | } | |
784 | else | |
785 | { | |
786 | /* We got here if the condition is false, i.e., SMALLER > LARGER. */ | |
787 | if (!operand_equal_for_phi_arg_p (lhs, arg_false)) | |
788 | return false; | |
789 | ||
790 | if (operand_equal_for_phi_arg_p (arg_true, larger)) | |
791 | { | |
792 | /* Case | |
793 | ||
794 | if (smaller > larger) | |
795 | { | |
796 | r' = MIN_EXPR (smaller, bound) | |
797 | } | |
798 | r = PHI <r', larger> --> to be turned to MAX_EXPR. */ | |
799 | if (ass_code != MIN_EXPR) | |
800 | return false; | |
801 | ||
802 | minmax = MAX_EXPR; | |
803 | if (operand_equal_for_phi_arg_p (op0, smaller)) | |
804 | bound = op1; | |
805 | else if (operand_equal_for_phi_arg_p (op1, smaller)) | |
806 | bound = op0; | |
807 | else | |
808 | return false; | |
809 | ||
810 | /* We need BOUND >= LARGER. */ | |
49d00087 | 811 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |
812 | bound, larger))) | |
194899bf | 813 | return false; |
814 | } | |
815 | else if (operand_equal_for_phi_arg_p (arg_true, smaller)) | |
816 | { | |
817 | /* Case | |
818 | ||
819 | if (smaller > larger) | |
820 | { | |
821 | r' = MAX_EXPR (larger, bound) | |
822 | } | |
823 | r = PHI <r', smaller> --> to be turned to MIN_EXPR. */ | |
824 | if (ass_code != MAX_EXPR) | |
825 | return false; | |
826 | ||
827 | minmax = MIN_EXPR; | |
828 | if (operand_equal_for_phi_arg_p (op0, larger)) | |
829 | bound = op1; | |
830 | else if (operand_equal_for_phi_arg_p (op1, larger)) | |
831 | bound = op0; | |
832 | else | |
833 | return false; | |
834 | ||
835 | /* We need BOUND <= SMALLER. */ | |
49d00087 | 836 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |
837 | bound, smaller))) | |
194899bf | 838 | return false; |
839 | } | |
840 | else | |
841 | return false; | |
842 | } | |
843 | ||
844 | /* Move the statement from the middle block. */ | |
75a70cf9 | 845 | gsi = gsi_last_bb (cond_bb); |
846 | gsi_from = gsi_last_bb (middle_bb); | |
847 | gsi_move_before (&gsi_from, &gsi); | |
194899bf | 848 | } |
849 | ||
850 | /* Emit the statement to compute min/max. */ | |
851 | result = duplicate_ssa_name (PHI_RESULT (phi), NULL); | |
75a70cf9 | 852 | new_stmt = gimple_build_assign_with_ops (minmax, result, arg0, arg1); |
853 | gsi = gsi_last_bb (cond_bb); | |
854 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
194899bf | 855 | |
a4844041 | 856 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |
194899bf | 857 | return true; |
858 | } | |
859 | ||
70512b93 | 860 | /* The function absolute_replacement does the main work of doing the absolute |
861 | replacement. Return true if the replacement is done. Otherwise return | |
862 | false. | |
863 | bb is the basic block where the replacement is going to be done on. arg0 | |
f7f07c95 | 864 | is argument 0 from the phi. Likewise for arg1. */ |
33784d89 | 865 | |
70512b93 | 866 | static bool |
33784d89 | 867 | abs_replacement (basic_block cond_bb, basic_block middle_bb, |
a4844041 | 868 | edge e0 ATTRIBUTE_UNUSED, edge e1, |
75a70cf9 | 869 | gimple phi, tree arg0, tree arg1) |
70512b93 | 870 | { |
871 | tree result; | |
75a70cf9 | 872 | gimple new_stmt, cond; |
873 | gimple_stmt_iterator gsi; | |
70512b93 | 874 | edge true_edge, false_edge; |
75a70cf9 | 875 | gimple assign; |
70512b93 | 876 | edge e; |
194899bf | 877 | tree rhs, lhs; |
70512b93 | 878 | bool negate; |
879 | enum tree_code cond_code; | |
880 | ||
881 | /* If the type says honor signed zeros we cannot do this | |
dac49aa5 | 882 | optimization. */ |
70512b93 | 883 | if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1)))) |
884 | return false; | |
885 | ||
