]>
Commit | Line | Data |
---|---|---|
4ee9c684 | 1 | /* Optimization of PHI nodes by converting them into straightline code. |
963aee26 | 2 | Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012 |
91cf53d5 | 3 | Free Software Foundation, 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" | |
0beac6fc | 27 | #include "flags.h" |
4ee9c684 | 28 | #include "tm_p.h" |
29 | #include "basic-block.h" | |
4ee9c684 | 30 | #include "tree-flow.h" |
31 | #include "tree-pass.h" | |
4ee9c684 | 32 | #include "langhooks.h" |
e6d0e152 | 33 | #include "pointer-set.h" |
34 | #include "domwalk.h" | |
ec611e12 | 35 | #include "cfgloop.h" |
36 | #include "tree-data-ref.h" | |
239e9670 | 37 | #include "gimple-pretty-print.h" |
38 | #include "insn-config.h" | |
39 | #include "expr.h" | |
40 | #include "optabs.h" | |
41 | ||
42 | #ifndef HAVE_conditional_move | |
43 | #define HAVE_conditional_move (0) | |
44 | #endif | |
4ee9c684 | 45 | |
75a70cf9 | 46 | static unsigned int tree_ssa_phiopt (void); |
239e9670 | 47 | static unsigned int tree_ssa_phiopt_worker (bool, bool); |
a4844041 | 48 | static bool conditional_replacement (basic_block, basic_block, |
75a70cf9 | 49 | edge, edge, gimple, tree, tree); |
fb9912ea | 50 | static int value_replacement (basic_block, basic_block, |
51 | edge, edge, gimple, tree, tree); | |
a4844041 | 52 | static bool minmax_replacement (basic_block, basic_block, |
75a70cf9 | 53 | edge, edge, gimple, tree, tree); |
a4844041 | 54 | static bool abs_replacement (basic_block, basic_block, |
75a70cf9 | 55 | edge, edge, gimple, tree, tree); |
e6d0e152 | 56 | static bool cond_store_replacement (basic_block, basic_block, edge, edge, |
57 | struct pointer_set_t *); | |
91cf53d5 | 58 | static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block); |
e6d0e152 | 59 | static struct pointer_set_t * get_non_trapping (void); |
75a70cf9 | 60 | static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree); |
239e9670 | 61 | static void hoist_adjacent_loads (basic_block, basic_block, |
62 | basic_block, basic_block); | |
63 | static bool gate_hoist_loads (void); | |
902929aa | 64 | |
caac37c2 | 65 | /* This pass tries to replaces an if-then-else block with an |
66 | assignment. We have four kinds of transformations. Some of these | |
67 | transformations are also performed by the ifcvt RTL optimizer. | |
68 | ||
69 | Conditional Replacement | |
70 | ----------------------- | |
71 | ||
30c5ffd2 | 72 | This transformation, implemented in conditional_replacement, |
caac37c2 | 73 | replaces |
4ee9c684 | 74 | |
75 | bb0: | |
76 | if (cond) goto bb2; else goto bb1; | |
77 | bb1: | |
78 | bb2: | |
caac37c2 | 79 | x = PHI <0 (bb1), 1 (bb0), ...>; |
4ee9c684 | 80 | |
caac37c2 | 81 | with |
20e5647c | 82 | |
2ab0a163 | 83 | bb0: |
caac37c2 | 84 | x' = cond; |
85 | goto bb2; | |
2ab0a163 | 86 | bb2: |
caac37c2 | 87 | x = PHI <x' (bb0), ...>; |
4ee9c684 | 88 | |
caac37c2 | 89 | We remove bb1 as it becomes unreachable. This occurs often due to |
90 | gimplification of conditionals. | |
20e5647c | 91 | |
caac37c2 | 92 | Value Replacement |
93 | ----------------- | |
94 | ||
95 | This transformation, implemented in value_replacement, replaces | |
0beac6fc | 96 | |
97 | bb0: | |
caac37c2 | 98 | if (a != b) goto bb2; else goto bb1; |
0beac6fc | 99 | bb1: |
100 | bb2: | |
caac37c2 | 101 | x = PHI <a (bb1), b (bb0), ...>; |
0beac6fc | 102 | |
caac37c2 | 103 | with |
0beac6fc | 104 | |
105 | bb0: | |
0beac6fc | 106 | bb2: |
caac37c2 | 107 | x = PHI <b (bb0), ...>; |
108 | ||
109 | This opportunity can sometimes occur as a result of other | |
110 | optimizations. | |
0beac6fc | 111 | |
caac37c2 | 112 | ABS Replacement |
113 | --------------- | |
70512b93 | 114 | |
caac37c2 | 115 | This transformation, implemented in abs_replacement, replaces |
70512b93 | 116 | |
117 | bb0: | |
caac37c2 | 118 | if (a >= 0) goto bb2; else goto bb1; |
70512b93 | 119 | bb1: |
caac37c2 | 120 | x = -a; |
70512b93 | 121 | bb2: |
caac37c2 | 122 | x = PHI <x (bb1), a (bb0), ...>; |
70512b93 | 123 | |
caac37c2 | 124 | with |
70512b93 | 125 | |
126 | bb0: | |
caac37c2 | 127 | x' = ABS_EXPR< a >; |
70512b93 | 128 | bb2: |
caac37c2 | 129 | x = PHI <x' (bb0), ...>; |
130 | ||
131 | MIN/MAX Replacement | |
132 | ------------------- | |
70512b93 | 133 | |
caac37c2 | 134 | This transformation, minmax_replacement replaces |
194899bf | 135 | |
136 | bb0: | |
caac37c2 | 137 | if (a <= b) goto bb2; else goto bb1; |
194899bf | 138 | bb1: |
194899bf | 139 | bb2: |
caac37c2 | 140 | x = PHI <b (bb1), a (bb0), ...>; |
194899bf | 141 | |
caac37c2 | 142 | with |
194899bf | 143 | |
caac37c2 | 144 | bb0: |
145 | x' = MIN_EXPR (a, b) | |
146 | bb2: | |
147 | x = PHI <x' (bb0), ...>; | |
194899bf | 148 | |
239e9670 | 149 | A similar transformation is done for MAX_EXPR. |
150 | ||
151 | ||
152 | This pass also performs a fifth transformation of a slightly different | |
153 | flavor. | |
154 | ||
155 | Adjacent Load Hoisting | |
156 | ---------------------- | |
157 | ||
158 | This transformation replaces | |
159 | ||
160 | bb0: | |
161 | if (...) goto bb2; else goto bb1; | |
162 | bb1: | |
163 | x1 = (<expr>).field1; | |
164 | goto bb3; | |
165 | bb2: | |
166 | x2 = (<expr>).field2; | |
167 | bb3: | |
168 | # x = PHI <x1, x2>; | |
169 | ||
170 | with | |
171 | ||
172 | bb0: | |
173 | x1 = (<expr>).field1; | |
174 | x2 = (<expr>).field2; | |
175 | if (...) goto bb2; else goto bb1; | |
176 | bb1: | |
177 | goto bb3; | |
178 | bb2: | |
179 | bb3: | |
180 | # x = PHI <x1, x2>; | |
181 | ||
182 | The purpose of this transformation is to enable generation of conditional | |
183 | move instructions such as Intel CMOVE or PowerPC ISEL. Because one of | |
184 | the loads is speculative, the transformation is restricted to very | |
185 | specific cases to avoid introducing a page fault. We are looking for | |
186 | the common idiom: | |
187 | ||
188 | if (...) | |
189 | x = y->left; | |
190 | else | |
191 | x = y->right; | |
192 | ||
193 | where left and right are typically adjacent pointers in a tree structure. */ | |
70512b93 | 194 | |
2a1990e9 | 195 | static unsigned int |
4ee9c684 | 196 | tree_ssa_phiopt (void) |
e6d0e152 | 197 | { |
239e9670 | 198 | return tree_ssa_phiopt_worker (false, gate_hoist_loads ()); |
e6d0e152 | 199 | } |
200 | ||
201 | /* This pass tries to transform conditional stores into unconditional | |
202 | ones, enabling further simplifications with the simpler then and else | |
203 | blocks. In particular it replaces this: | |
204 | ||
205 | bb0: | |
206 | if (cond) goto bb2; else goto bb1; | |
207 | bb1: | |
91cf53d5 | 208 | *p = RHS; |
e6d0e152 | 209 | bb2: |
210 | ||
211 | with | |
212 | ||
213 | bb0: | |
214 | if (cond) goto bb1; else goto bb2; | |
215 | bb1: | |
216 | condtmp' = *p; | |
217 | bb2: | |
218 | condtmp = PHI <RHS, condtmp'> | |
91cf53d5 | 219 | *p = condtmp; |
e6d0e152 | 220 | |
221 | This transformation can only be done under several constraints, | |
91cf53d5 | 222 | documented below. It also replaces: |
223 | ||
224 | bb0: | |
225 | if (cond) goto bb2; else goto bb1; | |
226 | bb1: | |
227 | *p = RHS1; | |
228 | goto bb3; | |
229 | bb2: | |
230 | *p = RHS2; | |
231 | bb3: | |
232 | ||
233 | with | |
234 | ||
235 | bb0: | |
236 | if (cond) goto bb3; else goto bb1; | |
237 | bb1: | |
238 | bb3: | |
239 | condtmp = PHI <RHS1, RHS2> | |
240 | *p = condtmp; */ | |
e6d0e152 | 241 | |
242 | static unsigned int | |
243 | tree_ssa_cs_elim (void) | |
244 | { | |
239e9670 | 245 | return tree_ssa_phiopt_worker (true, false); |
e6d0e152 | 246 | } |
247 | ||
c3597b05 | 248 | /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */ |
249 | ||
250 | static gimple | |
251 | single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1) | |
252 | { | |
253 | gimple_stmt_iterator i; | |
254 | gimple phi = NULL; | |
255 | if (gimple_seq_singleton_p (seq)) | |
256 | return gsi_stmt (gsi_start (seq)); | |
257 | for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i)) | |
258 | { | |
259 | gimple p = gsi_stmt (i); | |
260 | /* If the PHI arguments are equal then we can skip this PHI. */ | |
261 | if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx), | |
262 | gimple_phi_arg_def (p, e1->dest_idx))) | |
263 | continue; | |
264 | ||
265 | /* If we already have a PHI that has the two edge arguments are | |
266 | different, then return it is not a singleton for these PHIs. */ | |
267 | if (phi) | |
268 | return NULL; | |
269 | ||
270 | phi = p; | |
271 | } | |
272 | return phi; | |
273 | } | |
274 | ||
e6d0e152 | 275 | /* The core routine of conditional store replacement and normal |
276 | phi optimizations. Both share much of the infrastructure in how | |
277 | to match applicable basic block patterns. DO_STORE_ELIM is true | |
239e9670 | 278 | when we want to do conditional store replacement, false otherwise. |
279 | DO_HOIST_LOADS is true when we want to hoist adjacent loads out | |
280 | of diamond control flow patterns, false otherwise. */ | |
e6d0e152 | 281 | static unsigned int |
239e9670 | 282 | tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads) |
4ee9c684 | 283 | { |
284 | basic_block bb; | |
194899bf | 285 | basic_block *bb_order; |
286 | unsigned n, i; | |
1e4b21e3 | 287 | bool cfgchanged = false; |
e6d0e152 | 288 | struct pointer_set_t *nontrap = 0; |
289 | ||
290 | if (do_store_elim) | |
03d37e4e | 291 | /* Calculate the set of non-trapping memory accesses. */ |
292 | nontrap = get_non_trapping (); | |
194899bf | 293 | |
294 | /* Search every basic block for COND_EXPR we may be able to optimize. | |
295 | ||
296 | We walk the blocks in order that guarantees that a block with | |
297 | a single predecessor is processed before the predecessor. | |
298 | This ensures that we collapse inner ifs before visiting the | |
299 | outer ones, and also that we do not try to visit a removed | |
300 | block. */ | |
301 | bb_order = blocks_in_phiopt_order (); | |
4d2e5d52 | 302 | n = n_basic_blocks - NUM_FIXED_BLOCKS; |
