1 | /* Straight-line strength reduction. |
---|---|

2 | Copyright (C) 2012-2017 Free Software Foundation, Inc. |

3 | Contributed by Bill Schmidt, IBM <wschmidt@linux.ibm.com> |

4 | |

5 | This file is part of GCC. |

6 | |

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

8 | the terms of the GNU General Public License as published by the Free |

9 | Software Foundation; either version 3, or (at your option) any later |

10 | version. |

11 | |

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

13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |

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

15 | for more details. |

16 | |

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

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

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

20 | |

21 | /* There are many algorithms for performing strength reduction on |

22 | loops. This is not one of them. IVOPTS handles strength reduction |

23 | of induction variables just fine. This pass is intended to pick |

24 | up the crumbs it leaves behind, by considering opportunities for |

25 | strength reduction along dominator paths. |

26 | |

27 | Strength reduction addresses explicit multiplies, and certain |

28 | multiplies implicit in addressing expressions. It would also be |

29 | possible to apply strength reduction to divisions and modulos, |

30 | but such opportunities are relatively uncommon. |

31 | |

32 | Strength reduction is also currently restricted to integer operations. |

33 | If desired, it could be extended to floating-point operations under |

34 | control of something like -funsafe-math-optimizations. */ |

35 | |

36 | #include "config.h" |

37 | #include "system.h" |

38 | #include "coretypes.h" |

39 | #include "backend.h" |

40 | #include "rtl.h" |

41 | #include "tree.h" |

42 | #include "gimple.h" |

43 | #include "cfghooks.h" |

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

45 | #include "ssa.h" |

46 | #include "expmed.h" |

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

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

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

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

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

52 | #include "cfgloop.h" |

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

54 | #include "domwalk.h" |

55 | #include "params.h" |

56 | #include "tree-ssa-address.h" |

57 | #include "tree-affine.h" |

58 | #include "builtins.h" |

59 | |

60 | /* Information about a strength reduction candidate. Each statement |

61 | in the candidate table represents an expression of one of the |

62 | following forms (the special case of CAND_REF will be described |

63 | later): |

64 | |

65 | (CAND_MULT) S1: X = (B + i) * S |

66 | (CAND_ADD) S1: X = B + (i * S) |

67 | |

68 | Here X and B are SSA names, i is an integer constant, and S is |

69 | either an SSA name or a constant. We call B the "base," i the |

70 | "index", and S the "stride." |

71 | |

72 | Any statement S0 that dominates S1 and is of the form: |

73 | |

74 | (CAND_MULT) S0: Y = (B + i') * S |

75 | (CAND_ADD) S0: Y = B + (i' * S) |

76 | |

77 | is called a "basis" for S1. In both cases, S1 may be replaced by |

78 | |

79 | S1': X = Y + (i - i') * S, |

80 | |

81 | where (i - i') * S is folded to the extent possible. |

82 | |

83 | All gimple statements are visited in dominator order, and each |

84 | statement that may contribute to one of the forms of S1 above is |

85 | given at least one entry in the candidate table. Such statements |

86 | include addition, pointer addition, subtraction, multiplication, |

87 | negation, copies, and nontrivial type casts. If a statement may |

88 | represent more than one expression of the forms of S1 above, |

89 | multiple "interpretations" are stored in the table and chained |

90 | together. Examples: |

91 | |

92 | * An add of two SSA names may treat either operand as the base. |

93 | * A multiply of two SSA names, likewise. |

94 | * A copy or cast may be thought of as either a CAND_MULT with |

95 | i = 0 and S = 1, or as a CAND_ADD with i = 0 or S = 0. |

96 | |

97 | Candidate records are allocated from an obstack. They are addressed |

98 | both from a hash table keyed on S1, and from a vector of candidate |

99 | pointers arranged in predominator order. |

100 | |

101 | Opportunity note |

102 | ---------------- |

103 | Currently we don't recognize: |

104 | |

105 | S0: Y = (S * i') - B |

106 | S1: X = (S * i) - B |

107 | |

108 | as a strength reduction opportunity, even though this S1 would |

109 | also be replaceable by the S1' above. This can be added if it |

110 | comes up in practice. |

111 | |

112 | Strength reduction in addressing |

113 | -------------------------------- |

114 | There is another kind of candidate known as CAND_REF. A CAND_REF |

115 | describes a statement containing a memory reference having |

116 | complex addressing that might benefit from strength reduction. |

117 | Specifically, we are interested in references for which |

118 | get_inner_reference returns a base address, offset, and bitpos as |

119 | follows: |

120 | |

121 | base: MEM_REF (T1, C1) |

122 | offset: MULT_EXPR (PLUS_EXPR (T2, C2), C3) |

123 | bitpos: C4 * BITS_PER_UNIT |

124 | |

125 | Here T1 and T2 are arbitrary trees, and C1, C2, C3, C4 are |

126 | arbitrary integer constants. Note that C2 may be zero, in which |

127 | case the offset will be MULT_EXPR (T2, C3). |

128 | |

129 | When this pattern is recognized, the original memory reference |

130 | can be replaced with: |

131 | |

132 | MEM_REF (POINTER_PLUS_EXPR (T1, MULT_EXPR (T2, C3)), |

133 | C1 + (C2 * C3) + C4) |

134 | |

135 | which distributes the multiply to allow constant folding. When |

136 | two or more addressing expressions can be represented by MEM_REFs |

137 | of this form, differing only in the constants C1, C2, and C4, |

138 | making this substitution produces more efficient addressing during |

139 | the RTL phases. When there are not at least two expressions with |

140 | the same values of T1, T2, and C3, there is nothing to be gained |

141 | by the replacement. |

142 | |

143 | Strength reduction of CAND_REFs uses the same infrastructure as |

144 | that used by CAND_MULTs and CAND_ADDs. We record T1 in the base (B) |

145 | field, MULT_EXPR (T2, C3) in the stride (S) field, and |

146 | C1 + (C2 * C3) + C4 in the index (i) field. A basis for a CAND_REF |

147 | is thus another CAND_REF with the same B and S values. When at |

148 | least two CAND_REFs are chained together using the basis relation, |

149 | each of them is replaced as above, resulting in improved code |

150 | generation for addressing. |

151 | |

152 | Conditional candidates |

153 | ====================== |

154 | |

155 | Conditional candidates are best illustrated with an example. |

156 | Consider the code sequence: |

157 | |

158 | (1) x_0 = ...; |

159 | (2) a_0 = x_0 * 5; MULT (B: x_0; i: 0; S: 5) |

160 | if (...) |

161 | (3) x_1 = x_0 + 1; ADD (B: x_0, i: 1; S: 1) |

162 | (4) x_2 = PHI <x_0, x_1>; PHI (B: x_0, i: 0, S: 1) |

163 | (5) x_3 = x_2 + 1; ADD (B: x_2, i: 1, S: 1) |

164 | (6) a_1 = x_3 * 5; MULT (B: x_2, i: 1; S: 5) |

165 | |

166 | Here strength reduction is complicated by the uncertain value of x_2. |

167 | A legitimate transformation is: |

168 | |

169 | (1) x_0 = ...; |

170 | (2) a_0 = x_0 * 5; |

171 | if (...) |

172 | { |

173 | (3) [x_1 = x_0 + 1;] |

174 | (3a) t_1 = a_0 + 5; |

175 | } |

176 | (4) [x_2 = PHI <x_0, x_1>;] |

177 | (4a) t_2 = PHI <a_0, t_1>; |

178 | (5) [x_3 = x_2 + 1;] |

179 | (6r) a_1 = t_2 + 5; |

180 | |

181 | where the bracketed instructions may go dead. |

182 | |

183 | To recognize this opportunity, we have to observe that statement (6) |

184 | has a "hidden basis" (2). The hidden basis is unlike a normal basis |

185 | in that the statement and the hidden basis have different base SSA |

186 | names (x_2 and x_0, respectively). The relationship is established |

187 | when a statement's base name (x_2) is defined by a phi statement (4), |

188 | each argument of which (x_0, x_1) has an identical "derived base name." |

189 | If the argument is defined by a candidate (as x_1 is by (3)) that is a |

190 | CAND_ADD having a stride of 1, the derived base name of the argument is |

191 | the base name of the candidate (x_0). Otherwise, the argument itself |

192 | is its derived base name (as is the case with argument x_0). |

193 | |

194 | The hidden basis for statement (6) is the nearest dominating candidate |

195 | whose base name is the derived base name (x_0) of the feeding phi (4), |

196 | and whose stride is identical to that of the statement. We can then |

197 | create the new "phi basis" (4a) and feeding adds along incoming arcs (3a), |

198 | allowing the final replacement of (6) by the strength-reduced (6r). |

199 | |

200 | To facilitate this, a new kind of candidate (CAND_PHI) is introduced. |

201 | A CAND_PHI is not a candidate for replacement, but is maintained in the |

202 | candidate table to ease discovery of hidden bases. Any phi statement |

203 | whose arguments share a common derived base name is entered into the |

204 | table with the derived base name, an (arbitrary) index of zero, and a |

205 | stride of 1. A statement with a hidden basis can then be detected by |

206 | simply looking up its feeding phi definition in the candidate table, |

207 | extracting the derived base name, and searching for a basis in the |

208 | usual manner after substituting the derived base name. |

209 | |

210 | Note that the transformation is only valid when the original phi and |

211 | the statements that define the phi's arguments are all at the same |

212 | position in the loop hierarchy. */ |

213 | |

214 | |

215 | /* Index into the candidate vector, offset by 1. VECs are zero-based, |

216 | while cand_idx's are one-based, with zero indicating null. */ |

217 | typedef unsigned cand_idx; |

218 | |

219 | /* The kind of candidate. */ |

220 | enum cand_kind |

221 | { |

222 | CAND_MULT, |

223 | CAND_ADD, |

224 | CAND_REF, |

225 | CAND_PHI |

226 | }; |

227 | |

228 | struct slsr_cand_d |

229 | { |

230 | /* The candidate statement S1. */ |

231 | gimple *cand_stmt; |

232 | |

233 | /* The base expression B: often an SSA name, but not always. */ |

234 | tree base_expr; |

235 | |

236 | /* The stride S. */ |

237 | tree stride; |

238 | |

239 | /* The index constant i. */ |

240 | widest_int index; |

241 | |

242 | /* The type of the candidate. This is normally the type of base_expr, |

243 | but casts may have occurred when combining feeding instructions. |

244 | A candidate can only be a basis for candidates of the same final type. |

245 | (For CAND_REFs, this is the type to be used for operand 1 of the |

246 | replacement MEM_REF.) */ |

247 | tree cand_type; |

248 | |

249 | /* The type to be used to interpret the stride field when the stride |

250 | is not a constant. Normally the same as the type of the recorded |

251 | stride, but when the stride has been cast we need to maintain that |

252 | knowledge in order to make legal substitutions without losing |

253 | precision. When the stride is a constant, this will be sizetype. */ |

254 | tree stride_type; |

255 | |

256 | /* The kind of candidate (CAND_MULT, etc.). */ |

257 | enum cand_kind kind; |

258 | |

259 | /* Index of this candidate in the candidate vector. */ |

260 | cand_idx cand_num; |

261 | |

262 | /* Index of the next candidate record for the same statement. |

263 | A statement may be useful in more than one way (e.g., due to |

264 | commutativity). So we can have multiple "interpretations" |

265 | of a statement. */ |

266 | cand_idx next_interp; |

267 | |

268 | /* Index of the basis statement S0, if any, in the candidate vector. */ |

269 | cand_idx basis; |

270 | |

271 | /* First candidate for which this candidate is a basis, if one exists. */ |

272 | cand_idx dependent; |

273 | |

274 | /* Next candidate having the same basis as this one. */ |

275 | cand_idx sibling; |

276 | |

277 | /* If this is a conditional candidate, the CAND_PHI candidate |

278 | that defines the base SSA name B. */ |

279 | cand_idx def_phi; |

280 | |

281 | /* Savings that can be expected from eliminating dead code if this |

282 | candidate is replaced. */ |

283 | int dead_savings; |

284 | |

285 | /* For PHI candidates, use a visited flag to keep from processing the |

286 | same PHI twice from multiple paths. */ |

287 | int visited; |

288 | |

289 | /* We sometimes have to cache a phi basis with a phi candidate to |

290 | avoid processing it twice. Valid only if visited==1. */ |

291 | tree cached_basis; |

292 | }; |

293 | |

294 | typedef struct slsr_cand_d slsr_cand, *slsr_cand_t; |

295 | typedef const struct slsr_cand_d *const_slsr_cand_t; |

296 | |

297 | /* Pointers to candidates are chained together as part of a mapping |

298 | from base expressions to the candidates that use them. */ |

299 | |

300 | struct cand_chain_d |

301 | { |

302 | /* Base expression for the chain of candidates: often, but not |

303 | always, an SSA name. */ |

304 | tree base_expr; |

305 | |

306 | /* Pointer to a candidate. */ |

307 | slsr_cand_t cand; |

308 | |

309 | /* Chain pointer. */ |

310 | struct cand_chain_d *next; |

311 | |

312 | }; |

313 | |

314 | typedef struct cand_chain_d cand_chain, *cand_chain_t; |

315 | typedef const struct cand_chain_d *const_cand_chain_t; |

316 | |

317 | /* Information about a unique "increment" associated with candidates |

318 | having an SSA name for a stride. An increment is the difference |

319 | between the index of the candidate and the index of its basis, |

320 | i.e., (i - i') as discussed in the module commentary. |

321 | |

322 | When we are not going to generate address arithmetic we treat |

323 | increments that differ only in sign as the same, allowing sharing |

324 | of the cost of initializers. The absolute value of the increment |

325 | is stored in the incr_info. */ |

326 | |

327 | struct incr_info_d |

328 | { |

329 | /* The increment that relates a candidate to its basis. */ |

330 | widest_int incr; |

331 | |

332 | /* How many times the increment occurs in the candidate tree. */ |

333 | unsigned count; |

334 | |

335 | /* Cost of replacing candidates using this increment. Negative and |

336 | zero costs indicate replacement should be performed. */ |

337 | int cost; |

338 | |

339 | /* If this increment is profitable but is not -1, 0, or 1, it requires |

340 | an initializer T_0 = stride * incr to be found or introduced in the |

341 | nearest common dominator of all candidates. This field holds T_0 |

342 | for subsequent use. */ |

343 | tree initializer; |

344 | |

345 | /* If the initializer was found to already exist, this is the block |

346 | where it was found. */ |

347 | basic_block init_bb; |

348 | }; |

349 | |

350 | typedef struct incr_info_d incr_info, *incr_info_t; |

351 | |

352 | /* Candidates are maintained in a vector. If candidate X dominates |

353 | candidate Y, then X appears before Y in the vector; but the |

354 | converse does not necessarily hold. */ |

355 | static vec<slsr_cand_t> cand_vec; |

356 | |

357 | enum cost_consts |

358 | { |

359 | COST_NEUTRAL = 0, |

360 | COST_INFINITE = 1000 |

361 | }; |

362 | |

363 | enum stride_status |

364 | { |

365 | UNKNOWN_STRIDE = 0, |

366 | KNOWN_STRIDE = 1 |

367 | }; |

368 | |

369 | enum phi_adjust_status |

370 | { |

371 | NOT_PHI_ADJUST = 0, |

372 | PHI_ADJUST = 1 |

373 | }; |

374 | |

375 | enum count_phis_status |

376 | { |

377 | DONT_COUNT_PHIS = 0, |

378 | COUNT_PHIS = 1 |

379 | }; |

380 | |

381 | /* Constrain how many PHI nodes we will visit for a conditional |

382 | candidate (depth and breadth). */ |

383 | const int MAX_SPREAD = 16; |

384 | |

385 | /* Pointer map embodying a mapping from statements to candidates. */ |

386 | static hash_map<gimple *, slsr_cand_t> *stmt_cand_map; |

387 | |

388 | /* Obstack for candidates. */ |

389 | static struct obstack cand_obstack; |

390 | |

391 | /* Obstack for candidate chains. */ |

392 | static struct obstack chain_obstack; |

393 | |

394 | /* An array INCR_VEC of incr_infos is used during analysis of related |

395 | candidates having an SSA name for a stride. INCR_VEC_LEN describes |

396 | its current length. MAX_INCR_VEC_LEN is used to avoid costly |

397 | pathological cases. */ |

398 | static incr_info_t incr_vec; |

399 | static unsigned incr_vec_len; |

400 | const int MAX_INCR_VEC_LEN = 16; |

401 | |

402 | /* For a chain of candidates with unknown stride, indicates whether or not |

403 | we must generate pointer arithmetic when replacing statements. */ |

404 | static bool address_arithmetic_p; |

405 | |

406 | /* Forward function declarations. */ |

407 | static slsr_cand_t base_cand_from_table (tree); |

408 | static tree introduce_cast_before_cand (slsr_cand_t, tree, tree); |

409 | static bool legal_cast_p_1 (tree, tree); |

410 | |

411 | /* Produce a pointer to the IDX'th candidate in the candidate vector. */ |

412 | |

413 | static slsr_cand_t |

414 | lookup_cand (cand_idx idx) |

415 | { |

416 | return cand_vec[idx - 1]; |

417 | } |

418 | |

419 | /* Helper for hashing a candidate chain header. */ |

420 | |

421 | struct cand_chain_hasher : nofree_ptr_hash <cand_chain> |

422 | { |

423 | static inline hashval_t hash (const cand_chain *); |

424 | static inline bool equal (const cand_chain *, const cand_chain *); |

425 | }; |

426 | |

427 | inline hashval_t |

428 | cand_chain_hasher::hash (const cand_chain *p) |

429 | { |

430 | tree base_expr = p->base_expr; |

431 | return iterative_hash_expr (base_expr, 0); |

432 | } |

433 | |

434 | inline bool |

435 | cand_chain_hasher::equal (const cand_chain *chain1, const cand_chain *chain2) |

436 | { |

437 | return operand_equal_p (chain1->base_expr, chain2->base_expr, 0); |

438 | } |

439 | |

440 | /* Hash table embodying a mapping from base exprs to chains of candidates. */ |