70512b93 | 886 | /* OTHER_BLOCK must have only one executable statement which must have the |
887 | form arg0 = -arg1 or arg1 = -arg0. */ | |
70512b93 | 888 | |
194899bf | 889 | assign = last_and_only_stmt (middle_bb); |
70512b93 | 890 | /* If we did not find the proper negation assignment, then we can not |
891 | optimize. */ | |
892 | if (assign == NULL) | |
893 | return false; | |
194899bf | 894 | |
895 | /* If we got here, then we have found the only executable statement | |
896 | in OTHER_BLOCK. If it is anything other than arg = -arg1 or | |
897 | arg1 = -arg0, then we can not optimize. */ | |
75a70cf9 | 898 | if (gimple_code (assign) != GIMPLE_ASSIGN) |
194899bf | 899 | return false; |
900 | ||
75a70cf9 | 901 | lhs = gimple_assign_lhs (assign); |
194899bf | 902 | |
75a70cf9 | 903 | if (gimple_assign_rhs_code (assign) != NEGATE_EXPR) |
194899bf | 904 | return false; |
905 | ||
75a70cf9 | 906 | rhs = gimple_assign_rhs1 (assign); |
194899bf | 907 | |
908 | /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */ | |
909 | if (!(lhs == arg0 && rhs == arg1) | |
910 | && !(lhs == arg1 && rhs == arg0)) | |
911 | return false; | |
70512b93 | 912 | |
75a70cf9 | 913 | cond = last_stmt (cond_bb); |
70512b93 | 914 | result = PHI_RESULT (phi); |
915 | ||
916 | /* Only relationals comparing arg[01] against zero are interesting. */ | |
75a70cf9 | 917 | cond_code = gimple_cond_code (cond); |
70512b93 | 918 | if (cond_code != GT_EXPR && cond_code != GE_EXPR |
919 | && cond_code != LT_EXPR && cond_code != LE_EXPR) | |
920 | return false; | |
921 | ||
dac49aa5 | 922 | /* Make sure the conditional is arg[01] OP y. */ |
75a70cf9 | 923 | if (gimple_cond_lhs (cond) != rhs) |
70512b93 | 924 | return false; |
925 | ||
75a70cf9 | 926 | if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond))) |
927 | ? real_zerop (gimple_cond_rhs (cond)) | |
928 | : integer_zerop (gimple_cond_rhs (cond))) | |
70512b93 | 929 | ; |
930 | else | |
931 | return false; | |
932 | ||
933 | /* We need to know which is the true edge and which is the false | |
934 | edge so that we know if have abs or negative abs. */ | |
33784d89 | 935 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |
70512b93 | 936 | |
937 | /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we | |
938 | will need to negate the result. Similarly for LT_EXPR/LE_EXPR if | |
939 | the false edge goes to OTHER_BLOCK. */ | |
940 | if (cond_code == GT_EXPR || cond_code == GE_EXPR) | |
941 | e = true_edge; | |
942 | else | |
943 | e = false_edge; | |
20e5647c | 944 | |
33784d89 | 945 | if (e->dest == middle_bb) |
70512b93 | 946 | negate = true; |
947 | else | |
948 | negate = false; | |
20e5647c | 949 | |
33784d89 | 950 | result = duplicate_ssa_name (result, NULL); |
20e5647c | 951 | |
70512b93 | 952 | if (negate) |
b2a02a0e | 953 | { |
954 | tree tmp = create_tmp_var (TREE_TYPE (result), NULL); | |
987392e5 | 955 | add_referenced_var (tmp); |
b2a02a0e | 956 | lhs = make_ssa_name (tmp, NULL); |
957 | } | |
70512b93 | 958 | else |
959 | lhs = result; | |
960 | ||
dac49aa5 | 961 | /* Build the modify expression with abs expression. */ |
75a70cf9 | 962 | new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL); |
70512b93 | 963 | |
75a70cf9 | 964 | gsi = gsi_last_bb (cond_bb); |
965 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
70512b93 | 966 | |
967 | if (negate) | |
968 | { | |
75a70cf9 | 969 | /* Get the right GSI. We want to insert after the recently |
70512b93 | 970 | added ABS_EXPR statement (which we know is the first statement |
971 | in the block. */ | |
75a70cf9 | 972 | new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL); |
70512b93 | 973 | |