4ee9c684 | 303 | |
48e1416a | 304 | for (i = 0; i < n; i++) |
4ee9c684 | 305 | { |
75a70cf9 | 306 | gimple cond_stmt, phi; |
33784d89 | 307 | basic_block bb1, bb2; |
308 | edge e1, e2; | |
194899bf | 309 | tree arg0, arg1; |
310 | ||
311 | bb = bb_order[i]; | |
20e5647c | 312 | |
75a70cf9 | 313 | cond_stmt = last_stmt (bb); |
314 | /* Check to see if the last statement is a GIMPLE_COND. */ | |
315 | if (!cond_stmt | |
316 | || gimple_code (cond_stmt) != GIMPLE_COND) | |
33784d89 | 317 | continue; |
20e5647c | 318 | |
33784d89 | 319 | e1 = EDGE_SUCC (bb, 0); |
320 | bb1 = e1->dest; | |
321 | e2 = EDGE_SUCC (bb, 1); | |
322 | bb2 = e2->dest; | |
20e5647c | 323 | |
33784d89 | 324 | /* We cannot do the optimization on abnormal edges. */ |
325 | if ((e1->flags & EDGE_ABNORMAL) != 0 | |
326 | || (e2->flags & EDGE_ABNORMAL) != 0) | |
327 | continue; | |
20e5647c | 328 | |
33784d89 | 329 | /* If either bb1's succ or bb2 or bb2's succ is non NULL. */ |
ea091dfd | 330 | if (EDGE_COUNT (bb1->succs) == 0 |
33784d89 | 331 | || bb2 == NULL |
ea091dfd | 332 | || EDGE_COUNT (bb2->succs) == 0) |
33784d89 | 333 | continue; |
20e5647c | 334 | |
33784d89 | 335 | /* Find the bb which is the fall through to the other. */ |
336 | if (EDGE_SUCC (bb1, 0)->dest == bb2) | |
337 | ; | |
338 | else if (EDGE_SUCC (bb2, 0)->dest == bb1) | |
339 | { | |
340 | basic_block bb_tmp = bb1; | |
341 | edge e_tmp = e1; | |
342 | bb1 = bb2; | |
343 | bb2 = bb_tmp; | |
344 | e1 = e2; | |
345 | e2 = e_tmp; | |
346 | } | |
91cf53d5 | 347 | else if (do_store_elim |
348 | && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) | |
349 | { | |
350 | basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; | |
351 | ||
352 | if (!single_succ_p (bb1) | |
353 | || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0 | |
354 | || !single_succ_p (bb2) | |
355 | || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0 | |
356 | || EDGE_COUNT (bb3->preds) != 2) | |
357 | continue; | |
358 | if (cond_if_else_store_replacement (bb1, bb2, bb3)) | |
359 | cfgchanged = true; | |
360 | continue; | |
361 | } | |
239e9670 | 362 | else if (do_hoist_loads |
363 | && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) | |
364 | { | |
365 | basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; | |
366 | ||
367 | if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt))) | |
368 | && single_succ_p (bb1) | |
369 | && single_succ_p (bb2) | |
370 | && single_pred_p (bb1) | |
371 | && single_pred_p (bb2) | |
372 | && EDGE_COUNT (bb->succs) == 2 | |
373 | && EDGE_COUNT (bb3->preds) == 2 | |
374 | /* If one edge or the other is dominant, a conditional move | |
375 | is likely to perform worse than the well-predicted branch. */ | |
376 | && !predictable_edge_p (EDGE_SUCC (bb, 0)) | |
377 | && !predictable_edge_p (EDGE_SUCC (bb, 1))) | |
378 | hoist_adjacent_loads (bb, bb1, bb2, bb3); | |
379 | continue; | |
380 | } | |
33784d89 | 381 | else |
91cf53d5 | 382 | continue; |
20e5647c | 383 | |
33784d89 | 384 | e1 = EDGE_SUCC (bb1, 0); |
20e5647c | 385 | |
33784d89 | 386 | /* Make sure that bb1 is just a fall through. */ |
db5ba14c | 387 | if (!single_succ_p (bb1) |
33784d89 | 388 | || (e1->flags & EDGE_FALLTHRU) == 0) |
389 | continue; | |
20e5647c | 390 | |
3472707f | 391 | /* Also make sure that bb1 only have one predecessor and that it |
392 | is bb. */ | |
ea091dfd | 393 | if (!single_pred_p (bb1) |
394 | || single_pred (bb1) != bb) | |
33784d89 | 395 | continue; |
20e5647c | 396 | |
e6d0e152 | 397 | if (do_store_elim) |
398 | { | |
399 | /* bb1 is the middle block, bb2 the join block, bb the split block, | |
400 | e1 the fallthrough edge from bb1 to bb2. We can't do the | |
401 | optimization if the join block has more than two predecessors. */ | |
402 | if (EDGE_COUNT (bb2->preds) > 2) | |
403 | continue; | |
404 | if (cond_store_replacement (bb1, bb2, e1, e2, nontrap)) | |
405 | cfgchanged = true; | |
406 | } | |
407 | else | |
408 | { | |
75a70cf9 | 409 | gimple_seq phis = phi_nodes (bb2); |
2109076a | 410 | gimple_stmt_iterator gsi; |
fb9912ea | 411 | bool candorest = true; |
c3597b05 | 412 | |
fb9912ea | 413 | /* Value replacement can work with more than one PHI |
414 | so try that first. */ | |
415 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
416 | { | |
417 | phi = gsi_stmt (gsi); | |
418 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); | |
419 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); | |
420 | if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2) | |
421 | { | |
422 | candorest = false; | |
423 | cfgchanged = true; | |
424 | break; | |
425 | } | |
426 | } | |
e6d0e152 | 427 | |
fb9912ea | 428 | if (!candorest) |
429 | continue; | |
c3597b05 | 430 | |
431 | phi = single_non_singleton_phi_for_edges (phis, e1, e2); | |
2109076a | 432 | if (!phi) |
e6d0e152 | 433 | continue; |
434 | ||
75a70cf9 | 435 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); |
436 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); | |
e6d0e152 | 437 | |
438 | /* Something is wrong if we cannot find the arguments in the PHI | |
439 | node. */ | |
440 | gcc_assert (arg0 != NULL && arg1 != NULL); | |
441 | ||
442 | /* Do the replacement of conditional if it can be done. */ | |
443 | if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
444 | cfgchanged = true; | |
e6d0e152 | 445 | else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) |
446 | cfgchanged = true; | |
447 | else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
448 | cfgchanged = true; | |
449 | } | |
194899bf | 450 | } |
451 | ||
452 | free (bb_order); | |
48e1416a | 453 | |
e6d0e152 | 454 | if (do_store_elim) |
455 | pointer_set_destroy (nontrap); | |
456 | /* If the CFG has changed, we should cleanup the CFG. */ | |
457 | if (cfgchanged && do_store_elim) | |
458 | { | |
459 | /* In cond-store replacement we have added some loads on edges | |
460 | and new VOPS (as we moved the store, and created a load). */ | |
75a70cf9 | 461 | gsi_commit_edge_inserts (); |
e6d0e152 | 462 | return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals; |
463 | } | |
464 | else if (cfgchanged) | |
465 | return TODO_cleanup_cfg; | |
466 | return 0; | |
194899bf | 467 | } |
468 | ||
469 | /* Returns the list of basic blocks in the function in an order that guarantees | |
470 | that if a block X has just a single predecessor Y, then Y is after X in the | |
471 | ordering. */ | |
472 | ||
8530c7be | 473 | basic_block * |
194899bf | 474 | blocks_in_phiopt_order (void) |
475 | { | |
476 | basic_block x, y; | |
945865c5 | 477 | basic_block *order = XNEWVEC (basic_block, n_basic_blocks); |
48e1416a | 478 | unsigned n = n_basic_blocks - NUM_FIXED_BLOCKS; |
4d2e5d52 | 479 | unsigned np, i; |
48e1416a | 480 | sbitmap visited = sbitmap_alloc (last_basic_block); |
194899bf | 481 | |
48e1416a | 482 | #define MARK_VISITED(BB) (SET_BIT (visited, (BB)->index)) |
483 | #define VISITED_P(BB) (TEST_BIT (visited, (BB)->index)) | |
194899bf | 484 | |
485 | sbitmap_zero (visited); | |
486 | ||
487 | MARK_VISITED (ENTRY_BLOCK_PTR); | |
488 | FOR_EACH_BB (x) | |
489 | { | |
490 | if (VISITED_P (x)) | |
491 | continue; | |
492 | ||
493 | /* Walk the predecessors of x as long as they have precisely one | |
494 | predecessor and add them to the list, so that they get stored | |
495 | after x. */ | |
496 | for (y = x, np = 1; | |
497 | single_pred_p (y) && !VISITED_P (single_pred (y)); | |
498 | y = single_pred (y)) | |
499 | np++; | |
500 | for (y = x, i = n - np; | |
501 | single_pred_p (y) && !VISITED_P (single_pred (y)); | |
502 | y = single_pred (y), i++) | |
503 | { | |
504 | order[i] = y; | |
505 | MARK_VISITED (y); | |
2ab0a163 | 506 | } |
194899bf | 507 | order[i] = y; |
508 | MARK_VISITED (y); | |
509 | ||
510 | gcc_assert (i == n - 1); | |
511 | n -= np; | |
4ee9c684 | 512 | } |
194899bf | 513 | |
514 | sbitmap_free (visited); | |
515 | gcc_assert (n == 0); | |
516 | return order; | |
517 | ||
518 | #undef MARK_VISITED | |
519 | #undef VISITED_P | |
4ee9c684 | 520 | } |
521 | ||
47aaf6e6 | 522 | |
70512b93 | 523 | /* Return TRUE if block BB has no executable statements, otherwise return |
524 | FALSE. */ | |
47aaf6e6 | 525 | |
c91e8223 | 526 | bool |
47aaf6e6 | 527 | empty_block_p (basic_block bb) |
70512b93 | 528 | { |
70512b93 | 529 | /* BB must have no executable statements. */ |
9845d120 | 530 | gimple_stmt_iterator gsi = gsi_after_labels (bb); |
c3597b05 | 531 | if (phi_nodes (bb)) |
532 | return false; | |
9845d120 | 533 | if (gsi_end_p (gsi)) |
534 | return true; | |
535 | if (is_gimple_debug (gsi_stmt (gsi))) | |
536 | gsi_next_nondebug (&gsi); | |
537 | return gsi_end_p (gsi); | |
70512b93 | 538 | } |
539 | ||
fccee353 | 540 | /* Replace PHI node element whose edge is E in block BB with variable NEW. |
33784d89 | 541 | Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK |
902929aa | 542 | is known to have two edges, one of which must reach BB). */ |
543 | ||
544 | static void | |
a4844041 | 545 | replace_phi_edge_with_variable (basic_block cond_block, |
75a70cf9 | 546 | edge e, gimple phi, tree new_tree) |
902929aa | 547 | { |
75a70cf9 | 548 | basic_block bb = gimple_bb (phi); |
0e1a77e1 | 549 | basic_block block_to_remove; |
75a70cf9 | 550 | gimple_stmt_iterator gsi; |
33784d89 | 551 | |
20e5647c | 552 | /* Change the PHI argument to new. */ |
f0d6e81c | 553 | SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree); |
0e1a77e1 | 554 | |
0e1a77e1 | 555 | /* Remove the empty basic block. */ |
cd665a06 | 556 | if (EDGE_SUCC (cond_block, 0)->dest == bb) |
902929aa | 557 | { |
cd665a06 | 558 | EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU; |
559 | EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); | |
81c5be57 | 560 | EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE; |
561 | EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count; | |
0e1a77e1 | 562 | |
cd665a06 | 563 | block_to_remove = EDGE_SUCC (cond_block, 1)->dest; |