441 | static hash_table<cand_chain_hasher> *base_cand_map; |

442 | |

443 | /* Pointer map used by tree_to_aff_combination_expand. */ |

444 | static hash_map<tree, name_expansion *> *name_expansions; |

445 | /* Pointer map embodying a mapping from bases to alternative bases. */ |

446 | static hash_map<tree, tree> *alt_base_map; |

447 | |

448 | /* Given BASE, use the tree affine combiniation facilities to |

449 | find the underlying tree expression for BASE, with any |

450 | immediate offset excluded. |

451 | |

452 | N.B. we should eliminate this backtracking with better forward |

453 | analysis in a future release. */ |

454 | |

455 | static tree |

456 | get_alternative_base (tree base) |

457 | { |

458 | tree *result = alt_base_map->get (base); |

459 | |

460 | if (result == NULL) |

461 | { |

462 | tree expr; |

463 | aff_tree aff; |

464 | |

465 | tree_to_aff_combination_expand (base, TREE_TYPE (base), |

466 | &aff, &name_expansions); |

467 | aff.offset = 0; |

468 | expr = aff_combination_to_tree (&aff); |

469 | |

470 | gcc_assert (!alt_base_map->put (base, base == expr ? NULL : expr)); |

471 | |

472 | return expr == base ? NULL : expr; |

473 | } |

474 | |

475 | return *result; |

476 | } |

477 | |

478 | /* Look in the candidate table for a CAND_PHI that defines BASE and |

479 | return it if found; otherwise return NULL. */ |

480 | |

481 | static cand_idx |

482 | find_phi_def (tree base) |

483 | { |

484 | slsr_cand_t c; |

485 | |

486 | if (TREE_CODE (base) != SSA_NAME) |

487 | return 0; |

488 | |

489 | c = base_cand_from_table (base); |

490 | |

491 | if (!c || c->kind != CAND_PHI |

492 | || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (c->cand_stmt))) |

493 | return 0; |

494 | |

495 | return c->cand_num; |

496 | } |

497 | |

498 | /* Determine whether all uses of NAME are directly or indirectly |

499 | used by STMT. That is, we want to know whether if STMT goes |

500 | dead, the definition of NAME also goes dead. */ |

501 | static bool |

502 | uses_consumed_by_stmt (tree name, gimple *stmt, unsigned recurse = 0) |

503 | { |

504 | gimple *use_stmt; |

505 | imm_use_iterator iter; |

506 | bool retval = true; |

507 | |

508 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, name) |

509 | { |

510 | if (use_stmt == stmt || is_gimple_debug (use_stmt)) |

511 | continue; |

512 | |

513 | if (!is_gimple_assign (use_stmt) |

514 | || !gimple_get_lhs (use_stmt) |

515 | || !is_gimple_reg (gimple_get_lhs (use_stmt)) |

516 | || recurse >= 10 |

517 | || !uses_consumed_by_stmt (gimple_get_lhs (use_stmt), stmt, |

518 | recurse + 1)) |

519 | { |

520 | retval = false; |

521 | BREAK_FROM_IMM_USE_STMT (iter); |

522 | } |

523 | } |

524 | |

525 | return retval; |

526 | } |

527 | |

528 | /* Helper routine for find_basis_for_candidate. May be called twice: |

529 | once for the candidate's base expr, and optionally again either for |

530 | the candidate's phi definition or for a CAND_REF's alternative base |

531 | expression. */ |

532 | |

533 | static slsr_cand_t |

534 | find_basis_for_base_expr (slsr_cand_t c, tree base_expr) |

535 | { |

536 | cand_chain mapping_key; |

537 | cand_chain_t chain; |

538 | slsr_cand_t basis = NULL; |

539 | |

540 | // Limit potential of N^2 behavior for long candidate chains. |

541 | int iters = 0; |

542 | int max_iters = PARAM_VALUE (PARAM_MAX_SLSR_CANDIDATE_SCAN); |

543 | |

544 | mapping_key.base_expr = base_expr; |

545 | chain = base_cand_map->find (&mapping_key); |

546 | |

547 | for (; chain && iters < max_iters; chain = chain->next, ++iters) |

548 | { |

549 | slsr_cand_t one_basis = chain->cand; |

550 | |

551 | if (one_basis->kind != c->kind |

552 | || one_basis->cand_stmt == c->cand_stmt |

553 | || !operand_equal_p (one_basis->stride, c->stride, 0) |

554 | || !types_compatible_p (one_basis->cand_type, c->cand_type) |

555 | || !types_compatible_p (one_basis->stride_type, c->stride_type) |

556 | || !dominated_by_p (CDI_DOMINATORS, |

557 | gimple_bb (c->cand_stmt), |

558 | gimple_bb (one_basis->cand_stmt))) |

559 | continue; |

560 | |

561 | tree lhs = gimple_assign_lhs (one_basis->cand_stmt); |

562 | if (lhs && TREE_CODE (lhs) == SSA_NAME |

563 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) |

564 | continue; |

565 | |

566 | if (!basis || basis->cand_num < one_basis->cand_num) |

567 | basis = one_basis; |

568 | } |

569 | |

570 | return basis; |

571 | } |

572 | |

573 | /* Use the base expr from candidate C to look for possible candidates |

574 | that can serve as a basis for C. Each potential basis must also |

575 | appear in a block that dominates the candidate statement and have |

576 | the same stride and type. If more than one possible basis exists, |

577 | the one with highest index in the vector is chosen; this will be |

578 | the most immediately dominating basis. */ |

579 | |

580 | static int |

581 | find_basis_for_candidate (slsr_cand_t c) |

582 | { |

583 | slsr_cand_t basis = find_basis_for_base_expr (c, c->base_expr); |

584 | |

585 | /* If a candidate doesn't have a basis using its base expression, |

586 | it may have a basis hidden by one or more intervening phis. */ |

587 | if (!basis && c->def_phi) |

588 | { |

589 | basic_block basis_bb, phi_bb; |

590 | slsr_cand_t phi_cand = lookup_cand (c->def_phi); |

591 | basis = find_basis_for_base_expr (c, phi_cand->base_expr); |

592 | |

593 | if (basis) |

594 | { |

595 | /* A hidden basis must dominate the phi-definition of the |

596 | candidate's base name. */ |

597 | phi_bb = gimple_bb (phi_cand->cand_stmt); |

598 | basis_bb = gimple_bb (basis->cand_stmt); |

599 | |

600 | if (phi_bb == basis_bb |

601 | || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) |

602 | { |

603 | basis = NULL; |

604 | c->basis = 0; |

605 | } |

606 | |

607 | /* If we found a hidden basis, estimate additional dead-code |

608 | savings if the phi and its feeding statements can be removed. */ |

609 | tree feeding_var = gimple_phi_result (phi_cand->cand_stmt); |

610 | if (basis && uses_consumed_by_stmt (feeding_var, c->cand_stmt)) |

611 | c->dead_savings += phi_cand->dead_savings; |

612 | } |

613 | } |

614 | |

615 | if (flag_expensive_optimizations && !basis && c->kind == CAND_REF) |

616 | { |

617 | tree alt_base_expr = get_alternative_base (c->base_expr); |

618 | if (alt_base_expr) |

619 | basis = find_basis_for_base_expr (c, alt_base_expr); |

620 | } |

621 | |

622 | if (basis) |

623 | { |

624 | c->sibling = basis->dependent; |

625 | basis->dependent = c->cand_num; |

626 | return basis->cand_num; |

627 | } |

628 | |

629 | return 0; |

630 | } |

631 | |

632 | /* Record a mapping from BASE to C, indicating that C may potentially serve |

633 | as a basis using that base expression. BASE may be the same as |

634 | C->BASE_EXPR; alternatively BASE can be a different tree that share the |

635 | underlining expression of C->BASE_EXPR. */ |

636 | |

637 | static void |

638 | record_potential_basis (slsr_cand_t c, tree base) |

639 | { |

640 | cand_chain_t node; |

641 | cand_chain **slot; |

642 | |

643 | gcc_assert (base); |

644 | |

645 | node = (cand_chain_t) obstack_alloc (&chain_obstack, sizeof (cand_chain)); |

646 | node->base_expr = base; |

647 | node->cand = c; |

648 | node->next = NULL; |

649 | slot = base_cand_map->find_slot (node, INSERT); |

650 | |

651 | if (*slot) |

652 | { |

653 | cand_chain_t head = (cand_chain_t) (*slot); |

654 | node->next = head->next; |

655 | head->next = node; |

656 | } |

657 | else |

658 | *slot = node; |

659 | } |

660 | |

661 | /* Allocate storage for a new candidate and initialize its fields. |

662 | Attempt to find a basis for the candidate. |

663 | |

664 | For CAND_REF, an alternative base may also be recorded and used |

665 | to find a basis. This helps cases where the expression hidden |

666 | behind BASE (which is usually an SSA_NAME) has immediate offset, |

667 | e.g. |

668 | |

669 | a2[i][j] = 1; |

670 | a2[i + 20][j] = 2; */ |

671 | |

672 | static slsr_cand_t |

673 | alloc_cand_and_find_basis (enum cand_kind kind, gimple *gs, tree base, |

674 | const widest_int &index, tree stride, tree ctype, |

675 | tree stype, unsigned savings) |

676 | { |

677 | slsr_cand_t c = (slsr_cand_t) obstack_alloc (&cand_obstack, |

678 | sizeof (slsr_cand)); |

679 | c->cand_stmt = gs; |

680 | c->base_expr = base; |

681 | c->stride = stride; |

682 | c->index = index; |

683 | c->cand_type = ctype; |

684 | c->stride_type = stype; |

685 | c->kind = kind; |

686 | c->cand_num = cand_vec.length () + 1; |

687 | c->next_interp = 0; |

688 | c->dependent = 0; |

689 | c->sibling = 0; |

690 | c->def_phi = kind == CAND_MULT ? find_phi_def (base) : 0; |

691 | c->dead_savings = savings; |

692 | c->visited = 0; |

693 | c->cached_basis = NULL_TREE; |

694 | |

695 | cand_vec.safe_push (c); |

696 | |

697 | if (kind == CAND_PHI) |

698 | c->basis = 0; |

699 | else |

700 | c->basis = find_basis_for_candidate (c); |

701 | |

702 | record_potential_basis (c, base); |

703 | if (flag_expensive_optimizations && kind == CAND_REF) |

704 | { |

705 | tree alt_base = get_alternative_base (base); |

706 | if (alt_base) |

707 | record_potential_basis (c, alt_base); |

708 | } |

709 | |

710 | return c; |

711 | } |

712 | |

713 | /* Determine the target cost of statement GS when compiling according |

714 | to SPEED. */ |

715 | |

716 | static int |

717 | stmt_cost (gimple *gs, bool speed) |

718 | { |

719 | tree lhs, rhs1, rhs2; |

720 | machine_mode lhs_mode; |

721 | |

722 | gcc_assert (is_gimple_assign (gs)); |

723 | lhs = gimple_assign_lhs (gs); |

724 | rhs1 = gimple_assign_rhs1 (gs); |

725 | lhs_mode = TYPE_MODE (TREE_TYPE (lhs)); |

726 | |

727 | switch (gimple_assign_rhs_code (gs)) |

728 | { |

729 | case MULT_EXPR: |

730 | rhs2 = gimple_assign_rhs2 (gs); |

731 | |

732 | if (tree_fits_shwi_p (rhs2)) |

733 | return mult_by_coeff_cost (tree_to_shwi (rhs2), lhs_mode, speed); |

734 | |

735 | gcc_assert (TREE_CODE (rhs1) != INTEGER_CST); |

736 | return mul_cost (speed, lhs_mode); |

737 | |

738 | case PLUS_EXPR: |

739 | case POINTER_PLUS_EXPR: |

740 | case MINUS_EXPR: |

741 | return add_cost (speed, lhs_mode); |

742 | |

743 | case NEGATE_EXPR: |

744 | return neg_cost (speed, lhs_mode); |

745 | |

746 | CASE_CONVERT: |

747 | return convert_cost (lhs_mode, TYPE_MODE (TREE_TYPE (rhs1)), speed); |

748 | |

749 | /* Note that we don't assign costs to copies that in most cases |

750 | will go away. */ |

751 | case SSA_NAME: |

752 | return 0; |

753 | |

754 | default: |

755 | ; |

756 | } |

757 | |

758 | gcc_unreachable (); |

759 | return 0; |

760 | } |

761 | |

762 | /* Look up the defining statement for BASE_IN and return a pointer |

763 | to its candidate in the candidate table, if any; otherwise NULL. |

764 | Only CAND_ADD and CAND_MULT candidates are returned. */ |

765 | |

766 | static slsr_cand_t |

767 | base_cand_from_table (tree base_in) |

768 | { |

769 | slsr_cand_t *result; |

770 | |

771 | gimple *def = SSA_NAME_DEF_STMT (base_in); |

772 | if (!def) |

773 | return (slsr_cand_t) NULL; |

774 | |

775 | result = stmt_cand_map->get (def); |

776 | |

777 | if (result && (*result)->kind != CAND_REF) |

778 | return *result; |

779 | |

780 | return (slsr_cand_t) NULL; |

781 | } |

782 | |

783 | /* Add an entry to the statement-to-candidate mapping. */ |

784 | |

785 | static void |

786 | add_cand_for_stmt (gimple *gs, slsr_cand_t c) |

787 | { |

788 | gcc_assert (!stmt_cand_map->put (gs, c)); |

789 | } |

790 | |

791 | /* Given PHI which contains a phi statement, determine whether it |

792 | satisfies all the requirements of a phi candidate. If so, create |

793 | a candidate. Note that a CAND_PHI never has a basis itself, but |

794 | is used to help find a basis for subsequent candidates. */ |

795 | |

796 | static void |

797 | slsr_process_phi (gphi *phi, bool speed) |

798 | { |

799 | unsigned i; |

800 | tree arg0_base = NULL_TREE, base_type; |

801 | slsr_cand_t c; |

802 | struct loop *cand_loop = gimple_bb (phi)->loop_father; |

803 | unsigned savings = 0; |

804 | |

805 | /* A CAND_PHI requires each of its arguments to have the same |

806 | derived base name. (See the module header commentary for a |

807 | definition of derived base names.) Furthermore, all feeding |

808 | definitions must be in the same position in the loop hierarchy |

809 | as PHI. */ |

810 | |

811 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

812 | { |

813 | slsr_cand_t arg_cand; |

814 | tree arg = gimple_phi_arg_def (phi, i); |

815 | tree derived_base_name = NULL_TREE; |

816 | gimple *arg_stmt = NULL; |

817 | basic_block arg_bb = NULL; |

818 | |

819 | if (TREE_CODE (arg) != SSA_NAME) |

820 | return; |

821 | |

822 | arg_cand = base_cand_from_table (arg); |

823 | |

824 | if (arg_cand) |

825 | { |

826 | while (arg_cand->kind != CAND_ADD && arg_cand->kind != CAND_PHI) |

827 | { |

828 | if (!arg_cand->next_interp) |

829 | return; |

830 | |

831 | arg_cand = lookup_cand (arg_cand->next_interp); |

832 | } |

833 | |

834 | if (!integer_onep (arg_cand->stride)) |

835 | return; |

836 | |

837 | derived_base_name = arg_cand->base_expr; |

838 | arg_stmt = arg_cand->cand_stmt; |

839 | arg_bb = gimple_bb (arg_stmt); |

840 | |

841 | /* Gather potential dead code savings if the phi statement |

842 | can be removed later on. */ |

843 | if (uses_consumed_by_stmt (arg, phi)) |

844 | { |

845 | if (gimple_code (arg_stmt) == GIMPLE_PHI) |

846 | savings += arg_cand->dead_savings; |

847 | else |

848 | savings += stmt_cost (arg_stmt, speed); |

849 | } |

850 | } |

851 | else if (SSA_NAME_IS_DEFAULT_DEF (arg)) |

852 | { |

853 | derived_base_name = arg; |

854 | arg_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |

855 | } |

856 | |

857 | if (!arg_bb || arg_bb->loop_father != cand_loop) |

858 | return; |

859 | |

860 | if (i == 0) |

861 | arg0_base = derived_base_name; |

862 | else if (!operand_equal_p (derived_base_name, arg0_base, 0)) |

863 | return; |

864 | } |

865 | |

866 | /* Create the candidate. "alloc_cand_and_find_basis" is named |

867 | misleadingly for this case, as no basis will be sought for a |

868 | CAND_PHI. */ |

869 | base_type = TREE_TYPE (arg0_base); |

870 | |

871 | c = alloc_cand_and_find_basis (CAND_PHI, phi, arg0_base, |

872 | 0, integer_one_node, base_type, |

873 | sizetype, savings); |

874 | |

875 | /* Add the candidate to the statement-candidate mapping. */ |

876 | add_cand_for_stmt (phi, c); |

877 | } |

878 | |

879 | /* Given PBASE which is a pointer to tree, look up the defining |

880 | statement for it and check whether the candidate is in the |

881 | form of: |

882 | |

883 | X = B + (1 * S), S is integer constant |

884 | X = B + (i * S), S is integer one |

885 | |

886 | If so, set PBASE to the candidate's base_expr and return double |

887 | int (i * S). |

888 | Otherwise, just return double int zero. */ |

889 | |

890 | static widest_int |

891 | backtrace_base_for_ref (tree *pbase) |

892 | { |

893 | tree base_in = *pbase; |

894 | slsr_cand_t base_cand; |

895 | |

896 | STRIP_NOPS (base_in); |

897 | |

898 | /* Strip off widening conversion(s) to handle cases where |

899 | e.g. 'B' is widened from an 'int' in order to calculate |

900 | a 64-bit address. */ |

901 | if (CONVERT_EXPR_P (base_in) |

902 | && legal_cast_p_1 (TREE_TYPE (base_in), |

903 | TREE_TYPE (TREE_OPERAND (base_in, 0)))) |

904 | base_in = get_unwidened (base_in, NULL_TREE); |

905 | |

906 | if (TREE_CODE (base_in) != SSA_NAME) |

907 | return 0; |

908 | |

909 | base_cand = base_cand_from_table (base_in); |

910 | |

911 | while (base_cand && base_cand->kind != CAND_PHI) |

912 | { |

913 | if (base_cand->kind == CAND_ADD |

914 | && base_cand->index == 1 |

915 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |

916 | { |

917 | /* X = B + (1 * S), S is integer constant. */ |

918 | *pbase = base_cand->base_expr; |

919 | return wi::to_widest (base_cand->stride); |

920 | } |

921 | else if (base_cand->kind == CAND_ADD |

922 | && TREE_CODE (base_cand->stride) == INTEGER_CST |

923 | && integer_onep (base_cand->stride)) |

924 | { |

925 | /* X = B + (i * S), S is integer one. */ |

926 | *pbase = base_cand->base_expr; |

927 | return base_cand->index; |

928 | } |

929 | |

930 | if (base_cand->next_interp) |

931 | base_cand = lookup_cand (base_cand->next_interp); |

932 | else |

933 | base_cand = NULL; |

934 | } |

935 | |

936 | return 0; |

937 | } |

938 | |

939 | /* Look for the following pattern: |

940 | |

941 | *PBASE: MEM_REF (T1, C1) |

942 | |

943 | *POFFSET: MULT_EXPR (T2, C3) [C2 is zero] |

944 | or |

945 | MULT_EXPR (PLUS_EXPR (T2, C2), C3) |

946 | or |

947 | MULT_EXPR (MINUS_EXPR (T2, -C2), C3) |

948 | |

949 | *PINDEX: C4 * BITS_PER_UNIT |

950 | |

951 | If not present, leave the input values unchanged and return FALSE. |

952 | Otherwise, modify the input values as follows and return TRUE: |

953 | |

954 | *PBASE: T1 |

955 | *POFFSET: MULT_EXPR (T2, C3) |

956 | *PINDEX: C1 + (C2 * C3) + C4 |

957 | |

958 | When T2 is recorded by a CAND_ADD in the form of (T2' + C5), it |

959 | will be further restructured to: |

960 | |

961 | *PBASE: T1 |

962 | *POFFSET: MULT_EXPR (T2', C3) |

963 | *PINDEX: C1 + (C2 * C3) + C4 + (C5 * C3) */ |

964 | |

965 | static bool |

966 | restructure_reference (tree *pbase, tree *poffset, widest_int *pindex, |

967 | tree *ptype) |

968 | { |

969 | tree base = *pbase, offset = *poffset; |

970 | widest_int index = *pindex; |

971 | tree mult_op0, t1, t2, type; |

972 | widest_int c1, c2, c3, c4, c5; |

973 | |

974 | if (!base |

975 | || !offset |

976 | || TREE_CODE (base) != MEM_REF |

977 | || TREE_CODE (offset) != MULT_EXPR |

978 | || TREE_CODE (TREE_OPERAND (offset, 1)) != INTEGER_CST |

979 | || wi::umod_floor (index, BITS_PER_UNIT) != 0) |

980 | return false; |

981 | |

982 | t1 = TREE_OPERAND (base, 0); |

983 | c1 = widest_int::from (mem_ref_offset (base), SIGNED); |

984 | type = TREE_TYPE (TREE_OPERAND (base, 1)); |

985 | |

986 | mult_op0 = TREE_OPERAND (offset, 0); |

987 | c3 = wi::to_widest (TREE_OPERAND (offset, 1)); |

988 | |

989 | if (TREE_CODE (mult_op0) == PLUS_EXPR) |

990 | |

991 | if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST) |

992 | { |

993 | t2 = TREE_OPERAND (mult_op0, 0); |

994 | c2 = wi::to_widest (TREE_OPERAND (mult_op0, 1)); |

995 | } |

996 | else |

997 | return false; |

998 | |

999 | else if (TREE_CODE (mult_op0) == MINUS_EXPR) |

1000 | |

1001 | if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST) |

1002 | { |

1003 | t2 = TREE_OPERAND (mult_op0, 0); |

1004 | c2 = -wi::to_widest (TREE_OPERAND (mult_op0, 1)); |

1005 | } |

1006 | else |

1007 | return false; |

1008 | |

1009 | else |

1010 | { |

1011 | t2 = mult_op0; |

1012 | c2 = 0; |

1013 | } |

1014 | |

1015 | c4 = index >> LOG2_BITS_PER_UNIT; |

1016 | c5 = backtrace_base_for_ref (&t2); |

1017 | |

1018 | *pbase = t1; |

1019 | *poffset = fold_build2 (MULT_EXPR, sizetype, fold_convert (sizetype, t2), |

