1 | /* Optimization of PHI nodes by converting them into straightline code. |
---|---|

2 | Copyright (C) 2004-2017 Free Software Foundation, Inc. |

3 | |

4 | This file is part of GCC. |

5 | |

6 | GCC is free software; you can redistribute it and/or modify it |

7 | under the terms of the GNU General Public License as published by the |

8 | Free Software Foundation; either version 3, or (at your option) any |

9 | later version. |

10 | |

11 | GCC is distributed in the hope that it will be useful, but WITHOUT |

12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |

13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |

14 | for more details. |

15 | |

16 | You should have received a copy of the GNU General Public License |

17 | along with GCC; see the file COPYING3. If not see |

18 | <http://www.gnu.org/licenses/>. */ |

19 | |

20 | #include "config.h" |

21 | #include "system.h" |

22 | #include "coretypes.h" |

23 | #include "backend.h" |

24 | #include "insn-codes.h" |

25 | #include "rtl.h" |

26 | #include "tree.h" |

27 | #include "gimple.h" |

28 | #include "cfghooks.h" |

29 | #include "tree-pass.h" |

30 | #include "ssa.h" |

31 | #include "optabs-tree.h" |

32 | #include "insn-config.h" |

33 | #include "gimple-pretty-print.h" |

34 | #include "fold-const.h" |

35 | #include "stor-layout.h" |

36 | #include "cfganal.h" |

37 | #include "gimplify.h" |

38 | #include "gimple-iterator.h" |

39 | #include "gimplify-me.h" |

40 | #include "tree-cfg.h" |

41 | #include "tree-dfa.h" |

42 | #include "domwalk.h" |

43 | #include "cfgloop.h" |

44 | #include "tree-data-ref.h" |

45 | #include "tree-scalar-evolution.h" |

46 | #include "tree-inline.h" |

47 | #include "params.h" |

48 | |

49 | static unsigned int tree_ssa_phiopt_worker (bool, bool); |

50 | static bool conditional_replacement (basic_block, basic_block, |

51 | edge, edge, gphi *, tree, tree); |

52 | static gphi *factor_out_conditional_conversion (edge, edge, gphi *, tree, tree, |

53 | gimple *); |

54 | static int value_replacement (basic_block, basic_block, |

55 | edge, edge, gimple *, tree, tree); |

56 | static bool minmax_replacement (basic_block, basic_block, |

57 | edge, edge, gimple *, tree, tree); |

58 | static bool abs_replacement (basic_block, basic_block, |

59 | edge, edge, gimple *, tree, tree); |

60 | static bool cond_store_replacement (basic_block, basic_block, edge, edge, |

61 | hash_set<tree> *); |

62 | static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block); |

63 | static hash_set<tree> * get_non_trapping (); |

64 | static void replace_phi_edge_with_variable (basic_block, edge, gimple *, tree); |

65 | static void hoist_adjacent_loads (basic_block, basic_block, |

66 | basic_block, basic_block); |

67 | static bool gate_hoist_loads (void); |

68 | |

69 | /* This pass tries to transform conditional stores into unconditional |

70 | ones, enabling further simplifications with the simpler then and else |

71 | blocks. In particular it replaces this: |

72 | |

73 | bb0: |

74 | if (cond) goto bb2; else goto bb1; |

75 | bb1: |

76 | *p = RHS; |

77 | bb2: |

78 | |

79 | with |

80 | |

81 | bb0: |

82 | if (cond) goto bb1; else goto bb2; |

83 | bb1: |

84 | condtmp' = *p; |

85 | bb2: |

86 | condtmp = PHI <RHS, condtmp'> |

87 | *p = condtmp; |

88 | |

89 | This transformation can only be done under several constraints, |

90 | documented below. It also replaces: |

91 | |

92 | bb0: |

93 | if (cond) goto bb2; else goto bb1; |

94 | bb1: |

95 | *p = RHS1; |

96 | goto bb3; |

97 | bb2: |

98 | *p = RHS2; |

99 | bb3: |

100 | |

101 | with |

102 | |

103 | bb0: |

104 | if (cond) goto bb3; else goto bb1; |

105 | bb1: |

106 | bb3: |

107 | condtmp = PHI <RHS1, RHS2> |

108 | *p = condtmp; */ |

109 | |

110 | static unsigned int |

111 | tree_ssa_cs_elim (void) |

112 | { |

113 | unsigned todo; |

114 | /* ??? We are not interested in loop related info, but the following |

115 | will create it, ICEing as we didn't init loops with pre-headers. |

116 | An interfacing issue of find_data_references_in_bb. */ |

117 | loop_optimizer_init (LOOPS_NORMAL); |

118 | scev_initialize (); |

119 | todo = tree_ssa_phiopt_worker (true, false); |

120 | scev_finalize (); |

121 | loop_optimizer_finalize (); |

122 | return todo; |

123 | } |

124 | |

125 | /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */ |

126 | |

127 | static gphi * |

128 | single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1) |

129 | { |

130 | gimple_stmt_iterator i; |

131 | gphi *phi = NULL; |

132 | if (gimple_seq_singleton_p (seq)) |

133 | return as_a <gphi *> (gsi_stmt (gsi_start (seq))); |

134 | for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i)) |

135 | { |

136 | gphi *p = as_a <gphi *> (gsi_stmt (i)); |

137 | /* If the PHI arguments are equal then we can skip this PHI. */ |

138 | if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx), |

139 | gimple_phi_arg_def (p, e1->dest_idx))) |

140 | continue; |

141 | |

142 | /* If we already have a PHI that has the two edge arguments are |

143 | different, then return it is not a singleton for these PHIs. */ |

144 | if (phi) |

145 | return NULL; |

146 | |

147 | phi = p; |

148 | } |

149 | return phi; |

150 | } |

151 | |

152 | /* The core routine of conditional store replacement and normal |

153 | phi optimizations. Both share much of the infrastructure in how |

154 | to match applicable basic block patterns. DO_STORE_ELIM is true |

155 | when we want to do conditional store replacement, false otherwise. |

156 | DO_HOIST_LOADS is true when we want to hoist adjacent loads out |

157 | of diamond control flow patterns, false otherwise. */ |

158 | static unsigned int |

159 | tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads) |

160 | { |

161 | basic_block bb; |

162 | basic_block *bb_order; |

163 | unsigned n, i; |

164 | bool cfgchanged = false; |

165 | hash_set<tree> *nontrap = 0; |

166 | |

167 | if (do_store_elim) |

168 | /* Calculate the set of non-trapping memory accesses. */ |

169 | nontrap = get_non_trapping (); |

170 | |

171 | /* Search every basic block for COND_EXPR we may be able to optimize. |

172 | |

173 | We walk the blocks in order that guarantees that a block with |

174 | a single predecessor is processed before the predecessor. |

175 | This ensures that we collapse inner ifs before visiting the |

176 | outer ones, and also that we do not try to visit a removed |

177 | block. */ |

178 | bb_order = single_pred_before_succ_order (); |

179 | n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS; |

180 | |

181 | for (i = 0; i < n; i++) |

182 | { |

183 | gimple *cond_stmt; |

184 | gphi *phi; |

185 | basic_block bb1, bb2; |

186 | edge e1, e2; |

187 | tree arg0, arg1; |

188 | |

189 | bb = bb_order[i]; |

190 | |

191 | cond_stmt = last_stmt (bb); |

192 | /* Check to see if the last statement is a GIMPLE_COND. */ |

193 | if (!cond_stmt |

194 | || gimple_code (cond_stmt) != GIMPLE_COND) |

195 | continue; |

196 | |

197 | e1 = EDGE_SUCC (bb, 0); |

198 | bb1 = e1->dest; |

199 | e2 = EDGE_SUCC (bb, 1); |

200 | bb2 = e2->dest; |

201 | |

202 | /* We cannot do the optimization on abnormal edges. */ |

203 | if ((e1->flags & EDGE_ABNORMAL) != 0 |

204 | || (e2->flags & EDGE_ABNORMAL) != 0) |

205 | continue; |

206 | |

207 | /* If either bb1's succ or bb2 or bb2's succ is non NULL. */ |

208 | if (EDGE_COUNT (bb1->succs) == 0 |

209 | || bb2 == NULL |

210 | || EDGE_COUNT (bb2->succs) == 0) |

211 | continue; |

212 | |

213 | /* Find the bb which is the fall through to the other. */ |

214 | if (EDGE_SUCC (bb1, 0)->dest == bb2) |

215 | ; |

216 | else if (EDGE_SUCC (bb2, 0)->dest == bb1) |

217 | { |

218 | std::swap (bb1, bb2); |

219 | std::swap (e1, e2); |

220 | } |

221 | else if (do_store_elim |

222 | && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) |

223 | { |

224 | basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; |

225 | |

226 | if (!single_succ_p (bb1) |

227 | || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0 |

228 | || !single_succ_p (bb2) |

229 | || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0 |

230 | || EDGE_COUNT (bb3->preds) != 2) |

231 | continue; |

232 | if (cond_if_else_store_replacement (bb1, bb2, bb3)) |

233 | cfgchanged = true; |

234 | continue; |

235 | } |

236 | else if (do_hoist_loads |

237 | && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) |

238 | { |

239 | basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; |

240 | |

241 | if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt))) |

242 | && single_succ_p (bb1) |

243 | && single_succ_p (bb2) |

244 | && single_pred_p (bb1) |

245 | && single_pred_p (bb2) |

246 | && EDGE_COUNT (bb->succs) == 2 |

247 | && EDGE_COUNT (bb3->preds) == 2 |

248 | /* If one edge or the other is dominant, a conditional move |

249 | is likely to perform worse than the well-predicted branch. */ |

250 | && !predictable_edge_p (EDGE_SUCC (bb, 0)) |

251 | && !predictable_edge_p (EDGE_SUCC (bb, 1))) |

252 | hoist_adjacent_loads (bb, bb1, bb2, bb3); |

253 | continue; |

254 | } |

255 | else |

256 | continue; |

257 | |

258 | e1 = EDGE_SUCC (bb1, 0); |

259 | |

260 | /* Make sure that bb1 is just a fall through. */ |

261 | if (!single_succ_p (bb1) |

262 | || (e1->flags & EDGE_FALLTHRU) == 0) |

263 | continue; |

264 | |

265 | /* Also make sure that bb1 only have one predecessor and that it |

266 | is bb. */ |

267 | if (!single_pred_p (bb1) |

268 | || single_pred (bb1) != bb) |

269 | continue; |

270 | |

271 | if (do_store_elim) |

272 | { |

273 | /* bb1 is the middle block, bb2 the join block, bb the split block, |

274 | e1 the fallthrough edge from bb1 to bb2. We can't do the |

275 | optimization if the join block has more than two predecessors. */ |

276 | if (EDGE_COUNT (bb2->preds) > 2) |

277 | continue; |

278 | if (cond_store_replacement (bb1, bb2, e1, e2, nontrap)) |

279 | cfgchanged = true; |

280 | } |

281 | else |

282 | { |

283 | gimple_seq phis = phi_nodes (bb2); |

284 | gimple_stmt_iterator gsi; |

285 | bool candorest = true; |

286 | |

287 | /* Value replacement can work with more than one PHI |

288 | so try that first. */ |

289 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) |

290 | { |

291 | phi = as_a <gphi *> (gsi_stmt (gsi)); |

292 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); |

293 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); |

294 | if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2) |

295 | { |

296 | candorest = false; |

297 | cfgchanged = true; |

298 | break; |

299 | } |

300 | } |

301 | |

302 | if (!candorest) |

303 | continue; |

304 | |

305 | phi = single_non_singleton_phi_for_edges (phis, e1, e2); |

306 | if (!phi) |

307 | continue; |

308 | |

309 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); |

310 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); |

311 | |

312 | /* Something is wrong if we cannot find the arguments in the PHI |

313 | node. */ |

314 | gcc_assert (arg0 != NULL_TREE && arg1 != NULL_TREE); |

315 | |

316 | gphi *newphi = factor_out_conditional_conversion (e1, e2, phi, |

317 | arg0, arg1, |

318 | cond_stmt); |

319 | if (newphi != NULL) |

320 | { |

321 | phi = newphi; |

322 | /* factor_out_conditional_conversion may create a new PHI in |

323 | BB2 and eliminate an existing PHI in BB2. Recompute values |

324 | that may be affected by that change. */ |

325 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); |

326 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); |

327 | gcc_assert (arg0 != NULL_TREE && arg1 != NULL_TREE); |

328 | } |

329 | |

330 | /* Do the replacement of conditional if it can be done. */ |

331 | if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) |

332 | cfgchanged = true; |

333 | else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) |

334 | cfgchanged = true; |

335 | else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) |

336 | cfgchanged = true; |

337 | } |

338 | } |

339 | |

340 | free (bb_order); |

341 | |

342 | if (do_store_elim) |

343 | delete nontrap; |

344 | /* If the CFG has changed, we should cleanup the CFG. */ |

345 | if (cfgchanged && do_store_elim) |

346 | { |

347 | /* In cond-store replacement we have added some loads on edges |

348 | and new VOPS (as we moved the store, and created a load). */ |

349 | gsi_commit_edge_inserts (); |

350 | return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals; |

351 | } |

352 | else if (cfgchanged) |

353 | return TODO_cleanup_cfg; |

354 | return 0; |

355 | } |

356 | |

357 | /* Replace PHI node element whose edge is E in block BB with variable NEW. |