902929aa | 564 | } |
565 | else | |
566 | { | |
cd665a06 | 567 | EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU; |
568 | EDGE_SUCC (cond_block, 1)->flags | |
902929aa | 569 | &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |
81c5be57 | 570 | EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE; |
571 | EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count; | |
0e1a77e1 | 572 | |
cd665a06 | 573 | block_to_remove = EDGE_SUCC (cond_block, 0)->dest; |
902929aa | 574 | } |
0e1a77e1 | 575 | delete_basic_block (block_to_remove); |
20e5647c | 576 | |
902929aa | 577 | /* Eliminate the COND_EXPR at the end of COND_BLOCK. */ |
75a70cf9 | 578 | gsi = gsi_last_bb (cond_block); |
579 | gsi_remove (&gsi, true); | |
20e5647c | 580 | |
902929aa | 581 | if (dump_file && (dump_flags & TDF_DETAILS)) |
582 | fprintf (dump_file, | |
583 | "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n", | |
584 | cond_block->index, | |
585 | bb->index); | |
586 | } | |
587 | ||
588 | /* The function conditional_replacement does the main work of doing the | |
589 | conditional replacement. Return true if the replacement is done. | |
590 | Otherwise return false. | |
591 | BB is the basic block where the replacement is going to be done on. ARG0 | |
dac49aa5 | 592 | is argument 0 from PHI. Likewise for ARG1. */ |
902929aa | 593 | |
594 | static bool | |
33784d89 | 595 | conditional_replacement (basic_block cond_bb, basic_block middle_bb, |
75a70cf9 | 596 | edge e0, edge e1, gimple phi, |
33784d89 | 597 | tree arg0, tree arg1) |
902929aa | 598 | { |
599 | tree result; | |
75a70cf9 | 600 | gimple stmt, new_stmt; |
601 | tree cond; | |
602 | gimple_stmt_iterator gsi; | |
902929aa | 603 | edge true_edge, false_edge; |
75a70cf9 | 604 | tree new_var, new_var2; |
678919fd | 605 | bool neg; |
902929aa | 606 | |
435e1a75 | 607 | /* FIXME: Gimplification of complex type is too hard for now. */ |
47b88316 | 608 | /* We aren't prepared to handle vectors either (and it is a question |
609 | if it would be worthwhile anyway). */ | |
610 | if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0)) | |
611 | || POINTER_TYPE_P (TREE_TYPE (arg0))) | |
612 | || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1)) | |
613 | || POINTER_TYPE_P (TREE_TYPE (arg1)))) | |
435e1a75 | 614 | return false; |
615 | ||
678919fd | 616 | /* The PHI arguments have the constants 0 and 1, or 0 and -1, then |
617 | convert it to the conditional. */ | |
902929aa | 618 | if ((integer_zerop (arg0) && integer_onep (arg1)) |
619 | || (integer_zerop (arg1) && integer_onep (arg0))) | |
678919fd | 620 | neg = false; |
621 | else if ((integer_zerop (arg0) && integer_all_onesp (arg1)) | |
622 | || (integer_zerop (arg1) && integer_all_onesp (arg0))) | |
623 | neg = true; | |
902929aa | 624 | else |
625 | return false; | |
20e5647c | 626 | |
33784d89 | 627 | if (!empty_block_p (middle_bb)) |
902929aa | 628 | return false; |
20e5647c | 629 | |
75a70cf9 | 630 | /* At this point we know we have a GIMPLE_COND with two successors. |
2ab0a163 | 631 | One successor is BB, the other successor is an empty block which |
632 | falls through into BB. | |
20e5647c | 633 | |
2ab0a163 | 634 | There is a single PHI node at the join point (BB) and its arguments |
678919fd | 635 | are constants (0, 1) or (0, -1). |
20e5647c | 636 | |
2ab0a163 | 637 | So, given the condition COND, and the two PHI arguments, we can |
20e5647c | 638 | rewrite this PHI into non-branching code: |
639 | ||
2ab0a163 | 640 | dest = (COND) or dest = COND' |
20e5647c | 641 | |
2ab0a163 | 642 | We use the condition as-is if the argument associated with the |
643 | true edge has the value one or the argument associated with the | |
644 | false edge as the value zero. Note that those conditions are not | |
75a70cf9 | 645 | the same since only one of the outgoing edges from the GIMPLE_COND |
2ab0a163 | 646 | will directly reach BB and thus be associated with an argument. */ |
ae5a4794 | 647 | |
75a70cf9 | 648 | stmt = last_stmt (cond_bb); |
649 | result = PHI_RESULT (phi); | |
b2a02a0e | 650 | |
75a70cf9 | 651 | /* To handle special cases like floating point comparison, it is easier and |
652 | less error-prone to build a tree and gimplify it on the fly though it is | |
653 | less efficient. */ | |
6f9714b3 | 654 | cond = fold_build2_loc (gimple_location (stmt), |
655 | gimple_cond_code (stmt), boolean_type_node, | |
656 | gimple_cond_lhs (stmt), gimple_cond_rhs (stmt)); | |
4ee9c684 | 657 | |
75a70cf9 | 658 | /* We need to know which is the true edge and which is the false |
659 | edge so that we know when to invert the condition below. */ | |
660 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); | |
661 | if ((e0 == true_edge && integer_zerop (arg0)) | |
678919fd | 662 | || (e0 == false_edge && !integer_zerop (arg0)) |
75a70cf9 | 663 | || (e1 == true_edge && integer_zerop (arg1)) |
678919fd | 664 | || (e1 == false_edge && !integer_zerop (arg1))) |
6f9714b3 | 665 | cond = fold_build1_loc (gimple_location (stmt), |
678919fd | 666 | TRUTH_NOT_EXPR, TREE_TYPE (cond), cond); |
667 | ||
668 | if (neg) | |
669 | { | |
670 | cond = fold_convert_loc (gimple_location (stmt), | |
671 | TREE_TYPE (result), cond); | |
672 | cond = fold_build1_loc (gimple_location (stmt), | |
673 | NEGATE_EXPR, TREE_TYPE (cond), cond); | |
674 | } | |
75a70cf9 | 675 | |
676 | /* Insert our new statements at the end of conditional block before the | |
677 | COND_STMT. */ | |
678 | gsi = gsi_for_stmt (stmt); | |
679 | new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true, | |
680 | GSI_SAME_STMT); | |
681 | ||
682 | if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var))) | |
683 | { | |
efbcb6de | 684 | source_location locus_0, locus_1; |
685 | ||
03d37e4e | 686 | new_var2 = make_ssa_name (TREE_TYPE (result), NULL); |
75a70cf9 | 687 | new_stmt = gimple_build_assign_with_ops (CONVERT_EXPR, new_var2, |
688 | new_var, NULL); | |
75a70cf9 | 689 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |
690 | new_var = new_var2; | |
efbcb6de | 691 | |
692 | /* Set the locus to the first argument, unless is doesn't have one. */ | |
693 | locus_0 = gimple_phi_arg_location (phi, 0); | |
694 | locus_1 = gimple_phi_arg_location (phi, 1); | |
695 | if (locus_0 == UNKNOWN_LOCATION) | |
696 | locus_0 = locus_1; | |
697 | gimple_set_location (new_stmt, locus_0); | |
4ee9c684 | 698 | } |
20e5647c | 699 | |
75a70cf9 | 700 | replace_phi_edge_with_variable (cond_bb, e1, phi, new_var); |
902929aa | 701 | |
4ee9c684 | 702 | /* Note that we optimized this PHI. */ |
703 | return true; | |
704 | } | |
705 | ||
17b9476e | 706 | /* Update *ARG which is defined in STMT so that it contains the |
707 | computed value if that seems profitable. Return true if the | |
708 | statement is made dead by that rewriting. */ | |
709 | ||
710 | static bool | |
711 | jump_function_from_stmt (tree *arg, gimple stmt) | |
712 | { | |
713 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
714 | if (code == ADDR_EXPR) | |
715 | { | |
716 | /* For arg = &p->i transform it to p, if possible. */ | |
717 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
718 | HOST_WIDE_INT offset; | |
719 | tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0), | |
720 | &offset); | |
721 | if (tem | |
722 | && TREE_CODE (tem) == MEM_REF | |
723 | && double_int_zero_p | |
724 | (double_int_add (mem_ref_offset (tem), | |
725 | shwi_to_double_int (offset)))) | |
726 | { | |
727 | *arg = TREE_OPERAND (tem, 0); | |
728 | return true; | |
729 | } | |
730 | } | |
731 | /* TODO: Much like IPA-CP jump-functions we want to handle constant | |
732 | additions symbolically here, and we'd need to update the comparison | |
733 | code that compares the arg + cst tuples in our caller. For now the | |
734 | code above exactly handles the VEC_BASE pattern from vec.h. */ | |
735 | return false; | |
736 | } | |
737 | ||
0beac6fc | 738 | /* The function value_replacement does the main work of doing the value |
fb9912ea | 739 | replacement. Return non-zero if the replacement is done. Otherwise return |
740 | 0. If we remove the middle basic block, return 2. | |
0beac6fc | 741 | BB is the basic block where the replacement is going to be done on. ARG0 |
dac49aa5 | 742 | is argument 0 from the PHI. Likewise for ARG1. */ |
0beac6fc | 743 | |
fb9912ea | 744 | static int |
33784d89 | 745 | value_replacement (basic_block cond_bb, basic_block middle_bb, |
75a70cf9 | 746 | edge e0, edge e1, gimple phi, |
33784d89 | 747 | tree arg0, tree arg1) |
0beac6fc | 748 | { |
17b9476e | 749 | gimple_stmt_iterator gsi; |
75a70cf9 | 750 | gimple cond; |
0beac6fc | 751 | edge true_edge, false_edge; |
75a70cf9 | 752 | enum tree_code code; |
fb9912ea | 753 | bool emtpy_or_with_defined_p = true; |
0beac6fc | 754 | |
755 | /* If the type says honor signed zeros we cannot do this | |
dac49aa5 | 756 | optimization. */ |
0beac6fc | 757 | if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1)))) |
fb9912ea | 758 | return 0; |
0beac6fc | 759 | |
fb9912ea | 760 | /* If there is a statement in MIDDLE_BB that defines one of the PHI |
761 | arguments, then adjust arg0 or arg1. */ | |
17b9476e | 762 | gsi = gsi_after_labels (middle_bb); |
fb9912ea | 763 | if (!gsi_end_p (gsi) && is_gimple_debug (gsi_stmt (gsi))) |
764 | gsi_next_nondebug (&gsi); | |
765 | while (!gsi_end_p (gsi)) | |
17b9476e | 766 | { |
fb9912ea | 767 | gimple stmt = gsi_stmt (gsi); |
768 | tree lhs; | |
769 | gsi_next_nondebug (&gsi); | |
770 | if (!is_gimple_assign (stmt)) | |
17b9476e | 771 | { |
fb9912ea | 772 | emtpy_or_with_defined_p = false; |
773 | continue; | |
17b9476e | 774 | } |
fb9912ea | 775 | /* Now try to adjust arg0 or arg1 according to the computation |
776 | in the statement. */ | |
777 | lhs = gimple_assign_lhs (stmt); | |
778 | if (!(lhs == arg0 | |
779 | && jump_function_from_stmt (&arg0, stmt)) | |
780 | || (lhs == arg1 | |
781 | && jump_function_from_stmt (&arg1, stmt))) | |