1020 | wide_int_to_tree (sizetype, c3)); |

1021 | *pindex = c1 + c2 * c3 + c4 + c5 * c3; |

1022 | *ptype = type; |

1023 | |

1024 | return true; |

1025 | } |

1026 | |

1027 | /* Given GS which contains a data reference, create a CAND_REF entry in |

1028 | the candidate table and attempt to find a basis. */ |

1029 | |

1030 | static void |

1031 | slsr_process_ref (gimple *gs) |

1032 | { |

1033 | tree ref_expr, base, offset, type; |

1034 | HOST_WIDE_INT bitsize, bitpos; |

1035 | machine_mode mode; |

1036 | int unsignedp, reversep, volatilep; |

1037 | slsr_cand_t c; |

1038 | |

1039 | if (gimple_vdef (gs)) |

1040 | ref_expr = gimple_assign_lhs (gs); |

1041 | else |

1042 | ref_expr = gimple_assign_rhs1 (gs); |

1043 | |

1044 | if (!handled_component_p (ref_expr) |

1045 | || TREE_CODE (ref_expr) == BIT_FIELD_REF |

1046 | || (TREE_CODE (ref_expr) == COMPONENT_REF |

1047 | && DECL_BIT_FIELD (TREE_OPERAND (ref_expr, 1)))) |

1048 | return; |

1049 | |

1050 | base = get_inner_reference (ref_expr, &bitsize, &bitpos, &offset, &mode, |

1051 | &unsignedp, &reversep, &volatilep); |

1052 | if (reversep) |

1053 | return; |

1054 | widest_int index = bitpos; |

1055 | |

1056 | if (!restructure_reference (&base, &offset, &index, &type)) |

1057 | return; |

1058 | |

1059 | c = alloc_cand_and_find_basis (CAND_REF, gs, base, index, offset, |

1060 | type, sizetype, 0); |

1061 | |

1062 | /* Add the candidate to the statement-candidate mapping. */ |

1063 | add_cand_for_stmt (gs, c); |

1064 | } |

1065 | |

1066 | /* Create a candidate entry for a statement GS, where GS multiplies |

1067 | two SSA names BASE_IN and STRIDE_IN. Propagate any known information |

1068 | about the two SSA names into the new candidate. Return the new |

1069 | candidate. */ |

1070 | |

1071 | static slsr_cand_t |

1072 | create_mul_ssa_cand (gimple *gs, tree base_in, tree stride_in, bool speed) |

1073 | { |

1074 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; |

1075 | tree stype = NULL_TREE; |

1076 | widest_int index; |

1077 | unsigned savings = 0; |

1078 | slsr_cand_t c; |

1079 | slsr_cand_t base_cand = base_cand_from_table (base_in); |

1080 | |

1081 | /* Look at all interpretations of the base candidate, if necessary, |

1082 | to find information to propagate into this candidate. */ |

1083 | while (base_cand && !base && base_cand->kind != CAND_PHI) |

1084 | { |

1085 | |

1086 | if (base_cand->kind == CAND_MULT && integer_onep (base_cand->stride)) |

1087 | { |

1088 | /* Y = (B + i') * 1 |

1089 | X = Y * Z |

1090 | ================ |

1091 | X = (B + i') * Z */ |

1092 | base = base_cand->base_expr; |

1093 | index = base_cand->index; |

1094 | stride = stride_in; |

1095 | ctype = base_cand->cand_type; |

1096 | stype = TREE_TYPE (stride_in); |

1097 | if (has_single_use (base_in)) |

1098 | savings = (base_cand->dead_savings |

1099 | + stmt_cost (base_cand->cand_stmt, speed)); |

1100 | } |

1101 | else if (base_cand->kind == CAND_ADD |

1102 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |

1103 | { |

1104 | /* Y = B + (i' * S), S constant |

1105 | X = Y * Z |

1106 | ============================ |

1107 | X = B + ((i' * S) * Z) */ |

1108 | base = base_cand->base_expr; |

1109 | index = base_cand->index * wi::to_widest (base_cand->stride); |

1110 | stride = stride_in; |

1111 | ctype = base_cand->cand_type; |

1112 | stype = TREE_TYPE (stride_in); |

1113 | if (has_single_use (base_in)) |

1114 | savings = (base_cand->dead_savings |

1115 | + stmt_cost (base_cand->cand_stmt, speed)); |

1116 | } |

1117 | |

1118 | if (base_cand->next_interp) |

1119 | base_cand = lookup_cand (base_cand->next_interp); |

1120 | else |

1121 | base_cand = NULL; |

1122 | } |

1123 | |

1124 | if (!base) |

1125 | { |

1126 | /* No interpretations had anything useful to propagate, so |

1127 | produce X = (Y + 0) * Z. */ |

1128 | base = base_in; |

1129 | index = 0; |

1130 | stride = stride_in; |

1131 | ctype = TREE_TYPE (base_in); |

1132 | stype = TREE_TYPE (stride_in); |

1133 | } |

1134 | |

1135 | c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride, |

1136 | ctype, stype, savings); |

1137 | return c; |

1138 | } |

1139 | |

1140 | /* Create a candidate entry for a statement GS, where GS multiplies |

1141 | SSA name BASE_IN by constant STRIDE_IN. Propagate any known |

1142 | information about BASE_IN into the new candidate. Return the new |

1143 | candidate. */ |

1144 | |

1145 | static slsr_cand_t |

1146 | create_mul_imm_cand (gimple *gs, tree base_in, tree stride_in, bool speed) |

1147 | { |

1148 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; |

1149 | widest_int index, temp; |

1150 | unsigned savings = 0; |

1151 | slsr_cand_t c; |

1152 | slsr_cand_t base_cand = base_cand_from_table (base_in); |

1153 | |

1154 | /* Look at all interpretations of the base candidate, if necessary, |

1155 | to find information to propagate into this candidate. */ |

1156 | while (base_cand && !base && base_cand->kind != CAND_PHI) |

1157 | { |

1158 | if (base_cand->kind == CAND_MULT |

1159 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |

1160 | { |

1161 | /* Y = (B + i') * S, S constant |

1162 | X = Y * c |

1163 | ============================ |

1164 | X = (B + i') * (S * c) */ |

1165 | temp = wi::to_widest (base_cand->stride) * wi::to_widest (stride_in); |

1166 | if (wi::fits_to_tree_p (temp, TREE_TYPE (stride_in))) |

1167 | { |

1168 | base = base_cand->base_expr; |

1169 | index = base_cand->index; |

1170 | stride = wide_int_to_tree (TREE_TYPE (stride_in), temp); |

1171 | ctype = base_cand->cand_type; |

1172 | if (has_single_use (base_in)) |

1173 | savings = (base_cand->dead_savings |

1174 | + stmt_cost (base_cand->cand_stmt, speed)); |

1175 | } |

1176 | } |

1177 | else if (base_cand->kind == CAND_ADD && integer_onep (base_cand->stride)) |

1178 | { |

1179 | /* Y = B + (i' * 1) |

1180 | X = Y * c |

1181 | =========================== |

1182 | X = (B + i') * c */ |

1183 | base = base_cand->base_expr; |

1184 | index = base_cand->index; |

1185 | stride = stride_in; |

1186 | ctype = base_cand->cand_type; |

1187 | if (has_single_use (base_in)) |

1188 | savings = (base_cand->dead_savings |

1189 | + stmt_cost (base_cand->cand_stmt, speed)); |

1190 | } |

1191 | else if (base_cand->kind == CAND_ADD |

1192 | && base_cand->index == 1 |

1193 | && TREE_CODE (base_cand->stride) == INTEGER_CST) |

1194 | { |

1195 | /* Y = B + (1 * S), S constant |

1196 | X = Y * c |

1197 | =========================== |

1198 | X = (B + S) * c */ |

1199 | base = base_cand->base_expr; |

1200 | index = wi::to_widest (base_cand->stride); |

1201 | stride = stride_in; |

1202 | ctype = base_cand->cand_type; |

1203 | if (has_single_use (base_in)) |

1204 | savings = (base_cand->dead_savings |

1205 | + stmt_cost (base_cand->cand_stmt, speed)); |

1206 | } |

1207 | |

1208 | if (base_cand->next_interp) |

1209 | base_cand = lookup_cand (base_cand->next_interp); |

1210 | else |

1211 | base_cand = NULL; |

1212 | } |

1213 | |

1214 | if (!base) |

1215 | { |

1216 | /* No interpretations had anything useful to propagate, so |

1217 | produce X = (Y + 0) * c. */ |

1218 | base = base_in; |

1219 | index = 0; |

1220 | stride = stride_in; |

1221 | ctype = TREE_TYPE (base_in); |

1222 | } |

1223 | |

1224 | c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride, |

1225 | ctype, sizetype, savings); |

1226 | return c; |

1227 | } |

1228 | |

1229 | /* Given GS which is a multiply of scalar integers, make an appropriate |

1230 | entry in the candidate table. If this is a multiply of two SSA names, |

1231 | create two CAND_MULT interpretations and attempt to find a basis for |

1232 | each of them. Otherwise, create a single CAND_MULT and attempt to |

1233 | find a basis. */ |

1234 | |

1235 | static void |

1236 | slsr_process_mul (gimple *gs, tree rhs1, tree rhs2, bool speed) |

1237 | { |

1238 | slsr_cand_t c, c2; |

1239 | |

1240 | /* If this is a multiply of an SSA name with itself, it is highly |

1241 | unlikely that we will get a strength reduction opportunity, so |

1242 | don't record it as a candidate. This simplifies the logic for |

1243 | finding a basis, so if this is removed that must be considered. */ |

1244 | if (rhs1 == rhs2) |

1245 | return; |

1246 | |

1247 | if (TREE_CODE (rhs2) == SSA_NAME) |

1248 | { |

1249 | /* Record an interpretation of this statement in the candidate table |

1250 | assuming RHS1 is the base expression and RHS2 is the stride. */ |

1251 | c = create_mul_ssa_cand (gs, rhs1, rhs2, speed); |

1252 | |

1253 | /* Add the first interpretation to the statement-candidate mapping. */ |

1254 | add_cand_for_stmt (gs, c); |

1255 | |

1256 | /* Record another interpretation of this statement assuming RHS1 |

1257 | is the stride and RHS2 is the base expression. */ |

1258 | c2 = create_mul_ssa_cand (gs, rhs2, rhs1, speed); |

1259 | c->next_interp = c2->cand_num; |

1260 | } |

1261 | else |

1262 | { |

1263 | /* Record an interpretation for the multiply-immediate. */ |

1264 | c = create_mul_imm_cand (gs, rhs1, rhs2, speed); |

1265 | |

1266 | /* Add the interpretation to the statement-candidate mapping. */ |

1267 | add_cand_for_stmt (gs, c); |

1268 | } |

1269 | } |

1270 | |

1271 | /* Create a candidate entry for a statement GS, where GS adds two |

1272 | SSA names BASE_IN and ADDEND_IN if SUBTRACT_P is false, and |

1273 | subtracts ADDEND_IN from BASE_IN otherwise. Propagate any known |

1274 | information about the two SSA names into the new candidate. |

1275 | Return the new candidate. */ |

1276 | |

1277 | static slsr_cand_t |

1278 | create_add_ssa_cand (gimple *gs, tree base_in, tree addend_in, |

1279 | bool subtract_p, bool speed) |

1280 | { |

1281 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; |

1282 | tree stype = NULL_TREE; |

1283 | widest_int index; |

1284 | unsigned savings = 0; |

1285 | slsr_cand_t c; |

1286 | slsr_cand_t base_cand = base_cand_from_table (base_in); |

1287 | slsr_cand_t addend_cand = base_cand_from_table (addend_in); |

1288 | |

1289 | /* The most useful transformation is a multiply-immediate feeding |

1290 | an add or subtract. Look for that first. */ |

1291 | while (addend_cand && !base && addend_cand->kind != CAND_PHI) |

1292 | { |

1293 | if (addend_cand->kind == CAND_MULT |

1294 | && addend_cand->index == 0 |

1295 | && TREE_CODE (addend_cand->stride) == INTEGER_CST) |

1296 | { |

1297 | /* Z = (B + 0) * S, S constant |

1298 | X = Y +/- Z |

1299 | =========================== |

1300 | X = Y + ((+/-1 * S) * B) */ |

1301 | base = base_in; |

1302 | index = wi::to_widest (addend_cand->stride); |

1303 | if (subtract_p) |

1304 | index = -index; |

1305 | stride = addend_cand->base_expr; |

1306 | ctype = TREE_TYPE (base_in); |

1307 | stype = addend_cand->cand_type; |

1308 | if (has_single_use (addend_in)) |

1309 | savings = (addend_cand->dead_savings |

1310 | + stmt_cost (addend_cand->cand_stmt, speed)); |

1311 | } |

1312 | |

1313 | if (addend_cand->next_interp) |

1314 | addend_cand = lookup_cand (addend_cand->next_interp); |

1315 | else |

1316 | addend_cand = NULL; |

1317 | } |

1318 | |

1319 | while (base_cand && !base && base_cand->kind != CAND_PHI) |

1320 | { |

1321 | if (base_cand->kind == CAND_ADD |

1322 | && (base_cand->index == 0 |

1323 | || operand_equal_p (base_cand->stride, |

1324 | integer_zero_node, 0))) |

1325 | { |

1326 | /* Y = B + (i' * S), i' * S = 0 |

1327 | X = Y +/- Z |

1328 | ============================ |

1329 | X = B + (+/-1 * Z) */ |

1330 | base = base_cand->base_expr; |

1331 | index = subtract_p ? -1 : 1; |

1332 | stride = addend_in; |

1333 | ctype = base_cand->cand_type; |

1334 | stype = (TREE_CODE (addend_in) == INTEGER_CST ? sizetype |

1335 | : TREE_TYPE (addend_in)); |

1336 | if (has_single_use (base_in)) |

1337 | savings = (base_cand->dead_savings |

1338 | + stmt_cost (base_cand->cand_stmt, speed)); |

1339 | } |

1340 | else if (subtract_p) |

1341 | { |

1342 | slsr_cand_t subtrahend_cand = base_cand_from_table (addend_in); |

1343 | |

1344 | while (subtrahend_cand && !base && subtrahend_cand->kind != CAND_PHI) |

1345 | { |

1346 | if (subtrahend_cand->kind == CAND_MULT |

1347 | && subtrahend_cand->index == 0 |

1348 | && TREE_CODE (subtrahend_cand->stride) == INTEGER_CST) |

1349 | { |

1350 | /* Z = (B + 0) * S, S constant |

1351 | X = Y - Z |

1352 | =========================== |

1353 | Value: X = Y + ((-1 * S) * B) */ |

1354 | base = base_in; |

1355 | index = wi::to_widest (subtrahend_cand->stride); |

1356 | index = -index; |

1357 | stride = subtrahend_cand->base_expr; |

1358 | ctype = TREE_TYPE (base_in); |

1359 | stype = subtrahend_cand->cand_type; |

1360 | if (has_single_use (addend_in)) |

1361 | savings = (subtrahend_cand->dead_savings |

1362 | + stmt_cost (subtrahend_cand->cand_stmt, speed)); |

1363 | } |

1364 | |

1365 | if (subtrahend_cand->next_interp) |

1366 | subtrahend_cand = lookup_cand (subtrahend_cand->next_interp); |

1367 | else |

1368 | subtrahend_cand = NULL; |

1369 | } |

1370 | } |

1371 | |

1372 | if (base_cand->next_interp) |

1373 | base_cand = lookup_cand (base_cand->next_interp); |

1374 | else |

1375 | base_cand = NULL; |

1376 | } |

1377 | |

1378 | if (!base) |

1379 | { |

1380 | /* No interpretations had anything useful to propagate, so |

1381 | produce X = Y + (1 * Z). */ |

1382 | base = base_in; |

1383 | index = subtract_p ? -1 : 1; |

1384 | stride = addend_in; |

1385 | ctype = TREE_TYPE (base_in); |

1386 | stype = (TREE_CODE (addend_in) == INTEGER_CST ? sizetype |

1387 | : TREE_TYPE (addend_in)); |

1388 | } |

1389 | |

1390 | c = alloc_cand_and_find_basis (CAND_ADD, gs, base, index, stride, |

1391 | ctype, stype, savings); |

1392 | return c; |

1393 | } |

1394 | |

1395 | /* Create a candidate entry for a statement GS, where GS adds SSA |

1396 | name BASE_IN to constant INDEX_IN. Propagate any known information |

1397 | about BASE_IN into the new candidate. Return the new candidate. */ |

1398 | |

1399 | static slsr_cand_t |

1400 | create_add_imm_cand (gimple *gs, tree base_in, const widest_int &index_in, |

1401 | bool speed) |

1402 | { |

1403 | enum cand_kind kind = CAND_ADD; |

1404 | tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE; |

1405 | tree stype = NULL_TREE; |

1406 | widest_int index, multiple; |

1407 | unsigned savings = 0; |

1408 | slsr_cand_t c; |

1409 | slsr_cand_t base_cand = base_cand_from_table (base_in); |

1410 | |

1411 | while (base_cand && !base && base_cand->kind != CAND_PHI) |

1412 | { |

1413 | signop sign = TYPE_SIGN (TREE_TYPE (base_cand->stride)); |

1414 | |

1415 | if (TREE_CODE (base_cand->stride) == INTEGER_CST |

1416 | && wi::multiple_of_p (index_in, wi::to_widest (base_cand->stride), |

1417 | sign, &multiple)) |

1418 | { |

1419 | /* Y = (B + i') * S, S constant, c = kS for some integer k |

1420 | X = Y + c |

1421 | ============================ |

1422 | X = (B + (i'+ k)) * S |

1423 | OR |

1424 | Y = B + (i' * S), S constant, c = kS for some integer k |

1425 | X = Y + c |

1426 | ============================ |

1427 | X = (B + (i'+ k)) * S */ |

1428 | kind = base_cand->kind; |

1429 | base = base_cand->base_expr; |

1430 | index = base_cand->index + multiple; |

1431 | stride = base_cand->stride; |

1432 | ctype = base_cand->cand_type; |

1433 | stype = base_cand->stride_type; |

1434 | if (has_single_use (base_in)) |

1435 | savings = (base_cand->dead_savings |

1436 | + stmt_cost (base_cand->cand_stmt, speed)); |

1437 | } |

1438 | |

1439 | if (base_cand->next_interp) |

1440 | base_cand = lookup_cand (base_cand->next_interp); |

1441 | else |

1442 | base_cand = NULL; |

1443 | } |

1444 | |

1445 | if (!base) |

1446 | { |

1447 | /* No interpretations had anything useful to propagate, so |

1448 | produce X = Y + (c * 1). */ |

1449 | kind = CAND_ADD; |

1450 | base = base_in; |

1451 | index = index_in; |

1452 | stride = integer_one_node; |

1453 | ctype = TREE_TYPE (base_in); |

1454 | stype = sizetype; |

1455 | } |

1456 | |

1457 | c = alloc_cand_and_find_basis (kind, gs, base, index, stride, |

1458 | ctype, stype, savings); |

1459 | return c; |

1460 | } |

1461 | |

1462 | /* Given GS which is an add or subtract of scalar integers or pointers, |

1463 | make at least one appropriate entry in the candidate table. */ |

1464 | |

1465 | static void |

1466 | slsr_process_add (gimple *gs, tree rhs1, tree rhs2, bool speed) |

1467 | { |

1468 | bool subtract_p = gimple_assign_rhs_code (gs) == MINUS_EXPR; |

1469 | slsr_cand_t c = NULL, c2; |

1470 | |

1471 | if (TREE_CODE (rhs2) == SSA_NAME) |

1472 | { |

1473 | /* First record an interpretation assuming RHS1 is the base expression |

1474 | and RHS2 is the stride. But it doesn't make sense for the |

1475 | stride to be a pointer, so don't record a candidate in that case. */ |

1476 | if (!POINTER_TYPE_P (TREE_TYPE (rhs2))) |

1477 | { |

1478 | c = create_add_ssa_cand (gs, rhs1, rhs2, subtract_p, speed); |

1479 | |

1480 | /* Add the first interpretation to the statement-candidate |

1481 | mapping. */ |

1482 | add_cand_for_stmt (gs, c); |

1483 | } |

1484 | |

1485 | /* If the two RHS operands are identical, or this is a subtract, |

1486 | we're done. */ |

1487 | if (operand_equal_p (rhs1, rhs2, 0) || subtract_p) |

1488 | return; |

1489 | |

1490 | /* Otherwise, record another interpretation assuming RHS2 is the |

1491 | base expression and RHS1 is the stride, again provided that the |

1492 | stride is not a pointer. */ |

1493 | if (!POINTER_TYPE_P (TREE_TYPE (rhs1))) |

1494 | { |

1495 | c2 = create_add_ssa_cand (gs, rhs2, rhs1, false, speed); |

1496 | if (c) |

1497 | c->next_interp = c2->cand_num; |

1498 | else |

1499 | add_cand_for_stmt (gs, c2); |

1500 | } |

1501 | } |

1502 | else |

1503 | { |

1504 | /* Record an interpretation for the add-immediate. */ |

1505 | widest_int index = wi::to_widest (rhs2); |

1506 | if (subtract_p) |

1507 | index = -index; |

1508 | |

1509 | c = create_add_imm_cand (gs, rhs1, index, speed); |

1510 | |

1511 | /* Add the interpretation to the statement-candidate mapping. */ |

1512 | add_cand_for_stmt (gs, c); |

1513 | } |

1514 | } |

1515 | |

1516 | /* Given GS which is a negate of a scalar integer, make an appropriate |

1517 | entry in the candidate table. A negate is equivalent to a multiply |

1518 | by -1. */ |

1519 | |

1520 | static void |

1521 | slsr_process_neg (gimple *gs, tree rhs1, bool speed) |

1522 | { |

1523 | /* Record a CAND_MULT interpretation for the multiply by -1. */ |

1524 | slsr_cand_t c = create_mul_imm_cand (gs, rhs1, integer_minus_one_node, speed); |