358 | Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK |

359 | is known to have two edges, one of which must reach BB). */ |

360 | |

361 | static void |

362 | replace_phi_edge_with_variable (basic_block cond_block, |

363 | edge e, gimple *phi, tree new_tree) |

364 | { |

365 | basic_block bb = gimple_bb (phi); |

366 | basic_block block_to_remove; |

367 | gimple_stmt_iterator gsi; |

368 | |

369 | /* Change the PHI argument to new. */ |

370 | SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree); |

371 | |

372 | /* Remove the empty basic block. */ |

373 | if (EDGE_SUCC (cond_block, 0)->dest == bb) |

374 | { |

375 | EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU; |

376 | EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |

377 | EDGE_SUCC (cond_block, 0)->probability = profile_probability::always (); |

378 | |

379 | block_to_remove = EDGE_SUCC (cond_block, 1)->dest; |

380 | } |

381 | else |

382 | { |

383 | EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU; |

384 | EDGE_SUCC (cond_block, 1)->flags |

385 | &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |

386 | EDGE_SUCC (cond_block, 1)->probability = profile_probability::always (); |

387 | |

388 | block_to_remove = EDGE_SUCC (cond_block, 0)->dest; |

389 | } |

390 | delete_basic_block (block_to_remove); |

391 | |

392 | /* Eliminate the COND_EXPR at the end of COND_BLOCK. */ |

393 | gsi = gsi_last_bb (cond_block); |

394 | gsi_remove (&gsi, true); |

395 | |

396 | if (dump_file && (dump_flags & TDF_DETAILS)) |

397 | fprintf (dump_file, |

398 | "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n", |

399 | cond_block->index, |

400 | bb->index); |

401 | } |

402 | |

403 | /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI |

404 | stmt are CONVERT_STMT, factor out the conversion and perform the conversion |

405 | to the result of PHI stmt. COND_STMT is the controlling predicate. |

406 | Return the newly-created PHI, if any. */ |

407 | |

408 | static gphi * |

409 | factor_out_conditional_conversion (edge e0, edge e1, gphi *phi, |

410 | tree arg0, tree arg1, gimple *cond_stmt) |

411 | { |

412 | gimple *arg0_def_stmt = NULL, *arg1_def_stmt = NULL, *new_stmt; |

413 | tree new_arg0 = NULL_TREE, new_arg1 = NULL_TREE; |

414 | tree temp, result; |

415 | gphi *newphi; |

416 | gimple_stmt_iterator gsi, gsi_for_def; |

417 | source_location locus = gimple_location (phi); |

418 | enum tree_code convert_code; |

419 | |

420 | /* Handle only PHI statements with two arguments. TODO: If all |

421 | other arguments to PHI are INTEGER_CST or if their defining |

422 | statement have the same unary operation, we can handle more |

423 | than two arguments too. */ |

424 | if (gimple_phi_num_args (phi) != 2) |

425 | return NULL; |

426 | |

427 | /* First canonicalize to simplify tests. */ |

428 | if (TREE_CODE (arg0) != SSA_NAME) |

429 | { |

430 | std::swap (arg0, arg1); |

431 | std::swap (e0, e1); |

432 | } |

433 | |

434 | if (TREE_CODE (arg0) != SSA_NAME |

435 | || (TREE_CODE (arg1) != SSA_NAME |

436 | && TREE_CODE (arg1) != INTEGER_CST)) |

437 | return NULL; |

438 | |

439 | /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is |

440 | a conversion. */ |

441 | arg0_def_stmt = SSA_NAME_DEF_STMT (arg0); |

442 | if (!gimple_assign_cast_p (arg0_def_stmt)) |

443 | return NULL; |

444 | |

445 | /* Use the RHS as new_arg0. */ |

446 | convert_code = gimple_assign_rhs_code (arg0_def_stmt); |

447 | new_arg0 = gimple_assign_rhs1 (arg0_def_stmt); |

448 | if (convert_code == VIEW_CONVERT_EXPR) |

449 | { |

450 | new_arg0 = TREE_OPERAND (new_arg0, 0); |

451 | if (!is_gimple_reg_type (TREE_TYPE (new_arg0))) |

452 | return NULL; |

453 | } |

454 | |

455 | if (TREE_CODE (arg1) == SSA_NAME) |

456 | { |

457 | /* Check if arg1 is an SSA_NAME and the stmt which defines arg1 |

458 | is a conversion. */ |

459 | arg1_def_stmt = SSA_NAME_DEF_STMT (arg1); |

460 | if (!is_gimple_assign (arg1_def_stmt) |

461 | || gimple_assign_rhs_code (arg1_def_stmt) != convert_code) |

462 | return NULL; |

463 | |

464 | /* Use the RHS as new_arg1. */ |

465 | new_arg1 = gimple_assign_rhs1 (arg1_def_stmt); |

466 | if (convert_code == VIEW_CONVERT_EXPR) |

467 | new_arg1 = TREE_OPERAND (new_arg1, 0); |

468 | } |

469 | else |

470 | { |

471 | /* If arg1 is an INTEGER_CST, fold it to new type. */ |

472 | if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0)) |

473 | && int_fits_type_p (arg1, TREE_TYPE (new_arg0))) |

474 | { |

475 | if (gimple_assign_cast_p (arg0_def_stmt)) |

476 | { |

477 | /* For the INTEGER_CST case, we are just moving the |

478 | conversion from one place to another, which can often |

479 | hurt as the conversion moves further away from the |

480 | statement that computes the value. So, perform this |

481 | only if new_arg0 is an operand of COND_STMT, or |

482 | if arg0_def_stmt is the only non-debug stmt in |

483 | its basic block, because then it is possible this |

484 | could enable further optimizations (minmax replacement |

485 | etc.). See PR71016. */ |

486 | if (new_arg0 != gimple_cond_lhs (cond_stmt) |

487 | && new_arg0 != gimple_cond_rhs (cond_stmt) |

488 | && gimple_bb (arg0_def_stmt) == e0->src) |

489 | { |

490 | gsi = gsi_for_stmt (arg0_def_stmt); |

491 | gsi_prev_nondebug (&gsi); |

492 | if (!gsi_end_p (gsi)) |

493 | return NULL; |

494 | gsi = gsi_for_stmt (arg0_def_stmt); |

495 | gsi_next_nondebug (&gsi); |

496 | if (!gsi_end_p (gsi)) |

497 | return NULL; |

498 | } |

499 | new_arg1 = fold_convert (TREE_TYPE (new_arg0), arg1); |

500 | } |

501 | else |

502 | return NULL; |

503 | } |

504 | else |

505 | return NULL; |

506 | } |

507 | |

508 | /* If arg0/arg1 have > 1 use, then this transformation actually increases |

509 | the number of expressions evaluated at runtime. */ |

510 | if (!has_single_use (arg0) |

511 | || (arg1_def_stmt && !has_single_use (arg1))) |

512 | return NULL; |

513 | |

514 | /* If types of new_arg0 and new_arg1 are different bailout. */ |

515 | if (!types_compatible_p (TREE_TYPE (new_arg0), TREE_TYPE (new_arg1))) |

516 | return NULL; |

517 | |

518 | /* Create a new PHI stmt. */ |

519 | result = PHI_RESULT (phi); |

520 | temp = make_ssa_name (TREE_TYPE (new_arg0), NULL); |

521 | newphi = create_phi_node (temp, gimple_bb (phi)); |

522 | |

523 | if (dump_file && (dump_flags & TDF_DETAILS)) |

524 | { |

525 | fprintf (dump_file, "PHI "); |

526 | print_generic_expr (dump_file, gimple_phi_result (phi)); |

527 | fprintf (dump_file, |

528 | " changed to factor conversion out from COND_EXPR.\n"); |

529 | fprintf (dump_file, "New stmt with CAST that defines "); |

530 | print_generic_expr (dump_file, result); |

531 | fprintf (dump_file, ".\n"); |

532 | } |

533 | |

534 | /* Remove the old cast(s) that has single use. */ |

535 | gsi_for_def = gsi_for_stmt (arg0_def_stmt); |

536 | gsi_remove (&gsi_for_def, true); |

537 | release_defs (arg0_def_stmt); |

538 | |

539 | if (arg1_def_stmt) |

540 | { |

541 | gsi_for_def = gsi_for_stmt (arg1_def_stmt); |

542 | gsi_remove (&gsi_for_def, true); |

543 | release_defs (arg1_def_stmt); |

544 | } |

545 | |

546 | add_phi_arg (newphi, new_arg0, e0, locus); |

547 | add_phi_arg (newphi, new_arg1, e1, locus); |

548 | |

549 | /* Create the conversion stmt and insert it. */ |

550 | if (convert_code == VIEW_CONVERT_EXPR) |

551 | temp = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (result), temp); |

552 | new_stmt = gimple_build_assign (result, convert_code, temp); |

553 | gsi = gsi_after_labels (gimple_bb (phi)); |

554 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |

555 | |

556 | /* Remove the original PHI stmt. */ |

557 | gsi = gsi_for_stmt (phi); |

558 | gsi_remove (&gsi, true); |

559 | return newphi; |

560 | } |

561 | |

562 | /* The function conditional_replacement does the main work of doing the |

563 | conditional replacement. Return true if the replacement is done. |

564 | Otherwise return false. |

565 | BB is the basic block where the replacement is going to be done on. ARG0 |

566 | is argument 0 from PHI. Likewise for ARG1. */ |

567 | |

568 | static bool |

569 | conditional_replacement (basic_block cond_bb, basic_block middle_bb, |

570 | edge e0, edge e1, gphi *phi, |

571 | tree arg0, tree arg1) |

572 | { |

573 | tree result; |

574 | gimple *stmt; |

575 | gassign *new_stmt; |

576 | tree cond; |

577 | gimple_stmt_iterator gsi; |

578 | edge true_edge, false_edge; |

579 | tree new_var, new_var2; |

580 | bool neg; |

581 | |

582 | /* FIXME: Gimplification of complex type is too hard for now. */ |

583 | /* We aren't prepared to handle vectors either (and it is a question |

584 | if it would be worthwhile anyway). */ |

585 | if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0)) |

586 | || POINTER_TYPE_P (TREE_TYPE (arg0))) |

587 | || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1)) |

588 | || POINTER_TYPE_P (TREE_TYPE (arg1)))) |

589 | return false; |

590 | |

591 | /* The PHI arguments have the constants 0 and 1, or 0 and -1, then |

592 | convert it to the conditional. */ |

593 | if ((integer_zerop (arg0) && integer_onep (arg1)) |

594 | || (integer_zerop (arg1) && integer_onep (arg0))) |

595 | neg = false; |

596 | else if ((integer_zerop (arg0) && integer_all_onesp (arg1)) |

597 | || (integer_zerop (arg1) && integer_all_onesp (arg0))) |

598 | neg = true; |

599 | else |

600 | return false; |

601 | |

602 | if (!empty_block_p (middle_bb)) |

603 | return false; |

604 | |

605 | /* At this point we know we have a GIMPLE_COND with two successors. |

606 | One successor is BB, the other successor is an empty block which |

607 | falls through into BB. |

608 | |

609 | There is a single PHI node at the join point (BB) and its arguments |

610 | are constants (0, 1) or (0, -1). |

611 | |

612 | So, given the condition COND, and the two PHI arguments, we can |

613 | rewrite this PHI into non-branching code: |

614 | |

615 | dest = (COND) or dest = COND' |

616 | |

617 | We use the condition as-is if the argument associated with the |

618 | true edge has the value one or the argument associated with the |

619 | false edge as the value zero. Note that those conditions are not |

620 | the same since only one of the outgoing edges from the GIMPLE_COND |

621 | will directly reach BB and thus be associated with an argument. */ |

622 | |

623 | stmt = last_stmt (cond_bb); |

624 | result = PHI_RESULT (phi); |

625 | |

626 | /* To handle special cases like floating point comparison, it is easier and |