782 | emtpy_or_with_defined_p = false; | |
17b9476e | 783 | } |
0beac6fc | 784 | |
75a70cf9 | 785 | cond = last_stmt (cond_bb); |
786 | code = gimple_cond_code (cond); | |
0beac6fc | 787 | |
788 | /* This transformation is only valid for equality comparisons. */ | |
75a70cf9 | 789 | if (code != NE_EXPR && code != EQ_EXPR) |
fb9912ea | 790 | return 0; |
0beac6fc | 791 | |
792 | /* We need to know which is the true edge and which is the false | |
793 | edge so that we know if have abs or negative abs. */ | |
33784d89 | 794 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |
0beac6fc | 795 | |
796 | /* At this point we know we have a COND_EXPR with two successors. | |
797 | One successor is BB, the other successor is an empty block which | |
798 | falls through into BB. | |
799 | ||
800 | The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR. | |
801 | ||
802 | There is a single PHI node at the join point (BB) with two arguments. | |
803 | ||
804 | We now need to verify that the two arguments in the PHI node match | |
805 | the two arguments to the equality comparison. */ | |
20e5647c | 806 | |
75a70cf9 | 807 | if ((operand_equal_for_phi_arg_p (arg0, gimple_cond_lhs (cond)) |
808 | && operand_equal_for_phi_arg_p (arg1, gimple_cond_rhs (cond))) | |
809 | || (operand_equal_for_phi_arg_p (arg1, gimple_cond_lhs (cond)) | |
810 | && operand_equal_for_phi_arg_p (arg0, gimple_cond_rhs (cond)))) | |
0beac6fc | 811 | { |
812 | edge e; | |
813 | tree arg; | |
814 | ||
50737d20 | 815 | /* For NE_EXPR, we want to build an assignment result = arg where |
816 | arg is the PHI argument associated with the true edge. For | |
817 | EQ_EXPR we want the PHI argument associated with the false edge. */ | |
75a70cf9 | 818 | e = (code == NE_EXPR ? true_edge : false_edge); |
50737d20 | 819 | |
820 | /* Unfortunately, E may not reach BB (it may instead have gone to | |
821 | OTHER_BLOCK). If that is the case, then we want the single outgoing | |
822 | edge from OTHER_BLOCK which reaches BB and represents the desired | |
823 | path from COND_BLOCK. */ | |
33784d89 | 824 | if (e->dest == middle_bb) |
ea091dfd | 825 | e = single_succ_edge (e->dest); |
50737d20 | 826 | |
827 | /* Now we know the incoming edge to BB that has the argument for the | |
828 | RHS of our new assignment statement. */ | |
33784d89 | 829 | if (e0 == e) |
0beac6fc | 830 | arg = arg0; |
831 | else | |
832 | arg = arg1; | |
833 | ||
fb9912ea | 834 | /* If the middle basic block was empty or is defining the |
c3597b05 | 835 | PHI arguments and this is a single phi where the args are different |
836 | for the edges e0 and e1 then we can remove the middle basic block. */ | |
fb9912ea | 837 | if (emtpy_or_with_defined_p |
c3597b05 | 838 | && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), |
839 | e0, e1)) | |
fb9912ea | 840 | { |
841 | replace_phi_edge_with_variable (cond_bb, e1, phi, arg); | |
842 | /* Note that we optimized this PHI. */ | |
843 | return 2; | |
844 | } | |
845 | else | |
846 | { | |
847 | /* Replace the PHI arguments with arg. */ | |
848 | SET_PHI_ARG_DEF (phi, e0->dest_idx, arg); | |
849 | SET_PHI_ARG_DEF (phi, e1->dest_idx, arg); | |
850 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
851 | { | |
852 | fprintf (dump_file, "PHI "); | |
853 | print_generic_expr (dump_file, gimple_phi_result (phi), 0); | |
c3597b05 | 854 | fprintf (dump_file, " reduced for COND_EXPR in block %d to ", |
855 | cond_bb->index); | |
fb9912ea | 856 | print_generic_expr (dump_file, arg, 0); |
857 | fprintf (dump_file, ".\n"); | |
858 | } | |
859 | return 1; | |
860 | } | |
0beac6fc | 861 | |
0beac6fc | 862 | } |
fb9912ea | 863 | return 0; |
0beac6fc | 864 | } |
865 | ||
194899bf | 866 | /* The function minmax_replacement does the main work of doing the minmax |
867 | replacement. Return true if the replacement is done. Otherwise return | |
868 | false. | |
869 | BB is the basic block where the replacement is going to be done on. ARG0 | |
870 | is argument 0 from the PHI. Likewise for ARG1. */ | |
871 | ||
872 | static bool | |
873 | minmax_replacement (basic_block cond_bb, basic_block middle_bb, | |
75a70cf9 | 874 | edge e0, edge e1, gimple phi, |
194899bf | 875 | tree arg0, tree arg1) |
876 | { | |
877 | tree result, type; | |
75a70cf9 | 878 | gimple cond, new_stmt; |
194899bf | 879 | edge true_edge, false_edge; |
880 | enum tree_code cmp, minmax, ass_code; | |
881 | tree smaller, larger, arg_true, arg_false; | |
75a70cf9 | 882 | gimple_stmt_iterator gsi, gsi_from; |
194899bf | 883 | |
884 | type = TREE_TYPE (PHI_RESULT (phi)); | |
885 | ||
886 | /* The optimization may be unsafe due to NaNs. */ | |
887 | if (HONOR_NANS (TYPE_MODE (type))) | |
888 | return false; | |
889 | ||
75a70cf9 | 890 | cond = last_stmt (cond_bb); |
891 | cmp = gimple_cond_code (cond); | |
194899bf | 892 | |
893 | /* This transformation is only valid for order comparisons. Record which | |
894 | operand is smaller/larger if the result of the comparison is true. */ | |
895 | if (cmp == LT_EXPR || cmp == LE_EXPR) | |
896 | { | |
75a70cf9 | 897 | smaller = gimple_cond_lhs (cond); |
898 | larger = gimple_cond_rhs (cond); | |
194899bf | 899 | } |
900 | else if (cmp == GT_EXPR || cmp == GE_EXPR) | |
901 | { | |
75a70cf9 | 902 | smaller = gimple_cond_rhs (cond); |
903 | larger = gimple_cond_lhs (cond); | |
194899bf | 904 | } |
905 | else | |
906 | return false; | |
907 | ||
908 | /* We need to know which is the true edge and which is the false | |
909 | edge so that we know if have abs or negative abs. */ | |
910 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); | |
911 | ||
912 | /* Forward the edges over the middle basic block. */ | |
913 | if (true_edge->dest == middle_bb) | |
914 | true_edge = EDGE_SUCC (true_edge->dest, 0); | |
915 | if (false_edge->dest == middle_bb) | |
916 | false_edge = EDGE_SUCC (false_edge->dest, 0); | |
917 | ||
918 | if (true_edge == e0) | |
919 | { | |
920 | gcc_assert (false_edge == e1); | |
921 | arg_true = arg0; | |
922 | arg_false = arg1; | |
923 | } | |
924 | else | |
925 | { | |
926 | gcc_assert (false_edge == e0); | |
927 | gcc_assert (true_edge == e1); | |
928 | arg_true = arg1; | |
929 | arg_false = arg0; | |
930 | } | |
931 | ||
932 | if (empty_block_p (middle_bb)) | |
933 | { | |
934 | if (operand_equal_for_phi_arg_p (arg_true, smaller) | |
935 | && operand_equal_for_phi_arg_p (arg_false, larger)) | |
936 | { | |
937 | /* Case | |
48e1416a | 938 | |
194899bf | 939 | if (smaller < larger) |
940 | rslt = smaller; | |
941 | else | |
942 | rslt = larger; */ | |
943 | minmax = MIN_EXPR; | |
944 | } | |
945 | else if (operand_equal_for_phi_arg_p (arg_false, smaller) | |
946 | && operand_equal_for_phi_arg_p (arg_true, larger)) | |
947 | minmax = MAX_EXPR; | |
948 | else | |
949 | return false; | |
950 | } | |
951 | else | |
952 | { | |
953 | /* Recognize the following case, assuming d <= u: | |
954 | ||
955 | if (a <= u) | |
956 | b = MAX (a, d); | |
957 | x = PHI <b, u> | |
958 | ||
959 | This is equivalent to | |
960 | ||
961 | b = MAX (a, d); | |
962 | x = MIN (b, u); */ | |
963 | ||
75a70cf9 | 964 | gimple assign = last_and_only_stmt (middle_bb); |
965 | tree lhs, op0, op1, bound; | |
194899bf | 966 | |
967 | if (!assign | |
75a70cf9 | 968 | || gimple_code (assign) != GIMPLE_ASSIGN) |
194899bf | 969 | return false; |
970 | ||
75a70cf9 | 971 | lhs = gimple_assign_lhs (assign); |
972 | ass_code = gimple_assign_rhs_code (assign); | |
194899bf | 973 | if (ass_code != MAX_EXPR && ass_code != MIN_EXPR) |
974 | return false; | |
75a70cf9 | 975 | op0 = gimple_assign_rhs1 (assign); |
976 | op1 = gimple_assign_rhs2 (assign); | |
194899bf | 977 | |
978 | if (true_edge->src == middle_bb) | |
979 | { | |
980 | /* We got here if the condition is true, i.e., SMALLER < LARGER. */ | |
981 | if (!operand_equal_for_phi_arg_p (lhs, arg_true)) | |
982 | return false; | |
983 | ||
984 | if (operand_equal_for_phi_arg_p (arg_false, larger)) | |
985 | { | |
986 | /* Case | |
987 | ||
988 | if (smaller < larger) | |
989 | { | |
990 | r' = MAX_EXPR (smaller, bound) | |
991 | } | |
992 | r = PHI <r', larger> --> to be turned to MIN_EXPR. */ | |
993 | if (ass_code != MAX_EXPR) | |
994 | return false; | |
995 | ||
996 | minmax = MIN_EXPR; | |
997 | if (operand_equal_for_phi_arg_p (op0, smaller)) | |
998 | bound = op1; | |
999 | else if (operand_equal_for_phi_arg_p (op1, smaller)) | |
1000 | bound = op0; | |
1001 | else | |
1002 | return false; | |
1003 | ||
1004 | /* We need BOUND <= LARGER. */ | |
49d00087 | 1005 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |
1006 | bound, larger))) | |
194899bf | 1007 | return false; |
1008 | } | |
1009 | else if (operand_equal_for_phi_arg_p (arg_false, smaller)) | |
1010 | { | |
1011 | /* Case | |
1012 | ||
1013 | if (smaller < larger) | |
1014 | { | |
1015 | r' = MIN_EXPR (larger, bound) | |
1016 | } | |
1017 | r = PHI <r', smaller> --> to be turned to MAX_EXPR. */ | |
1018 | if (ass_code != MIN_EXPR) | |
1019 | return false; | |
1020 | ||
1021 | minmax = MAX_EXPR; | |
1022 | if (operand_equal_for_phi_arg_p (op0, larger)) | |
1023 | bound = op1; | |
1024 | else if (operand_equal_for_phi_arg_p (op1, larger)) | |
1025 | bound = op0; | |
1026 | else | |
1027 | return false; | |
1028 | ||
1029 | /* We need BOUND >= SMALLER. */ | |
49d00087 | 1030 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |
1031 | bound, smaller))) | |
194899bf | 1032 | return false; |
1033 | } | |
1034 | else | |
1035 | return false; | |
1036 | } | |
1037 | else | |
1038 | { | |
1039 | /* We got here if the condition is false, i.e., SMALLER > LARGER. */ | |
1040 | if (!operand_equal_for_phi_arg_p (lhs, arg_false)) | |
1041 | return false; | |
1042 | ||
1043 | if (operand_equal_for_phi_arg_p (arg_true, larger)) | |
1044 | { | |
1045 | /* Case | |
1046 | ||
1047 | if (smaller > larger) | |
1048 | { | |
1049 | r' = MIN_EXPR (smaller, bound) | |
1050 | } | |
1051 | r = PHI <r', larger> --> to be turned to MAX_EXPR. */ | |