1525 | |

1526 | /* Add the interpretation to the statement-candidate mapping. */ |

1527 | add_cand_for_stmt (gs, c); |

1528 | } |

1529 | |

1530 | /* Help function for legal_cast_p, operating on two trees. Checks |

1531 | whether it's allowable to cast from RHS to LHS. See legal_cast_p |

1532 | for more details. */ |

1533 | |

1534 | static bool |

1535 | legal_cast_p_1 (tree lhs_type, tree rhs_type) |

1536 | { |

1537 | unsigned lhs_size, rhs_size; |

1538 | bool lhs_wraps, rhs_wraps; |

1539 | |

1540 | lhs_size = TYPE_PRECISION (lhs_type); |

1541 | rhs_size = TYPE_PRECISION (rhs_type); |

1542 | lhs_wraps = ANY_INTEGRAL_TYPE_P (lhs_type) && TYPE_OVERFLOW_WRAPS (lhs_type); |

1543 | rhs_wraps = ANY_INTEGRAL_TYPE_P (rhs_type) && TYPE_OVERFLOW_WRAPS (rhs_type); |

1544 | |

1545 | if (lhs_size < rhs_size |

1546 | || (rhs_wraps && !lhs_wraps) |

1547 | || (rhs_wraps && lhs_wraps && rhs_size != lhs_size)) |

1548 | return false; |

1549 | |

1550 | return true; |

1551 | } |

1552 | |

1553 | /* Return TRUE if GS is a statement that defines an SSA name from |

1554 | a conversion and is legal for us to combine with an add and multiply |

1555 | in the candidate table. For example, suppose we have: |

1556 | |

1557 | A = B + i; |

1558 | C = (type) A; |

1559 | D = C * S; |

1560 | |

1561 | Without the type-cast, we would create a CAND_MULT for D with base B, |

1562 | index i, and stride S. We want to record this candidate only if it |

1563 | is equivalent to apply the type cast following the multiply: |

1564 | |

1565 | A = B + i; |

1566 | E = A * S; |

1567 | D = (type) E; |

1568 | |

1569 | We will record the type with the candidate for D. This allows us |

1570 | to use a similar previous candidate as a basis. If we have earlier seen |

1571 | |

1572 | A' = B + i'; |

1573 | C' = (type) A'; |

1574 | D' = C' * S; |

1575 | |

1576 | we can replace D with |

1577 | |

1578 | D = D' + (i - i') * S; |

1579 | |

1580 | But if moving the type-cast would change semantics, we mustn't do this. |

1581 | |

1582 | This is legitimate for casts from a non-wrapping integral type to |

1583 | any integral type of the same or larger size. It is not legitimate |

1584 | to convert a wrapping type to a non-wrapping type, or to a wrapping |

1585 | type of a different size. I.e., with a wrapping type, we must |

1586 | assume that the addition B + i could wrap, in which case performing |

1587 | the multiply before or after one of the "illegal" type casts will |

1588 | have different semantics. */ |

1589 | |

1590 | static bool |

1591 | legal_cast_p (gimple *gs, tree rhs) |

1592 | { |

1593 | if (!is_gimple_assign (gs) |

1594 | || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (gs))) |

1595 | return false; |

1596 | |

1597 | return legal_cast_p_1 (TREE_TYPE (gimple_assign_lhs (gs)), TREE_TYPE (rhs)); |

1598 | } |

1599 | |

1600 | /* Given GS which is a cast to a scalar integer type, determine whether |

1601 | the cast is legal for strength reduction. If so, make at least one |

1602 | appropriate entry in the candidate table. */ |

1603 | |

1604 | static void |

1605 | slsr_process_cast (gimple *gs, tree rhs1, bool speed) |

1606 | { |

1607 | tree lhs, ctype; |

1608 | slsr_cand_t base_cand, c = NULL, c2; |

1609 | unsigned savings = 0; |

1610 | |

1611 | if (!legal_cast_p (gs, rhs1)) |

1612 | return; |

1613 | |

1614 | lhs = gimple_assign_lhs (gs); |

1615 | base_cand = base_cand_from_table (rhs1); |

1616 | ctype = TREE_TYPE (lhs); |

1617 | |

1618 | if (base_cand && base_cand->kind != CAND_PHI) |

1619 | { |

1620 | while (base_cand) |

1621 | { |

1622 | /* Propagate all data from the base candidate except the type, |

1623 | which comes from the cast, and the base candidate's cast, |

1624 | which is no longer applicable. */ |

1625 | if (has_single_use (rhs1)) |

1626 | savings = (base_cand->dead_savings |

1627 | + stmt_cost (base_cand->cand_stmt, speed)); |

1628 | |

1629 | c = alloc_cand_and_find_basis (base_cand->kind, gs, |

1630 | base_cand->base_expr, |

1631 | base_cand->index, base_cand->stride, |

1632 | ctype, base_cand->stride_type, |

1633 | savings); |

1634 | if (base_cand->next_interp) |

1635 | base_cand = lookup_cand (base_cand->next_interp); |

1636 | else |

1637 | base_cand = NULL; |

1638 | } |

1639 | } |

1640 | else |

1641 | { |

1642 | /* If nothing is known about the RHS, create fresh CAND_ADD and |

1643 | CAND_MULT interpretations: |

1644 | |

1645 | X = Y + (0 * 1) |

1646 | X = (Y + 0) * 1 |

1647 | |

1648 | The first of these is somewhat arbitrary, but the choice of |

1649 | 1 for the stride simplifies the logic for propagating casts |

1650 | into their uses. */ |

1651 | c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1, 0, |

1652 | integer_one_node, ctype, sizetype, 0); |

1653 | c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1, 0, |

1654 | integer_one_node, ctype, sizetype, 0); |

1655 | c->next_interp = c2->cand_num; |

1656 | } |

1657 | |

1658 | /* Add the first (or only) interpretation to the statement-candidate |

1659 | mapping. */ |

1660 | add_cand_for_stmt (gs, c); |

1661 | } |

1662 | |

1663 | /* Given GS which is a copy of a scalar integer type, make at least one |

1664 | appropriate entry in the candidate table. |

1665 | |

1666 | This interface is included for completeness, but is unnecessary |

1667 | if this pass immediately follows a pass that performs copy |

1668 | propagation, such as DOM. */ |

1669 | |

1670 | static void |

1671 | slsr_process_copy (gimple *gs, tree rhs1, bool speed) |

1672 | { |

1673 | slsr_cand_t base_cand, c = NULL, c2; |

1674 | unsigned savings = 0; |

1675 | |

1676 | base_cand = base_cand_from_table (rhs1); |

1677 | |

1678 | if (base_cand && base_cand->kind != CAND_PHI) |

1679 | { |

1680 | while (base_cand) |

1681 | { |

1682 | /* Propagate all data from the base candidate. */ |

1683 | if (has_single_use (rhs1)) |

1684 | savings = (base_cand->dead_savings |

1685 | + stmt_cost (base_cand->cand_stmt, speed)); |

1686 | |

1687 | c = alloc_cand_and_find_basis (base_cand->kind, gs, |

1688 | base_cand->base_expr, |

1689 | base_cand->index, base_cand->stride, |

1690 | base_cand->cand_type, |

1691 | base_cand->stride_type, savings); |

1692 | if (base_cand->next_interp) |

1693 | base_cand = lookup_cand (base_cand->next_interp); |

1694 | else |

1695 | base_cand = NULL; |

1696 | } |

1697 | } |

1698 | else |

1699 | { |

1700 | /* If nothing is known about the RHS, create fresh CAND_ADD and |

1701 | CAND_MULT interpretations: |

1702 | |

1703 | X = Y + (0 * 1) |

1704 | X = (Y + 0) * 1 |

1705 | |

1706 | The first of these is somewhat arbitrary, but the choice of |

1707 | 1 for the stride simplifies the logic for propagating casts |

1708 | into their uses. */ |

1709 | c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1, 0, |

1710 | integer_one_node, TREE_TYPE (rhs1), |

1711 | sizetype, 0); |

1712 | c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1, 0, |

1713 | integer_one_node, TREE_TYPE (rhs1), |

1714 | sizetype, 0); |

1715 | c->next_interp = c2->cand_num; |

1716 | } |

1717 | |

1718 | /* Add the first (or only) interpretation to the statement-candidate |

1719 | mapping. */ |

1720 | add_cand_for_stmt (gs, c); |

1721 | } |

1722 | |

1723 | class find_candidates_dom_walker : public dom_walker |

1724 | { |

1725 | public: |

1726 | find_candidates_dom_walker (cdi_direction direction) |

1727 | : dom_walker (direction) {} |

1728 | virtual edge before_dom_children (basic_block); |

1729 | }; |

1730 | |

1731 | /* Find strength-reduction candidates in block BB. */ |

1732 | |

1733 | edge |

1734 | find_candidates_dom_walker::before_dom_children (basic_block bb) |

1735 | { |

1736 | bool speed = optimize_bb_for_speed_p (bb); |

1737 | |

1738 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); |

1739 | gsi_next (&gsi)) |

1740 | slsr_process_phi (gsi.phi (), speed); |

1741 | |

1742 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); |

1743 | gsi_next (&gsi)) |

1744 | { |

1745 | gimple *gs = gsi_stmt (gsi); |

1746 | |

1747 | if (gimple_vuse (gs) && gimple_assign_single_p (gs)) |

1748 | slsr_process_ref (gs); |

1749 | |

1750 | else if (is_gimple_assign (gs) |

1751 | && (INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (gs))) |

1752 | || POINTER_TYPE_P (TREE_TYPE (gimple_assign_lhs (gs))))) |

1753 | { |

1754 | tree rhs1 = NULL_TREE, rhs2 = NULL_TREE; |

1755 | |

1756 | switch (gimple_assign_rhs_code (gs)) |

1757 | { |

1758 | case MULT_EXPR: |

1759 | case PLUS_EXPR: |

1760 | rhs1 = gimple_assign_rhs1 (gs); |

1761 | rhs2 = gimple_assign_rhs2 (gs); |

1762 | /* Should never happen, but currently some buggy situations |

1763 | in earlier phases put constants in rhs1. */ |

1764 | if (TREE_CODE (rhs1) != SSA_NAME) |

1765 | continue; |

1766 | break; |

1767 | |

1768 | /* Possible future opportunity: rhs1 of a ptr+ can be |

1769 | an ADDR_EXPR. */ |

1770 | case POINTER_PLUS_EXPR: |

1771 | case MINUS_EXPR: |

1772 | rhs2 = gimple_assign_rhs2 (gs); |

1773 | gcc_fallthrough (); |

1774 | |

1775 | CASE_CONVERT: |

1776 | case SSA_NAME: |

1777 | case NEGATE_EXPR: |

1778 | rhs1 = gimple_assign_rhs1 (gs); |

1779 | if (TREE_CODE (rhs1) != SSA_NAME) |

1780 | continue; |

1781 | break; |

1782 | |

1783 | default: |

1784 | ; |

1785 | } |

1786 | |

1787 | switch (gimple_assign_rhs_code (gs)) |

1788 | { |

1789 | case MULT_EXPR: |

1790 | slsr_process_mul (gs, rhs1, rhs2, speed); |

1791 | break; |

1792 | |

1793 | case PLUS_EXPR: |

1794 | case POINTER_PLUS_EXPR: |

1795 | case MINUS_EXPR: |

1796 | slsr_process_add (gs, rhs1, rhs2, speed); |

1797 | break; |

1798 | |

1799 | case NEGATE_EXPR: |

1800 | slsr_process_neg (gs, rhs1, speed); |

1801 | break; |

1802 | |

1803 | CASE_CONVERT: |

1804 | slsr_process_cast (gs, rhs1, speed); |

1805 | break; |

1806 | |

1807 | case SSA_NAME: |

1808 | slsr_process_copy (gs, rhs1, speed); |

1809 | break; |

1810 | |

1811 | default: |

1812 | ; |

1813 | } |

1814 | } |

1815 | } |

1816 | return NULL; |

1817 | } |

1818 | |

1819 | /* Dump a candidate for debug. */ |

1820 | |

1821 | static void |

1822 | dump_candidate (slsr_cand_t c) |

1823 | { |

1824 | fprintf (dump_file, "%3d [%d] ", c->cand_num, |

1825 | gimple_bb (c->cand_stmt)->index); |

1826 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |

1827 | switch (c->kind) |

1828 | { |

1829 | case CAND_MULT: |

1830 | fputs (" MULT : (", dump_file); |

1831 | print_generic_expr (dump_file, c->base_expr); |

1832 | fputs (" + ", dump_file); |

1833 | print_decs (c->index, dump_file); |

1834 | fputs (") * ", dump_file); |

1835 | if (TREE_CODE (c->stride) != INTEGER_CST |

1836 | && c->stride_type != TREE_TYPE (c->stride)) |

1837 | { |

1838 | fputs ("(", dump_file); |

1839 | print_generic_expr (dump_file, c->stride_type); |

1840 | fputs (")", dump_file); |

1841 | } |

1842 | print_generic_expr (dump_file, c->stride); |

1843 | fputs (" : ", dump_file); |

1844 | break; |

1845 | case CAND_ADD: |

1846 | fputs (" ADD : ", dump_file); |

1847 | print_generic_expr (dump_file, c->base_expr); |

1848 | fputs (" + (", dump_file); |

1849 | print_decs (c->index, dump_file); |

1850 | fputs (" * ", dump_file); |

1851 | if (TREE_CODE (c->stride) != INTEGER_CST |

1852 | && c->stride_type != TREE_TYPE (c->stride)) |

1853 | { |

1854 | fputs ("(", dump_file); |

1855 | print_generic_expr (dump_file, c->stride_type); |

1856 | fputs (")", dump_file); |

1857 | } |

1858 | print_generic_expr (dump_file, c->stride); |

1859 | fputs (") : ", dump_file); |

1860 | break; |

1861 | case CAND_REF: |

1862 | fputs (" REF : ", dump_file); |

1863 | print_generic_expr (dump_file, c->base_expr); |

1864 | fputs (" + (", dump_file); |

1865 | print_generic_expr (dump_file, c->stride); |

1866 | fputs (") + ", dump_file); |

1867 | print_decs (c->index, dump_file); |

1868 | fputs (" : ", dump_file); |

1869 | break; |

1870 | case CAND_PHI: |

1871 | fputs (" PHI : ", dump_file); |

1872 | print_generic_expr (dump_file, c->base_expr); |

1873 | fputs (" + (unknown * ", dump_file); |

1874 | print_generic_expr (dump_file, c->stride); |

1875 | fputs (") : ", dump_file); |

1876 | break; |

1877 | default: |

1878 | gcc_unreachable (); |

1879 | } |

1880 | print_generic_expr (dump_file, c->cand_type); |

1881 | fprintf (dump_file, "\n basis: %d dependent: %d sibling: %d\n", |

1882 | c->basis, c->dependent, c->sibling); |

1883 | fprintf (dump_file, " next-interp: %d dead-savings: %d\n", |

1884 | c->next_interp, c->dead_savings); |

1885 | if (c->def_phi) |

1886 | fprintf (dump_file, " phi: %d\n", c->def_phi); |

1887 | fputs ("\n", dump_file); |

1888 | } |

1889 | |

1890 | /* Dump the candidate vector for debug. */ |

1891 | |

1892 | static void |

1893 | dump_cand_vec (void) |

1894 | { |

1895 | unsigned i; |

1896 | slsr_cand_t c; |

1897 | |

1898 | fprintf (dump_file, "\nStrength reduction candidate vector:\n\n"); |

1899 | |

1900 | FOR_EACH_VEC_ELT (cand_vec, i, c) |

1901 | dump_candidate (c); |

1902 | } |

1903 | |

1904 | /* Callback used to dump the candidate chains hash table. */ |

1905 | |

1906 | int |

1907 | ssa_base_cand_dump_callback (cand_chain **slot, void *ignored ATTRIBUTE_UNUSED) |