627 | less error-prone to build a tree and gimplify it on the fly though it is |

628 | less efficient. */ |

629 | cond = fold_build2_loc (gimple_location (stmt), |

630 | gimple_cond_code (stmt), boolean_type_node, |

631 | gimple_cond_lhs (stmt), gimple_cond_rhs (stmt)); |

632 | |

633 | /* We need to know which is the true edge and which is the false |

634 | edge so that we know when to invert the condition below. */ |

635 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |

636 | if ((e0 == true_edge && integer_zerop (arg0)) |

637 | || (e0 == false_edge && !integer_zerop (arg0)) |

638 | || (e1 == true_edge && integer_zerop (arg1)) |

639 | || (e1 == false_edge && !integer_zerop (arg1))) |

640 | cond = fold_build1_loc (gimple_location (stmt), |

641 | TRUTH_NOT_EXPR, TREE_TYPE (cond), cond); |

642 | |

643 | if (neg) |

644 | { |

645 | cond = fold_convert_loc (gimple_location (stmt), |

646 | TREE_TYPE (result), cond); |

647 | cond = fold_build1_loc (gimple_location (stmt), |

648 | NEGATE_EXPR, TREE_TYPE (cond), cond); |

649 | } |

650 | |

651 | /* Insert our new statements at the end of conditional block before the |

652 | COND_STMT. */ |

653 | gsi = gsi_for_stmt (stmt); |

654 | new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true, |

655 | GSI_SAME_STMT); |

656 | |

657 | if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var))) |

658 | { |

659 | source_location locus_0, locus_1; |

660 | |

661 | new_var2 = make_ssa_name (TREE_TYPE (result)); |

662 | new_stmt = gimple_build_assign (new_var2, CONVERT_EXPR, new_var); |

663 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |

664 | new_var = new_var2; |

665 | |

666 | /* Set the locus to the first argument, unless is doesn't have one. */ |

667 | locus_0 = gimple_phi_arg_location (phi, 0); |

668 | locus_1 = gimple_phi_arg_location (phi, 1); |

669 | if (locus_0 == UNKNOWN_LOCATION) |

670 | locus_0 = locus_1; |

671 | gimple_set_location (new_stmt, locus_0); |

672 | } |

673 | |

674 | replace_phi_edge_with_variable (cond_bb, e1, phi, new_var); |

675 | |

676 | /* Note that we optimized this PHI. */ |

677 | return true; |

678 | } |

679 | |

680 | /* Update *ARG which is defined in STMT so that it contains the |

681 | computed value if that seems profitable. Return true if the |

682 | statement is made dead by that rewriting. */ |

683 | |

684 | static bool |

685 | jump_function_from_stmt (tree *arg, gimple *stmt) |

686 | { |

687 | enum tree_code code = gimple_assign_rhs_code (stmt); |

688 | if (code == ADDR_EXPR) |

689 | { |

690 | /* For arg = &p->i transform it to p, if possible. */ |

691 | tree rhs1 = gimple_assign_rhs1 (stmt); |

692 | HOST_WIDE_INT offset; |

693 | tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0), |

694 | &offset); |

695 | if (tem |

696 | && TREE_CODE (tem) == MEM_REF |

697 | && (mem_ref_offset (tem) + offset) == 0) |

698 | { |

699 | *arg = TREE_OPERAND (tem, 0); |

700 | return true; |

701 | } |

702 | } |

703 | /* TODO: Much like IPA-CP jump-functions we want to handle constant |

704 | additions symbolically here, and we'd need to update the comparison |

705 | code that compares the arg + cst tuples in our caller. For now the |

706 | code above exactly handles the VEC_BASE pattern from vec.h. */ |

707 | return false; |

708 | } |

709 | |

710 | /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional |

711 | of the form SSA_NAME NE 0. |

712 | |

713 | If RHS is fed by a simple EQ_EXPR comparison of two values, see if |

714 | the two input values of the EQ_EXPR match arg0 and arg1. |

715 | |

716 | If so update *code and return TRUE. Otherwise return FALSE. */ |

717 | |

718 | static bool |

719 | rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1, |

720 | enum tree_code *code, const_tree rhs) |

721 | { |

722 | /* Obviously if RHS is not an SSA_NAME, we can't look at the defining |

723 | statement. */ |

724 | if (TREE_CODE (rhs) == SSA_NAME) |

725 | { |

726 | gimple *def1 = SSA_NAME_DEF_STMT (rhs); |

727 | |

728 | /* Verify the defining statement has an EQ_EXPR on the RHS. */ |

729 | if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR) |

730 | { |

731 | /* Finally verify the source operands of the EQ_EXPR are equal |

732 | to arg0 and arg1. */ |

733 | tree op0 = gimple_assign_rhs1 (def1); |

734 | tree op1 = gimple_assign_rhs2 (def1); |

735 | if ((operand_equal_for_phi_arg_p (arg0, op0) |

736 | && operand_equal_for_phi_arg_p (arg1, op1)) |

737 | || (operand_equal_for_phi_arg_p (arg0, op1) |

738 | && operand_equal_for_phi_arg_p (arg1, op0))) |

739 | { |

740 | /* We will perform the optimization. */ |

741 | *code = gimple_assign_rhs_code (def1); |

742 | return true; |

743 | } |

744 | } |

745 | } |

746 | return false; |

747 | } |

748 | |

749 | /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND. |

750 | |

751 | Also return TRUE if arg0/arg1 are equal to the source arguments of a |

752 | an EQ comparison feeding a BIT_AND_EXPR which feeds COND. |

753 | |

754 | Return FALSE otherwise. */ |

755 | |

756 | static bool |

757 | operand_equal_for_value_replacement (const_tree arg0, const_tree arg1, |

758 | enum tree_code *code, gimple *cond) |

759 | { |

760 | gimple *def; |

761 | tree lhs = gimple_cond_lhs (cond); |

762 | tree rhs = gimple_cond_rhs (cond); |

763 | |

764 | if ((operand_equal_for_phi_arg_p (arg0, lhs) |

765 | && operand_equal_for_phi_arg_p (arg1, rhs)) |

766 | || (operand_equal_for_phi_arg_p (arg1, lhs) |

767 | && operand_equal_for_phi_arg_p (arg0, rhs))) |

768 | return true; |

769 | |

770 | /* Now handle more complex case where we have an EQ comparison |

771 | which feeds a BIT_AND_EXPR which feeds COND. |

772 | |

773 | First verify that COND is of the form SSA_NAME NE 0. */ |

774 | if (*code != NE_EXPR || !integer_zerop (rhs) |

775 | || TREE_CODE (lhs) != SSA_NAME) |

776 | return false; |

777 | |

778 | /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */ |

779 | def = SSA_NAME_DEF_STMT (lhs); |

780 | if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR) |

781 | return false; |

782 | |

783 | /* Now verify arg0/arg1 correspond to the source arguments of an |

784 | EQ comparison feeding the BIT_AND_EXPR. */ |

785 | |

786 | tree tmp = gimple_assign_rhs1 (def); |

787 | if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp)) |

788 | return true; |

789 | |

790 | tmp = gimple_assign_rhs2 (def); |

791 | if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp)) |

792 | return true; |

793 | |

794 | return false; |

795 | } |

796 | |

797 | /* Returns true if ARG is a neutral element for operation CODE |

798 | on the RIGHT side. */ |

799 | |

800 | static bool |

801 | neutral_element_p (tree_code code, tree arg, bool right) |

802 | { |

803 | switch (code) |

804 | { |

805 | case PLUS_EXPR: |

806 | case BIT_IOR_EXPR: |

807 | case BIT_XOR_EXPR: |

808 | return integer_zerop (arg); |

809 | |

810 | case LROTATE_EXPR: |

811 | case RROTATE_EXPR: |

812 | case LSHIFT_EXPR: |

813 | case RSHIFT_EXPR: |

814 | case MINUS_EXPR: |

815 | case POINTER_PLUS_EXPR: |

816 | return right && integer_zerop (arg); |

817 | |

818 | case MULT_EXPR: |

819 | return integer_onep (arg); |

820 | |

821 | case TRUNC_DIV_EXPR: |

822 | case CEIL_DIV_EXPR: |

823 | case FLOOR_DIV_EXPR: |

824 | case ROUND_DIV_EXPR: |

825 | case EXACT_DIV_EXPR: |

826 | return right && integer_onep (arg); |

827 | |

828 | case BIT_AND_EXPR: |

829 | return integer_all_onesp (arg); |

830 | |

831 | default: |

832 | return false; |

833 | } |

834 | } |

835 | |

836 | /* Returns true if ARG is an absorbing element for operation CODE. */ |

837 | |

838 | static bool |

839 | absorbing_element_p (tree_code code, tree arg, bool right, tree rval) |

840 | { |

841 | switch (code) |

842 | { |

843 | case BIT_IOR_EXPR: |

844 | return integer_all_onesp (arg); |

845 | |

846 | case MULT_EXPR: |

847 | case BIT_AND_EXPR: |

848 | return integer_zerop (arg); |

849 | |

850 | case LSHIFT_EXPR: |

851 | case RSHIFT_EXPR: |

852 | case LROTATE_EXPR: |

853 | case RROTATE_EXPR: |

854 | return !right && integer_zerop (arg); |

855 | |

856 | case TRUNC_DIV_EXPR: |

857 | case CEIL_DIV_EXPR: |

858 | case FLOOR_DIV_EXPR: |

859 | case ROUND_DIV_EXPR: |

860 | case EXACT_DIV_EXPR: |

861 | case TRUNC_MOD_EXPR: |

862 | case CEIL_MOD_EXPR: |

863 | case FLOOR_MOD_EXPR: |

864 | case ROUND_MOD_EXPR: |

865 | return (!right |

866 | && integer_zerop (arg) |

867 | && tree_single_nonzero_warnv_p (rval, NULL)); |

868 | |

869 | default: |

870 | return false; |

871 | } |

872 | } |

873 | |

874 | /* The function value_replacement does the main work of doing the value |

875 | replacement. Return non-zero if the replacement is done. Otherwise return |