1052 | if (ass_code != MIN_EXPR) | |
1053 | return false; | |
1054 | ||
1055 | minmax = MAX_EXPR; | |
1056 | if (operand_equal_for_phi_arg_p (op0, smaller)) | |
1057 | bound = op1; | |
1058 | else if (operand_equal_for_phi_arg_p (op1, smaller)) | |
1059 | bound = op0; | |
1060 | else | |
1061 | return false; | |
1062 | ||
1063 | /* We need BOUND >= LARGER. */ | |
49d00087 | 1064 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |
1065 | bound, larger))) | |
194899bf | 1066 | return false; |
1067 | } | |
1068 | else if (operand_equal_for_phi_arg_p (arg_true, smaller)) | |
1069 | { | |
1070 | /* Case | |
1071 | ||
1072 | if (smaller > larger) | |
1073 | { | |
1074 | r' = MAX_EXPR (larger, bound) | |
1075 | } | |
1076 | r = PHI <r', smaller> --> to be turned to MIN_EXPR. */ | |
1077 | if (ass_code != MAX_EXPR) | |
1078 | return false; | |
1079 | ||
1080 | minmax = MIN_EXPR; | |
1081 | if (operand_equal_for_phi_arg_p (op0, larger)) | |
1082 | bound = op1; | |
1083 | else if (operand_equal_for_phi_arg_p (op1, larger)) | |
1084 | bound = op0; | |
1085 | else | |
1086 | return false; | |
1087 | ||
1088 | /* We need BOUND <= SMALLER. */ | |
49d00087 | 1089 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |
1090 | bound, smaller))) | |
194899bf | 1091 | return false; |
1092 | } | |
1093 | else | |
1094 | return false; | |
1095 | } | |
1096 | ||
1097 | /* Move the statement from the middle block. */ | |
75a70cf9 | 1098 | gsi = gsi_last_bb (cond_bb); |
445a6ba5 | 1099 | gsi_from = gsi_last_nondebug_bb (middle_bb); |
75a70cf9 | 1100 | gsi_move_before (&gsi_from, &gsi); |
194899bf | 1101 | } |
1102 | ||
1103 | /* Emit the statement to compute min/max. */ | |
1104 | result = duplicate_ssa_name (PHI_RESULT (phi), NULL); | |
75a70cf9 | 1105 | new_stmt = gimple_build_assign_with_ops (minmax, result, arg0, arg1); |
1106 | gsi = gsi_last_bb (cond_bb); | |
1107 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
194899bf | 1108 | |
a4844041 | 1109 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |
194899bf | 1110 | return true; |
1111 | } | |
1112 | ||
70512b93 | 1113 | /* The function absolute_replacement does the main work of doing the absolute |
1114 | replacement. Return true if the replacement is done. Otherwise return | |
1115 | false. | |
1116 | bb is the basic block where the replacement is going to be done on. arg0 | |
f7f07c95 | 1117 | is argument 0 from the phi. Likewise for arg1. */ |
33784d89 | 1118 | |
70512b93 | 1119 | static bool |
33784d89 | 1120 | abs_replacement (basic_block cond_bb, basic_block middle_bb, |
a4844041 | 1121 | edge e0 ATTRIBUTE_UNUSED, edge e1, |
75a70cf9 | 1122 | gimple phi, tree arg0, tree arg1) |
70512b93 | 1123 | { |
1124 | tree result; | |
75a70cf9 | 1125 | gimple new_stmt, cond; |
1126 | gimple_stmt_iterator gsi; | |
70512b93 | 1127 | edge true_edge, false_edge; |
75a70cf9 | 1128 | gimple assign; |
70512b93 | 1129 | edge e; |
194899bf | 1130 | tree rhs, lhs; |
70512b93 | 1131 | bool negate; |
1132 | enum tree_code cond_code; | |
1133 | ||
1134 | /* If the type says honor signed zeros we cannot do this | |
dac49aa5 | 1135 | optimization. */ |
70512b93 | 1136 | if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1)))) |
1137 | return false; | |
1138 | ||
70512b93 | 1139 | /* OTHER_BLOCK must have only one executable statement which must have the |
1140 | form arg0 = -arg1 or arg1 = -arg0. */ | |
70512b93 | 1141 | |
194899bf | 1142 | assign = last_and_only_stmt (middle_bb); |
70512b93 | 1143 | /* If we did not find the proper negation assignment, then we can not |
1144 | optimize. */ | |
1145 | if (assign == NULL) | |
1146 | return false; | |
48e1416a | 1147 | |
194899bf | 1148 | /* If we got here, then we have found the only executable statement |
1149 | in OTHER_BLOCK. If it is anything other than arg = -arg1 or | |
1150 | arg1 = -arg0, then we can not optimize. */ | |
75a70cf9 | 1151 | if (gimple_code (assign) != GIMPLE_ASSIGN) |
194899bf | 1152 | return false; |
1153 | ||
75a70cf9 | 1154 | lhs = gimple_assign_lhs (assign); |
194899bf | 1155 | |
75a70cf9 | 1156 | if (gimple_assign_rhs_code (assign) != NEGATE_EXPR) |
194899bf | 1157 | return false; |
1158 | ||
75a70cf9 | 1159 | rhs = gimple_assign_rhs1 (assign); |
48e1416a | 1160 | |
194899bf | 1161 | /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */ |
1162 | if (!(lhs == arg0 && rhs == arg1) | |
1163 | && !(lhs == arg1 && rhs == arg0)) | |
1164 | return false; | |
70512b93 | 1165 | |
75a70cf9 | 1166 | cond = last_stmt (cond_bb); |
70512b93 | 1167 | result = PHI_RESULT (phi); |
1168 | ||
1169 | /* Only relationals comparing arg[01] against zero are interesting. */ | |
75a70cf9 | 1170 | cond_code = gimple_cond_code (cond); |
70512b93 | 1171 | if (cond_code != GT_EXPR && cond_code != GE_EXPR |
1172 | && cond_code != LT_EXPR && cond_code != LE_EXPR) | |
1173 | return false; | |
1174 | ||
dac49aa5 | 1175 | /* Make sure the conditional is arg[01] OP y. */ |
75a70cf9 | 1176 | if (gimple_cond_lhs (cond) != rhs) |
70512b93 | 1177 | return false; |
1178 | ||
75a70cf9 | 1179 | if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond))) |
1180 | ? real_zerop (gimple_cond_rhs (cond)) | |
1181 | : integer_zerop (gimple_cond_rhs (cond))) | |
70512b93 | 1182 | ; |
1183 | else | |
1184 | return false; | |
1185 | ||
1186 | /* We need to know which is the true edge and which is the false | |
1187 | edge so that we know if have abs or negative abs. */ | |
33784d89 | 1188 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |
70512b93 | 1189 | |
1190 | /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we | |
1191 | will need to negate the result. Similarly for LT_EXPR/LE_EXPR if | |
1192 | the false edge goes to OTHER_BLOCK. */ | |
1193 | if (cond_code == GT_EXPR || cond_code == GE_EXPR) | |
1194 | e = true_edge; | |
1195 | else | |
1196 | e = false_edge; | |
20e5647c | 1197 | |
33784d89 | 1198 | if (e->dest == middle_bb) |
70512b93 | 1199 | negate = true; |
1200 | else | |
1201 | negate = false; | |
20e5647c | 1202 | |
33784d89 | 1203 | result = duplicate_ssa_name (result, NULL); |
20e5647c | 1204 | |
70512b93 | 1205 | if (negate) |
03d37e4e | 1206 | lhs = make_ssa_name (TREE_TYPE (result), NULL); |
70512b93 | 1207 | else |
1208 | lhs = result; | |
1209 | ||
dac49aa5 | 1210 | /* Build the modify expression with abs expression. */ |
75a70cf9 | 1211 | new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL); |
70512b93 | 1212 | |
75a70cf9 | 1213 | gsi = gsi_last_bb (cond_bb); |
1214 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
70512b93 | 1215 | |
1216 | if (negate) | |
1217 | { | |
75a70cf9 | 1218 | /* Get the right GSI. We want to insert after the recently |
70512b93 | 1219 | added ABS_EXPR statement (which we know is the first statement |
1220 | in the block. */ | |
75a70cf9 | 1221 | new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL); |
70512b93 | 1222 | |
75a70cf9 | 1223 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); |
70512b93 | 1224 | } |
20e5647c | 1225 | |
a4844041 | 1226 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |
70512b93 | 1227 | |
1228 | /* Note that we optimized this PHI. */ | |
1229 | return true; | |
1230 | } | |
1231 | ||
e6d0e152 | 1232 | /* Auxiliary functions to determine the set of memory accesses which |
1233 | can't trap because they are preceded by accesses to the same memory | |
182cf5a9 | 1234 | portion. We do that for MEM_REFs, so we only need to track |
e6d0e152 | 1235 | the SSA_NAME of the pointer indirectly referenced. The algorithm |
1236 | simply is a walk over all instructions in dominator order. When | |
182cf5a9 | 1237 | we see an MEM_REF we determine if we've already seen a same |
e6d0e152 | 1238 | ref anywhere up to the root of the dominator tree. If we do the |
af4f74fa | 1239 | current access can't trap. If we don't see any dominating access |
e6d0e152 | 1240 | the current access might trap, but might also make later accesses |
af4f74fa | 1241 | non-trapping, so we remember it. We need to be careful with loads |
1242 | or stores, for instance a load might not trap, while a store would, | |
1243 | so if we see a dominating read access this doesn't mean that a later | |
1244 | write access would not trap. Hence we also need to differentiate the | |
1245 | type of access(es) seen. | |
1246 | ||
1247 | ??? We currently are very conservative and assume that a load might | |
1248 | trap even if a store doesn't (write-only memory). This probably is | |
1249 | overly conservative. */ | |
e6d0e152 | 1250 | |
182cf5a9 | 1251 | /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF |
e6d0e152 | 1252 | through it was seen, which would constitute a no-trap region for |
1253 | same accesses. */ | |
1254 | struct name_to_bb | |
1255 | { | |
963aee26 | 1256 | unsigned int ssa_name_ver; |
1257 | bool store; | |
1258 | HOST_WIDE_INT offset, size; | |
e6d0e152 | 1259 | basic_block bb; |
1260 | }; | |
1261 | ||
1262 | /* The hash table for remembering what we've seen. */ | |
1263 | static htab_t seen_ssa_names; | |
1264 | ||
182cf5a9 | 1265 | /* The set of MEM_REFs which can't trap. */ |
e6d0e152 | 1266 | static struct pointer_set_t *nontrap_set; |
1267 | ||
963aee26 | 1268 | /* The hash function. */ |
e6d0e152 | 1269 | static hashval_t |
1270 | name_to_bb_hash (const void *p) | |
1271 | { | |
963aee26 | 1272 | const struct name_to_bb *n = (const struct name_to_bb *) p; |
1273 | return n->ssa_name_ver ^ (((hashval_t) n->store) << 31) | |
1274 | ^ (n->offset << 6) ^ (n->size << 3); | |
e6d0e152 | 1275 | } |
1276 | ||
963aee26 | 1277 | /* The equality function of *P1 and *P2. */ |
e6d0e152 | 1278 | static int |
1279 | name_to_bb_eq (const void *p1, const void *p2) | |
1280 | { | |
af4f74fa | 1281 | const struct name_to_bb *n1 = (const struct name_to_bb *)p1; |
1282 | const struct name_to_bb *n2 = (const struct name_to_bb *)p2; | |
e6d0e152 | 1283 | |
963aee26 | 1284 | return n1->ssa_name_ver == n2->ssa_name_ver |