75a70cf9 | 974 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); |
70512b93 | 975 | } |
20e5647c | 976 | |
a4844041 | 977 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |
70512b93 | 978 | |
979 | /* Note that we optimized this PHI. */ | |
980 | return true; | |
981 | } | |
982 | ||
e6d0e152 | 983 | /* Auxiliary functions to determine the set of memory accesses which |
984 | can't trap because they are preceded by accesses to the same memory | |
985 | portion. We do that for INDIRECT_REFs, so we only need to track | |
986 | the SSA_NAME of the pointer indirectly referenced. The algorithm | |
987 | simply is a walk over all instructions in dominator order. When | |
988 | we see an INDIRECT_REF we determine if we've already seen a same | |
989 | ref anywhere up to the root of the dominator tree. If we do the | |
af4f74fa | 990 | current access can't trap. If we don't see any dominating access |
e6d0e152 | 991 | the current access might trap, but might also make later accesses |
af4f74fa | 992 | non-trapping, so we remember it. We need to be careful with loads |
993 | or stores, for instance a load might not trap, while a store would, | |
994 | so if we see a dominating read access this doesn't mean that a later | |
995 | write access would not trap. Hence we also need to differentiate the | |
996 | type of access(es) seen. | |
997 | ||
998 | ??? We currently are very conservative and assume that a load might | |
999 | trap even if a store doesn't (write-only memory). This probably is | |
1000 | overly conservative. */ | |
e6d0e152 | 1001 | |
1002 | /* A hash-table of SSA_NAMEs, and in which basic block an INDIRECT_REF | |
1003 | through it was seen, which would constitute a no-trap region for | |
1004 | same accesses. */ | |
1005 | struct name_to_bb | |
1006 | { | |
1007 | tree ssa_name; | |
1008 | basic_block bb; | |
af4f74fa | 1009 | unsigned store : 1; |
e6d0e152 | 1010 | }; |
1011 | ||
1012 | /* The hash table for remembering what we've seen. */ | |
1013 | static htab_t seen_ssa_names; | |
1014 | ||
1015 | /* The set of INDIRECT_REFs which can't trap. */ | |
1016 | static struct pointer_set_t *nontrap_set; | |
1017 | ||
1018 | /* The hash function, based on the pointer to the pointer SSA_NAME. */ | |
1019 | static hashval_t | |
1020 | name_to_bb_hash (const void *p) | |
1021 | { | |
f7f3687c | 1022 | const_tree n = ((const struct name_to_bb *)p)->ssa_name; |
1023 | return htab_hash_pointer (n) ^ ((const struct name_to_bb *)p)->store; | |
e6d0e152 | 1024 | } |
1025 | ||
1026 | /* The equality function of *P1 and *P2. SSA_NAMEs are shared, so | |
1027 | it's enough to simply compare them for equality. */ | |
1028 | static int | |
1029 | name_to_bb_eq (const void *p1, const void *p2) | |
1030 | { | |
af4f74fa | 1031 | const struct name_to_bb *n1 = (const struct name_to_bb *)p1; |
1032 | const struct name_to_bb *n2 = (const struct name_to_bb *)p2; | |
e6d0e152 | 1033 | |
af4f74fa | 1034 | return n1->ssa_name == n2->ssa_name && n1->store == n2->store; |
e6d0e152 | 1035 | } |
1036 | ||
f0b5f617 | 1037 | /* We see the expression EXP in basic block BB. If it's an interesting |
e6d0e152 | 1038 | expression (an INDIRECT_REF through an SSA_NAME) possibly insert the |
af4f74fa | 1039 | expression into the set NONTRAP or the hash table of seen expressions. |
1040 | STORE is true if this expression is on the LHS, otherwise it's on | |
1041 | the RHS. */ | |
e6d0e152 | 1042 | static void |
af4f74fa | 1043 | add_or_mark_expr (basic_block bb, tree exp, |
1044 | struct pointer_set_t *nontrap, bool store) | |
e6d0e152 | 1045 | { |
1046 | if (INDIRECT_REF_P (exp) | |