876 | 0. If we remove the middle basic block, return 2. |

877 | BB is the basic block where the replacement is going to be done on. ARG0 |

878 | is argument 0 from the PHI. Likewise for ARG1. */ |

879 | |

880 | static int |

881 | value_replacement (basic_block cond_bb, basic_block middle_bb, |

882 | edge e0, edge e1, gimple *phi, |

883 | tree arg0, tree arg1) |

884 | { |

885 | gimple_stmt_iterator gsi; |

886 | gimple *cond; |

887 | edge true_edge, false_edge; |

888 | enum tree_code code; |

889 | bool emtpy_or_with_defined_p = true; |

890 | |

891 | /* If the type says honor signed zeros we cannot do this |

892 | optimization. */ |

893 | if (HONOR_SIGNED_ZEROS (arg1)) |

894 | return 0; |

895 | |

896 | /* If there is a statement in MIDDLE_BB that defines one of the PHI |

897 | arguments, then adjust arg0 or arg1. */ |

898 | gsi = gsi_start_nondebug_after_labels_bb (middle_bb); |

899 | while (!gsi_end_p (gsi)) |

900 | { |

901 | gimple *stmt = gsi_stmt (gsi); |

902 | tree lhs; |

903 | gsi_next_nondebug (&gsi); |

904 | if (!is_gimple_assign (stmt)) |

905 | { |

906 | emtpy_or_with_defined_p = false; |

907 | continue; |

908 | } |

909 | /* Now try to adjust arg0 or arg1 according to the computation |

910 | in the statement. */ |

911 | lhs = gimple_assign_lhs (stmt); |

912 | if (!(lhs == arg0 |

913 | && jump_function_from_stmt (&arg0, stmt)) |

914 | || (lhs == arg1 |

915 | && jump_function_from_stmt (&arg1, stmt))) |

916 | emtpy_or_with_defined_p = false; |

917 | } |

918 | |

919 | cond = last_stmt (cond_bb); |

920 | code = gimple_cond_code (cond); |

921 | |

922 | /* This transformation is only valid for equality comparisons. */ |

923 | if (code != NE_EXPR && code != EQ_EXPR) |

924 | return 0; |

925 | |

926 | /* We need to know which is the true edge and which is the false |

927 | edge so that we know if have abs or negative abs. */ |

928 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |

929 | |

930 | /* At this point we know we have a COND_EXPR with two successors. |

931 | One successor is BB, the other successor is an empty block which |

932 | falls through into BB. |

933 | |

934 | The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR. |

935 | |

936 | There is a single PHI node at the join point (BB) with two arguments. |

937 | |

938 | We now need to verify that the two arguments in the PHI node match |

939 | the two arguments to the equality comparison. */ |

940 | |

941 | if (operand_equal_for_value_replacement (arg0, arg1, &code, cond)) |

942 | { |

943 | edge e; |

944 | tree arg; |

945 | |

946 | /* For NE_EXPR, we want to build an assignment result = arg where |

947 | arg is the PHI argument associated with the true edge. For |

948 | EQ_EXPR we want the PHI argument associated with the false edge. */ |

949 | e = (code == NE_EXPR ? true_edge : false_edge); |

950 | |

951 | /* Unfortunately, E may not reach BB (it may instead have gone to |

952 | OTHER_BLOCK). If that is the case, then we want the single outgoing |

953 | edge from OTHER_BLOCK which reaches BB and represents the desired |

954 | path from COND_BLOCK. */ |

955 | if (e->dest == middle_bb) |

956 | e = single_succ_edge (e->dest); |

957 | |

958 | /* Now we know the incoming edge to BB that has the argument for the |

959 | RHS of our new assignment statement. */ |

960 | if (e0 == e) |

961 | arg = arg0; |

962 | else |

963 | arg = arg1; |

964 | |

965 | /* If the middle basic block was empty or is defining the |

966 | PHI arguments and this is a single phi where the args are different |

967 | for the edges e0 and e1 then we can remove the middle basic block. */ |

968 | if (emtpy_or_with_defined_p |

969 | && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), |

970 | e0, e1) == phi) |

971 | { |

972 | replace_phi_edge_with_variable (cond_bb, e1, phi, arg); |

973 | /* Note that we optimized this PHI. */ |

974 | return 2; |

975 | } |

976 | else |

977 | { |

978 | /* Replace the PHI arguments with arg. */ |

979 | SET_PHI_ARG_DEF (phi, e0->dest_idx, arg); |

980 | SET_PHI_ARG_DEF (phi, e1->dest_idx, arg); |

981 | if (dump_file && (dump_flags & TDF_DETAILS)) |

982 | { |

983 | fprintf (dump_file, "PHI "); |

984 | print_generic_expr (dump_file, gimple_phi_result (phi)); |

985 | fprintf (dump_file, " reduced for COND_EXPR in block %d to ", |

986 | cond_bb->index); |

987 | print_generic_expr (dump_file, arg); |

988 | fprintf (dump_file, ".\n"); |

989 | } |

990 | return 1; |

991 | } |

992 | |

993 | } |

994 | |

995 | /* Now optimize (x != 0) ? x + y : y to just x + y. */ |

996 | gsi = gsi_last_nondebug_bb (middle_bb); |

997 | if (gsi_end_p (gsi)) |

998 | return 0; |

999 | |

1000 | gimple *assign = gsi_stmt (gsi); |

1001 | if (!is_gimple_assign (assign) |

1002 | || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS |

1003 | || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)) |

1004 | && !POINTER_TYPE_P (TREE_TYPE (arg0)))) |

1005 | return 0; |

1006 | |

1007 | /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */ |

1008 | if (!gimple_seq_empty_p (phi_nodes (middle_bb))) |

1009 | return 0; |

1010 | |

1011 | /* Allow up to 2 cheap preparation statements that prepare argument |

1012 | for assign, e.g.: |

1013 | if (y_4 != 0) |

1014 | goto <bb 3>; |

1015 | else |

1016 | goto <bb 4>; |

1017 | <bb 3>: |

1018 | _1 = (int) y_4; |

1019 | iftmp.0_6 = x_5(D) r<< _1; |

1020 | <bb 4>: |

1021 | # iftmp.0_2 = PHI <iftmp.0_6(3), x_5(D)(2)> |

1022 | or: |

1023 | if (y_3(D) == 0) |

1024 | goto <bb 4>; |

1025 | else |

1026 | goto <bb 3>; |

1027 | <bb 3>: |

1028 | y_4 = y_3(D) & 31; |

1029 | _1 = (int) y_4; |

1030 | _6 = x_5(D) r<< _1; |

1031 | <bb 4>: |

1032 | # _2 = PHI <x_5(D)(2), _6(3)> */ |

1033 | gimple *prep_stmt[2] = { NULL, NULL }; |

1034 | int prep_cnt; |

1035 | for (prep_cnt = 0; ; prep_cnt++) |

1036 | { |

1037 | gsi_prev_nondebug (&gsi); |

1038 | if (gsi_end_p (gsi)) |

1039 | break; |

1040 | |

1041 | gimple *g = gsi_stmt (gsi); |

1042 | if (gimple_code (g) == GIMPLE_LABEL) |

1043 | break; |

1044 | |

1045 | if (prep_cnt == 2 || !is_gimple_assign (g)) |

1046 | return 0; |

1047 | |

1048 | tree lhs = gimple_assign_lhs (g); |

1049 | tree rhs1 = gimple_assign_rhs1 (g); |

1050 | use_operand_p use_p; |

1051 | gimple *use_stmt; |

1052 | if (TREE_CODE (lhs) != SSA_NAME |

1053 | || TREE_CODE (rhs1) != SSA_NAME |

1054 | || !INTEGRAL_TYPE_P (TREE_TYPE (lhs)) |

1055 | || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1)) |

1056 | || !single_imm_use (lhs, &use_p, &use_stmt) |

1057 | || use_stmt != (prep_cnt ? prep_stmt[prep_cnt - 1] : assign)) |

1058 | return 0; |

1059 | switch (gimple_assign_rhs_code (g)) |

1060 | { |

1061 | CASE_CONVERT: |

1062 | break; |

1063 | case PLUS_EXPR: |

1064 | case BIT_AND_EXPR: |

1065 | case BIT_IOR_EXPR: |

1066 | case BIT_XOR_EXPR: |

1067 | if (TREE_CODE (gimple_assign_rhs2 (g)) != INTEGER_CST) |

1068 | return 0; |

1069 | break; |

1070 | default: |

1071 | return 0; |

1072 | } |

1073 | prep_stmt[prep_cnt] = g; |

1074 | } |

1075 | |

1076 | /* Only transform if it removes the condition. */ |

1077 | if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1)) |

1078 | return 0; |

1079 | |

1080 | /* Size-wise, this is always profitable. */ |

1081 | if (optimize_bb_for_speed_p (cond_bb) |

1082 | /* The special case is useless if it has a low probability. */ |

1083 | && profile_status_for_fn (cfun) != PROFILE_ABSENT |

1084 | && EDGE_PRED (middle_bb, 0)->probability < profile_probability::even () |

1085 | /* If assign is cheap, there is no point avoiding it. */ |

1086 | && estimate_num_insns (bb_seq (middle_bb), &eni_time_weights) |

1087 | >= 3 * estimate_num_insns (cond, &eni_time_weights)) |

1088 | return 0; |

1089 | |

1090 | tree lhs = gimple_assign_lhs (assign); |

1091 | tree rhs1 = gimple_assign_rhs1 (assign); |

1092 | tree rhs2 = gimple_assign_rhs2 (assign); |

1093 | enum tree_code code_def = gimple_assign_rhs_code (assign); |

1094 | tree cond_lhs = gimple_cond_lhs (cond); |

1095 | tree cond_rhs = gimple_cond_rhs (cond); |

1096 | |

1097 | /* Propagate the cond_rhs constant through preparation stmts, |

1098 | make sure UB isn't invoked while doing that. */ |

1099 | for (int i = prep_cnt - 1; i >= 0; --i) |

1100 | { |

1101 | gimple *g = prep_stmt[i]; |

1102 | tree grhs1 = gimple_assign_rhs1 (g); |

1103 | if (!operand_equal_for_phi_arg_p (cond_lhs, grhs1)) |

1104 | return 0; |

1105 | cond_lhs = gimple_assign_lhs (g); |

1106 | cond_rhs = fold_convert (TREE_TYPE (grhs1), cond_rhs); |

1107 | if (TREE_CODE (cond_rhs) != INTEGER_CST |

1108 | || TREE_OVERFLOW (cond_rhs)) |

1109 | return 0; |

1110 | if (gimple_assign_rhs_class (g) == GIMPLE_BINARY_RHS) |

1111 | { |

1112 | cond_rhs = int_const_binop (gimple_assign_rhs_code (g), cond_rhs, |

1113 | gimple_assign_rhs2 (g)); |

1114 | if (TREE_OVERFLOW (cond_rhs)) |

1115 | return 0; |

1116 | } |

1117 | cond_rhs = fold_convert (TREE_TYPE (cond_lhs), cond_rhs); |

1118 | if (TREE_CODE (cond_rhs) != INTEGER_CST |

1119 | || TREE_OVERFLOW (cond_rhs)) |

1120 | return 0; |

1121 | } |

1122 | |

1123 | if (((code == NE_EXPR && e1 == false_edge) |

1124 | || (code == EQ_EXPR && e1 == true_edge)) |

1125 | && arg0 == lhs |

1126 | && ((arg1 == rhs1 |

1127 | && operand_equal_for_phi_arg_p (rhs2, cond_lhs) |

1128 | && neutral_element_p (code_def, cond_rhs, true)) |

1129 | || (arg1 == rhs2 |

1130 | && operand_equal_for_phi_arg_p (rhs1, cond_lhs) |

1131 | && neutral_element_p (code_def, cond_rhs, false)) |

1132 | || (operand_equal_for_phi_arg_p (arg1, cond_rhs) |

1133 | && ((operand_equal_for_phi_arg_p (rhs2, cond_lhs) |

1134 | && absorbing_element_p (code_def, cond_rhs, true, rhs2)) |

1135 | || (operand_equal_for_phi_arg_p (rhs1, cond_lhs) |

1136 | && absorbing_element_p (code_def, |

1137 | cond_rhs, false, rhs2)))))) |

1138 | { |

1139 | gsi = gsi_for_stmt (cond); |

1140 | /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6 |

1141 | def-stmt in: |

1142 | if (n_5 != 0) |

1143 | goto <bb 3>; |

1144 | else |

1145 | goto <bb 4>; |

1146 | |

1147 | <bb 3>: |

1148 | # RANGE [0, 4294967294] |

1149 | u_6 = n_5 + 4294967295; |

1150 | |

1151 | <bb 4>: |

1152 | # u_3 = PHI <u_6(3), 4294967295(2)> */ |

1153 | reset_flow_sensitive_info (lhs); |

1154 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))) |

1155 | { |

1156 | /* If available, we can use VR of phi result at least. */ |

1157 | tree phires = gimple_phi_result (phi); |

1158 | struct range_info_def *phires_range_info |

1159 | = SSA_NAME_RANGE_INFO (phires); |

1160 | if (phires_range_info) |

1161 | duplicate_ssa_name_range_info (lhs, SSA_NAME_RANGE_TYPE (phires), |

1162 | phires_range_info); |

1163 | } |

1164 | gimple_stmt_iterator gsi_from; |

1165 | for (int i = prep_cnt - 1; i >= 0; --i) |

1166 | { |

1167 | tree plhs = gimple_assign_lhs (prep_stmt[i]); |

1168 | reset_flow_sensitive_info (plhs); |

1169 | gsi_from = gsi_for_stmt (prep_stmt[i]); |

1170 | gsi_move_before (&gsi_from, &gsi); |

1171 | } |

1172 | gsi_from = gsi_for_stmt (assign); |

1173 | gsi_move_before (&gsi_from, &gsi); |

1174 | replace_phi_edge_with_variable (cond_bb, e1, phi, lhs); |

1175 | return 2; |

1176 | } |

1177 | |

1178 | return 0; |

1179 | } |

1180 | |

1181 | /* The function minmax_replacement does the main work of doing the minmax |

1182 | replacement. Return true if the replacement is done. Otherwise return |

1183 | false. |

1184 | BB is the basic block where the replacement is going to be done on. ARG0 |

1185 | is argument 0 from the PHI. Likewise for ARG1. */ |

1186 | |

1187 | static bool |

1188 | minmax_replacement (basic_block cond_bb, basic_block middle_bb, |

1189 | edge e0, edge e1, gimple *phi, |

1190 | tree arg0, tree arg1) |

1191 | { |

1192 | tree result, type; |

1193 | gcond *cond; |

1194 | gassign *new_stmt; |

1195 | edge true_edge, false_edge; |

1196 | enum tree_code cmp, minmax, ass_code; |

1197 | tree smaller, alt_smaller, larger, alt_larger, arg_true, arg_false; |

1198 | gimple_stmt_iterator gsi, gsi_from; |

1199 | |

1200 | type = TREE_TYPE (PHI_RESULT (phi)); |

1201 | |

1202 | /* The optimization may be unsafe due to NaNs. */ |

1203 | if (HONOR_NANS (type) || HONOR_SIGNED_ZEROS (type)) |

1204 | return false; |

1205 | |

1206 | cond = as_a <gcond *> (last_stmt (cond_bb)); |

1207 | cmp = gimple_cond_code (cond); |

1208 | |

1209 | /* This transformation is only valid for order comparisons. Record which |