1285 | && n1->store == n2->store | |
1286 | && n1->offset == n2->offset | |
1287 | && n1->size == n2->size; | |
e6d0e152 | 1288 | } |
1289 | ||
f0b5f617 | 1290 | /* We see the expression EXP in basic block BB. If it's an interesting |
182cf5a9 | 1291 | expression (an MEM_REF through an SSA_NAME) possibly insert the |
af4f74fa | 1292 | expression into the set NONTRAP or the hash table of seen expressions. |
1293 | STORE is true if this expression is on the LHS, otherwise it's on | |
1294 | the RHS. */ | |
e6d0e152 | 1295 | static void |
af4f74fa | 1296 | add_or_mark_expr (basic_block bb, tree exp, |
1297 | struct pointer_set_t *nontrap, bool store) | |
e6d0e152 | 1298 | { |
963aee26 | 1299 | HOST_WIDE_INT size; |
1300 | ||
182cf5a9 | 1301 | if (TREE_CODE (exp) == MEM_REF |
963aee26 | 1302 | && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME |
1303 | && host_integerp (TREE_OPERAND (exp, 1), 0) | |
1304 | && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0) | |
e6d0e152 | 1305 | { |
1306 | tree name = TREE_OPERAND (exp, 0); | |
1307 | struct name_to_bb map; | |
1308 | void **slot; | |
af4f74fa | 1309 | struct name_to_bb *n2bb; |
e6d0e152 | 1310 | basic_block found_bb = 0; |
1311 | ||
182cf5a9 | 1312 | /* Try to find the last seen MEM_REF through the same |
e6d0e152 | 1313 | SSA_NAME, which can trap. */ |
963aee26 | 1314 | map.ssa_name_ver = SSA_NAME_VERSION (name); |
e6d0e152 | 1315 | map.bb = 0; |
af4f74fa | 1316 | map.store = store; |
963aee26 | 1317 | map.offset = tree_low_cst (TREE_OPERAND (exp, 1), 0); |
1318 | map.size = size; | |
1319 | ||
e6d0e152 | 1320 | slot = htab_find_slot (seen_ssa_names, &map, INSERT); |
af4f74fa | 1321 | n2bb = (struct name_to_bb *) *slot; |
1322 | if (n2bb) | |
1323 | found_bb = n2bb->bb; | |
e6d0e152 | 1324 | |
182cf5a9 | 1325 | /* If we've found a trapping MEM_REF, _and_ it dominates EXP |
e6d0e152 | 1326 | (it's in a basic block on the path from us to the dominator root) |
1327 | then we can't trap. */ | |
1328 | if (found_bb && found_bb->aux == (void *)1) | |
1329 | { | |
1330 | pointer_set_insert (nontrap, exp); | |
1331 | } | |
1332 | else | |
1333 | { | |
1334 | /* EXP might trap, so insert it into the hash table. */ | |
af4f74fa | 1335 | if (n2bb) |
e6d0e152 | 1336 | { |
af4f74fa | 1337 | n2bb->bb = bb; |
e6d0e152 | 1338 | } |
1339 | else | |
1340 | { | |
af4f74fa | 1341 | n2bb = XNEW (struct name_to_bb); |
963aee26 | 1342 | n2bb->ssa_name_ver = SSA_NAME_VERSION (name); |
af4f74fa | 1343 | n2bb->bb = bb; |
1344 | n2bb->store = store; | |
963aee26 | 1345 | n2bb->offset = map.offset; |
1346 | n2bb->size = size; | |
af4f74fa | 1347 | *slot = n2bb; |
e6d0e152 | 1348 | } |
1349 | } | |
1350 | } | |
1351 | } | |
1352 | ||
1353 | /* Called by walk_dominator_tree, when entering the block BB. */ | |
1354 | static void | |
1355 | nt_init_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb) | |
1356 | { | |
75a70cf9 | 1357 | gimple_stmt_iterator gsi; |
e6d0e152 | 1358 | /* Mark this BB as being on the path to dominator root. */ |
1359 | bb->aux = (void*)1; | |
1360 | ||
1361 | /* And walk the statements in order. */ | |
75a70cf9 | 1362 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
e6d0e152 | 1363 | { |
75a70cf9 | 1364 | gimple stmt = gsi_stmt (gsi); |
e6d0e152 | 1365 | |
963aee26 | 1366 | if (gimple_assign_single_p (stmt)) |
e6d0e152 | 1367 | { |
75a70cf9 | 1368 | add_or_mark_expr (bb, gimple_assign_lhs (stmt), nontrap_set, true); |
1369 | add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), nontrap_set, false); | |
e6d0e152 | 1370 | } |
1371 | } | |
1372 | } | |
1373 | ||
1374 | /* Called by walk_dominator_tree, when basic block BB is exited. */ | |
1375 | static void | |
1376 | nt_fini_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb) | |
1377 | { | |
1378 | /* This BB isn't on the path to dominator root anymore. */ | |
1379 | bb->aux = NULL; | |
1380 | } | |
1381 | ||
1382 | /* This is the entry point of gathering non trapping memory accesses. | |
1383 | It will do a dominator walk over the whole function, and it will | |
1384 | make use of the bb->aux pointers. It returns a set of trees | |
182cf5a9 | 1385 | (the MEM_REFs itself) which can't trap. */ |
e6d0e152 | 1386 | static struct pointer_set_t * |
1387 | get_non_trapping (void) | |
1388 | { | |
1389 | struct pointer_set_t *nontrap; | |
1390 | struct dom_walk_data walk_data; | |
1391 | ||
1392 | nontrap = pointer_set_create (); | |
1393 | seen_ssa_names = htab_create (128, name_to_bb_hash, name_to_bb_eq, | |
1394 | free); | |
1395 | /* We're going to do a dominator walk, so ensure that we have | |
1396 | dominance information. */ | |
1397 | calculate_dominance_info (CDI_DOMINATORS); | |
1398 | ||
1399 | /* Setup callbacks for the generic dominator tree walker. */ | |
1400 | nontrap_set = nontrap; | |
e6d0e152 | 1401 | walk_data.dom_direction = CDI_DOMINATORS; |
1402 | walk_data.initialize_block_local_data = NULL; | |
6bf320fb | 1403 | walk_data.before_dom_children = nt_init_block; |
1404 | walk_data.after_dom_children = nt_fini_block; | |
e6d0e152 | 1405 | walk_data.global_data = NULL; |
1406 | walk_data.block_local_data_size = 0; | |
e6d0e152 | 1407 | |
1408 | init_walk_dominator_tree (&walk_data); | |
1409 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
1410 | fini_walk_dominator_tree (&walk_data); | |
1411 | htab_delete (seen_ssa_names); | |
1412 | ||
1413 | return nontrap; | |
1414 | } | |
1415 | ||
1416 | /* Do the main work of conditional store replacement. We already know | |
1417 | that the recognized pattern looks like so: | |
1418 | ||
1419 | split: | |
1420 | if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1) | |
1421 | MIDDLE_BB: | |
1422 | something | |
1423 | fallthrough (edge E0) | |
1424 | JOIN_BB: | |
1425 | some more | |
1426 | ||
1427 | We check that MIDDLE_BB contains only one store, that that store | |
1428 | doesn't trap (not via NOTRAP, but via checking if an access to the same | |
1429 | memory location dominates us) and that the store has a "simple" RHS. */ | |
1430 | ||
1431 | static bool | |
1432 | cond_store_replacement (basic_block middle_bb, basic_block join_bb, | |
1433 | edge e0, edge e1, struct pointer_set_t *nontrap) | |
1434 | { | |
75a70cf9 | 1435 | gimple assign = last_and_only_stmt (middle_bb); |
03d37e4e | 1436 | tree lhs, rhs, name, name2; |
75a70cf9 | 1437 | gimple newphi, new_stmt; |
1438 | gimple_stmt_iterator gsi; | |
efbcb6de | 1439 | source_location locus; |
e6d0e152 | 1440 | |
1441 | /* Check if middle_bb contains of only one store. */ | |
1442 | if (!assign | |
91cf53d5 | 1443 | || !gimple_assign_single_p (assign)) |
e6d0e152 | 1444 | return false; |
1445 | ||
efbcb6de | 1446 | locus = gimple_location (assign); |
75a70cf9 | 1447 | lhs = gimple_assign_lhs (assign); |
1448 | rhs = gimple_assign_rhs1 (assign); | |
182cf5a9 | 1449 | if (TREE_CODE (lhs) != MEM_REF |
91cf53d5 | 1450 | || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME |
3211fa0a | 1451 | || !is_gimple_reg_type (TREE_TYPE (lhs))) |
e6d0e152 | 1452 | return false; |
91cf53d5 | 1453 | |
e6d0e152 | 1454 | /* Prove that we can move the store down. We could also check |
1455 | TREE_THIS_NOTRAP here, but in that case we also could move stores, | |
1456 | whose value is not available readily, which we want to avoid. */ | |
1457 | if (!pointer_set_contains (nontrap, lhs)) | |
1458 | return false; | |
1459 | ||
1460 | /* Now we've checked the constraints, so do the transformation: | |
1461 | 1) Remove the single store. */ | |
75a70cf9 | 1462 | gsi = gsi_for_stmt (assign); |
3211fa0a | 1463 | unlink_stmt_vdef (assign); |
75a70cf9 | 1464 | gsi_remove (&gsi, true); |
91cf53d5 | 1465 | release_defs (assign); |
e6d0e152 | 1466 | |
03d37e4e | 1467 | /* 2) Insert a load from the memory of the store to the temporary |
e6d0e152 | 1468 | on the edge which did not contain the store. */ |
1469 | lhs = unshare_expr (lhs); | |
03d37e4e | 1470 | name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |
1471 | new_stmt = gimple_build_assign (name, lhs); | |
efbcb6de | 1472 | gimple_set_location (new_stmt, locus); |
75a70cf9 | 1473 | gsi_insert_on_edge (e1, new_stmt); |
e6d0e152 | 1474 | |
03d37e4e | 1475 | /* 3) Create a PHI node at the join block, with one argument |
e6d0e152 | 1476 | holding the old RHS, and the other holding the temporary |
1477 | where we stored the old memory contents. */ | |
03d37e4e | 1478 | name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |
1479 | newphi = create_phi_node (name2, join_bb); | |
60d535d2 | 1480 | add_phi_arg (newphi, rhs, e0, locus); |
1481 | add_phi_arg (newphi, name, e1, locus); | |
e6d0e152 | 1482 | |
1483 | lhs = unshare_expr (lhs); | |
75a70cf9 | 1484 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); |
e6d0e152 | 1485 | |
03d37e4e | 1486 | /* 4) Insert that PHI node. */ |
75a70cf9 | 1487 | gsi = gsi_after_labels (join_bb); |
1488 | if (gsi_end_p (gsi)) | |
e6d0e152 | 1489 | { |
75a70cf9 | 1490 | gsi = gsi_last_bb (join_bb); |
1491 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); | |
e6d0e152 | 1492 | } |
1493 | else | |
75a70cf9 | 1494 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |
e6d0e152 | 1495 | |
1496 | return true; | |
1497 | } | |
4ee9c684 | 1498 | |
ec611e12 | 1499 | /* Do the main work of conditional store replacement. */ |
91cf53d5 | 1500 | |
1501 | static bool | |
ec611e12 | 1502 | cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb, |
1503 | basic_block join_bb, gimple then_assign, | |
1504 | gimple else_assign) | |
91cf53d5 | 1505 | { |
03d37e4e | 1506 | tree lhs_base, lhs, then_rhs, else_rhs, name; |
91cf53d5 | 1507 | source_location then_locus, else_locus; |
1508 | gimple_stmt_iterator gsi; | |
1509 | gimple newphi, new_stmt; | |
1510 | ||
91cf53d5 | 1511 | if (then_assign == NULL |
1512 | || !gimple_assign_single_p (then_assign) | |
3c25489e | 1513 | || gimple_clobber_p (then_assign) |
91cf53d5 | 1514 | || else_assign == NULL |