1047 | && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME) | |
1048 | { | |
1049 | tree name = TREE_OPERAND (exp, 0); | |
1050 | struct name_to_bb map; | |
1051 | void **slot; | |
af4f74fa | 1052 | struct name_to_bb *n2bb; |
e6d0e152 | 1053 | basic_block found_bb = 0; |
1054 | ||
1055 | /* Try to find the last seen INDIRECT_REF through the same | |
1056 | SSA_NAME, which can trap. */ | |
1057 | map.ssa_name = name; | |
1058 | map.bb = 0; | |
af4f74fa | 1059 | map.store = store; |
e6d0e152 | 1060 | slot = htab_find_slot (seen_ssa_names, &map, INSERT); |
af4f74fa | 1061 | n2bb = (struct name_to_bb *) *slot; |
1062 | if (n2bb) | |
1063 | found_bb = n2bb->bb; | |
e6d0e152 | 1064 | |
1065 | /* If we've found a trapping INDIRECT_REF, _and_ it dominates EXP | |
1066 | (it's in a basic block on the path from us to the dominator root) | |
1067 | then we can't trap. */ | |
1068 | if (found_bb && found_bb->aux == (void *)1) | |
1069 | { | |
1070 | pointer_set_insert (nontrap, exp); | |
1071 | } | |
1072 | else | |
1073 | { | |
1074 | /* EXP might trap, so insert it into the hash table. */ | |
af4f74fa | 1075 | if (n2bb) |
e6d0e152 | 1076 | { |
af4f74fa | 1077 | n2bb->bb = bb; |
e6d0e152 | 1078 | } |
1079 | else | |
1080 | { | |
af4f74fa | 1081 | n2bb = XNEW (struct name_to_bb); |
1082 | n2bb->ssa_name = name; | |
1083 | n2bb->bb = bb; | |
1084 | n2bb->store = store; | |
1085 | *slot = n2bb; | |
e6d0e152 | 1086 | } |
1087 | } | |
1088 | } | |
1089 | } | |
1090 | ||
1091 | /* Called by walk_dominator_tree, when entering the block BB. */ | |
1092 | static void | |
1093 | nt_init_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb) | |
1094 | { | |
75a70cf9 | 1095 | gimple_stmt_iterator gsi; |
e6d0e152 | 1096 | /* Mark this BB as being on the path to dominator root. */ |
1097 | bb->aux = (void*)1; | |
1098 | ||
1099 | /* And walk the statements in order. */ | |
75a70cf9 | 1100 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
e6d0e152 | 1101 | { |
75a70cf9 | 1102 | gimple stmt = gsi_stmt (gsi); |
e6d0e152 | 1103 | |
75a70cf9 | 1104 | if (is_gimple_assign (stmt)) |
e6d0e152 | 1105 | { |
75a70cf9 | 1106 | add_or_mark_expr (bb, gimple_assign_lhs (stmt), nontrap_set, true); |
1107 | add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), nontrap_set, false); | |
1108 | if (get_gimple_rhs_num_ops (gimple_assign_rhs_code (stmt)) > 1) | |
1109 | add_or_mark_expr (bb, gimple_assign_rhs2 (stmt), nontrap_set, | |
1110 | false); | |
e6d0e152 | 1111 | } |
1112 | } | |
1113 | } | |
1114 | ||
1115 | /* Called by walk_dominator_tree, when basic block BB is exited. */ | |
1116 | static void | |
1117 | nt_fini_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb) | |
1118 | { | |
1119 | /* This BB isn't on the path to dominator root anymore. */ | |
1120 | bb->aux = NULL; | |
1121 | } | |
1122 | ||
1123 | /* This is the entry point of gathering non trapping memory accesses. | |
1124 | It will do a dominator walk over the whole function, and it will | |
1125 | make use of the bb->aux pointers. It returns a set of trees | |
1126 | (the INDIRECT_REFs itself) which can't trap. */ | |
1127 | static struct pointer_set_t * | |
1128 | get_non_trapping (void) | |
1129 | { | |
1130 | struct pointer_set_t *nontrap; | |
1131 | struct dom_walk_data walk_data; | |
1132 | ||
1133 | nontrap = pointer_set_create (); | |
1134 | seen_ssa_names = htab_create (128, name_to_bb_hash, name_to_bb_eq, | |
1135 | free); | |
1136 | /* We're going to do a dominator walk, so ensure that we have | |
1137 | dominance information. */ | |
1138 | calculate_dominance_info (CDI_DOMINATORS); | |
1139 | ||
1140 | /* Setup callbacks for the generic dominator tree walker. */ | |