1210 | operand is smaller/larger if the result of the comparison is true. */ |

1211 | alt_smaller = NULL_TREE; |

1212 | alt_larger = NULL_TREE; |

1213 | if (cmp == LT_EXPR || cmp == LE_EXPR) |

1214 | { |

1215 | smaller = gimple_cond_lhs (cond); |

1216 | larger = gimple_cond_rhs (cond); |

1217 | /* If we have smaller < CST it is equivalent to smaller <= CST-1. |

1218 | Likewise smaller <= CST is equivalent to smaller < CST+1. */ |

1219 | if (TREE_CODE (larger) == INTEGER_CST) |

1220 | { |

1221 | if (cmp == LT_EXPR) |

1222 | { |

1223 | bool overflow; |

1224 | wide_int alt = wi::sub (wi::to_wide (larger), 1, |

1225 | TYPE_SIGN (TREE_TYPE (larger)), |

1226 | &overflow); |

1227 | if (! overflow) |

1228 | alt_larger = wide_int_to_tree (TREE_TYPE (larger), alt); |

1229 | } |

1230 | else |

1231 | { |

1232 | bool overflow; |

1233 | wide_int alt = wi::add (wi::to_wide (larger), 1, |

1234 | TYPE_SIGN (TREE_TYPE (larger)), |

1235 | &overflow); |

1236 | if (! overflow) |

1237 | alt_larger = wide_int_to_tree (TREE_TYPE (larger), alt); |

1238 | } |

1239 | } |

1240 | } |

1241 | else if (cmp == GT_EXPR || cmp == GE_EXPR) |

1242 | { |

1243 | smaller = gimple_cond_rhs (cond); |

1244 | larger = gimple_cond_lhs (cond); |

1245 | /* If we have larger > CST it is equivalent to larger >= CST+1. |

1246 | Likewise larger >= CST is equivalent to larger > CST-1. */ |

1247 | if (TREE_CODE (smaller) == INTEGER_CST) |

1248 | { |

1249 | if (cmp == GT_EXPR) |

1250 | { |

1251 | bool overflow; |

1252 | wide_int alt = wi::add (wi::to_wide (smaller), 1, |

1253 | TYPE_SIGN (TREE_TYPE (smaller)), |

1254 | &overflow); |

1255 | if (! overflow) |

1256 | alt_smaller = wide_int_to_tree (TREE_TYPE (smaller), alt); |

1257 | } |

1258 | else |

1259 | { |

1260 | bool overflow; |

1261 | wide_int alt = wi::sub (wi::to_wide (smaller), 1, |

1262 | TYPE_SIGN (TREE_TYPE (smaller)), |

1263 | &overflow); |

1264 | if (! overflow) |

1265 | alt_smaller = wide_int_to_tree (TREE_TYPE (smaller), alt); |

1266 | } |

1267 | } |

1268 | } |

1269 | else |

1270 | return false; |

1271 | |

1272 | /* We need to know which is the true edge and which is the false |

1273 | edge so that we know if have abs or negative abs. */ |

1274 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |

1275 | |

1276 | /* Forward the edges over the middle basic block. */ |

1277 | if (true_edge->dest == middle_bb) |

1278 | true_edge = EDGE_SUCC (true_edge->dest, 0); |

1279 | if (false_edge->dest == middle_bb) |

1280 | false_edge = EDGE_SUCC (false_edge->dest, 0); |

1281 | |

1282 | if (true_edge == e0) |

1283 | { |

1284 | gcc_assert (false_edge == e1); |

1285 | arg_true = arg0; |

1286 | arg_false = arg1; |

1287 | } |

1288 | else |

1289 | { |

1290 | gcc_assert (false_edge == e0); |

1291 | gcc_assert (true_edge == e1); |

1292 | arg_true = arg1; |

1293 | arg_false = arg0; |

1294 | } |

1295 | |

1296 | if (empty_block_p (middle_bb)) |

1297 | { |

1298 | if ((operand_equal_for_phi_arg_p (arg_true, smaller) |

1299 | || (alt_smaller |

1300 | && operand_equal_for_phi_arg_p (arg_true, alt_smaller))) |

1301 | && (operand_equal_for_phi_arg_p (arg_false, larger) |

1302 | || (alt_larger |

1303 | && operand_equal_for_phi_arg_p (arg_true, alt_larger)))) |

1304 | { |

1305 | /* Case |

1306 | |

1307 | if (smaller < larger) |

1308 | rslt = smaller; |

1309 | else |

1310 | rslt = larger; */ |

1311 | minmax = MIN_EXPR; |

1312 | } |

1313 | else if ((operand_equal_for_phi_arg_p (arg_false, smaller) |

1314 | || (alt_smaller |

1315 | && operand_equal_for_phi_arg_p (arg_false, alt_smaller))) |

1316 | && (operand_equal_for_phi_arg_p (arg_true, larger) |

1317 | || (alt_larger |

1318 | && operand_equal_for_phi_arg_p (arg_true, alt_larger)))) |

1319 | minmax = MAX_EXPR; |

1320 | else |

1321 | return false; |

1322 | } |

1323 | else |

1324 | { |

1325 | /* Recognize the following case, assuming d <= u: |

1326 | |

1327 | if (a <= u) |

1328 | b = MAX (a, d); |

1329 | x = PHI <b, u> |

1330 | |

1331 | This is equivalent to |

1332 | |

1333 | b = MAX (a, d); |

1334 | x = MIN (b, u); */ |

1335 | |

1336 | gimple *assign = last_and_only_stmt (middle_bb); |

1337 | tree lhs, op0, op1, bound; |

1338 | |

1339 | if (!assign |

1340 | || gimple_code (assign) != GIMPLE_ASSIGN) |

1341 | return false; |

1342 | |

1343 | lhs = gimple_assign_lhs (assign); |

1344 | ass_code = gimple_assign_rhs_code (assign); |

1345 | if (ass_code != MAX_EXPR && ass_code != MIN_EXPR) |

1346 | return false; |

1347 | op0 = gimple_assign_rhs1 (assign); |

1348 | op1 = gimple_assign_rhs2 (assign); |

1349 | |

1350 | if (true_edge->src == middle_bb) |

1351 | { |

1352 | /* We got here if the condition is true, i.e., SMALLER < LARGER. */ |

1353 | if (!operand_equal_for_phi_arg_p (lhs, arg_true)) |

1354 | return false; |

1355 | |

1356 | if (operand_equal_for_phi_arg_p (arg_false, larger) |

1357 | || (alt_larger |

1358 | && operand_equal_for_phi_arg_p (arg_false, alt_larger))) |

1359 | { |

1360 | /* Case |

1361 | |

1362 | if (smaller < larger) |

1363 | { |

1364 | r' = MAX_EXPR (smaller, bound) |

1365 | } |

1366 | r = PHI <r', larger> --> to be turned to MIN_EXPR. */ |

1367 | if (ass_code != MAX_EXPR) |

1368 | return false; |

1369 | |

1370 | minmax = MIN_EXPR; |

1371 | if (operand_equal_for_phi_arg_p (op0, smaller) |

1372 | || (alt_smaller |

1373 | && operand_equal_for_phi_arg_p (op0, alt_smaller))) |

1374 | bound = op1; |

1375 | else if (operand_equal_for_phi_arg_p (op1, smaller) |

1376 | || (alt_smaller |

1377 | && operand_equal_for_phi_arg_p (op1, alt_smaller))) |

1378 | bound = op0; |

1379 | else |

1380 | return false; |

1381 | |

1382 | /* We need BOUND <= LARGER. */ |

1383 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |

1384 | bound, larger))) |

1385 | return false; |

1386 | } |

1387 | else if (operand_equal_for_phi_arg_p (arg_false, smaller) |

1388 | || (alt_smaller |

1389 | && operand_equal_for_phi_arg_p (arg_false, alt_smaller))) |

1390 | { |

1391 | /* Case |

1392 | |

1393 | if (smaller < larger) |

1394 | { |

1395 | r' = MIN_EXPR (larger, bound) |

1396 | } |

1397 | r = PHI <r', smaller> --> to be turned to MAX_EXPR. */ |

1398 | if (ass_code != MIN_EXPR) |

1399 | return false; |

1400 | |

1401 | minmax = MAX_EXPR; |

1402 | if (operand_equal_for_phi_arg_p (op0, larger) |

1403 | || (alt_larger |

1404 | && operand_equal_for_phi_arg_p (op0, alt_larger))) |

1405 | bound = op1; |

1406 | else if (operand_equal_for_phi_arg_p (op1, larger) |

1407 | || (alt_larger |

1408 | && operand_equal_for_phi_arg_p (op1, alt_larger))) |

1409 | bound = op0; |

1410 | else |

1411 | return false; |

1412 | |

1413 | /* We need BOUND >= SMALLER. */ |

1414 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |

1415 | bound, smaller))) |

1416 | return false; |

1417 | } |

1418 | else |

1419 | return false; |

1420 | } |

1421 | else |

1422 | { |

1423 | /* We got here if the condition is false, i.e., SMALLER > LARGER. */ |

1424 | if (!operand_equal_for_phi_arg_p (lhs, arg_false)) |

1425 | return false; |

1426 | |

1427 | if (operand_equal_for_phi_arg_p (arg_true, larger) |

1428 | || (alt_larger |

1429 | && operand_equal_for_phi_arg_p (arg_true, alt_larger))) |

1430 | { |

1431 | /* Case |

1432 | |

1433 | if (smaller > larger) |

1434 | { |

1435 | r' = MIN_EXPR (smaller, bound) |

1436 | } |

1437 | r = PHI <r', larger> --> to be turned to MAX_EXPR. */ |

1438 | if (ass_code != MIN_EXPR) |

1439 | return false; |

1440 | |

1441 | minmax = MAX_EXPR; |

1442 | if (operand_equal_for_phi_arg_p (op0, smaller) |

1443 | || (alt_smaller |

1444 | && operand_equal_for_phi_arg_p (op0, alt_smaller))) |

1445 | bound = op1; |

1446 | else if (operand_equal_for_phi_arg_p (op1, smaller) |

1447 | || (alt_smaller |

1448 | && operand_equal_for_phi_arg_p (op1, alt_smaller))) |

1449 | bound = op0; |

1450 | else |

1451 | return false; |

1452 | |

1453 | /* We need BOUND >= LARGER. */ |

1454 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |

1455 | bound, larger))) |

1456 | return false; |

1457 | } |

1458 | else if (operand_equal_for_phi_arg_p (arg_true, smaller) |

1459 | || (alt_smaller |

1460 | && operand_equal_for_phi_arg_p (arg_true, alt_smaller))) |

1461 | { |

1462 | /* Case |

1463 | |

1464 | if (smaller > larger) |

1465 | { |

1466 | r' = MAX_EXPR (larger, bound) |

1467 | } |

1468 | r = PHI <r', smaller> --> to be turned to MIN_EXPR. */ |

1469 | if (ass_code != MAX_EXPR) |

1470 | return false; |

1471 | |

1472 | minmax = MIN_EXPR; |

1473 | if (operand_equal_for_phi_arg_p (op0, larger)) |

1474 | bound = op1; |

1475 | else if (operand_equal_for_phi_arg_p (op1, larger)) |

1476 | bound = op0; |

1477 | else |

1478 | return false; |

1479 | |

1480 | /* We need BOUND <= SMALLER. */ |

1481 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |

1482 | bound, smaller))) |

1483 | return false; |

1484 | } |

1485 | else |

1486 | return false; |

1487 | } |

1488 | |

1489 | /* Move the statement from the middle block. */ |

1490 | gsi = gsi_last_bb (cond_bb); |

1491 | gsi_from = gsi_last_nondebug_bb (middle_bb); |

1492 | reset_flow_sensitive_info (SINGLE_SSA_TREE_OPERAND (gsi_stmt (gsi_from), |

1493 | SSA_OP_DEF)); |

1494 | gsi_move_before (&gsi_from, &gsi); |

1495 | } |

1496 | |

1497 | /* Create an SSA var to hold the min/max result. If we're the only |

1498 | things setting the target PHI, then we can clone the PHI |

1499 | variable. Otherwise we must create a new one. */ |

1500 | result = PHI_RESULT (phi); |

1501 | if (EDGE_COUNT (gimple_bb (phi)->preds) == 2) |

1502 | result = duplicate_ssa_name (result, NULL); |

1503 | else |

1504 | result = make_ssa_name (TREE_TYPE (result)); |

1505 | |

1506 | /* Emit the statement to compute min/max. */ |

1507 | new_stmt = gimple_build_assign (result, minmax, arg0, arg1); |

1508 | gsi = gsi_last_bb (cond_bb); |

1509 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |

1510 | |

1511 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |

1512 | |

1513 | return true; |

1514 | } |

1515 | |

1516 | /* The function absolute_replacement does the main work of doing the absolute |

1517 | replacement. Return true if the replacement is done. Otherwise return |

1518 | false. |

1519 | bb is the basic block where the replacement is going to be done on. arg0 |

1520 | is argument 0 from the phi. Likewise for arg1. */ |

1521 | |

1522 | static bool |

1523 | abs_replacement (basic_block cond_bb, basic_block middle_bb, |

1524 | edge e0 ATTRIBUTE_UNUSED, edge e1, |

1525 | gimple *phi, tree arg0, tree arg1) |

1526 | { |

1527 | tree result; |

1528 | gassign *new_stmt; |

1529 | gimple *cond; |

1530 | gimple_stmt_iterator gsi; |

1531 | edge true_edge, false_edge; |

1532 | gimple *assign; |

1533 | edge e; |

1534 | tree rhs, lhs; |

1535 | bool negate; |

1536 | enum tree_code cond_code; |

1537 | |

1538 | /* If the type says honor signed zeros we cannot do this |

1539 | optimization. */ |

1540 | if (HONOR_SIGNED_ZEROS (arg1)) |

1541 | return false; |

1542 | |

1543 | /* OTHER_BLOCK must have only one executable statement which must have the |