3c25489e | 1515 | || !gimple_assign_single_p (else_assign) |
1516 | || gimple_clobber_p (else_assign)) | |
91cf53d5 | 1517 | return false; |
1518 | ||
1519 | lhs = gimple_assign_lhs (then_assign); | |
1520 | if (!is_gimple_reg_type (TREE_TYPE (lhs)) | |
1521 | || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0)) | |
1522 | return false; | |
1523 | ||
1524 | lhs_base = get_base_address (lhs); | |
1525 | if (lhs_base == NULL_TREE | |
1526 | || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF)) | |
1527 | return false; | |
1528 | ||
1529 | then_rhs = gimple_assign_rhs1 (then_assign); | |
1530 | else_rhs = gimple_assign_rhs1 (else_assign); | |
1531 | then_locus = gimple_location (then_assign); | |
1532 | else_locus = gimple_location (else_assign); | |
1533 | ||
1534 | /* Now we've checked the constraints, so do the transformation: | |
1535 | 1) Remove the stores. */ | |
1536 | gsi = gsi_for_stmt (then_assign); | |
1537 | unlink_stmt_vdef (then_assign); | |
1538 | gsi_remove (&gsi, true); | |
1539 | release_defs (then_assign); | |
1540 | ||
1541 | gsi = gsi_for_stmt (else_assign); | |
1542 | unlink_stmt_vdef (else_assign); | |
1543 | gsi_remove (&gsi, true); | |
1544 | release_defs (else_assign); | |
1545 | ||
03d37e4e | 1546 | /* 2) Create a PHI node at the join block, with one argument |
91cf53d5 | 1547 | holding the old RHS, and the other holding the temporary |
1548 | where we stored the old memory contents. */ | |
03d37e4e | 1549 | name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |
1550 | newphi = create_phi_node (name, join_bb); | |
60d535d2 | 1551 | add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus); |
1552 | add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus); | |
91cf53d5 | 1553 | |
1554 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); | |
1555 | ||
03d37e4e | 1556 | /* 3) Insert that PHI node. */ |
91cf53d5 | 1557 | gsi = gsi_after_labels (join_bb); |
1558 | if (gsi_end_p (gsi)) | |
1559 | { | |
1560 | gsi = gsi_last_bb (join_bb); | |
1561 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); | |
1562 | } | |
1563 | else | |
1564 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
1565 | ||
1566 | return true; | |
1567 | } | |
1568 | ||
ec611e12 | 1569 | /* Conditional store replacement. We already know |
1570 | that the recognized pattern looks like so: | |
1571 | ||
1572 | split: | |
1573 | if (cond) goto THEN_BB; else goto ELSE_BB (edge E1) | |
1574 | THEN_BB: | |
1575 | ... | |
1576 | X = Y; | |
1577 | ... | |
1578 | goto JOIN_BB; | |
1579 | ELSE_BB: | |
1580 | ... | |
1581 | X = Z; | |
1582 | ... | |
1583 | fallthrough (edge E0) | |
1584 | JOIN_BB: | |
1585 | some more | |
1586 | ||
1587 | We check that it is safe to sink the store to JOIN_BB by verifying that | |
1588 | there are no read-after-write or write-after-write dependencies in | |
1589 | THEN_BB and ELSE_BB. */ | |
1590 | ||
1591 | static bool | |
1592 | cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb, | |
1593 | basic_block join_bb) | |
1594 | { | |
1595 | gimple then_assign = last_and_only_stmt (then_bb); | |
1596 | gimple else_assign = last_and_only_stmt (else_bb); | |
1597 | VEC (data_reference_p, heap) *then_datarefs, *else_datarefs; | |
1598 | VEC (ddr_p, heap) *then_ddrs, *else_ddrs; | |
1599 | gimple then_store, else_store; | |
1600 | bool found, ok = false, res; | |
1601 | struct data_dependence_relation *ddr; | |
1602 | data_reference_p then_dr, else_dr; | |
1603 | int i, j; | |
1604 | tree then_lhs, else_lhs; | |
1605 | VEC (gimple, heap) *then_stores, *else_stores; | |
1606 | basic_block blocks[3]; | |
1607 | ||
1608 | if (MAX_STORES_TO_SINK == 0) | |
1609 | return false; | |
1610 | ||
1611 | /* Handle the case with single statement in THEN_BB and ELSE_BB. */ | |
1612 | if (then_assign && else_assign) | |
1613 | return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, | |
1614 | then_assign, else_assign); | |
1615 | ||
1616 | /* Find data references. */ | |
1617 | then_datarefs = VEC_alloc (data_reference_p, heap, 1); | |
1618 | else_datarefs = VEC_alloc (data_reference_p, heap, 1); | |
1619 | if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs) | |
1620 | == chrec_dont_know) | |
1621 | || !VEC_length (data_reference_p, then_datarefs) | |
1622 | || (find_data_references_in_bb (NULL, else_bb, &else_datarefs) | |
1623 | == chrec_dont_know) | |
1624 | || !VEC_length (data_reference_p, else_datarefs)) | |
1625 | { | |
1626 | free_data_refs (then_datarefs); | |
1627 | free_data_refs (else_datarefs); | |
1628 | return false; | |
1629 | } | |
1630 | ||
1631 | /* Find pairs of stores with equal LHS. */ | |
1632 | then_stores = VEC_alloc (gimple, heap, 1); | |
1633 | else_stores = VEC_alloc (gimple, heap, 1); | |
1634 | FOR_EACH_VEC_ELT (data_reference_p, then_datarefs, i, then_dr) | |
1635 | { | |
1636 | if (DR_IS_READ (then_dr)) | |
1637 | continue; | |
1638 | ||
1639 | then_store = DR_STMT (then_dr); | |
728dcc71 | 1640 | then_lhs = gimple_get_lhs (then_store); |
ec611e12 | 1641 | found = false; |
1642 | ||
1643 | FOR_EACH_VEC_ELT (data_reference_p, else_datarefs, j, else_dr) | |
1644 | { | |
1645 | if (DR_IS_READ (else_dr)) | |
1646 | continue; | |
1647 | ||
1648 | else_store = DR_STMT (else_dr); | |
728dcc71 | 1649 | else_lhs = gimple_get_lhs (else_store); |
ec611e12 | 1650 | |
1651 | if (operand_equal_p (then_lhs, else_lhs, 0)) | |
1652 | { | |
1653 | found = true; | |
1654 | break; | |
1655 | } | |
1656 | } | |
1657 | ||
1658 | if (!found) | |
1659 | continue; | |
1660 | ||
1661 | VEC_safe_push (gimple, heap, then_stores, then_store); | |
1662 | VEC_safe_push (gimple, heap, else_stores, else_store); | |
1663 | } | |
1664 | ||
1665 | /* No pairs of stores found. */ | |
1666 | if (!VEC_length (gimple, then_stores) | |
1667 | || VEC_length (gimple, then_stores) > (unsigned) MAX_STORES_TO_SINK) | |
1668 | { | |
1669 | free_data_refs (then_datarefs); | |
1670 | free_data_refs (else_datarefs); | |
1671 | VEC_free (gimple, heap, then_stores); | |
1672 | VEC_free (gimple, heap, else_stores); | |
1673 | return false; | |
1674 | } | |
1675 | ||
1676 | /* Compute and check data dependencies in both basic blocks. */ | |
1677 | then_ddrs = VEC_alloc (ddr_p, heap, 1); | |
1678 | else_ddrs = VEC_alloc (ddr_p, heap, 1); | |
8b3fb720 | 1679 | if (!compute_all_dependences (then_datarefs, &then_ddrs, NULL, false) |
1680 | || !compute_all_dependences (else_datarefs, &else_ddrs, NULL, false)) | |
1681 | { | |
1682 | free_dependence_relations (then_ddrs); | |
1683 | free_dependence_relations (else_ddrs); | |
1684 | free_data_refs (then_datarefs); | |
1685 | free_data_refs (else_datarefs); | |
1686 | VEC_free (gimple, heap, then_stores); | |
1687 | VEC_free (gimple, heap, else_stores); | |
1688 | return false; | |
1689 | } | |
ec611e12 | 1690 | blocks[0] = then_bb; |
1691 | blocks[1] = else_bb; | |
1692 | blocks[2] = join_bb; | |
1693 | renumber_gimple_stmt_uids_in_blocks (blocks, 3); | |
1694 | ||
1695 | /* Check that there are no read-after-write or write-after-write dependencies | |
1696 | in THEN_BB. */ | |
1697 | FOR_EACH_VEC_ELT (ddr_p, then_ddrs, i, ddr) | |
1698 | { | |
1699 | struct data_reference *dra = DDR_A (ddr); | |
1700 | struct data_reference *drb = DDR_B (ddr); | |
1701 | ||
1702 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known | |
1703 | && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) | |
1704 | && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) | |
1705 | || (DR_IS_READ (drb) && DR_IS_WRITE (dra) | |
1706 | && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) | |
1707 | || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) | |
1708 | { | |
1709 | free_dependence_relations (then_ddrs); | |
1710 | free_dependence_relations (else_ddrs); | |
2473bfb7 | 1711 | free_data_refs (then_datarefs); |
1712 | free_data_refs (else_datarefs); | |
ec611e12 | 1713 | VEC_free (gimple, heap, then_stores); |
1714 | VEC_free (gimple, heap, else_stores); | |
1715 | return false; | |
1716 | } | |
1717 | } | |
1718 | ||
1719 | /* Check that there are no read-after-write or write-after-write dependencies | |
1720 | in ELSE_BB. */ | |
1721 | FOR_EACH_VEC_ELT (ddr_p, else_ddrs, i, ddr) | |
1722 | { | |
1723 | struct data_reference *dra = DDR_A (ddr); | |
1724 | struct data_reference *drb = DDR_B (ddr); | |
1725 | ||
1726 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known | |
1727 | && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) | |
1728 | && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) | |
1729 | || (DR_IS_READ (drb) && DR_IS_WRITE (dra) | |
1730 | && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) | |
1731 | || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) | |
1732 | { | |
1733 | free_dependence_relations (then_ddrs); | |
1734 | free_dependence_relations (else_ddrs); | |
2473bfb7 | 1735 | free_data_refs (then_datarefs); |
1736 | free_data_refs (else_datarefs); | |
ec611e12 | 1737 | VEC_free (gimple, heap, then_stores); |
1738 | VEC_free (gimple, heap, else_stores); | |
1739 | return false; | |
1740 | } | |
1741 | } | |
1742 | ||
1743 | /* Sink stores with same LHS. */ | |
1744 | FOR_EACH_VEC_ELT (gimple, then_stores, i, then_store) | |
1745 | { | |
1746 | else_store = VEC_index (gimple, else_stores, i); | |
1747 | res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, | |
1748 | then_store, else_store); | |
1749 | ok = ok || res; | |
1750 | } | |
1751 | ||
1752 | free_dependence_relations (then_ddrs); | |
1753 | free_dependence_relations (else_ddrs); | |
2473bfb7 | 1754 | free_data_refs (then_datarefs); |
1755 | free_data_refs (else_datarefs); | |
ec611e12 | 1756 | VEC_free (gimple, heap, then_stores); |
1757 | VEC_free (gimple, heap, else_stores); | |
1758 | ||
1759 | return ok; | |
1760 | } | |
1761 | ||
239e9670 | 1762 | /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */ |
1763 | ||
1764 | static bool | |
1765 | local_mem_dependence (gimple stmt, basic_block bb) | |
1766 | { | |