1141 | nontrap_set = nontrap; | |
1142 | walk_data.walk_stmts_backward = false; | |
1143 | walk_data.dom_direction = CDI_DOMINATORS; | |
1144 | walk_data.initialize_block_local_data = NULL; | |
1145 | walk_data.before_dom_children_before_stmts = nt_init_block; | |
1146 | walk_data.before_dom_children_walk_stmts = NULL; | |
1147 | walk_data.before_dom_children_after_stmts = NULL; | |
1148 | walk_data.after_dom_children_before_stmts = NULL; | |
1149 | walk_data.after_dom_children_walk_stmts = NULL; | |
1150 | walk_data.after_dom_children_after_stmts = nt_fini_block; | |
1151 | walk_data.global_data = NULL; | |
1152 | walk_data.block_local_data_size = 0; | |
1153 | walk_data.interesting_blocks = NULL; | |
1154 | ||
1155 | init_walk_dominator_tree (&walk_data); | |
1156 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
1157 | fini_walk_dominator_tree (&walk_data); | |
1158 | htab_delete (seen_ssa_names); | |
1159 | ||
1160 | return nontrap; | |
1161 | } | |
1162 | ||
1163 | /* Do the main work of conditional store replacement. We already know | |
1164 | that the recognized pattern looks like so: | |
1165 | ||
1166 | split: | |
1167 | if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1) | |
1168 | MIDDLE_BB: | |
1169 | something | |
1170 | fallthrough (edge E0) | |
1171 | JOIN_BB: | |
1172 | some more | |
1173 | ||
1174 | We check that MIDDLE_BB contains only one store, that that store | |
1175 | doesn't trap (not via NOTRAP, but via checking if an access to the same | |
1176 | memory location dominates us) and that the store has a "simple" RHS. */ | |
1177 | ||
1178 | static bool | |
1179 | cond_store_replacement (basic_block middle_bb, basic_block join_bb, | |
1180 | edge e0, edge e1, struct pointer_set_t *nontrap) | |
1181 | { | |
75a70cf9 | 1182 | gimple assign = last_and_only_stmt (middle_bb); |
1183 | tree lhs, rhs, name; | |
1184 | gimple newphi, new_stmt; | |
1185 | gimple_stmt_iterator gsi; | |
1186 | enum tree_code code; | |
e6d0e152 | 1187 | |
1188 | /* Check if middle_bb contains of only one store. */ | |
1189 | if (!assign | |
75a70cf9 | 1190 | || gimple_code (assign) != GIMPLE_ASSIGN) |
e6d0e152 | 1191 | return false; |
1192 | ||
75a70cf9 | 1193 | lhs = gimple_assign_lhs (assign); |
1194 | rhs = gimple_assign_rhs1 (assign); | |
e6d0e152 | 1195 | if (!INDIRECT_REF_P (lhs)) |
1196 | return false; | |
75a70cf9 | 1197 | |
1198 | /* RHS is either a single SSA_NAME or a constant. */ | |
1199 | code = gimple_assign_rhs_code (assign); | |
1200 | if (get_gimple_rhs_class (code) != GIMPLE_SINGLE_RHS | |
1201 | || (code != SSA_NAME && !is_gimple_min_invariant (rhs))) | |
e6d0e152 | 1202 | return false; |
1203 | /* Prove that we can move the store down. We could also check | |
1204 | TREE_THIS_NOTRAP here, but in that case we also could move stores, | |
1205 | whose value is not available readily, which we want to avoid. */ | |
1206 | if (!pointer_set_contains (nontrap, lhs)) | |
1207 | return false; | |
1208 | ||
1209 | /* Now we've checked the constraints, so do the transformation: | |
1210 | 1) Remove the single store. */ | |
1211 | mark_symbols_for_renaming (assign); | |
75a70cf9 | 1212 | gsi = gsi_for_stmt (assign); |
1213 | gsi_remove (&gsi, true); | |
e6d0e152 | 1214 | |
1215 | /* 2) Create a temporary where we can store the old content | |
1216 | of the memory touched by the store, if we need to. */ | |
1217 | if (!condstoretemp || TREE_TYPE (lhs) != TREE_TYPE (condstoretemp)) | |
1218 | { | |
1219 | condstoretemp = create_tmp_var (TREE_TYPE (lhs), "cstore"); | |
1220 | get_var_ann (condstoretemp); | |
0aa073df | 1221 | if (TREE_CODE (TREE_TYPE (lhs)) == COMPLEX_TYPE |