1544 | form arg0 = -arg1 or arg1 = -arg0. */ |

1545 | |

1546 | assign = last_and_only_stmt (middle_bb); |

1547 | /* If we did not find the proper negation assignment, then we can not |

1548 | optimize. */ |

1549 | if (assign == NULL) |

1550 | return false; |

1551 | |

1552 | /* If we got here, then we have found the only executable statement |

1553 | in OTHER_BLOCK. If it is anything other than arg = -arg1 or |

1554 | arg1 = -arg0, then we can not optimize. */ |

1555 | if (gimple_code (assign) != GIMPLE_ASSIGN) |

1556 | return false; |

1557 | |

1558 | lhs = gimple_assign_lhs (assign); |

1559 | |

1560 | if (gimple_assign_rhs_code (assign) != NEGATE_EXPR) |

1561 | return false; |

1562 | |

1563 | rhs = gimple_assign_rhs1 (assign); |

1564 | |

1565 | /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */ |

1566 | if (!(lhs == arg0 && rhs == arg1) |

1567 | && !(lhs == arg1 && rhs == arg0)) |

1568 | return false; |

1569 | |

1570 | cond = last_stmt (cond_bb); |

1571 | result = PHI_RESULT (phi); |

1572 | |

1573 | /* Only relationals comparing arg[01] against zero are interesting. */ |

1574 | cond_code = gimple_cond_code (cond); |

1575 | if (cond_code != GT_EXPR && cond_code != GE_EXPR |

1576 | && cond_code != LT_EXPR && cond_code != LE_EXPR) |

1577 | return false; |

1578 | |

1579 | /* Make sure the conditional is arg[01] OP y. */ |

1580 | if (gimple_cond_lhs (cond) != rhs) |

1581 | return false; |

1582 | |

1583 | if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond))) |

1584 | ? real_zerop (gimple_cond_rhs (cond)) |

1585 | : integer_zerop (gimple_cond_rhs (cond))) |

1586 | ; |

1587 | else |

1588 | return false; |

1589 | |

1590 | /* We need to know which is the true edge and which is the false |

1591 | edge so that we know if have abs or negative abs. */ |

1592 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |

1593 | |

1594 | /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we |

1595 | will need to negate the result. Similarly for LT_EXPR/LE_EXPR if |

1596 | the false edge goes to OTHER_BLOCK. */ |

1597 | if (cond_code == GT_EXPR || cond_code == GE_EXPR) |

1598 | e = true_edge; |

1599 | else |

1600 | e = false_edge; |

1601 | |

1602 | if (e->dest == middle_bb) |

1603 | negate = true; |

1604 | else |

1605 | negate = false; |

1606 | |

1607 | /* If the code negates only iff positive then make sure to not |

1608 | introduce undefined behavior when negating or computing the absolute. |

1609 | ??? We could use range info if present to check for arg1 == INT_MIN. */ |

1610 | if (negate |

1611 | && (ANY_INTEGRAL_TYPE_P (TREE_TYPE (arg1)) |

1612 | && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (arg1)))) |

1613 | return false; |

1614 | |

1615 | result = duplicate_ssa_name (result, NULL); |

1616 | |

1617 | if (negate) |

1618 | lhs = make_ssa_name (TREE_TYPE (result)); |

1619 | else |

1620 | lhs = result; |

1621 | |

1622 | /* Build the modify expression with abs expression. */ |

1623 | new_stmt = gimple_build_assign (lhs, ABS_EXPR, rhs); |

1624 | |

1625 | gsi = gsi_last_bb (cond_bb); |

1626 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |

1627 | |

1628 | if (negate) |

1629 | { |

1630 | /* Get the right GSI. We want to insert after the recently |

1631 | added ABS_EXPR statement (which we know is the first statement |

1632 | in the block. */ |

1633 | new_stmt = gimple_build_assign (result, NEGATE_EXPR, lhs); |

1634 | |

1635 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); |

1636 | } |

1637 | |

1638 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |

1639 | |

1640 | /* Note that we optimized this PHI. */ |

1641 | return true; |

1642 | } |

1643 | |

1644 | /* Auxiliary functions to determine the set of memory accesses which |

1645 | can't trap because they are preceded by accesses to the same memory |

1646 | portion. We do that for MEM_REFs, so we only need to track |

1647 | the SSA_NAME of the pointer indirectly referenced. The algorithm |

1648 | simply is a walk over all instructions in dominator order. When |

1649 | we see an MEM_REF we determine if we've already seen a same |

1650 | ref anywhere up to the root of the dominator tree. If we do the |

1651 | current access can't trap. If we don't see any dominating access |

1652 | the current access might trap, but might also make later accesses |

1653 | non-trapping, so we remember it. We need to be careful with loads |

1654 | or stores, for instance a load might not trap, while a store would, |

1655 | so if we see a dominating read access this doesn't mean that a later |

1656 | write access would not trap. Hence we also need to differentiate the |

1657 | type of access(es) seen. |

1658 | |

1659 | ??? We currently are very conservative and assume that a load might |

1660 | trap even if a store doesn't (write-only memory). This probably is |

1661 | overly conservative. */ |

1662 | |

1663 | /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF |

1664 | through it was seen, which would constitute a no-trap region for |

1665 | same accesses. */ |

1666 | struct name_to_bb |

1667 | { |

1668 | unsigned int ssa_name_ver; |

1669 | unsigned int phase; |

1670 | bool store; |

1671 | HOST_WIDE_INT offset, size; |

1672 | basic_block bb; |

1673 | }; |

1674 | |

1675 | /* Hashtable helpers. */ |

1676 | |

1677 | struct ssa_names_hasher : free_ptr_hash <name_to_bb> |

1678 | { |

1679 | static inline hashval_t hash (const name_to_bb *); |

1680 | static inline bool equal (const name_to_bb *, const name_to_bb *); |

1681 | }; |

1682 | |

1683 | /* Used for quick clearing of the hash-table when we see calls. |

1684 | Hash entries with phase < nt_call_phase are invalid. */ |

1685 | static unsigned int nt_call_phase; |

1686 | |

1687 | /* The hash function. */ |

1688 | |

1689 | inline hashval_t |

1690 | ssa_names_hasher::hash (const name_to_bb *n) |

1691 | { |

1692 | return n->ssa_name_ver ^ (((hashval_t) n->store) << 31) |

1693 | ^ (n->offset << 6) ^ (n->size << 3); |

1694 | } |

1695 | |

1696 | /* The equality function of *P1 and *P2. */ |

1697 | |

1698 | inline bool |

1699 | ssa_names_hasher::equal (const name_to_bb *n1, const name_to_bb *n2) |

1700 | { |

1701 | return n1->ssa_name_ver == n2->ssa_name_ver |

1702 | && n1->store == n2->store |

1703 | && n1->offset == n2->offset |

1704 | && n1->size == n2->size; |

1705 | } |

1706 | |

1707 | class nontrapping_dom_walker : public dom_walker |

1708 | { |

1709 | public: |

1710 | nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps) |

1711 | : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {} |

1712 | |

1713 | virtual edge before_dom_children (basic_block); |

1714 | virtual void after_dom_children (basic_block); |

1715 | |

1716 | private: |

1717 | |

1718 | /* We see the expression EXP in basic block BB. If it's an interesting |

1719 | expression (an MEM_REF through an SSA_NAME) possibly insert the |

1720 | expression into the set NONTRAP or the hash table of seen expressions. |

1721 | STORE is true if this expression is on the LHS, otherwise it's on |

1722 | the RHS. */ |

1723 | void add_or_mark_expr (basic_block, tree, bool); |

1724 | |

1725 | hash_set<tree> *m_nontrapping; |

1726 | |

1727 | /* The hash table for remembering what we've seen. */ |

1728 | hash_table<ssa_names_hasher> m_seen_ssa_names; |

1729 | }; |

1730 | |

1731 | /* Called by walk_dominator_tree, when entering the block BB. */ |

1732 | edge |

1733 | nontrapping_dom_walker::before_dom_children (basic_block bb) |

1734 | { |

1735 | edge e; |

1736 | edge_iterator ei; |

1737 | gimple_stmt_iterator gsi; |

1738 | |

1739 | /* If we haven't seen all our predecessors, clear the hash-table. */ |

1740 | FOR_EACH_EDGE (e, ei, bb->preds) |

1741 | if ((((size_t)e->src->aux) & 2) == 0) |

1742 | { |

1743 | nt_call_phase++; |

1744 | break; |

1745 | } |

1746 | |

1747 | /* Mark this BB as being on the path to dominator root and as visited. */ |