1767 | tree vuse = gimple_vuse (stmt); | |
1768 | gimple def; | |
1769 | ||
1770 | if (!vuse) | |
1771 | return false; | |
1772 | ||
1773 | def = SSA_NAME_DEF_STMT (vuse); | |
1774 | return (def && gimple_bb (def) == bb); | |
1775 | } | |
1776 | ||
1777 | /* Given a "diamond" control-flow pattern where BB0 tests a condition, | |
1778 | BB1 and BB2 are "then" and "else" blocks dependent on this test, | |
1779 | and BB3 rejoins control flow following BB1 and BB2, look for | |
1780 | opportunities to hoist loads as follows. If BB3 contains a PHI of | |
1781 | two loads, one each occurring in BB1 and BB2, and the loads are | |
1782 | provably of adjacent fields in the same structure, then move both | |
1783 | loads into BB0. Of course this can only be done if there are no | |
1784 | dependencies preventing such motion. | |
1785 | ||
1786 | One of the hoisted loads will always be speculative, so the | |
1787 | transformation is currently conservative: | |
1788 | ||
1789 | - The fields must be strictly adjacent. | |
1790 | - The two fields must occupy a single memory block that is | |
1791 | guaranteed to not cross a page boundary. | |
1792 | ||
1793 | The last is difficult to prove, as such memory blocks should be | |
1794 | aligned on the minimum of the stack alignment boundary and the | |
1795 | alignment guaranteed by heap allocation interfaces. Thus we rely | |
1796 | on a parameter for the alignment value. | |
1797 | ||
1798 | Provided a good value is used for the last case, the first | |
1799 | restriction could possibly be relaxed. */ | |
1800 | ||
1801 | static void | |
1802 | hoist_adjacent_loads (basic_block bb0, basic_block bb1, | |
1803 | basic_block bb2, basic_block bb3) | |
1804 | { | |
1805 | int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE); | |
1806 | unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT); | |
1807 | gimple_stmt_iterator gsi; | |
1808 | ||
1809 | /* Walk the phis in bb3 looking for an opportunity. We are looking | |
1810 | for phis of two SSA names, one each of which is defined in bb1 and | |
1811 | bb2. */ | |
1812 | for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1813 | { | |
1814 | gimple phi_stmt = gsi_stmt (gsi); | |
1815 | gimple def1, def2, defswap; | |
1816 | tree arg1, arg2, ref1, ref2, field1, field2, fieldswap; | |
1817 | tree tree_offset1, tree_offset2, tree_size2, next; | |
1818 | int offset1, offset2, size2; | |
1819 | unsigned align1; | |
1820 | gimple_stmt_iterator gsi2; | |
1821 | basic_block bb_for_def1, bb_for_def2; | |
1822 | ||
7c782c9b | 1823 | if (gimple_phi_num_args (phi_stmt) != 2 |
1824 | || virtual_operand_p (gimple_phi_result (phi_stmt))) | |
239e9670 | 1825 | continue; |
1826 | ||
1827 | arg1 = gimple_phi_arg_def (phi_stmt, 0); | |
1828 | arg2 = gimple_phi_arg_def (phi_stmt, 1); | |
1829 | ||
1830 | if (TREE_CODE (arg1) != SSA_NAME | |
1831 | || TREE_CODE (arg2) != SSA_NAME | |
1832 | || SSA_NAME_IS_DEFAULT_DEF (arg1) | |
7c782c9b | 1833 | || SSA_NAME_IS_DEFAULT_DEF (arg2)) |
239e9670 | 1834 | continue; |
1835 | ||
1836 | def1 = SSA_NAME_DEF_STMT (arg1); | |
1837 | def2 = SSA_NAME_DEF_STMT (arg2); | |
1838 | ||
1839 | if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2) | |
1840 | && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2)) | |
1841 | continue; | |
1842 | ||
1843 | /* Check the mode of the arguments to be sure a conditional move | |
1844 | can be generated for it. */ | |
935611bc | 1845 | if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1))) |
1846 | == CODE_FOR_nothing) | |
239e9670 | 1847 | continue; |
1848 | ||
1849 | /* Both statements must be assignments whose RHS is a COMPONENT_REF. */ | |
1850 | if (!gimple_assign_single_p (def1) | |
1851 | || !gimple_assign_single_p (def2)) | |
1852 | continue; | |
1853 | ||
1854 | ref1 = gimple_assign_rhs1 (def1); | |
1855 | ref2 = gimple_assign_rhs1 (def2); | |
1856 | ||
1857 | if (TREE_CODE (ref1) != COMPONENT_REF | |
1858 | || TREE_CODE (ref2) != COMPONENT_REF) | |
1859 | continue; | |
1860 | ||
1861 | /* The zeroth operand of the two component references must be | |
1862 | identical. It is not sufficient to compare get_base_address of | |
1863 | the two references, because this could allow for different | |
1864 | elements of the same array in the two trees. It is not safe to | |
1865 | assume that the existence of one array element implies the | |
1866 | existence of a different one. */ | |
1867 | if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0)) | |
1868 | continue; | |
1869 | ||
1870 | field1 = TREE_OPERAND (ref1, 1); | |
1871 | field2 = TREE_OPERAND (ref2, 1); | |
1872 | ||
1873 | /* Check for field adjacency, and ensure field1 comes first. */ | |
1874 | for (next = DECL_CHAIN (field1); | |
1875 | next && TREE_CODE (next) != FIELD_DECL; | |
1876 | next = DECL_CHAIN (next)) | |
1877 | ; | |
1878 | ||
1879 | if (next != field2) | |
1880 | { | |
1881 | for (next = DECL_CHAIN (field2); | |
1882 | next && TREE_CODE (next) != FIELD_DECL; | |
1883 | next = DECL_CHAIN (next)) | |
1884 | ; | |
1885 | ||
1886 | if (next != field1) | |
1887 | continue; | |
1888 | ||
1889 | fieldswap = field1; | |
1890 | field1 = field2; | |
1891 | field2 = fieldswap; | |
1892 | defswap = def1; | |
1893 | def1 = def2; | |
1894 | def2 = defswap; | |
239e9670 | 1895 | } |
1896 | ||
7c74ee50 | 1897 | bb_for_def1 = gimple_bb (def1); |
1898 | bb_for_def2 = gimple_bb (def2); | |
1899 | ||
239e9670 | 1900 | /* Check for proper alignment of the first field. */ |
1901 | tree_offset1 = bit_position (field1); | |
1902 | tree_offset2 = bit_position (field2); | |
1903 | tree_size2 = DECL_SIZE (field2); | |
1904 | ||
1905 | if (!host_integerp (tree_offset1, 1) | |
1906 | || !host_integerp (tree_offset2, 1) | |
1907 | || !host_integerp (tree_size2, 1)) | |
1908 | continue; | |
1909 | ||
1910 | offset1 = TREE_INT_CST_LOW (tree_offset1); | |
1911 | offset2 = TREE_INT_CST_LOW (tree_offset2); | |
1912 | size2 = TREE_INT_CST_LOW (tree_size2); | |
1913 | align1 = DECL_ALIGN (field1) % param_align_bits; | |
1914 | ||
1915 | if (offset1 % BITS_PER_UNIT != 0) | |
1916 | continue; | |
1917 | ||
1918 | /* For profitability, the two field references should fit within | |
1919 | a single cache line. */ | |
1920 | if (align1 + offset2 - offset1 + size2 > param_align_bits) | |
1921 | continue; | |
1922 | ||
1923 | /* The two expressions cannot be dependent upon vdefs defined | |
1924 | in bb1/bb2. */ | |
1925 | if (local_mem_dependence (def1, bb_for_def1) | |
1926 | || local_mem_dependence (def2, bb_for_def2)) | |
1927 | continue; | |
1928 | ||
1929 | /* The conditions are satisfied; hoist the loads from bb1 and bb2 into | |
1930 | bb0. We hoist the first one first so that a cache miss is handled | |
1931 | efficiently regardless of hardware cache-fill policy. */ | |
1932 | gsi2 = gsi_for_stmt (def1); | |
1933 | gsi_move_to_bb_end (&gsi2, bb0); | |
1934 | gsi2 = gsi_for_stmt (def2); | |
1935 | gsi_move_to_bb_end (&gsi2, bb0); | |
1936 | ||
1937 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1938 | { | |
1939 | fprintf (dump_file, | |
1940 | "\nHoisting adjacent loads from %d and %d into %d: \n", | |
1941 | bb_for_def1->index, bb_for_def2->index, bb0->index); | |
1942 | print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS); | |
1943 | print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS); | |
1944 | } | |
1945 | } | |
1946 | } | |
1947 | ||
1948 | /* Determine whether we should attempt to hoist adjacent loads out of | |
1949 | diamond patterns in pass_phiopt. Always hoist loads if | |
1950 | -fhoist-adjacent-loads is specified and the target machine has | |
6f0ddab1 | 1951 | both a conditional move instruction and a defined cache line size. */ |
239e9670 | 1952 | |
1953 | static bool | |
1954 | gate_hoist_loads (void) | |
1955 | { | |
6f0ddab1 | 1956 | return (flag_hoist_adjacent_loads == 1 |
1957 | && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE) | |
1958 | && HAVE_conditional_move); | |
239e9670 | 1959 | } |
1960 | ||
4ee9c684 | 1961 | /* Always do these optimizations if we have SSA |
20e5647c | 1962 | trees to work on. */ |
4ee9c684 | 1963 | static bool |
1964 | gate_phiopt (void) | |
1965 | { | |
1966 | return 1; | |
1967 | } | |
20e5647c | 1968 | |
20099e35 | 1969 | struct gimple_opt_pass pass_phiopt = |
4ee9c684 | 1970 | { |
20099e35 | 1971 | { |
1972 | GIMPLE_PASS, | |
4ee9c684 | 1973 | "phiopt", /* name */ |
1974 | gate_phiopt, /* gate */ | |
1975 | tree_ssa_phiopt, /* execute */ | |
1976 | NULL, /* sub */ | |
1977 | NULL, /* next */ | |
1978 | 0, /* static_pass_number */ | |
1979 | TV_TREE_PHIOPT, /* tv_id */ | |
2f8eb909 | 1980 | PROP_cfg | PROP_ssa, /* properties_required */ |
4ee9c684 | 1981 | 0, /* properties_provided */ |
1982 | 0, /* properties_destroyed */ | |
1983 | 0, /* todo_flags_start */ | |
771e2890 | 1984 | TODO_ggc_collect |
88dbf20f | 1985 | | TODO_verify_ssa |
88dbf20f | 1986 | | TODO_verify_flow |
20099e35 | 1987 | | TODO_verify_stmts /* todo_flags_finish */ |
1988 | } | |
4ee9c684 | 1989 | }; |
e6d0e152 | 1990 | |
1991 | static bool | |
1992 | gate_cselim (void) | |
1993 | { | |
1994 | return flag_tree_cselim; | |
1995 | } | |
1996 | ||
20099e35 | 1997 | struct gimple_opt_pass pass_cselim = |
e6d0e152 | 1998 | { |
20099e35 | 1999 | { |
2000 | GIMPLE_PASS, | |
e6d0e152 | 2001 | "cselim", /* name */ |
2002 | gate_cselim, /* gate */ | |
2003 | tree_ssa_cs_elim, /* execute */ | |
2004 | NULL, /* sub */ | |
2005 | NULL, /* next */ | |
2006 | 0, /* static_pass_number */ | |
2007 | TV_TREE_PHIOPT, /* tv_id */ | |
2f8eb909 | 2008 | PROP_cfg | PROP_ssa, /* properties_required */ |
e6d0e152 | 2009 | 0, /* properties_provided */ |
2010 | 0, /* properties_destroyed */ | |
2011 | 0, /* todo_flags_start */ | |
771e2890 | 2012 | TODO_ggc_collect |
e6d0e152 | 2013 | | TODO_verify_ssa |
2014 | | TODO_verify_flow | |
20099e35 | 2015 | | TODO_verify_stmts /* todo_flags_finish */ |
2016 | } | |
e6d0e152 | 2017 | }; |