1222 | || TREE_CODE (TREE_TYPE (lhs)) == VECTOR_TYPE) | |
1223 | DECL_GIMPLE_REG_P (condstoretemp) = 1; | |
e6d0e152 | 1224 | } |
1225 | add_referenced_var (condstoretemp); | |
1226 | ||
1227 | /* 3) Insert a load from the memory of the store to the temporary | |
1228 | on the edge which did not contain the store. */ | |
1229 | lhs = unshare_expr (lhs); | |
75a70cf9 | 1230 | new_stmt = gimple_build_assign (condstoretemp, lhs); |
1231 | name = make_ssa_name (condstoretemp, new_stmt); | |
1232 | gimple_assign_set_lhs (new_stmt, name); | |
1233 | mark_symbols_for_renaming (new_stmt); | |
1234 | gsi_insert_on_edge (e1, new_stmt); | |
e6d0e152 | 1235 | |
1236 | /* 4) Create a PHI node at the join block, with one argument | |
1237 | holding the old RHS, and the other holding the temporary | |
1238 | where we stored the old memory contents. */ | |
1239 | newphi = create_phi_node (condstoretemp, join_bb); | |
1240 | add_phi_arg (newphi, rhs, e0); | |
1241 | add_phi_arg (newphi, name, e1); | |
1242 | ||
1243 | lhs = unshare_expr (lhs); | |
75a70cf9 | 1244 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); |
1245 | mark_symbols_for_renaming (new_stmt); | |
e6d0e152 | 1246 | |
1247 | /* 5) Insert that PHI node. */ | |
75a70cf9 | 1248 | gsi = gsi_after_labels (join_bb); |
1249 | if (gsi_end_p (gsi)) | |
e6d0e152 | 1250 | { |
75a70cf9 | 1251 | gsi = gsi_last_bb (join_bb); |
1252 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); | |
e6d0e152 | 1253 | } |
1254 | else | |
75a70cf9 | 1255 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |
e6d0e152 | 1256 | |
1257 | return true; | |
1258 | } | |
4ee9c684 | 1259 | |
1260 | /* Always do these optimizations if we have SSA | |
20e5647c | 1261 | trees to work on. */ |
4ee9c684 | 1262 | static bool |
1263 | gate_phiopt (void) | |
1264 | { | |
1265 | return 1; | |
1266 | } | |
20e5647c | 1267 | |
20099e35 | 1268 | struct gimple_opt_pass pass_phiopt = |
4ee9c684 | 1269 | { |
20099e35 | 1270 | { |
1271 | GIMPLE_PASS, | |
4ee9c684 | 1272 | "phiopt", /* name */ |
1273 | gate_phiopt, /* gate */ | |
1274 | tree_ssa_phiopt, /* execute */ | |
1275 | NULL, /* sub */ | |
1276 | NULL, /* next */ | |
1277 | 0, /* static_pass_number */ | |
1278 | TV_TREE_PHIOPT, /* tv_id */ | |
2f8eb909 | 1279 | PROP_cfg | PROP_ssa, /* properties_required */ |
4ee9c684 | 1280 | 0, /* properties_provided */ |
1281 | 0, /* properties_destroyed */ | |
1282 | 0, /* todo_flags_start */ | |
1e4b21e3 | 1283 | TODO_dump_func |
88dbf20f | 1284 | | TODO_ggc_collect |
1285 | | TODO_verify_ssa | |
88dbf20f | 1286 | | TODO_verify_flow |
20099e35 | 1287 | | TODO_verify_stmts /* todo_flags_finish */ |
1288 | } | |
4ee9c684 | 1289 | }; |
e6d0e152 | 1290 | |
1291 | static bool | |
1292 | gate_cselim (void) | |
1293 | { | |
1294 | return flag_tree_cselim; | |
1295 | } | |
1296 | ||
20099e35 | 1297 | struct gimple_opt_pass pass_cselim = |
e6d0e152 | 1298 | { |
20099e35 | 1299 | { |
1300 | GIMPLE_PASS, | |
e6d0e152 | 1301 | "cselim", /* name */ |
1302 | gate_cselim, /* gate */ | |
1303 | tree_ssa_cs_elim, /* execute */ | |
1304 | NULL, /* sub */ | |
1305 | NULL, /* next */ | |
1306 | 0, /* static_pass_number */ | |
1307 | TV_TREE_PHIOPT, /* tv_id */ | |
2f8eb909 | 1308 | PROP_cfg | PROP_ssa, /* properties_required */ |
e6d0e152 | 1309 | 0, /* properties_provided */ |
1310 | 0, /* properties_destroyed */ | |
1311 | 0, /* todo_flags_start */ | |
1312 | TODO_dump_func | |
1313 | | TODO_ggc_collect | |
1314 | | TODO_verify_ssa | |
1315 | | TODO_verify_flow | |
20099e35 | 1316 | | TODO_verify_stmts /* todo_flags_finish */ |
1317 | } | |
e6d0e152 | 1318 | }; |