1748 | bb->aux = (void*)(1 | 2); |

1749 | |

1750 | /* And walk the statements in order. */ |

1751 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |

1752 | { |

1753 | gimple *stmt = gsi_stmt (gsi); |

1754 | |

1755 | if ((gimple_code (stmt) == GIMPLE_ASM && gimple_vdef (stmt)) |

1756 | || (is_gimple_call (stmt) |

1757 | && (!nonfreeing_call_p (stmt) || !nonbarrier_call_p (stmt)))) |

1758 | nt_call_phase++; |

1759 | else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt)) |

1760 | { |

1761 | add_or_mark_expr (bb, gimple_assign_lhs (stmt), true); |

1762 | add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false); |

1763 | } |

1764 | } |

1765 | return NULL; |

1766 | } |

1767 | |

1768 | /* Called by walk_dominator_tree, when basic block BB is exited. */ |

1769 | void |

1770 | nontrapping_dom_walker::after_dom_children (basic_block bb) |

1771 | { |

1772 | /* This BB isn't on the path to dominator root anymore. */ |

1773 | bb->aux = (void*)2; |

1774 | } |

1775 | |

1776 | /* We see the expression EXP in basic block BB. If it's an interesting |

1777 | expression (an MEM_REF through an SSA_NAME) possibly insert the |

1778 | expression into the set NONTRAP or the hash table of seen expressions. |

1779 | STORE is true if this expression is on the LHS, otherwise it's on |

1780 | the RHS. */ |

1781 | void |

1782 | nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store) |

1783 | { |

1784 | HOST_WIDE_INT size; |

1785 | |

1786 | if (TREE_CODE (exp) == MEM_REF |

1787 | && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME |

1788 | && tree_fits_shwi_p (TREE_OPERAND (exp, 1)) |

1789 | && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0) |

1790 | { |

1791 | tree name = TREE_OPERAND (exp, 0); |

1792 | struct name_to_bb map; |

1793 | name_to_bb **slot; |

1794 | struct name_to_bb *n2bb; |

1795 | basic_block found_bb = 0; |

1796 | |

1797 | /* Try to find the last seen MEM_REF through the same |

1798 | SSA_NAME, which can trap. */ |

1799 | map.ssa_name_ver = SSA_NAME_VERSION (name); |

1800 | map.phase = 0; |

1801 | map.bb = 0; |

1802 | map.store = store; |

1803 | map.offset = tree_to_shwi (TREE_OPERAND (exp, 1)); |

1804 | map.size = size; |

1805 | |

1806 | slot = m_seen_ssa_names.find_slot (&map, INSERT); |

1807 | n2bb = *slot; |

1808 | if (n2bb && n2bb->phase >= nt_call_phase) |

1809 | found_bb = n2bb->bb; |

1810 | |

1811 | /* If we've found a trapping MEM_REF, _and_ it dominates EXP |

1812 | (it's in a basic block on the path from us to the dominator root) |

1813 | then we can't trap. */ |

1814 | if (found_bb && (((size_t)found_bb->aux) & 1) == 1) |

1815 | { |

1816 | m_nontrapping->add (exp); |

1817 | } |

1818 | else |

1819 | { |

1820 | /* EXP might trap, so insert it into the hash table. */ |

1821 | if (n2bb) |

1822 | { |

1823 | n2bb->phase = nt_call_phase; |

1824 | n2bb->bb = bb; |

1825 | } |

1826 | else |

1827 | { |

1828 | n2bb = XNEW (struct name_to_bb); |

1829 | n2bb->ssa_name_ver = SSA_NAME_VERSION (name); |

1830 | n2bb->phase = nt_call_phase; |

1831 | n2bb->bb = bb; |

1832 | n2bb->store = store; |

1833 | n2bb->offset = map.offset; |

1834 | n2bb->size = size; |

1835 | *slot = n2bb; |

1836 | } |

1837 | } |

1838 | } |

1839 | } |

1840 | |

1841 | /* This is the entry point of gathering non trapping memory accesses. |

1842 | It will do a dominator walk over the whole function, and it will |

1843 | make use of the bb->aux pointers. It returns a set of trees |

1844 | (the MEM_REFs itself) which can't trap. */ |

1845 | static hash_set<tree> * |

1846 | get_non_trapping (void) |

1847 | { |

1848 | nt_call_phase = 0; |

1849 | hash_set<tree> *nontrap = new hash_set<tree>; |

1850 | /* We're going to do a dominator walk, so ensure that we have |

1851 | dominance information. */ |

1852 | calculate_dominance_info (CDI_DOMINATORS); |

1853 | |

1854 | nontrapping_dom_walker (CDI_DOMINATORS, nontrap) |

1855 | .walk (cfun->cfg->x_entry_block_ptr); |

1856 | |

1857 | clear_aux_for_blocks (); |

1858 | return nontrap; |

1859 | } |

1860 | |

1861 | /* Do the main work of conditional store replacement. We already know |

1862 | that the recognized pattern looks like so: |

1863 | |

1864 | split: |

1865 | if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1) |

1866 | MIDDLE_BB: |

1867 | something |

1868 | fallthrough (edge E0) |

1869 | JOIN_BB: |

1870 | some more |

1871 | |

1872 | We check that MIDDLE_BB contains only one store, that that store |

1873 | doesn't trap (not via NOTRAP, but via checking if an access to the same |

1874 | memory location dominates us) and that the store has a "simple" RHS. */ |

1875 | |

1876 | static bool |

1877 | cond_store_replacement (basic_block middle_bb, basic_block join_bb, |

1878 | edge e0, edge e1, hash_set<tree> *nontrap) |

1879 | { |

1880 | gimple *assign = last_and_only_stmt (middle_bb); |

1881 | tree lhs, rhs, name, name2; |

1882 | gphi *newphi; |

1883 | gassign *new_stmt; |

1884 | gimple_stmt_iterator gsi; |

1885 | source_location locus; |

1886 | |

1887 | /* Check if middle_bb contains of only one store. */ |

1888 | if (!assign |

1889 | || !gimple_assign_single_p (assign) |

1890 | || gimple_has_volatile_ops (assign)) |

1891 | return false; |

1892 | |

1893 | locus = gimple_location (assign); |

1894 | lhs = gimple_assign_lhs (assign); |

1895 | rhs = gimple_assign_rhs1 (assign); |

1896 | if (TREE_CODE (lhs) != MEM_REF |

1897 | || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME |

1898 | || !is_gimple_reg_type (TREE_TYPE (lhs))) |

1899 | return false; |

1900 | |

1901 | /* Prove that we can move the store down. We could also check |

1902 | TREE_THIS_NOTRAP here, but in that case we also could move stores, |

1903 | whose value is not available readily, which we want to avoid. */ |

1904 | if (!nontrap->contains (lhs)) |

1905 | return false; |

1906 | |

1907 | /* Now we've checked the constraints, so do the transformation: |

1908 | 1) Remove the single store. */ |

1909 | gsi = gsi_for_stmt (assign); |

1910 | unlink_stmt_vdef (assign); |

1911 | gsi_remove (&gsi, true); |

1912 | release_defs (assign); |

1913 | |

1914 | /* Make both store and load use alias-set zero as we have to |

1915 | deal with the case of the store being a conditional change |

1916 | of the dynamic type. */ |

1917 | lhs = unshare_expr (lhs); |

1918 | tree *basep = &lhs; |

1919 | while (handled_component_p (*basep)) |

1920 | basep = &TREE_OPERAND (*basep, 0); |

1921 | if (TREE_CODE (*basep) == MEM_REF |

1922 | || TREE_CODE (*basep) == TARGET_MEM_REF) |

1923 | TREE_OPERAND (*basep, 1) |

1924 | = fold_convert (ptr_type_node, TREE_OPERAND (*basep, 1)); |

1925 | else |

1926 | *basep = build2 (MEM_REF, TREE_TYPE (*basep), |

1927 | build_fold_addr_expr (*basep), |

1928 | build_zero_cst (ptr_type_node)); |

1929 | |

1930 | /* 2) Insert a load from the memory of the store to the temporary |

1931 | on the edge which did not contain the store. */ |

1932 | name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |

1933 | new_stmt = gimple_build_assign (name, lhs); |

1934 | gimple_set_location (new_stmt, locus); |

1935 | gsi_insert_on_edge (e1, new_stmt); |

1936 | |

1937 | /* 3) Create a PHI node at the join block, with one argument |

1938 | holding the old RHS, and the other holding the temporary |

1939 | where we stored the old memory contents. */ |

1940 | name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |

1941 | newphi = create_phi_node (name2, join_bb); |

1942 | add_phi_arg (newphi, rhs, e0, locus); |

1943 | add_phi_arg (newphi, name, e1, locus); |

1944 | |

1945 | lhs = unshare_expr (lhs); |

1946 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); |

1947 | |

1948 | /* 4) Insert that PHI node. */ |

1949 | gsi = gsi_after_labels (join_bb); |

1950 | if (gsi_end_p (gsi)) |

1951 | { |

1952 | gsi = gsi_last_bb (join_bb); |

1953 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); |

1954 | } |

1955 | else |

1956 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |

1957 | |

1958 | return true; |

1959 | } |

1960 | |

1961 | /* Do the main work of conditional store replacement. */ |

1962 | |

1963 | static bool |

1964 | cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb, |

1965 | basic_block join_bb, gimple *then_assign, |

1966 | gimple *else_assign) |

1967 | { |

1968 | tree lhs_base, lhs, then_rhs, else_rhs, name; |

1969 | source_location then_locus, else_locus; |

1970 | gimple_stmt_iterator gsi; |

1971 | gphi *newphi; |

1972 | gassign *new_stmt; |

1973 | |

1974 | if (then_assign == NULL |

1975 | || !gimple_assign_single_p (then_assign) |

1976 | || gimple_clobber_p (then_assign) |

1977 | || gimple_has_volatile_ops (then_assign) |

1978 | || else_assign == NULL |

1979 | || !gimple_assign_single_p (else_assign) |

1980 | || gimple_clobber_p (else_assign) |

1981 | || gimple_has_volatile_ops (else_assign)) |

1982 | return false; |

1983 | |

1984 | lhs = gimple_assign_lhs (then_assign); |

1985 | if (!is_gimple_reg_type (TREE_TYPE (lhs)) |

1986 | || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0)) |

1987 | return false; |

1988 | |

1989 | lhs_base = get_base_address (lhs); |

1990 | if (lhs_base == NULL_TREE |

1991 | || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF)) |

1992 | return false; |

1993 | |

1994 | then_rhs = gimple_assign_rhs1 (then_assign); |

1995 | else_rhs = gimple_assign_rhs1 (else_assign); |

1996 | then_locus = gimple_location (then_assign); |

1997 | else_locus = gimple_location (else_assign); |

1998 | |

1999 | /* Now we've checked the constraints, so do the transformation: |

2000 | 1) Remove the stores. */ |

2001 | gsi = gsi_for_stmt (then_assign); |

2002 | unlink_stmt_vdef (then_assign); |

2003 | gsi_remove (&gsi, true); |

2004 | release_defs (then_assign); |

2005 | |

2006 | gsi = gsi_for_stmt (else_assign); |

2007 | unlink_stmt_vdef (else_assign); |

2008 | gsi_remove (&gsi, true); |

2009 | release_defs (else_assign); |

2010 | |

2011 | /* 2) Create a PHI node at the join block, with one argument |

2012 | holding the old RHS, and the other holding the temporary |

2013 | where we stored the old memory contents. */ |

2014 | name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |

2015 | newphi = create_phi_node (name, join_bb); |

2016 | add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus); |

2017 | add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus); |

2018 | |

2019 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); |

2020 | |

2021 | /* 3) Insert that PHI node. */ |

2022 | gsi = gsi_after_labels (join_bb); |

2023 | if (gsi_end_p (gsi)) |

2024 | { |

2025 | gsi = gsi_last_bb (join_bb); |

2026 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); |

2027 | } |

2028 | else |

2029 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |

2030 | |

2031 | return true; |

2032 | } |

2033 | |

2034 | /* Conditional store replacement. We already know |

2035 | that the recognized pattern looks like so: |

2036 | |

2037 | split: |

2038 | if (cond) goto THEN_BB; else goto ELSE_BB (edge E1) |

2039 | THEN_BB: |

2040 | ... |

2041 | X = Y; |

2042 | ... |

2043 | goto JOIN_BB; |

2044 | ELSE_BB: |

2045 | ... |

2046 | X = Z; |

2047 | ... |

2048 | fallthrough (edge E0) |

2049 | JOIN_BB: |

2050 | some more |

2051 | |

2052 | We check that it is safe to sink the store to JOIN_BB by verifying that |

2053 | there are no read-after-write or write-after-write dependencies in |

2054 | THEN_BB and ELSE_BB. */ |

2055 | |

2056 | static bool |

2057 | cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb, |

2058 | basic_block join_bb) |

2059 | { |

2060 | gimple *then_assign = last_and_only_stmt (then_bb); |

2061 | gimple *else_assign = last_and_only_stmt (else_bb); |

2062 | vec<data_reference_p> then_datarefs, else_datarefs; |

2063 | vec<ddr_p> then_ddrs, else_ddrs; |

2064 | gimple *then_store, *else_store; |

2065 | bool found, ok = false, res; |

2066 | struct data_dependence_relation *ddr; |

2067 | data_reference_p then_dr, else_dr; |

2068 | int i, j; |

2069 | tree then_lhs, else_lhs; |

2070 | basic_block blocks[3]; |

2071 | |

2072 | if (MAX_STORES_TO_SINK == 0) |

2073 | return false; |

2074 | |

2075 | /* Handle the case with single statement in THEN_BB and ELSE_BB. */ |

2076 | if (then_assign && else_assign) |

2077 | return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, |

2078 | then_assign, else_assign); |

2079 | |

2080 | /* Find data references. */ |

2081 | then_datarefs.create (1); |

2082 | else_datarefs.create (1); |

2083 | if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs) |

2084 | == chrec_dont_know) |

2085 | || !then_datarefs.length () |

2086 | || (find_data_references_in_bb (NULL, else_bb, &else_datarefs) |

2087 | == chrec_dont_know) |

2088 | || !else_datarefs.length ()) |

2089 | { |

2090 | free_data_refs (then_datarefs); |

2091 | free_data_refs (else_datarefs); |

2092 | return false; |

2093 | } |

2094 | |

2095 | /* Find pairs of stores with equal LHS. */ |

2096 | auto_vec<gimple *, 1> then_stores, else_stores; |

2097 | FOR_EACH_VEC_ELT (then_datarefs, i, then_dr) |

2098 | { |

2099 | if (DR_IS_READ (then_dr)) |

2100 | continue; |

2101 | |

2102 | then_store = DR_STMT (then_dr); |

2103 | then_lhs = gimple_get_lhs (then_store); |

2104 | if (then_lhs == NULL_TREE) |

2105 | continue; |

2106 | found = false; |

2107 | |

2108 | FOR_EACH_VEC_ELT (else_datarefs, j, else_dr) |

2109 | { |

2110 | if (DR_IS_READ (else_dr)) |

2111 | continue; |

2112 | |

2113 | else_store = DR_STMT (else_dr); |

2114 | else_lhs = gimple_get_lhs (else_store); |

2115 | if (else_lhs == NULL_TREE) |

2116 | continue; |

2117 | |

2118 | if (operand_equal_p (then_lhs, else_lhs, 0)) |

2119 | { |

2120 | found = true; |

2121 | break; |

2122 | } |

2123 | } |

2124 | |

2125 | if (!found) |

2126 | continue; |

2127 | |

2128 | then_stores.safe_push (then_store); |

2129 | else_stores.safe_push (else_store); |

2130 | } |

2131 | |

2132 | /* No pairs of stores found. */ |

2133 | if (!then_stores.length () |

2134 | || then_stores.length () > (unsigned) MAX_STORES_TO_SINK) |

2135 | { |

2136 | free_data_refs (then_datarefs); |

2137 | free_data_refs (else_datarefs); |

2138 | return false; |

2139 | } |

2140 | |

2141 | /* Compute and check data dependencies in both basic blocks. */ |

2142 | then_ddrs.create (1); |

2143 | else_ddrs.create (1); |

2144 | if (!compute_all_dependences (then_datarefs, &then_ddrs, |

2145 | vNULL, false) |

2146 | || !compute_all_dependences (else_datarefs, &else_ddrs, |

2147 | vNULL, false)) |

2148 | { |

2149 | free_dependence_relations (then_ddrs); |

2150 | free_dependence_relations (else_ddrs); |

2151 | free_data_refs (then_datarefs); |

2152 | free_data_refs (else_datarefs); |

2153 | return false; |

2154 | } |

2155 | blocks[0] = then_bb; |

2156 | blocks[1] = else_bb; |

2157 | blocks[2] = join_bb; |

2158 | renumber_gimple_stmt_uids_in_blocks (blocks, 3); |

2159 | |

2160 | /* Check that there are no read-after-write or write-after-write dependencies |

2161 | in THEN_BB. */ |

2162 | FOR_EACH_VEC_ELT (then_ddrs, i, ddr) |

2163 | { |

2164 | struct data_reference *dra = DDR_A (ddr); |

2165 | struct data_reference *drb = DDR_B (ddr); |

2166 | |

2167 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known |

2168 | && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) |

2169 | && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) |

2170 | || (DR_IS_READ (drb) && DR_IS_WRITE (dra) |

2171 | && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) |

2172 | || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) |

2173 | { |

2174 | free_dependence_relations (then_ddrs); |

2175 | free_dependence_relations (else_ddrs); |

2176 | free_data_refs (then_datarefs); |

2177 | free_data_refs (else_datarefs); |

2178 | return false; |

2179 | } |

2180 | } |

2181 | |

2182 | /* Check that there are no read-after-write or write-after-write dependencies |