1908 | { |

1909 | const_cand_chain_t chain = *slot; |

1910 | cand_chain_t p; |

1911 | |

1912 | print_generic_expr (dump_file, chain->base_expr); |

1913 | fprintf (dump_file, " -> %d", chain->cand->cand_num); |

1914 | |

1915 | for (p = chain->next; p; p = p->next) |

1916 | fprintf (dump_file, " -> %d", p->cand->cand_num); |

1917 | |

1918 | fputs ("\n", dump_file); |

1919 | return 1; |

1920 | } |

1921 | |

1922 | /* Dump the candidate chains. */ |

1923 | |

1924 | static void |

1925 | dump_cand_chains (void) |

1926 | { |

1927 | fprintf (dump_file, "\nStrength reduction candidate chains:\n\n"); |

1928 | base_cand_map->traverse_noresize <void *, ssa_base_cand_dump_callback> |

1929 | (NULL); |

1930 | fputs ("\n", dump_file); |

1931 | } |

1932 | |

1933 | /* Dump the increment vector for debug. */ |

1934 | |

1935 | static void |

1936 | dump_incr_vec (void) |

1937 | { |

1938 | if (dump_file && (dump_flags & TDF_DETAILS)) |

1939 | { |

1940 | unsigned i; |

1941 | |

1942 | fprintf (dump_file, "\nIncrement vector:\n\n"); |

1943 | |

1944 | for (i = 0; i < incr_vec_len; i++) |

1945 | { |

1946 | fprintf (dump_file, "%3d increment: ", i); |

1947 | print_decs (incr_vec[i].incr, dump_file); |

1948 | fprintf (dump_file, "\n count: %d", incr_vec[i].count); |

1949 | fprintf (dump_file, "\n cost: %d", incr_vec[i].cost); |

1950 | fputs ("\n initializer: ", dump_file); |

1951 | print_generic_expr (dump_file, incr_vec[i].initializer); |

1952 | fputs ("\n\n", dump_file); |

1953 | } |

1954 | } |

1955 | } |

1956 | |

1957 | /* Replace *EXPR in candidate C with an equivalent strength-reduced |

1958 | data reference. */ |

1959 | |

1960 | static void |

1961 | replace_ref (tree *expr, slsr_cand_t c) |

1962 | { |

1963 | tree add_expr, mem_ref, acc_type = TREE_TYPE (*expr); |

1964 | unsigned HOST_WIDE_INT misalign; |

1965 | unsigned align; |

1966 | |

1967 | /* Ensure the memory reference carries the minimum alignment |

1968 | requirement for the data type. See PR58041. */ |

1969 | get_object_alignment_1 (*expr, &align, &misalign); |

1970 | if (misalign != 0) |

1971 | align = least_bit_hwi (misalign); |

1972 | if (align < TYPE_ALIGN (acc_type)) |

1973 | acc_type = build_aligned_type (acc_type, align); |

1974 | |

1975 | add_expr = fold_build2 (POINTER_PLUS_EXPR, c->cand_type, |

1976 | c->base_expr, c->stride); |

1977 | mem_ref = fold_build2 (MEM_REF, acc_type, add_expr, |

1978 | wide_int_to_tree (c->cand_type, c->index)); |

1979 | |

1980 | /* Gimplify the base addressing expression for the new MEM_REF tree. */ |

1981 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

1982 | TREE_OPERAND (mem_ref, 0) |

1983 | = force_gimple_operand_gsi (&gsi, TREE_OPERAND (mem_ref, 0), |

1984 | /*simple_p=*/true, NULL, |

1985 | /*before=*/true, GSI_SAME_STMT); |

1986 | copy_ref_info (mem_ref, *expr); |

1987 | *expr = mem_ref; |

1988 | update_stmt (c->cand_stmt); |

1989 | } |

1990 | |

1991 | /* Replace CAND_REF candidate C, each sibling of candidate C, and each |

1992 | dependent of candidate C with an equivalent strength-reduced data |

1993 | reference. */ |

1994 | |

1995 | static void |

1996 | replace_refs (slsr_cand_t c) |

1997 | { |

1998 | if (dump_file && (dump_flags & TDF_DETAILS)) |

1999 | { |

2000 | fputs ("Replacing reference: ", dump_file); |

2001 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |

2002 | } |

2003 | |

2004 | if (gimple_vdef (c->cand_stmt)) |

2005 | { |

2006 | tree *lhs = gimple_assign_lhs_ptr (c->cand_stmt); |

2007 | replace_ref (lhs, c); |

2008 | } |

2009 | else |

2010 | { |

2011 | tree *rhs = gimple_assign_rhs1_ptr (c->cand_stmt); |

2012 | replace_ref (rhs, c); |

2013 | } |

2014 | |

2015 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2016 | { |

2017 | fputs ("With: ", dump_file); |

2018 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |

2019 | fputs ("\n", dump_file); |

2020 | } |

2021 | |

2022 | if (c->sibling) |

2023 | replace_refs (lookup_cand (c->sibling)); |

2024 | |

2025 | if (c->dependent) |

2026 | replace_refs (lookup_cand (c->dependent)); |

2027 | } |

2028 | |

2029 | /* Return TRUE if candidate C is dependent upon a PHI. */ |

2030 | |

2031 | static bool |

2032 | phi_dependent_cand_p (slsr_cand_t c) |

2033 | { |

2034 | /* A candidate is not necessarily dependent upon a PHI just because |

2035 | it has a phi definition for its base name. It may have a basis |

2036 | that relies upon the same phi definition, in which case the PHI |

2037 | is irrelevant to this candidate. */ |

2038 | return (c->def_phi |

2039 | && c->basis |

2040 | && lookup_cand (c->basis)->def_phi != c->def_phi); |

2041 | } |

2042 | |

2043 | /* Calculate the increment required for candidate C relative to |

2044 | its basis. */ |

2045 | |

2046 | static widest_int |

2047 | cand_increment (slsr_cand_t c) |

2048 | { |

2049 | slsr_cand_t basis; |

2050 | |

2051 | /* If the candidate doesn't have a basis, just return its own |

2052 | index. This is useful in record_increments to help us find |

2053 | an existing initializer. Also, if the candidate's basis is |

2054 | hidden by a phi, then its own index will be the increment |

2055 | from the newly introduced phi basis. */ |

2056 | if (!c->basis || phi_dependent_cand_p (c)) |

2057 | return c->index; |

2058 | |

2059 | basis = lookup_cand (c->basis); |

2060 | gcc_assert (operand_equal_p (c->base_expr, basis->base_expr, 0)); |

2061 | return c->index - basis->index; |

2062 | } |

2063 | |

2064 | /* Calculate the increment required for candidate C relative to |

2065 | its basis. If we aren't going to generate pointer arithmetic |

2066 | for this candidate, return the absolute value of that increment |

2067 | instead. */ |

2068 | |

2069 | static inline widest_int |

2070 | cand_abs_increment (slsr_cand_t c) |

2071 | { |

2072 | widest_int increment = cand_increment (c); |

2073 | |

2074 | if (!address_arithmetic_p && wi::neg_p (increment)) |

2075 | increment = -increment; |

2076 | |

2077 | return increment; |

2078 | } |

2079 | |

2080 | /* Return TRUE iff candidate C has already been replaced under |

2081 | another interpretation. */ |

2082 | |

2083 | static inline bool |

2084 | cand_already_replaced (slsr_cand_t c) |

2085 | { |

2086 | return (gimple_bb (c->cand_stmt) == 0); |

2087 | } |

2088 | |

2089 | /* Common logic used by replace_unconditional_candidate and |

2090 | replace_conditional_candidate. */ |

2091 | |

2092 | static void |

2093 | replace_mult_candidate (slsr_cand_t c, tree basis_name, widest_int bump) |

2094 | { |

2095 | tree target_type = TREE_TYPE (gimple_assign_lhs (c->cand_stmt)); |

2096 | enum tree_code cand_code = gimple_assign_rhs_code (c->cand_stmt); |

2097 | |

2098 | /* It is not useful to replace casts, copies, negates, or adds of |

2099 | an SSA name and a constant. */ |

2100 | if (cand_code == SSA_NAME |

2101 | || CONVERT_EXPR_CODE_P (cand_code) |

2102 | || cand_code == PLUS_EXPR |

2103 | || cand_code == POINTER_PLUS_EXPR |

2104 | || cand_code == MINUS_EXPR |

2105 | || cand_code == NEGATE_EXPR) |

2106 | return; |

2107 | |

2108 | enum tree_code code = PLUS_EXPR; |

2109 | tree bump_tree; |

2110 | gimple *stmt_to_print = NULL; |

2111 | |

2112 | if (wi::neg_p (bump)) |

2113 | { |

2114 | code = MINUS_EXPR; |

2115 | bump = -bump; |

2116 | } |

2117 | |

2118 | /* It is possible that the resulting bump doesn't fit in target_type. |

2119 | Abandon the replacement in this case. This does not affect |

2120 | siblings or dependents of C. */ |

2121 | if (bump != wi::ext (bump, TYPE_PRECISION (target_type), |

2122 | TYPE_SIGN (target_type))) |

2123 | return; |

2124 | |

2125 | bump_tree = wide_int_to_tree (target_type, bump); |

2126 | |

2127 | /* If the basis name and the candidate's LHS have incompatible types, |

2128 | introduce a cast. */ |

2129 | if (!useless_type_conversion_p (target_type, TREE_TYPE (basis_name))) |

2130 | basis_name = introduce_cast_before_cand (c, target_type, basis_name); |

2131 | |

2132 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2133 | { |

2134 | fputs ("Replacing: ", dump_file); |

2135 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |

2136 | } |

2137 | |

2138 | if (bump == 0) |

2139 | { |

2140 | tree lhs = gimple_assign_lhs (c->cand_stmt); |

2141 | gassign *copy_stmt = gimple_build_assign (lhs, basis_name); |

2142 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

2143 | slsr_cand_t cc = c; |

2144 | gimple_set_location (copy_stmt, gimple_location (c->cand_stmt)); |

2145 | gsi_replace (&gsi, copy_stmt, false); |

2146 | c->cand_stmt = copy_stmt; |

2147 | while (cc->next_interp) |

2148 | { |

2149 | cc = lookup_cand (cc->next_interp); |

2150 | cc->cand_stmt = copy_stmt; |

2151 | } |

2152 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2153 | stmt_to_print = copy_stmt; |

2154 | } |

2155 | else |

2156 | { |

2157 | tree rhs1, rhs2; |

2158 | if (cand_code != NEGATE_EXPR) { |

2159 | rhs1 = gimple_assign_rhs1 (c->cand_stmt); |

2160 | rhs2 = gimple_assign_rhs2 (c->cand_stmt); |

2161 | } |

2162 | if (cand_code != NEGATE_EXPR |

2163 | && ((operand_equal_p (rhs1, basis_name, 0) |

2164 | && operand_equal_p (rhs2, bump_tree, 0)) |

2165 | || (operand_equal_p (rhs1, bump_tree, 0) |

2166 | && operand_equal_p (rhs2, basis_name, 0)))) |

2167 | { |

2168 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2169 | { |

2170 | fputs ("(duplicate, not actually replacing)", dump_file); |

2171 | stmt_to_print = c->cand_stmt; |

2172 | } |

2173 | } |

2174 | else |

2175 | { |

2176 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

2177 | slsr_cand_t cc = c; |

2178 | gimple_assign_set_rhs_with_ops (&gsi, code, basis_name, bump_tree); |

2179 | update_stmt (gsi_stmt (gsi)); |

2180 | c->cand_stmt = gsi_stmt (gsi); |

2181 | while (cc->next_interp) |

2182 | { |

2183 | cc = lookup_cand (cc->next_interp); |

2184 | cc->cand_stmt = gsi_stmt (gsi); |

2185 | } |

2186 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2187 | stmt_to_print = gsi_stmt (gsi); |

2188 | } |

2189 | } |

2190 | |

2191 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2192 | { |

2193 | fputs ("With: ", dump_file); |

2194 | print_gimple_stmt (dump_file, stmt_to_print, 0); |

2195 | fputs ("\n", dump_file); |

2196 | } |

2197 | } |

2198 | |

2199 | /* Replace candidate C with an add or subtract. Note that we only |

2200 | operate on CAND_MULTs with known strides, so we will never generate |

2201 | a POINTER_PLUS_EXPR. Each candidate X = (B + i) * S is replaced by |

2202 | X = Y + ((i - i') * S), as described in the module commentary. The |

2203 | folded value ((i - i') * S) is referred to here as the "bump." */ |

2204 | |

2205 | static void |

2206 | replace_unconditional_candidate (slsr_cand_t c) |

2207 | { |

2208 | slsr_cand_t basis; |

2209 | |

2210 | if (cand_already_replaced (c)) |

2211 | return; |

2212 | |

2213 | basis = lookup_cand (c->basis); |

2214 | widest_int bump = cand_increment (c) * wi::to_widest (c->stride); |

2215 | |

2216 | replace_mult_candidate (c, gimple_assign_lhs (basis->cand_stmt), bump); |

2217 | } |

2218 | |

2219 | /* Return the index in the increment vector of the given INCREMENT, |

2220 | or -1 if not found. The latter can occur if more than |

2221 | MAX_INCR_VEC_LEN increments have been found. */ |

2222 | |

2223 | static inline int |

2224 | incr_vec_index (const widest_int &increment) |

2225 | { |

2226 | unsigned i; |

2227 | |

2228 | for (i = 0; i < incr_vec_len && increment != incr_vec[i].incr; i++) |

2229 | ; |

2230 | |

2231 | if (i < incr_vec_len) |

2232 | return i; |

2233 | else |

2234 | return -1; |

2235 | } |

2236 | |

2237 | /* Create a new statement along edge E to add BASIS_NAME to the product |

2238 | of INCREMENT and the stride of candidate C. Create and return a new |

2239 | SSA name from *VAR to be used as the LHS of the new statement. |

2240 | KNOWN_STRIDE is true iff C's stride is a constant. */ |

2241 | |

2242 | static tree |

2243 | create_add_on_incoming_edge (slsr_cand_t c, tree basis_name, |

2244 | widest_int increment, edge e, location_t loc, |

2245 | bool known_stride) |

2246 | { |

2247 | tree lhs, basis_type; |

2248 | gassign *new_stmt, *cast_stmt = NULL; |

2249 | |

2250 | /* If the add candidate along this incoming edge has the same |

2251 | index as C's hidden basis, the hidden basis represents this |

2252 | edge correctly. */ |

2253 | if (increment == 0) |

2254 | return basis_name; |

2255 | |

2256 | basis_type = TREE_TYPE (basis_name); |

2257 | lhs = make_temp_ssa_name (basis_type, NULL, "slsr"); |

2258 | |

2259 | /* Occasionally people convert integers to pointers without a |

2260 | cast, leading us into trouble if we aren't careful. */ |

2261 | enum tree_code plus_code |

2262 | = POINTER_TYPE_P (basis_type) ? POINTER_PLUS_EXPR : PLUS_EXPR; |

2263 | |

2264 | if (known_stride) |

2265 | { |

2266 | tree bump_tree; |

2267 | enum tree_code code = plus_code; |

2268 | widest_int bump = increment * wi::to_widest (c->stride); |

2269 | if (wi::neg_p (bump) && !POINTER_TYPE_P (basis_type)) |

2270 | { |

2271 | code = MINUS_EXPR; |

2272 | bump = -bump; |

2273 | } |

2274 | |

2275 | tree stride_type = POINTER_TYPE_P (basis_type) ? sizetype : basis_type; |

2276 | bump_tree = wide_int_to_tree (stride_type, bump); |

2277 | new_stmt = gimple_build_assign (lhs, code, basis_name, bump_tree); |

2278 | } |

2279 | else |

2280 | { |

2281 | int i; |

2282 | bool negate_incr = !POINTER_TYPE_P (basis_type) && wi::neg_p (increment); |

2283 | i = incr_vec_index (negate_incr ? -increment : increment); |

2284 | gcc_assert (i >= 0); |

2285 | |

2286 | if (incr_vec[i].initializer) |

2287 | { |

2288 | enum tree_code code = negate_incr ? MINUS_EXPR : plus_code; |

2289 | new_stmt = gimple_build_assign (lhs, code, basis_name, |

2290 | incr_vec[i].initializer); |

2291 | } |

2292 | else { |

2293 | tree stride; |

2294 | |

2295 | if (!types_compatible_p (TREE_TYPE (c->stride), c->stride_type)) |

2296 | { |

2297 | tree cast_stride = make_temp_ssa_name (c->stride_type, NULL, |

2298 | "slsr"); |

2299 | cast_stmt = gimple_build_assign (cast_stride, NOP_EXPR, |

2300 | c->stride); |

2301 | stride = cast_stride; |

2302 | } |

2303 | else |

2304 | stride = c->stride; |

2305 | |

2306 | if (increment == 1) |

2307 | new_stmt = gimple_build_assign (lhs, plus_code, basis_name, stride); |

2308 | else if (increment == -1) |

2309 | new_stmt = gimple_build_assign (lhs, MINUS_EXPR, basis_name, stride); |

2310 | else |

2311 | gcc_unreachable (); |

2312 | } |

2313 | } |

2314 | |

2315 | if (cast_stmt) |

2316 | { |

2317 | gimple_set_location (cast_stmt, loc); |

2318 | gsi_insert_on_edge (e, cast_stmt); |

2319 | } |

2320 | |

2321 | gimple_set_location (new_stmt, loc); |

2322 | gsi_insert_on_edge (e, new_stmt); |

2323 | |

2324 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2325 | { |

2326 | if (cast_stmt) |

2327 | { |

2328 | fprintf (dump_file, "Inserting cast on edge %d->%d: ", |

2329 | e->src->index, e->dest->index); |

2330 | print_gimple_stmt (dump_file, cast_stmt, 0); |

2331 | } |

2332 | fprintf (dump_file, "Inserting on edge %d->%d: ", e->src->index, |

2333 | e->dest->index); |

2334 | print_gimple_stmt (dump_file, new_stmt, 0); |

2335 | } |

2336 | |

2337 | return lhs; |

2338 | } |

2339 | |

2340 | /* Clear the visited field for a tree of PHI candidates. */ |

2341 | |

2342 | static void |

2343 | clear_visited (gphi *phi) |

2344 | { |

2345 | unsigned i; |

2346 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |

2347 | |

2348 | if (phi_cand->visited) |

2349 | { |

2350 | phi_cand->visited = 0; |

2351 | |

2352 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

2353 | { |

2354 | tree arg = gimple_phi_arg_def (phi, i); |

2355 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

2356 | if (gimple_code (arg_def) == GIMPLE_PHI) |

2357 | clear_visited (as_a <gphi *> (arg_def)); |

2358 | } |

2359 | } |

2360 | } |

2361 | |

2362 | /* Recursive helper function for create_phi_basis. */ |

2363 | |

2364 | static tree |

2365 | create_phi_basis_1 (slsr_cand_t c, gimple *from_phi, tree basis_name, |

2366 | location_t loc, bool known_stride) |

2367 | { |

2368 | int i; |

2369 | tree name, phi_arg; |

2370 | gphi *phi; |

2371 | slsr_cand_t basis = lookup_cand (c->basis); |

2372 | int nargs = gimple_phi_num_args (from_phi); |

2373 | basic_block phi_bb = gimple_bb (from_phi); |

2374 | slsr_cand_t phi_cand = *stmt_cand_map->get (from_phi); |

2375 | auto_vec<tree> phi_args (nargs); |

2376 | |

2377 | if (phi_cand->visited) |

2378 | return phi_cand->cached_basis; |

2379 | phi_cand->visited = 1; |

2380 | |

2381 | /* Process each argument of the existing phi that represents |

2382 | conditionally-executed add candidates. */ |

2383 | for (i = 0; i < nargs; i++) |

2384 | { |

2385 | edge e = (*phi_bb->preds)[i]; |

2386 | tree arg = gimple_phi_arg_def (from_phi, i); |

2387 | tree feeding_def; |

2388 | |

2389 | /* If the phi argument is the base name of the CAND_PHI, then |

2390 | this incoming arc should use the hidden basis. */ |

2391 | if (operand_equal_p (arg, phi_cand->base_expr, 0)) |

2392 | if (basis->index == 0) |

2393 | feeding_def = gimple_assign_lhs (basis->cand_stmt); |

2394 | else |

2395 | { |

2396 | widest_int incr = -basis->index; |

2397 | feeding_def = create_add_on_incoming_edge (c, basis_name, incr, |

2398 | e, loc, known_stride); |

2399 | } |

2400 | else |

2401 | { |

2402 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

2403 | |

2404 | /* If there is another phi along this incoming edge, we must |

2405 | process it in the same fashion to ensure that all basis |

2406 | adjustments are made along its incoming edges. */ |

2407 | if (gimple_code (arg_def) == GIMPLE_PHI) |

2408 | feeding_def = create_phi_basis_1 (c, arg_def, basis_name, |

2409 | loc, known_stride); |

2410 | else |

2411 | { |

2412 | slsr_cand_t arg_cand = base_cand_from_table (arg); |

2413 | widest_int diff = arg_cand->index - basis->index; |

2414 | feeding_def = create_add_on_incoming_edge (c, basis_name, diff, |

2415 | e, loc, known_stride); |

2416 | } |

2417 | } |

2418 | |

2419 | /* Because of recursion, we need to save the arguments in a vector |

2420 | so we can create the PHI statement all at once. Otherwise the |

2421 | storage for the half-created PHI can be reclaimed. */ |

2422 | phi_args.safe_push (feeding_def); |

2423 | } |

2424 | |

2425 | /* Create the new phi basis. */ |

2426 | name = make_temp_ssa_name (TREE_TYPE (basis_name), NULL, "slsr"); |

2427 | phi = create_phi_node (name, phi_bb); |

2428 | SSA_NAME_DEF_STMT (name) = phi; |

2429 | |

2430 | FOR_EACH_VEC_ELT (phi_args, i, phi_arg) |

2431 | { |

2432 | edge e = (*phi_bb->preds)[i]; |

2433 | add_phi_arg (phi, phi_arg, e, loc); |

2434 | } |

2435 | |

2436 | update_stmt (phi); |

2437 | |

2438 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2439 | { |

2440 | fputs ("Introducing new phi basis: ", dump_file); |

2441 | print_gimple_stmt (dump_file, phi, 0); |

2442 | } |

2443 | |

2444 | phi_cand->cached_basis = name; |

2445 | return name; |

2446 | } |

2447 | |

2448 | /* Given a candidate C with BASIS_NAME being the LHS of C's basis which |

2449 | is hidden by the phi node FROM_PHI, create a new phi node in the same |

2450 | block as FROM_PHI. The new phi is suitable for use as a basis by C, |

2451 | with its phi arguments representing conditional adjustments to the |

2452 | hidden basis along conditional incoming paths. Those adjustments are |

2453 | made by creating add statements (and sometimes recursively creating |

2454 | phis) along those incoming paths. LOC is the location to attach to |

2455 | the introduced statements. KNOWN_STRIDE is true iff C's stride is a |

2456 | constant. */ |

2457 | |

2458 | static tree |

2459 | create_phi_basis (slsr_cand_t c, gimple *from_phi, tree basis_name, |

2460 | location_t loc, bool known_stride) |

2461 | { |

2462 | tree retval = create_phi_basis_1 (c, from_phi, basis_name, loc, |

2463 | known_stride); |

2464 | gcc_assert (retval); |

2465 | clear_visited (as_a <gphi *> (from_phi)); |

2466 | return retval; |

2467 | } |

2468 | |

2469 | /* Given a candidate C whose basis is hidden by at least one intervening |