2183 | in ELSE_BB. */ |

2184 | FOR_EACH_VEC_ELT (else_ddrs, i, ddr) |

2185 | { |

2186 | struct data_reference *dra = DDR_A (ddr); |

2187 | struct data_reference *drb = DDR_B (ddr); |

2188 | |

2189 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known |

2190 | && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) |

2191 | && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) |

2192 | || (DR_IS_READ (drb) && DR_IS_WRITE (dra) |

2193 | && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) |

2194 | || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) |

2195 | { |

2196 | free_dependence_relations (then_ddrs); |

2197 | free_dependence_relations (else_ddrs); |

2198 | free_data_refs (then_datarefs); |

2199 | free_data_refs (else_datarefs); |

2200 | return false; |

2201 | } |

2202 | } |

2203 | |

2204 | /* Sink stores with same LHS. */ |

2205 | FOR_EACH_VEC_ELT (then_stores, i, then_store) |

2206 | { |

2207 | else_store = else_stores[i]; |

2208 | res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, |

2209 | then_store, else_store); |

2210 | ok = ok || res; |

2211 | } |

2212 | |

2213 | free_dependence_relations (then_ddrs); |

2214 | free_dependence_relations (else_ddrs); |

2215 | free_data_refs (then_datarefs); |

2216 | free_data_refs (else_datarefs); |

2217 | |

2218 | return ok; |

2219 | } |

2220 | |

2221 | /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */ |

2222 | |

2223 | static bool |

2224 | local_mem_dependence (gimple *stmt, basic_block bb) |

2225 | { |

2226 | tree vuse = gimple_vuse (stmt); |

2227 | gimple *def; |

2228 | |

2229 | if (!vuse) |

2230 | return false; |

2231 | |

2232 | def = SSA_NAME_DEF_STMT (vuse); |

2233 | return (def && gimple_bb (def) == bb); |

2234 | } |

2235 | |

2236 | /* Given a "diamond" control-flow pattern where BB0 tests a condition, |

2237 | BB1 and BB2 are "then" and "else" blocks dependent on this test, |

2238 | and BB3 rejoins control flow following BB1 and BB2, look for |

2239 | opportunities to hoist loads as follows. If BB3 contains a PHI of |

2240 | two loads, one each occurring in BB1 and BB2, and the loads are |

2241 | provably of adjacent fields in the same structure, then move both |

2242 | loads into BB0. Of course this can only be done if there are no |

2243 | dependencies preventing such motion. |

2244 | |

2245 | One of the hoisted loads will always be speculative, so the |

2246 | transformation is currently conservative: |

2247 | |

2248 | - The fields must be strictly adjacent. |

2249 | - The two fields must occupy a single memory block that is |

2250 | guaranteed to not cross a page boundary. |

2251 | |

2252 | The last is difficult to prove, as such memory blocks should be |

2253 | aligned on the minimum of the stack alignment boundary and the |

2254 | alignment guaranteed by heap allocation interfaces. Thus we rely |

2255 | on a parameter for the alignment value. |

2256 | |

2257 | Provided a good value is used for the last case, the first |

2258 | restriction could possibly be relaxed. */ |

2259 | |

2260 | static void |

2261 | hoist_adjacent_loads (basic_block bb0, basic_block bb1, |

2262 | basic_block bb2, basic_block bb3) |

2263 | { |

2264 | int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE); |

2265 | unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT); |

2266 | gphi_iterator gsi; |

2267 | |

2268 | /* Walk the phis in bb3 looking for an opportunity. We are looking |

2269 | for phis of two SSA names, one each of which is defined in bb1 and |

2270 | bb2. */ |

2271 | for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi)) |

2272 | { |

2273 | gphi *phi_stmt = gsi.phi (); |

2274 | gimple *def1, *def2; |

2275 | tree arg1, arg2, ref1, ref2, field1, field2; |

2276 | tree tree_offset1, tree_offset2, tree_size2, next; |

2277 | int offset1, offset2, size2; |

2278 | unsigned align1; |

2279 | gimple_stmt_iterator gsi2; |

2280 | basic_block bb_for_def1, bb_for_def2; |

2281 | |

2282 | if (gimple_phi_num_args (phi_stmt) != 2 |

2283 | || virtual_operand_p (gimple_phi_result (phi_stmt))) |

2284 | continue; |

2285 | |

2286 | arg1 = gimple_phi_arg_def (phi_stmt, 0); |

2287 | arg2 = gimple_phi_arg_def (phi_stmt, 1); |

2288 | |

2289 | if (TREE_CODE (arg1) != SSA_NAME |

2290 | || TREE_CODE (arg2) != SSA_NAME |

2291 | || SSA_NAME_IS_DEFAULT_DEF (arg1) |

2292 | || SSA_NAME_IS_DEFAULT_DEF (arg2)) |

2293 | continue; |

2294 | |

2295 | def1 = SSA_NAME_DEF_STMT (arg1); |

2296 | def2 = SSA_NAME_DEF_STMT (arg2); |

2297 | |

2298 | if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2) |

2299 | && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2)) |

2300 | continue; |

2301 | |

2302 | /* Check the mode of the arguments to be sure a conditional move |

2303 | can be generated for it. */ |

2304 | if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1))) |

2305 | == CODE_FOR_nothing) |

2306 | continue; |

2307 | |

2308 | /* Both statements must be assignments whose RHS is a COMPONENT_REF. */ |

2309 | if (!gimple_assign_single_p (def1) |

2310 | || !gimple_assign_single_p (def2) |

2311 | || gimple_has_volatile_ops (def1) |

2312 | || gimple_has_volatile_ops (def2)) |

2313 | continue; |

2314 | |

2315 | ref1 = gimple_assign_rhs1 (def1); |

2316 | ref2 = gimple_assign_rhs1 (def2); |

2317 | |

2318 | if (TREE_CODE (ref1) != COMPONENT_REF |

2319 | || TREE_CODE (ref2) != COMPONENT_REF) |

2320 | continue; |

2321 | |

2322 | /* The zeroth operand of the two component references must be |

2323 | identical. It is not sufficient to compare get_base_address of |

2324 | the two references, because this could allow for different |

2325 | elements of the same array in the two trees. It is not safe to |

2326 | assume that the existence of one array element implies the |

2327 | existence of a different one. */ |

2328 | if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0)) |

2329 | continue; |

2330 | |

2331 | field1 = TREE_OPERAND (ref1, 1); |

2332 | field2 = TREE_OPERAND (ref2, 1); |

2333 | |

2334 | /* Check for field adjacency, and ensure field1 comes first. */ |

2335 | for (next = DECL_CHAIN (field1); |

2336 | next && TREE_CODE (next) != FIELD_DECL; |

2337 | next = DECL_CHAIN (next)) |

2338 | ; |

2339 | |

2340 | if (next != field2) |

2341 | { |

2342 | for (next = DECL_CHAIN (field2); |

2343 | next && TREE_CODE (next) != FIELD_DECL; |

2344 | next = DECL_CHAIN (next)) |

2345 | ; |

2346 | |

2347 | if (next != field1) |

2348 | continue; |

2349 | |

2350 | std::swap (field1, field2); |

2351 | std::swap (def1, def2); |

2352 | } |

2353 | |

2354 | bb_for_def1 = gimple_bb (def1); |

2355 | bb_for_def2 = gimple_bb (def2); |

2356 | |

2357 | /* Check for proper alignment of the first field. */ |

2358 | tree_offset1 = bit_position (field1); |

2359 | tree_offset2 = bit_position (field2); |

2360 | tree_size2 = DECL_SIZE (field2); |

2361 | |

2362 | if (!tree_fits_uhwi_p (tree_offset1) |

2363 | || !tree_fits_uhwi_p (tree_offset2) |

2364 | || !tree_fits_uhwi_p (tree_size2)) |

2365 | continue; |

2366 | |

2367 | offset1 = tree_to_uhwi (tree_offset1); |

2368 | offset2 = tree_to_uhwi (tree_offset2); |

2369 | size2 = tree_to_uhwi (tree_size2); |

2370 | align1 = DECL_ALIGN (field1) % param_align_bits; |

2371 | |

2372 | if (offset1 % BITS_PER_UNIT != 0) |

2373 | continue; |

2374 | |

2375 | /* For profitability, the two field references should fit within |

2376 | a single cache line. */ |

2377 | if (align1 + offset2 - offset1 + size2 > param_align_bits) |

2378 | continue; |

2379 | |

2380 | /* The two expressions cannot be dependent upon vdefs defined |

2381 | in bb1/bb2. */ |

2382 | if (local_mem_dependence (def1, bb_for_def1) |

2383 | || local_mem_dependence (def2, bb_for_def2)) |

2384 | continue; |

2385 | |

2386 | /* The conditions are satisfied; hoist the loads from bb1 and bb2 into |

2387 | bb0. We hoist the first one first so that a cache miss is handled |

2388 | efficiently regardless of hardware cache-fill policy. */ |

2389 | gsi2 = gsi_for_stmt (def1); |

2390 | gsi_move_to_bb_end (&gsi2, bb0); |

2391 | gsi2 = gsi_for_stmt (def2); |

2392 | gsi_move_to_bb_end (&gsi2, bb0); |

2393 | |

2394 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2395 | { |

2396 | fprintf (dump_file, |

2397 | "\nHoisting adjacent loads from %d and %d into %d: \n", |

2398 | bb_for_def1->index, bb_for_def2->index, bb0->index); |

2399 | print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS); |

2400 | print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS); |

2401 | } |

2402 | } |

2403 | } |

2404 | |

2405 | /* Determine whether we should attempt to hoist adjacent loads out of |

2406 | diamond patterns in pass_phiopt. Always hoist loads if |

2407 | -fhoist-adjacent-loads is specified and the target machine has |

2408 | both a conditional move instruction and a defined cache line size. */ |

2409 | |

2410 | static bool |

2411 | gate_hoist_loads (void) |

2412 | { |

2413 | return (flag_hoist_adjacent_loads == 1 |

2414 | && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE) |

2415 | && HAVE_conditional_move); |

2416 | } |

2417 | |

2418 | /* This pass tries to replaces an if-then-else block with an |

2419 | assignment. We have four kinds of transformations. Some of these |

2420 | transformations are also performed by the ifcvt RTL optimizer. |

2421 | |

2422 | Conditional Replacement |

2423 | ----------------------- |

2424 | |

2425 | This transformation, implemented in conditional_replacement, |

2426 | replaces |

2427 | |

2428 | bb0: |

2429 | if (cond) goto bb2; else goto bb1; |

2430 | bb1: |

2431 | bb2: |

2432 | x = PHI <0 (bb1), 1 (bb0), ...>; |

2433 | |

2434 | with |

2435 | |

2436 | bb0: |

2437 | x' = cond; |

2438 | goto bb2; |

2439 | bb2: |

2440 | x = PHI <x' (bb0), ...>; |

2441 | |

2442 | We remove bb1 as it becomes unreachable. This occurs often due to |

2443 | gimplification of conditionals. |

2444 | |

2445 | Value Replacement |

2446 | ----------------- |

2447 | |

2448 | This transformation, implemented in value_replacement, replaces |

2449 | |

2450 | bb0: |

2451 | if (a != b) goto bb2; else goto bb1; |

2452 | bb1: |

2453 | bb2: |

2454 | x = PHI <a (bb1), b (bb0), ...>; |

2455 | |

2456 | with |

2457 | |

2458 | bb0: |

2459 | bb2: |

2460 | x = PHI <b (bb0), ...>; |

2461 | |

2462 | This opportunity can sometimes occur as a result of other |

2463 | optimizations. |

2464 | |

2465 | |

2466 | Another case caught by value replacement looks like this: |

2467 | |

2468 | bb0: |

2469 | t1 = a == CONST; |

2470 | t2 = b > c; |

2471 | t3 = t1 & t2; |

2472 | if (t3 != 0) goto bb1; else goto bb2; |

2473 | bb1: |

2474 | bb2: |

2475 | x = PHI (CONST, a) |

2476 | |

2477 | Gets replaced with: |

2478 | bb0: |

2479 | bb2: |

2480 | t1 = a == CONST; |

2481 | t2 = b > c; |

2482 | t3 = t1 & t2; |

2483 | x = a; |

2484 | |

2485 | ABS Replacement |

2486 | --------------- |

2487 | |

2488 | This transformation, implemented in abs_replacement, replaces |

2489 | |

2490 | bb0: |

2491 | if (a >= 0) goto bb2; else goto bb1; |

2492 | bb1: |

2493 | x = -a; |

2494 | bb2: |

2495 | x = PHI <x (bb1), a (bb0), ...>; |

2496 | |

2497 | with |

2498 | |

2499 | bb0: |

2500 | x' = ABS_EXPR< a >; |

2501 | bb2: |

2502 | x = PHI <x' (bb0), ...>; |

2503 | |

2504 | MIN/MAX Replacement |

2505 | ------------------- |

2506 | |

2507 | This transformation, minmax_replacement replaces |

2508 | |

2509 | bb0: |

2510 | if (a <= b) goto bb2; else goto bb1; |

2511 | bb1: |

2512 | bb2: |

2513 | x = PHI <b (bb1), a (bb0), ...>; |

2514 | |

2515 | with |

2516 | |

2517 | bb0: |

2518 | x' = MIN_EXPR (a, b) |

2519 | bb2: |

2520 | x = PHI <x' (bb0), ...>; |

2521 | |

2522 | A similar transformation is done for MAX_EXPR. |

2523 | |

2524 | |

2525 | This pass also performs a fifth transformation of a slightly different |

2526 | flavor. |

2527 | |

2528 | Factor conversion in COND_EXPR |

2529 | ------------------------------ |

2530 | |

2531 | This transformation factors the conversion out of COND_EXPR with |

2532 | factor_out_conditional_conversion. |

2533 | |

2534 | For example: |

2535 | if (a <= CST) goto <bb 3>; else goto <bb 4>; |

2536 | <bb 3>: |

2537 | tmp = (int) a; |

2538 | <bb 4>: |

2539 | tmp = PHI <tmp, CST> |

2540 | |

2541 | Into: |

2542 | if (a <= CST) goto <bb 3>; else goto <bb 4>; |

2543 | <bb 3>: |

2544 | <bb 4>: |

2545 | a = PHI <a, CST> |

2546 | tmp = (int) a; |

2547 | |

2548 | Adjacent Load Hoisting |

2549 | ---------------------- |

2550 | |

2551 | This transformation replaces |

2552 | |

2553 | bb0: |

2554 | if (...) goto bb2; else goto bb1; |

2555 | bb1: |

2556 | x1 = (<expr>).field1; |

2557 | goto bb3; |

2558 | bb2: |

2559 | x2 = (<expr>).field2; |

2560 | bb3: |

2561 | # x = PHI <x1, x2>; |

2562 | |

2563 | with |

2564 | |

2565 | bb0: |

2566 | x1 = (<expr>).field1; |

2567 | x2 = (<expr>).field2; |

2568 | if (...) goto bb2; else goto bb1; |

2569 | bb1: |

2570 | goto bb3; |

2571 | bb2: |

2572 | bb3: |

2573 | # x = PHI <x1, x2>; |

2574 | |

2575 | The purpose of this transformation is to enable generation of conditional |

2576 | move instructions such as Intel CMOVE or PowerPC ISEL. Because one of |

2577 | the loads is speculative, the transformation is restricted to very |

2578 | specific cases to avoid introducing a page fault. We are looking for |

2579 | the common idiom: |

2580 | |

2581 | if (...) |

2582 | x = y->left; |

2583 | else |

2584 | x = y->right; |

2585 | |

2586 | where left and right are typically adjacent pointers in a tree structure. */ |

2587 | |

2588 | namespace { |

2589 | |

2590 | const pass_data pass_data_phiopt = |

2591 | { |

2592 | GIMPLE_PASS, /* type */ |

2593 | "phiopt", /* name */ |

2594 | OPTGROUP_NONE, /* optinfo_flags */ |

2595 | TV_TREE_PHIOPT, /* tv_id */ |

2596 | ( PROP_cfg | PROP_ssa ), /* properties_required */ |

2597 | 0, /* properties_provided */ |

2598 | 0, /* properties_destroyed */ |

2599 | 0, /* todo_flags_start */ |

2600 | 0, /* todo_flags_finish */ |

2601 | }; |

2602 | |

2603 | class pass_phiopt : public gimple_opt_pass |

2604 | { |

2605 | public: |

2606 | pass_phiopt (gcc::context *ctxt) |

2607 | : gimple_opt_pass (pass_data_phiopt, ctxt) |

2608 | {} |

2609 | |

2610 | /* opt_pass methods: */ |

2611 | opt_pass * clone () { return new pass_phiopt (m_ctxt); } |

2612 | virtual bool gate (function *) { return flag_ssa_phiopt; } |

2613 | virtual unsigned int execute (function *) |

2614 | { |

2615 | return tree_ssa_phiopt_worker (false, gate_hoist_loads ()); |

2616 | } |

2617 | |

2618 | }; // class pass_phiopt |

2619 | |

2620 | } // anon namespace |

2621 | |

2622 | gimple_opt_pass * |

2623 | make_pass_phiopt (gcc::context *ctxt) |

2624 | { |

2625 | return new pass_phiopt (ctxt); |

2626 | } |

2627 | |

2628 | namespace { |

2629 | |

2630 | const pass_data pass_data_cselim = |

2631 | { |

2632 | GIMPLE_PASS, /* type */ |

2633 | "cselim", /* name */ |

2634 | OPTGROUP_NONE, /* optinfo_flags */ |

2635 | TV_TREE_PHIOPT, /* tv_id */ |

2636 | ( PROP_cfg | PROP_ssa ), /* properties_required */ |

2637 | 0, /* properties_provided */ |

2638 | 0, /* properties_destroyed */ |

2639 | 0, /* todo_flags_start */ |

2640 | 0, /* todo_flags_finish */ |

2641 | }; |

2642 | |

2643 | class pass_cselim : public gimple_opt_pass |

2644 | { |

2645 | public: |

2646 | pass_cselim (gcc::context *ctxt) |

2647 | : gimple_opt_pass (pass_data_cselim, ctxt) |

2648 | {} |

2649 | |

2650 | /* opt_pass methods: */ |

2651 | virtual bool gate (function *) { return flag_tree_cselim; } |

2652 | virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); } |

2653 | |

2654 | }; // class pass_cselim |

2655 | |

2656 | } // anon namespace |

2657 | |

2658 | gimple_opt_pass * |

2659 | make_pass_cselim (gcc::context *ctxt) |

2660 | { |

2661 | return new pass_cselim (ctxt); |

2662 | } |

2663 |