2470 | phi, introduce a matching number of new phis to represent its basis |

2471 | adjusted by conditional increments along possible incoming paths. Then |

2472 | replace C as though it were an unconditional candidate, using the new |

2473 | basis. */ |

2474 | |

2475 | static void |

2476 | replace_conditional_candidate (slsr_cand_t c) |

2477 | { |

2478 | tree basis_name, name; |

2479 | slsr_cand_t basis; |

2480 | location_t loc; |

2481 | |

2482 | /* Look up the LHS SSA name from C's basis. This will be the |

2483 | RHS1 of the adds we will introduce to create new phi arguments. */ |

2484 | basis = lookup_cand (c->basis); |

2485 | basis_name = gimple_assign_lhs (basis->cand_stmt); |

2486 | |

2487 | /* Create a new phi statement which will represent C's true basis |

2488 | after the transformation is complete. */ |

2489 | loc = gimple_location (c->cand_stmt); |

2490 | name = create_phi_basis (c, lookup_cand (c->def_phi)->cand_stmt, |

2491 | basis_name, loc, KNOWN_STRIDE); |

2492 | |

2493 | /* Replace C with an add of the new basis phi and a constant. */ |

2494 | widest_int bump = c->index * wi::to_widest (c->stride); |

2495 | |

2496 | replace_mult_candidate (c, name, bump); |

2497 | } |

2498 | |

2499 | /* Recursive helper function for phi_add_costs. SPREAD is a measure of |

2500 | how many PHI nodes we have visited at this point in the tree walk. */ |

2501 | |

2502 | static int |

2503 | phi_add_costs_1 (gimple *phi, slsr_cand_t c, int one_add_cost, int *spread) |

2504 | { |

2505 | unsigned i; |

2506 | int cost = 0; |

2507 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |

2508 | |

2509 | if (phi_cand->visited) |

2510 | return 0; |

2511 | |

2512 | phi_cand->visited = 1; |

2513 | (*spread)++; |

2514 | |

2515 | /* If we work our way back to a phi that isn't dominated by the hidden |

2516 | basis, this isn't a candidate for replacement. Indicate this by |

2517 | returning an unreasonably high cost. It's not easy to detect |

2518 | these situations when determining the basis, so we defer the |

2519 | decision until now. */ |

2520 | basic_block phi_bb = gimple_bb (phi); |

2521 | slsr_cand_t basis = lookup_cand (c->basis); |

2522 | basic_block basis_bb = gimple_bb (basis->cand_stmt); |

2523 | |

2524 | if (phi_bb == basis_bb || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) |

2525 | return COST_INFINITE; |

2526 | |

2527 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

2528 | { |

2529 | tree arg = gimple_phi_arg_def (phi, i); |

2530 | |

2531 | if (arg != phi_cand->base_expr) |

2532 | { |

2533 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

2534 | |

2535 | if (gimple_code (arg_def) == GIMPLE_PHI) |

2536 | { |

2537 | cost += phi_add_costs_1 (arg_def, c, one_add_cost, spread); |

2538 | |

2539 | if (cost >= COST_INFINITE || *spread > MAX_SPREAD) |

2540 | return COST_INFINITE; |

2541 | } |

2542 | else |

2543 | { |

2544 | slsr_cand_t arg_cand = base_cand_from_table (arg); |

2545 | |

2546 | if (arg_cand->index != c->index) |

2547 | cost += one_add_cost; |

2548 | } |

2549 | } |

2550 | } |

2551 | |

2552 | return cost; |

2553 | } |

2554 | |

2555 | /* Compute the expected costs of inserting basis adjustments for |

2556 | candidate C with phi-definition PHI. The cost of inserting |

2557 | one adjustment is given by ONE_ADD_COST. If PHI has arguments |

2558 | which are themselves phi results, recursively calculate costs |

2559 | for those phis as well. */ |

2560 | |

2561 | static int |

2562 | phi_add_costs (gimple *phi, slsr_cand_t c, int one_add_cost) |

2563 | { |

2564 | int spread = 0; |

2565 | int retval = phi_add_costs_1 (phi, c, one_add_cost, &spread); |

2566 | clear_visited (as_a <gphi *> (phi)); |

2567 | return retval; |

2568 | } |

2569 | /* For candidate C, each sibling of candidate C, and each dependent of |

2570 | candidate C, determine whether the candidate is dependent upon a |

2571 | phi that hides its basis. If not, replace the candidate unconditionally. |

2572 | Otherwise, determine whether the cost of introducing compensation code |

2573 | for the candidate is offset by the gains from strength reduction. If |

2574 | so, replace the candidate and introduce the compensation code. */ |

2575 | |

2576 | static void |

2577 | replace_uncond_cands_and_profitable_phis (slsr_cand_t c) |

2578 | { |

2579 | if (phi_dependent_cand_p (c)) |

2580 | { |

2581 | /* A multiply candidate with a stride of 1 is just an artifice |

2582 | of a copy or cast; there is no value in replacing it. */ |

2583 | if (c->kind == CAND_MULT && wi::to_widest (c->stride) != 1) |

2584 | { |

2585 | /* A candidate dependent upon a phi will replace a multiply by |

2586 | a constant with an add, and will insert at most one add for |

2587 | each phi argument. Add these costs with the potential dead-code |

2588 | savings to determine profitability. */ |

2589 | bool speed = optimize_bb_for_speed_p (gimple_bb (c->cand_stmt)); |

2590 | int mult_savings = stmt_cost (c->cand_stmt, speed); |

2591 | gimple *phi = lookup_cand (c->def_phi)->cand_stmt; |

2592 | tree phi_result = gimple_phi_result (phi); |

2593 | int one_add_cost = add_cost (speed, |

2594 | TYPE_MODE (TREE_TYPE (phi_result))); |

2595 | int add_costs = one_add_cost + phi_add_costs (phi, c, one_add_cost); |

2596 | int cost = add_costs - mult_savings - c->dead_savings; |

2597 | |

2598 | if (dump_file && (dump_flags & TDF_DETAILS)) |

2599 | { |

2600 | fprintf (dump_file, " Conditional candidate %d:\n", c->cand_num); |

2601 | fprintf (dump_file, " add_costs = %d\n", add_costs); |

2602 | fprintf (dump_file, " mult_savings = %d\n", mult_savings); |

2603 | fprintf (dump_file, " dead_savings = %d\n", c->dead_savings); |

2604 | fprintf (dump_file, " cost = %d\n", cost); |

2605 | if (cost <= COST_NEUTRAL) |

2606 | fputs (" Replacing...\n", dump_file); |

2607 | else |

2608 | fputs (" Not replaced.\n", dump_file); |

2609 | } |

2610 | |

2611 | if (cost <= COST_NEUTRAL) |

2612 | replace_conditional_candidate (c); |

2613 | } |

2614 | } |

2615 | else |

2616 | replace_unconditional_candidate (c); |

2617 | |

2618 | if (c->sibling) |

2619 | replace_uncond_cands_and_profitable_phis (lookup_cand (c->sibling)); |

2620 | |

2621 | if (c->dependent) |

2622 | replace_uncond_cands_and_profitable_phis (lookup_cand (c->dependent)); |

2623 | } |

2624 | |

2625 | /* Count the number of candidates in the tree rooted at C that have |

2626 | not already been replaced under other interpretations. */ |

2627 | |

2628 | static int |

2629 | count_candidates (slsr_cand_t c) |

2630 | { |

2631 | unsigned count = cand_already_replaced (c) ? 0 : 1; |

2632 | |

2633 | if (c->sibling) |

2634 | count += count_candidates (lookup_cand (c->sibling)); |

2635 | |

2636 | if (c->dependent) |

2637 | count += count_candidates (lookup_cand (c->dependent)); |

2638 | |

2639 | return count; |

2640 | } |

2641 | |

2642 | /* Increase the count of INCREMENT by one in the increment vector. |

2643 | INCREMENT is associated with candidate C. If INCREMENT is to be |

2644 | conditionally executed as part of a conditional candidate replacement, |

2645 | IS_PHI_ADJUST is true, otherwise false. If an initializer |

2646 | T_0 = stride * I is provided by a candidate that dominates all |

2647 | candidates with the same increment, also record T_0 for subsequent use. */ |

2648 | |

2649 | static void |

2650 | record_increment (slsr_cand_t c, widest_int increment, bool is_phi_adjust) |

2651 | { |

2652 | bool found = false; |

2653 | unsigned i; |

2654 | |

2655 | /* Treat increments that differ only in sign as identical so as to |

2656 | share initializers, unless we are generating pointer arithmetic. */ |

2657 | if (!address_arithmetic_p && wi::neg_p (increment)) |

2658 | increment = -increment; |

2659 | |

2660 | for (i = 0; i < incr_vec_len; i++) |

2661 | { |

2662 | if (incr_vec[i].incr == increment) |

2663 | { |

2664 | incr_vec[i].count++; |

2665 | found = true; |

2666 | |

2667 | /* If we previously recorded an initializer that doesn't |

2668 | dominate this candidate, it's not going to be useful to |

2669 | us after all. */ |

2670 | if (incr_vec[i].initializer |

2671 | && !dominated_by_p (CDI_DOMINATORS, |

2672 | gimple_bb (c->cand_stmt), |

2673 | incr_vec[i].init_bb)) |

2674 | { |

2675 | incr_vec[i].initializer = NULL_TREE; |

2676 | incr_vec[i].init_bb = NULL; |

2677 | } |

2678 | |

2679 | break; |

2680 | } |

2681 | } |

2682 | |

2683 | if (!found && incr_vec_len < MAX_INCR_VEC_LEN - 1) |

2684 | { |

2685 | /* The first time we see an increment, create the entry for it. |

2686 | If this is the root candidate which doesn't have a basis, set |

2687 | the count to zero. We're only processing it so it can possibly |

2688 | provide an initializer for other candidates. */ |

2689 | incr_vec[incr_vec_len].incr = increment; |

2690 | incr_vec[incr_vec_len].count = c->basis || is_phi_adjust ? 1 : 0; |

2691 | incr_vec[incr_vec_len].cost = COST_INFINITE; |

2692 | |

2693 | /* Optimistically record the first occurrence of this increment |

2694 | as providing an initializer (if it does); we will revise this |

2695 | opinion later if it doesn't dominate all other occurrences. |

2696 | Exception: increments of 0, 1 never need initializers; |

2697 | and phi adjustments don't ever provide initializers. */ |

2698 | if (c->kind == CAND_ADD |

2699 | && !is_phi_adjust |

2700 | && c->index == increment |

2701 | && (increment > 1 || increment < 0) |

2702 | && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR |

2703 | || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR)) |

2704 | { |

2705 | tree t0 = NULL_TREE; |

2706 | tree rhs1 = gimple_assign_rhs1 (c->cand_stmt); |

2707 | tree rhs2 = gimple_assign_rhs2 (c->cand_stmt); |

2708 | if (operand_equal_p (rhs1, c->base_expr, 0)) |

2709 | t0 = rhs2; |

2710 | else if (operand_equal_p (rhs2, c->base_expr, 0)) |

2711 | t0 = rhs1; |

2712 | if (t0 |

2713 | && SSA_NAME_DEF_STMT (t0) |

2714 | && gimple_bb (SSA_NAME_DEF_STMT (t0))) |

2715 | { |

2716 | incr_vec[incr_vec_len].initializer = t0; |

2717 | incr_vec[incr_vec_len++].init_bb |

2718 | = gimple_bb (SSA_NAME_DEF_STMT (t0)); |

2719 | } |

2720 | else |

2721 | { |

2722 | incr_vec[incr_vec_len].initializer = NULL_TREE; |

2723 | incr_vec[incr_vec_len++].init_bb = NULL; |

2724 | } |

2725 | } |

2726 | else |

2727 | { |

2728 | incr_vec[incr_vec_len].initializer = NULL_TREE; |

2729 | incr_vec[incr_vec_len++].init_bb = NULL; |

2730 | } |

2731 | } |

2732 | } |

2733 | |

2734 | /* Recursive helper function for record_phi_increments. */ |

2735 | |

2736 | static void |

2737 | record_phi_increments_1 (slsr_cand_t basis, gimple *phi) |

2738 | { |

2739 | unsigned i; |

2740 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |

2741 | |

2742 | if (phi_cand->visited) |

2743 | return; |

2744 | phi_cand->visited = 1; |

2745 | |

2746 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

2747 | { |

2748 | tree arg = gimple_phi_arg_def (phi, i); |

2749 | |

2750 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) |

2751 | { |

2752 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

2753 | |

2754 | if (gimple_code (arg_def) == GIMPLE_PHI) |

2755 | record_phi_increments_1 (basis, arg_def); |

2756 | else |

2757 | { |

2758 | slsr_cand_t arg_cand = base_cand_from_table (arg); |

2759 | widest_int diff = arg_cand->index - basis->index; |

2760 | record_increment (arg_cand, diff, PHI_ADJUST); |

2761 | } |

2762 | } |

2763 | } |

2764 | } |

2765 | |

2766 | /* Given phi statement PHI that hides a candidate from its BASIS, find |

2767 | the increments along each incoming arc (recursively handling additional |

2768 | phis that may be present) and record them. These increments are the |

2769 | difference in index between the index-adjusting statements and the |

2770 | index of the basis. */ |

2771 | |

2772 | static void |

2773 | record_phi_increments (slsr_cand_t basis, gimple *phi) |

2774 | { |

2775 | record_phi_increments_1 (basis, phi); |

2776 | clear_visited (as_a <gphi *> (phi)); |

2777 | } |

2778 | |

2779 | /* Determine how many times each unique increment occurs in the set |

2780 | of candidates rooted at C's parent, recording the data in the |

2781 | increment vector. For each unique increment I, if an initializer |

2782 | T_0 = stride * I is provided by a candidate that dominates all |

2783 | candidates with the same increment, also record T_0 for subsequent |

2784 | use. */ |

2785 | |

2786 | static void |

2787 | record_increments (slsr_cand_t c) |

2788 | { |

2789 | if (!cand_already_replaced (c)) |

2790 | { |

2791 | if (!phi_dependent_cand_p (c)) |

2792 | record_increment (c, cand_increment (c), NOT_PHI_ADJUST); |

2793 | else |

2794 | { |

2795 | /* A candidate with a basis hidden by a phi will have one |

2796 | increment for its relationship to the index represented by |

2797 | the phi, and potentially additional increments along each |

2798 | incoming edge. For the root of the dependency tree (which |

2799 | has no basis), process just the initial index in case it has |

2800 | an initializer that can be used by subsequent candidates. */ |

2801 | record_increment (c, c->index, NOT_PHI_ADJUST); |

2802 | |

2803 | if (c->basis) |

2804 | record_phi_increments (lookup_cand (c->basis), |

2805 | lookup_cand (c->def_phi)->cand_stmt); |

2806 | } |

2807 | } |

2808 | |

2809 | if (c->sibling) |

2810 | record_increments (lookup_cand (c->sibling)); |

2811 | |

2812 | if (c->dependent) |

2813 | record_increments (lookup_cand (c->dependent)); |

2814 | } |

2815 | |

2816 | /* Recursive helper function for phi_incr_cost. */ |

2817 | |

2818 | static int |

2819 | phi_incr_cost_1 (slsr_cand_t c, const widest_int &incr, gimple *phi, |

2820 | int *savings) |

2821 | { |

2822 | unsigned i; |

2823 | int cost = 0; |

2824 | slsr_cand_t basis = lookup_cand (c->basis); |

2825 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |

2826 | |

2827 | if (phi_cand->visited) |

2828 | return 0; |

2829 | phi_cand->visited = 1; |

2830 | |

2831 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

2832 | { |

2833 | tree arg = gimple_phi_arg_def (phi, i); |

2834 | |

2835 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) |

2836 | { |

2837 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

2838 | |

2839 | if (gimple_code (arg_def) == GIMPLE_PHI) |

2840 | { |

2841 | int feeding_savings = 0; |

2842 | tree feeding_var = gimple_phi_result (arg_def); |

2843 | cost += phi_incr_cost_1 (c, incr, arg_def, &feeding_savings); |

2844 | if (uses_consumed_by_stmt (feeding_var, phi)) |

2845 | *savings += feeding_savings; |

2846 | } |

2847 | else |

2848 | { |

2849 | slsr_cand_t arg_cand = base_cand_from_table (arg); |

2850 | widest_int diff = arg_cand->index - basis->index; |

2851 | |

2852 | if (incr == diff) |

2853 | { |

2854 | tree basis_lhs = gimple_assign_lhs (basis->cand_stmt); |

2855 | tree lhs = gimple_assign_lhs (arg_cand->cand_stmt); |

2856 | cost += add_cost (true, TYPE_MODE (TREE_TYPE (basis_lhs))); |

2857 | if (uses_consumed_by_stmt (lhs, phi)) |

2858 | *savings += stmt_cost (arg_cand->cand_stmt, true); |

2859 | } |

2860 | } |

2861 | } |

2862 | } |

2863 | |

2864 | return cost; |

2865 | } |

2866 | |

2867 | /* Add up and return the costs of introducing add statements that |

2868 | require the increment INCR on behalf of candidate C and phi |

2869 | statement PHI. Accumulate into *SAVINGS the potential savings |

2870 | from removing existing statements that feed PHI and have no other |

2871 | uses. */ |

2872 | |

2873 | static int |

2874 | phi_incr_cost (slsr_cand_t c, const widest_int &incr, gimple *phi, |

2875 | int *savings) |

2876 | { |

2877 | int retval = phi_incr_cost_1 (c, incr, phi, savings); |

2878 | clear_visited (as_a <gphi *> (phi)); |

2879 | return retval; |

2880 | } |

2881 | |

2882 | /* Return the first candidate in the tree rooted at C that has not |

2883 | already been replaced, favoring siblings over dependents. */ |

2884 | |

2885 | static slsr_cand_t |

2886 | unreplaced_cand_in_tree (slsr_cand_t c) |

2887 | { |

2888 | if (!cand_already_replaced (c)) |

2889 | return c; |

2890 | |

2891 | if (c->sibling) |

2892 | { |

2893 | slsr_cand_t sib = unreplaced_cand_in_tree (lookup_cand (c->sibling)); |

2894 | if (sib) |

2895 | return sib; |

2896 | } |

2897 | |

2898 | if (c->dependent) |

2899 | { |

2900 | slsr_cand_t dep = unreplaced_cand_in_tree (lookup_cand (c->dependent)); |

2901 | if (dep) |

2902 | return dep; |

2903 | } |

2904 | |

2905 | return NULL; |

2906 | } |

2907 | |

2908 | /* Return TRUE if the candidates in the tree rooted at C should be |

2909 | optimized for speed, else FALSE. We estimate this based on the block |

2910 | containing the most dominant candidate in the tree that has not yet |

2911 | been replaced. */ |

2912 | |

2913 | static bool |

2914 | optimize_cands_for_speed_p (slsr_cand_t c) |

2915 | { |

2916 | slsr_cand_t c2 = unreplaced_cand_in_tree (c); |

2917 | gcc_assert (c2); |

2918 | return optimize_bb_for_speed_p (gimple_bb (c2->cand_stmt)); |

2919 | } |

2920 | |

2921 | /* Add COST_IN to the lowest cost of any dependent path starting at |

2922 | candidate C or any of its siblings, counting only candidates along |

2923 | such paths with increment INCR. Assume that replacing a candidate |

2924 | reduces cost by REPL_SAVINGS. Also account for savings from any |

2925 | statements that would go dead. If COUNT_PHIS is true, include |

2926 | costs of introducing feeding statements for conditional candidates. */ |

2927 | |

2928 | static int |

2929 | lowest_cost_path (int cost_in, int repl_savings, slsr_cand_t c, |

2930 | const widest_int &incr, bool count_phis) |

2931 | { |

2932 | int local_cost, sib_cost, savings = 0; |

2933 | widest_int cand_incr = cand_abs_increment (c); |

2934 | |

2935 | if (cand_already_replaced (c)) |

2936 | local_cost = cost_in; |

2937 | else if (incr == cand_incr) |

2938 | local_cost = cost_in - repl_savings - c->dead_savings; |

2939 | else |

2940 | local_cost = cost_in - c->dead_savings; |

2941 | |

2942 | if (count_phis |

2943 | && phi_dependent_cand_p (c) |

2944 | && !cand_already_replaced (c)) |

2945 | { |

2946 | gimple *phi = lookup_cand (c->def_phi)->cand_stmt; |

2947 | local_cost += phi_incr_cost (c, incr, phi, &savings); |

2948 | |

2949 | if (uses_consumed_by_stmt (gimple_phi_result (phi), c->cand_stmt)) |

2950 | local_cost -= savings; |

2951 | } |

2952 | |

2953 | if (c->dependent) |

2954 | local_cost = lowest_cost_path (local_cost, repl_savings, |

2955 | lookup_cand (c->dependent), incr, |

2956 | count_phis); |

2957 | |

2958 | if (c->sibling) |

2959 | { |

2960 | sib_cost = lowest_cost_path (cost_in, repl_savings, |

2961 | lookup_cand (c->sibling), incr, |

2962 | count_phis); |

2963 | local_cost = MIN (local_cost, sib_cost); |

2964 | } |

2965 | |

2966 | return local_cost; |

2967 | } |

2968 | |

2969 | /* Compute the total savings that would accrue from all replacements |

2970 | in the candidate tree rooted at C, counting only candidates with |

2971 | increment INCR. Assume that replacing a candidate reduces cost |

2972 | by REPL_SAVINGS. Also account for savings from statements that |

2973 | would go dead. */ |

2974 | |

2975 | static int |

2976 | total_savings (int repl_savings, slsr_cand_t c, const widest_int &incr, |

2977 | bool count_phis) |

2978 | { |

2979 | int savings = 0; |

2980 | widest_int cand_incr = cand_abs_increment (c); |

2981 | |

2982 | if (incr == cand_incr && !cand_already_replaced (c)) |

2983 | savings += repl_savings + c->dead_savings; |

2984 | |

2985 | if (count_phis |

2986 | && phi_dependent_cand_p (c) |

2987 | && !cand_already_replaced (c)) |

2988 | { |

2989 | int phi_savings = 0; |

2990 | gimple *phi = lookup_cand (c->def_phi)->cand_stmt; |

2991 | savings -= phi_incr_cost (c, incr, phi, &phi_savings); |

2992 | |

2993 | if (uses_consumed_by_stmt (gimple_phi_result (phi), c->cand_stmt)) |

2994 | savings += phi_savings; |

2995 | } |

2996 | |

2997 | if (c->dependent) |

2998 | savings += total_savings (repl_savings, lookup_cand (c->dependent), incr, |

2999 | count_phis); |

3000 | |

3001 | if (c->sibling) |

3002 | savings += total_savings (repl_savings, lookup_cand (c->sibling), incr, |

3003 | count_phis); |

3004 | |

3005 | return savings; |

3006 | } |

3007 | |

3008 | /* Use target-specific costs to determine and record which increments |

3009 | in the current candidate tree are profitable to replace, assuming |

3010 | MODE and SPEED. FIRST_DEP is the first dependent of the root of |

3011 | the candidate tree. |

3012 | |

3013 | One slight limitation here is that we don't account for the possible |

3014 | introduction of casts in some cases. See replace_one_candidate for |

3015 | the cases where these are introduced. This should probably be cleaned |

3016 | up sometime. */ |

3017 | |

3018 | static void |

3019 | analyze_increments (slsr_cand_t first_dep, machine_mode mode, bool speed) |

3020 | { |

3021 | unsigned i; |

3022 | |

3023 | for (i = 0; i < incr_vec_len; i++) |

3024 | { |

3025 | HOST_WIDE_INT incr = incr_vec[i].incr.to_shwi (); |

3026 | |

3027 | /* If somehow this increment is bigger than a HWI, we won't |

3028 | be optimizing candidates that use it. And if the increment |

3029 | has a count of zero, nothing will be done with it. */ |

3030 | if (!wi::fits_shwi_p (incr_vec[i].incr) || !incr_vec[i].count) |

3031 | incr_vec[i].cost = COST_INFINITE; |

3032 | |

3033 | /* Increments of 0, 1, and -1 are always profitable to replace, |

3034 | because they always replace a multiply or add with an add or |

3035 | copy, and may cause one or more existing instructions to go |

3036 | dead. Exception: -1 can't be assumed to be profitable for |

3037 | pointer addition. */ |

3038 | else if (incr == 0 |

3039 | || incr == 1 |

3040 | || (incr == -1 |

3041 | && !POINTER_TYPE_P (first_dep->cand_type))) |

3042 | incr_vec[i].cost = COST_NEUTRAL; |

3043 | |

3044 | /* If we need to add an initializer, give up if a cast from the |

3045 | candidate's type to its stride's type can lose precision. |

3046 | Note that this already takes into account that the stride may |

3047 | have been cast to a wider type, in which case this test won't |

3048 | fire. Example: |

3049 | |

3050 | short int _1; |

3051 | _2 = (int) _1; |

3052 | _3 = _2 * 10; |

3053 | _4 = x + _3; ADD: x + (10 * (int)_1) : int |

3054 | _5 = _2 * 15; |

3055 | _6 = x + _5; ADD: x + (15 * (int)_1) : int |

3056 | |

3057 | Although the stride was a short int initially, the stride |

3058 | used in the analysis has been widened to an int, and such |

3059 | widening will be done in the initializer as well. */ |

3060 | else if (!incr_vec[i].initializer |

3061 | && TREE_CODE (first_dep->stride) != INTEGER_CST |

3062 | && !legal_cast_p_1 (first_dep->stride_type, |

3063 | TREE_TYPE (gimple_assign_lhs |

3064 | (first_dep->cand_stmt)))) |

3065 | incr_vec[i].cost = COST_INFINITE; |

3066 | |

3067 | /* If we need to add an initializer, make sure we don't introduce |

3068 | a multiply by a pointer type, which can happen in certain cast |

3069 | scenarios. */ |

3070 | else if (!incr_vec[i].initializer |

3071 | && TREE_CODE (first_dep->stride) != INTEGER_CST |

3072 | && POINTER_TYPE_P (first_dep->stride_type)) |

3073 | incr_vec[i].cost = COST_INFINITE; |

3074 | |

3075 | /* For any other increment, if this is a multiply candidate, we |

3076 | must introduce a temporary T and initialize it with |

3077 | T_0 = stride * increment. When optimizing for speed, walk the |

3078 | candidate tree to calculate the best cost reduction along any |

3079 | path; if it offsets the fixed cost of inserting the initializer, |

3080 | replacing the increment is profitable. When optimizing for |

3081 | size, instead calculate the total cost reduction from replacing |

3082 | all candidates with this increment. */ |

3083 | else if (first_dep->kind == CAND_MULT) |

3084 | { |

3085 | int cost = mult_by_coeff_cost (incr, mode, speed); |

3086 | int repl_savings = mul_cost (speed, mode) - add_cost (speed, mode); |

3087 | if (speed) |

3088 | cost = lowest_cost_path (cost, repl_savings, first_dep, |

3089 | incr_vec[i].incr, COUNT_PHIS); |

3090 | else |

3091 | cost -= total_savings (repl_savings, first_dep, incr_vec[i].incr, |

3092 | COUNT_PHIS); |

3093 | |

3094 | incr_vec[i].cost = cost; |

3095 | } |

3096 | |

3097 | /* If this is an add candidate, the initializer may already |

3098 | exist, so only calculate the cost of the initializer if it |

3099 | doesn't. We are replacing one add with another here, so the |

3100 | known replacement savings is zero. We will account for removal |

3101 | of dead instructions in lowest_cost_path or total_savings. */ |

3102 | else |

3103 | { |

3104 | int cost = 0; |

3105 | if (!incr_vec[i].initializer) |

3106 | cost = mult_by_coeff_cost (incr, mode, speed); |

3107 | |

3108 | if (speed) |

3109 | cost = lowest_cost_path (cost, 0, first_dep, incr_vec[i].incr, |

3110 | DONT_COUNT_PHIS); |

3111 | else |

3112 | cost -= total_savings (0, first_dep, incr_vec[i].incr, |

3113 | DONT_COUNT_PHIS); |

3114 | |

3115 | incr_vec[i].cost = cost; |

3116 | } |

3117 | } |

3118 | } |

3119 | |

3120 | /* Return the nearest common dominator of BB1 and BB2. If the blocks |

3121 | are identical, return the earlier of C1 and C2 in *WHERE. Otherwise, |

3122 | if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2, |

3123 | return C2 in *WHERE; and if the NCD matches neither, return NULL in |

3124 | *WHERE. Note: It is possible for one of C1 and C2 to be NULL. */ |

3125 | |

3126 | static basic_block |

3127 | ncd_for_two_cands (basic_block bb1, basic_block bb2, |

3128 | slsr_cand_t c1, slsr_cand_t c2, slsr_cand_t *where) |

3129 | { |

3130 | basic_block ncd; |

3131 | |

3132 | if (!bb1) |

3133 | { |

3134 | *where = c2; |

3135 | return bb2; |

3136 | } |

3137 | |

3138 | if (!bb2) |

3139 | { |

3140 | *where = c1; |

3141 | return bb1; |

3142 | } |

3143 | |

3144 | ncd = nearest_common_dominator (CDI_DOMINATORS, bb1, bb2); |

3145 | |

3146 | /* If both candidates are in the same block, the earlier |

3147 | candidate wins. */ |

3148 | if (bb1 == ncd && bb2 == ncd) |

3149 | { |

3150 | if (!c1 || (c2 && c2->cand_num < c1->cand_num)) |

3151 | *where = c2; |

3152 | else |

3153 | *where = c1; |

3154 | } |

3155 | |

3156 | /* Otherwise, if one of them produced a candidate in the |

3157 | dominator, that one wins. */ |

3158 | else if (bb1 == ncd) |

3159 | *where = c1; |

3160 | |

3161 | else if (bb2 == ncd) |

3162 | *where = c2; |

3163 | |

3164 | /* If neither matches the dominator, neither wins. */ |

3165 | else |

3166 | *where = NULL; |

3167 | |

3168 | return ncd; |

3169 | } |

3170 | |

3171 | /* Consider all candidates that feed PHI. Find the nearest common |

3172 | dominator of those candidates requiring the given increment INCR. |

3173 | Further find and return the nearest common dominator of this result |

3174 | with block NCD. If the returned block contains one or more of the |

3175 | candidates, return the earliest candidate in the block in *WHERE. */ |

3176 | |

3177 | static basic_block |

3178 | ncd_with_phi (slsr_cand_t c, const widest_int &incr, gphi *phi, |

3179 | basic_block ncd, slsr_cand_t *where) |

3180 | { |

3181 | unsigned i; |

3182 | slsr_cand_t basis = lookup_cand (c->basis); |

3183 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |

3184 | |

3185 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

3186 | { |

3187 | tree arg = gimple_phi_arg_def (phi, i); |

3188 | |

3189 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) |

3190 | { |

3191 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

3192 | |

3193 | if (gimple_code (arg_def) == GIMPLE_PHI) |

3194 | ncd = ncd_with_phi (c, incr, as_a <gphi *> (arg_def), ncd, |

3195 | where); |

3196 | else |

3197 | { |

3198 | slsr_cand_t arg_cand = base_cand_from_table (arg); |

3199 | widest_int diff = arg_cand->index - basis->index; |

3200 | basic_block pred = gimple_phi_arg_edge (phi, i)->src; |

3201 | |

3202 | if ((incr == diff) || (!address_arithmetic_p && incr == -diff)) |

3203 | ncd = ncd_for_two_cands (ncd, pred, *where, NULL, where); |

3204 | } |

3205 | } |

3206 | } |

3207 | |

3208 | return ncd; |

3209 | } |

3210 | |

3211 | /* Consider the candidate C together with any candidates that feed |

3212 | C's phi dependence (if any). Find and return the nearest common |

3213 | dominator of those candidates requiring the given increment INCR. |

3214 | If the returned block contains one or more of the candidates, |

3215 | return the earliest candidate in the block in *WHERE. */ |

3216 | |

3217 | static basic_block |

3218 | ncd_of_cand_and_phis (slsr_cand_t c, const widest_int &incr, slsr_cand_t *where) |

3219 | { |

3220 | basic_block ncd = NULL; |

3221 | |

3222 | if (cand_abs_increment (c) == incr) |

3223 | { |

3224 | ncd = gimple_bb (c->cand_stmt); |

3225 | *where = c; |

3226 | } |

3227 | |

3228 | if (phi_dependent_cand_p (c)) |

3229 | ncd = ncd_with_phi (c, incr, |

3230 | as_a <gphi *> (lookup_cand (c->def_phi)->cand_stmt), |

3231 | ncd, where); |

3232 | |

3233 | return ncd; |

3234 | } |

3235 | |

3236 | /* Consider all candidates in the tree rooted at C for which INCR |

3237 | represents the required increment of C relative to its basis. |

3238 | Find and return the basic block that most nearly dominates all |

3239 | such candidates. If the returned block contains one or more of |

3240 | the candidates, return the earliest candidate in the block in |

3241 | *WHERE. */ |

3242 | |

3243 | static basic_block |

3244 | nearest_common_dominator_for_cands (slsr_cand_t c, const widest_int &incr, |

3245 | slsr_cand_t *where) |

3246 | { |

3247 | basic_block sib_ncd = NULL, dep_ncd = NULL, this_ncd = NULL, ncd; |

3248 | slsr_cand_t sib_where = NULL, dep_where = NULL, this_where = NULL, new_where; |

3249 | |

3250 | /* First find the NCD of all siblings and dependents. */ |

3251 | if (c->sibling) |

3252 | sib_ncd = nearest_common_dominator_for_cands (lookup_cand (c->sibling), |

3253 | incr, &sib_where); |

3254 | if (c->dependent) |

3255 | dep_ncd = nearest_common_dominator_for_cands (lookup_cand (c->dependent), |

3256 | incr, &dep_where); |

3257 | if (!sib_ncd && !dep_ncd) |

3258 | { |

3259 | new_where = NULL; |

3260 | ncd = NULL; |

3261 | } |

3262 | else if (sib_ncd && !dep_ncd) |

3263 | { |

3264 | new_where = sib_where; |

3265 | ncd = sib_ncd; |

3266 | } |

3267 | else if (dep_ncd && !sib_ncd) |

3268 | { |

3269 | new_where = dep_where; |

3270 | ncd = dep_ncd; |

3271 | } |

3272 | else |

3273 | ncd = ncd_for_two_cands (sib_ncd, dep_ncd, sib_where, |

3274 | dep_where, &new_where); |

3275 | |

3276 | /* If the candidate's increment doesn't match the one we're interested |

3277 | in (and nor do any increments for feeding defs of a phi-dependence), |

3278 | then the result depends only on siblings and dependents. */ |

3279 | this_ncd = ncd_of_cand_and_phis (c, incr, &this_where); |

3280 | |

3281 | if (!this_ncd || cand_already_replaced (c)) |

3282 | { |

3283 | *where = new_where; |

3284 | return ncd; |

3285 | } |

3286 | |

3287 | /* Otherwise, compare this candidate with the result from all siblings |

3288 | and dependents. */ |

3289 | ncd = ncd_for_two_cands (ncd, this_ncd, new_where, this_where, where); |

3290 | |

3291 | return ncd; |

3292 | } |

3293 | |

3294 | /* Return TRUE if the increment indexed by INDEX is profitable to replace. */ |

3295 | |

3296 | static inline bool |

3297 | profitable_increment_p (unsigned index) |

3298 | { |

3299 | return (incr_vec[index].cost <= COST_NEUTRAL); |

3300 | } |

3301 | |

3302 | /* For each profitable increment in the increment vector not equal to |

3303 | 0 or 1 (or -1, for non-pointer arithmetic), find the nearest common |

3304 | dominator of all statements in the candidate chain rooted at C |

3305 | that require that increment, and insert an initializer |

3306 | T_0 = stride * increment at that location. Record T_0 with the |

3307 | increment record. */ |

3308 | |

3309 | static void |

3310 | insert_initializers (slsr_cand_t c) |

3311 | { |

3312 | unsigned i; |

3313 | |

3314 | for (i = 0; i < incr_vec_len; i++) |

3315 | { |

3316 | basic_block bb; |

3317 | slsr_cand_t where = NULL; |

3318 | gassign *init_stmt; |

3319 | gassign *cast_stmt = NULL; |

3320 | tree new_name, incr_tree, init_stride; |

3321 | widest_int incr = incr_vec[i].incr; |

3322 | |

3323 | if (!profitable_increment_p (i) |

3324 | || incr == 1 |

3325 | || (incr == -1 |

3326 | && (!POINTER_TYPE_P (lookup_cand (c->basis)->cand_type))) |

3327 | || incr == 0) |

3328 | continue; |

3329 | |

3330 | /* We may have already identified an existing initializer that |

3331 | will suffice. */ |

3332 | if (incr_vec[i].initializer) |

3333 | { |

3334 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3335 | { |

3336 | fputs ("Using existing initializer: ", dump_file); |

3337 | print_gimple_stmt (dump_file, |

3338 | SSA_NAME_DEF_STMT (incr_vec[i].initializer), |

3339 | 0, 0); |

3340 | } |

3341 | continue; |

3342 | } |

3343 | |

3344 | /* Find the block that most closely dominates all candidates |

3345 | with this increment. If there is at least one candidate in |

3346 | that block, the earliest one will be returned in WHERE. */ |

3347 | bb = nearest_common_dominator_for_cands (c, incr, &where); |

3348 | |

3349 | /* If the NCD is not dominated by the block containing the |

3350 | definition of the stride, we can't legally insert a |

3351 | single initializer. Mark the increment as unprofitable |

3352 | so we don't make any replacements. FIXME: Multiple |

3353 | initializers could be placed with more analysis. */ |

3354 | gimple *stride_def = SSA_NAME_DEF_STMT (c->stride); |

3355 | basic_block stride_bb = gimple_bb (stride_def); |

3356 | |

3357 | if (stride_bb && !dominated_by_p (CDI_DOMINATORS, bb, stride_bb)) |

3358 | { |

3359 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3360 | fprintf (dump_file, |

3361 | "Initializer #%d cannot be legally placed\n", i); |

3362 | incr_vec[i].cost = COST_INFINITE; |

3363 | continue; |

3364 | } |

3365 | |

3366 | /* If the nominal stride has a different type than the recorded |

3367 | stride type, build a cast from the nominal stride to that type. */ |

3368 | if (!types_compatible_p (TREE_TYPE (c->stride), c->stride_type)) |

3369 | { |

3370 | init_stride = make_temp_ssa_name (c->stride_type, NULL, "slsr"); |

3371 | cast_stmt = gimple_build_assign (init_stride, NOP_EXPR, c->stride); |

3372 | } |

3373 | else |

3374 | init_stride = c->stride; |

3375 | |

3376 | /* Create a new SSA name to hold the initializer's value. */ |

3377 | new_name = make_temp_ssa_name (c->stride_type, NULL, "slsr"); |

3378 | incr_vec[i].initializer = new_name; |

3379 | |

3380 | /* Create the initializer and insert it in the latest possible |

3381 | dominating position. */ |

3382 | incr_tree = wide_int_to_tree (c->stride_type, incr); |

3383 | init_stmt = gimple_build_assign (new_name, MULT_EXPR, |

3384 | init_stride, incr_tree); |

3385 | if (where) |

3386 | { |

3387 | gimple_stmt_iterator gsi = gsi_for_stmt (where->cand_stmt); |

3388 | location_t loc = gimple_location (where->cand_stmt); |

3389 | |

3390 | if (cast_stmt) |

3391 | { |

3392 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); |

3393 | gimple_set_location (cast_stmt, loc); |

3394 | } |

3395 | |

3396 | gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT); |

3397 | gimple_set_location (init_stmt, loc); |

3398 | } |

3399 | else |

3400 | { |

3401 | gimple_stmt_iterator gsi = gsi_last_bb (bb); |

3402 | gimple *basis_stmt = lookup_cand (c->basis)->cand_stmt; |

3403 | location_t loc = gimple_location (basis_stmt); |

3404 | |

3405 | if (!gsi_end_p (gsi) && stmt_ends_bb_p (gsi_stmt (gsi))) |

3406 | { |

3407 | if (cast_stmt) |

3408 | { |

3409 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); |

3410 | gimple_set_location (cast_stmt, loc); |

3411 | } |

3412 | gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT); |

3413 | } |

3414 | else |

3415 | { |

3416 | if (cast_stmt) |

3417 | { |

3418 | gsi_insert_after (&gsi, cast_stmt, GSI_NEW_STMT); |

3419 | gimple_set_location (cast_stmt, loc); |

3420 | } |

3421 | gsi_insert_after (&gsi, init_stmt, GSI_NEW_STMT); |

3422 | } |

3423 | |

3424 | gimple_set_location (init_stmt, gimple_location (basis_stmt)); |

3425 | } |

3426 | |

3427 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3428 | { |

3429 | if (cast_stmt) |

3430 | { |

3431 | fputs ("Inserting stride cast: ", dump_file); |

3432 | print_gimple_stmt (dump_file, cast_stmt, 0); |

3433 | } |

3434 | fputs ("Inserting initializer: ", dump_file); |

3435 | print_gimple_stmt (dump_file, init_stmt, 0); |

3436 | } |

3437 | } |

3438 | } |

3439 | |

3440 | /* Recursive helper function for all_phi_incrs_profitable. */ |

3441 | |

3442 | static bool |

3443 | all_phi_incrs_profitable_1 (slsr_cand_t c, gphi *phi, int *spread) |

3444 | { |

3445 | unsigned i; |

3446 | slsr_cand_t basis = lookup_cand (c->basis); |

3447 | slsr_cand_t phi_cand = *stmt_cand_map->get (phi); |

3448 | |

3449 | if (phi_cand->visited) |

3450 | return true; |

3451 | |

3452 | phi_cand->visited = 1; |

3453 | (*spread)++; |

3454 | |

3455 | /* If the basis doesn't dominate the PHI (including when the PHI is |

3456 | in the same block as the basis), we won't be able to create a PHI |

3457 | using the basis here. */ |

3458 | basic_block basis_bb = gimple_bb (basis->cand_stmt); |

3459 | basic_block phi_bb = gimple_bb (phi); |

3460 | |

3461 | if (phi_bb == basis_bb |

3462 | || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb)) |

3463 | return false; |

3464 | |

3465 | for (i = 0; i < gimple_phi_num_args (phi); i++) |

3466 | { |

3467 | /* If the PHI arg resides in a block not dominated by the basis, |

3468 | we won't be able to create a PHI using the basis here. */ |

3469 | basic_block pred_bb = gimple_phi_arg_edge (phi, i)->src; |

3470 | |

3471 | if (!dominated_by_p (CDI_DOMINATORS, pred_bb, basis_bb)) |

3472 | return false; |

3473 | |

3474 | tree arg = gimple_phi_arg_def (phi, i); |

3475 | |

3476 | if (!operand_equal_p (arg, phi_cand->base_expr, 0)) |

3477 | { |

3478 | gimple *arg_def = SSA_NAME_DEF_STMT (arg); |

3479 | |

3480 | if (gimple_code (arg_def) == GIMPLE_PHI) |

3481 | { |

3482 | if (!all_phi_incrs_profitable_1 (c, as_a <gphi *> (arg_def), |

3483 | spread) |

3484 | || *spread > MAX_SPREAD) |

3485 | return false; |

3486 | } |

3487 | else |

3488 | { |

3489 | int j; |

3490 | slsr_cand_t arg_cand = base_cand_from_table (arg); |

3491 | widest_int increment = arg_cand->index - basis->index; |

3492 | |

3493 | if (!address_arithmetic_p && wi::neg_p (increment)) |

3494 | increment = -increment; |

3495 | |

3496 | j = incr_vec_index (increment); |

3497 | |

3498 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3499 | { |

3500 | fprintf (dump_file, " Conditional candidate %d, phi: ", |

3501 | c->cand_num); |

3502 | print_gimple_stmt (dump_file, phi, 0); |

3503 | fputs (" increment: ", dump_file); |

3504 | print_decs (increment, dump_file); |

3505 | if (j < 0) |

3506 | fprintf (dump_file, |

3507 | "\n Not replaced; incr_vec overflow.\n"); |

3508 | else { |

3509 | fprintf (dump_file, "\n cost: %d\n", incr_vec[j].cost); |

3510 | if (profitable_increment_p (j)) |

3511 | fputs (" Replacing...\n", dump_file); |

3512 | else |

3513 | fputs (" Not replaced.\n", dump_file); |

3514 | } |

3515 | } |

3516 | |

3517 | if (j < 0 || !profitable_increment_p (j)) |

3518 | return false; |

3519 | } |

3520 | } |

3521 | } |

3522 | |

3523 | return true; |

3524 | } |

3525 | |

3526 | /* Return TRUE iff all required increments for candidates feeding PHI |

3527 | are profitable (and legal!) to replace on behalf of candidate C. */ |

3528 | |

3529 | static bool |

3530 | all_phi_incrs_profitable (slsr_cand_t c, gphi *phi) |

3531 | { |

3532 | int spread = 0; |

3533 | bool retval = all_phi_incrs_profitable_1 (c, phi, &spread); |

3534 | clear_visited (phi); |

3535 | return retval; |

3536 | } |

3537 | |

3538 | /* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of |

3539 | type TO_TYPE, and insert it in front of the statement represented |

3540 | by candidate C. Use *NEW_VAR to create the new SSA name. Return |

3541 | the new SSA name. */ |

3542 | |

3543 | static tree |

3544 | introduce_cast_before_cand (slsr_cand_t c, tree to_type, tree from_expr) |

3545 | { |

3546 | tree cast_lhs; |

3547 | gassign *cast_stmt; |

3548 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

3549 | |

3550 | cast_lhs = make_temp_ssa_name (to_type, NULL, "slsr"); |

3551 | cast_stmt = gimple_build_assign (cast_lhs, NOP_EXPR, from_expr); |

3552 | gimple_set_location (cast_stmt, gimple_location (c->cand_stmt)); |

3553 | gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT); |

3554 | |

3555 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3556 | { |

3557 | fputs (" Inserting: ", dump_file); |

3558 | print_gimple_stmt (dump_file, cast_stmt, 0); |

3559 | } |

3560 | |

3561 | return cast_lhs; |

3562 | } |

3563 | |

3564 | /* Replace the RHS of the statement represented by candidate C with |

3565 | NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't |

3566 | leave C unchanged or just interchange its operands. The original |

3567 | operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2. |

3568 | If the replacement was made and we are doing a details dump, |

3569 | return the revised statement, else NULL. */ |

3570 | |

3571 | static gimple * |

3572 | replace_rhs_if_not_dup (enum tree_code new_code, tree new_rhs1, tree new_rhs2, |

3573 | enum tree_code old_code, tree old_rhs1, tree old_rhs2, |

3574 | slsr_cand_t c) |

3575 | { |

3576 | if (new_code != old_code |

3577 | || ((!operand_equal_p (new_rhs1, old_rhs1, 0) |

3578 | || !operand_equal_p (new_rhs2, old_rhs2, 0)) |

3579 | && (!operand_equal_p (new_rhs1, old_rhs2, 0) |

3580 | || !operand_equal_p (new_rhs2, old_rhs1, 0)))) |

3581 | { |

3582 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

3583 | slsr_cand_t cc = c; |

3584 | gimple_assign_set_rhs_with_ops (&gsi, new_code, new_rhs1, new_rhs2); |

3585 | update_stmt (gsi_stmt (gsi)); |

3586 | c->cand_stmt = gsi_stmt (gsi); |

3587 | while (cc->next_interp) |

3588 | { |

3589 | cc = lookup_cand (cc->next_interp); |

3590 | cc->cand_stmt = gsi_stmt (gsi); |

3591 | } |

3592 | |

3593 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3594 | return gsi_stmt (gsi); |

3595 | } |

3596 | |

3597 | else if (dump_file && (dump_flags & TDF_DETAILS)) |

3598 | fputs (" (duplicate, not actually replacing)\n", dump_file); |

3599 | |

3600 | return NULL; |

3601 | } |

3602 | |

3603 | /* Strength-reduce the statement represented by candidate C by replacing |

3604 | it with an equivalent addition or subtraction. I is the index into |

3605 | the increment vector identifying C's increment. NEW_VAR is used to |

3606 | create a new SSA name if a cast needs to be introduced. BASIS_NAME |

3607 | is the rhs1 to use in creating the add/subtract. */ |

3608 | |

3609 | static void |

3610 | replace_one_candidate (slsr_cand_t c, unsigned i, tree basis_name) |

3611 | { |

3612 | gimple *stmt_to_print = NULL; |

3613 | tree orig_rhs1, orig_rhs2; |

3614 | tree rhs2; |

3615 | enum tree_code orig_code, repl_code; |

3616 | widest_int cand_incr; |

3617 | |

3618 | orig_code = gimple_assign_rhs_code (c->cand_stmt); |

3619 | orig_rhs1 = gimple_assign_rhs1 (c->cand_stmt); |

3620 | orig_rhs2 = gimple_assign_rhs2 (c->cand_stmt); |

3621 | cand_incr = cand_increment (c); |

3622 | |

3623 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3624 | { |

3625 | fputs ("Replacing: ", dump_file); |

3626 | print_gimple_stmt (dump_file, c->cand_stmt, 0); |

3627 | stmt_to_print = c->cand_stmt; |

3628 | } |

3629 | |

3630 | if (address_arithmetic_p) |

3631 | repl_code = POINTER_PLUS_EXPR; |

3632 | else |

3633 | repl_code = PLUS_EXPR; |

3634 | |

3635 | /* If the increment has an initializer T_0, replace the candidate |

3636 | statement with an add of the basis name and the initializer. */ |

3637 | if (incr_vec[i].initializer) |

3638 | { |

3639 | tree init_type = TREE_TYPE (incr_vec[i].initializer); |

3640 | tree orig_type = TREE_TYPE (orig_rhs2); |

3641 | |

3642 | if (types_compatible_p (orig_type, init_type)) |

3643 | rhs2 = incr_vec[i].initializer; |

3644 | else |

3645 | rhs2 = introduce_cast_before_cand (c, orig_type, |

3646 | incr_vec[i].initializer); |

3647 | |

3648 | if (incr_vec[i].incr != cand_incr) |

3649 | { |

3650 | gcc_assert (repl_code == PLUS_EXPR); |

3651 | repl_code = MINUS_EXPR; |

3652 | } |

3653 | |

3654 | stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2, |

3655 | orig_code, orig_rhs1, orig_rhs2, |

3656 | c); |

3657 | } |

3658 | |

3659 | /* Otherwise, the increment is one of -1, 0, and 1. Replace |

3660 | with a subtract of the stride from the basis name, a copy |

3661 | from the basis name, or an add of the stride to the basis |

3662 | name, respectively. It may be necessary to introduce a |

3663 | cast (or reuse an existing cast). */ |

3664 | else if (cand_incr == 1) |

3665 | { |

3666 | tree stride_type = TREE_TYPE (c->stride); |

3667 | tree orig_type = TREE_TYPE (orig_rhs2); |

3668 | |

3669 | if (types_compatible_p (orig_type, stride_type)) |

3670 | rhs2 = c->stride; |

3671 | else |

3672 | rhs2 = introduce_cast_before_cand (c, orig_type, c->stride); |

3673 | |

3674 | stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2, |

3675 | orig_code, orig_rhs1, orig_rhs2, |

3676 | c); |

3677 | } |

3678 | |

3679 | else if (cand_incr == -1) |

3680 | { |

3681 | tree stride_type = TREE_TYPE (c->stride); |

3682 | tree orig_type = TREE_TYPE (orig_rhs2); |

3683 | gcc_assert (repl_code != POINTER_PLUS_EXPR); |

3684 | |

3685 | if (types_compatible_p (orig_type, stride_type)) |

3686 | rhs2 = c->stride; |

3687 | else |

3688 | rhs2 = introduce_cast_before_cand (c, orig_type, c->stride); |

3689 | |

3690 | if (orig_code != MINUS_EXPR |

3691 | || !operand_equal_p (basis_name, orig_rhs1, 0) |

3692 | || !operand_equal_p (rhs2, orig_rhs2, 0)) |

3693 | { |

3694 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

3695 | slsr_cand_t cc = c; |

3696 | gimple_assign_set_rhs_with_ops (&gsi, MINUS_EXPR, basis_name, rhs2); |

3697 | update_stmt (gsi_stmt (gsi)); |

3698 | c->cand_stmt = gsi_stmt (gsi); |

3699 | while (cc->next_interp) |

3700 | { |

3701 | cc = lookup_cand (cc->next_interp); |

3702 | cc->cand_stmt = gsi_stmt (gsi); |

3703 | } |

3704 | |

3705 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3706 | stmt_to_print = gsi_stmt (gsi); |

3707 | } |

3708 | else if (dump_file && (dump_flags & TDF_DETAILS)) |

3709 | fputs (" (duplicate, not actually replacing)\n", dump_file); |

3710 | } |

3711 | |

3712 | else if (cand_incr == 0) |

3713 | { |

3714 | tree lhs = gimple_assign_lhs (c->cand_stmt); |

3715 | tree lhs_type = TREE_TYPE (lhs); |

3716 | tree basis_type = TREE_TYPE (basis_name); |

3717 | |

3718 | if (types_compatible_p (lhs_type, basis_type)) |

3719 | { |

3720 | gassign *copy_stmt = gimple_build_assign (lhs, basis_name); |

3721 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

3722 | slsr_cand_t cc = c; |

3723 | gimple_set_location (copy_stmt, gimple_location (c->cand_stmt)); |

3724 | gsi_replace (&gsi, copy_stmt, false); |

3725 | c->cand_stmt = copy_stmt; |

3726 | while (cc->next_interp) |

3727 | { |

3728 | cc = lookup_cand (cc->next_interp); |

3729 | cc->cand_stmt = copy_stmt; |

3730 | } |

3731 | |

3732 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3733 | stmt_to_print = copy_stmt; |

3734 | } |

3735 | else |

3736 | { |

3737 | gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt); |

3738 | gassign *cast_stmt = gimple_build_assign (lhs, NOP_EXPR, basis_name); |

3739 | slsr_cand_t cc = c; |

3740 | gimple_set_location (cast_stmt, gimple_location (c->cand_stmt)); |

3741 | gsi_replace (&gsi, cast_stmt, false); |

3742 | c->cand_stmt = cast_stmt; |

3743 | while (cc->next_interp) |

3744 | { |

3745 | cc = lookup_cand (cc->next_interp); |

3746 | cc->cand_stmt = cast_stmt; |

3747 | } |

3748 | |

3749 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3750 | stmt_to_print = cast_stmt; |

3751 | } |

3752 | } |

3753 | else |

3754 | gcc_unreachable (); |

3755 | |

3756 | if (dump_file && (dump_flags & TDF_DETAILS) && stmt_to_print) |

3757 | { |

3758 | fputs ("With: ", dump_file); |

3759 | print_gimple_stmt (dump_file, stmt_to_print, 0); |

3760 | fputs ("\n", dump_file); |

3761 | } |

3762 | } |

3763 | |

3764 | /* For each candidate in the tree rooted at C, replace it with |

3765 | an increment if such has been shown to be profitable. */ |

3766 | |

3767 | static void |

3768 | replace_profitable_candidates (slsr_cand_t c) |

3769 | { |

3770 | if (!cand_already_replaced (c)) |

3771 | { |

3772 | widest_int increment = cand_abs_increment (c); |

3773 | enum tree_code orig_code = gimple_assign_rhs_code (c->cand_stmt); |

3774 | int i; |

3775 | |

3776 | i = incr_vec_index (increment); |

3777 | |

3778 | /* Only process profitable increments. Nothing useful can be done |

3779 | to a cast or copy. */ |

3780 | if (i >= 0 |

3781 | && profitable_increment_p (i) |

3782 | && orig_code != SSA_NAME |

3783 | && !CONVERT_EXPR_CODE_P (orig_code)) |

3784 | { |

3785 | if (phi_dependent_cand_p (c)) |

3786 | { |

3787 | gphi *phi = as_a <gphi *> (lookup_cand (c->def_phi)->cand_stmt); |

3788 | |

3789 | if (all_phi_incrs_profitable (c, phi)) |

3790 | { |

3791 | /* Look up the LHS SSA name from C's basis. This will be |

3792 | the RHS1 of the adds we will introduce to create new |

3793 | phi arguments. */ |

3794 | slsr_cand_t basis = lookup_cand (c->basis); |

3795 | tree basis_name = gimple_assign_lhs (basis->cand_stmt); |

3796 | |

3797 | /* Create a new phi statement that will represent C's true |

3798 | basis after the transformation is complete. */ |

3799 | location_t loc = gimple_location (c->cand_stmt); |

3800 | tree name = create_phi_basis (c, phi, basis_name, |

3801 | loc, UNKNOWN_STRIDE); |

3802 | |

3803 | /* Replace C with an add of the new basis phi and the |

3804 | increment. */ |

3805 | replace_one_candidate (c, i, name); |

3806 | } |

3807 | } |

3808 | else |

3809 | { |

3810 | slsr_cand_t basis = lookup_cand (c->basis); |

3811 | tree basis_name = gimple_assign_lhs (basis->cand_stmt); |

3812 | replace_one_candidate (c, i, basis_name); |

3813 | } |

3814 | } |

3815 | } |

3816 | |

3817 | if (c->sibling) |

3818 | replace_profitable_candidates (lookup_cand (c->sibling)); |

3819 | |

3820 | if (c->dependent) |

3821 | replace_profitable_candidates (lookup_cand (c->dependent)); |

3822 | } |

3823 | |

3824 | /* Analyze costs of related candidates in the candidate vector, |

3825 | and make beneficial replacements. */ |

3826 | |

3827 | static void |

3828 | analyze_candidates_and_replace (void) |

3829 | { |

3830 | unsigned i; |

3831 | slsr_cand_t c; |

3832 | |

3833 | /* Each candidate that has a null basis and a non-null |

3834 | dependent is the root of a tree of related statements. |

3835 | Analyze each tree to determine a subset of those |

3836 | statements that can be replaced with maximum benefit. */ |

3837 | FOR_EACH_VEC_ELT (cand_vec, i, c) |

3838 | { |

3839 | slsr_cand_t first_dep; |

3840 | |

3841 | if (c->basis != 0 || c->dependent == 0) |

3842 | continue; |

3843 | |

3844 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3845 | fprintf (dump_file, "\nProcessing dependency tree rooted at %d.\n", |

3846 | c->cand_num); |

3847 | |

3848 | first_dep = lookup_cand (c->dependent); |

3849 | |

3850 | /* If this is a chain of CAND_REFs, unconditionally replace |

3851 | each of them with a strength-reduced data reference. */ |

3852 | if (c->kind == CAND_REF) |

3853 | replace_refs (c); |

3854 | |

3855 | /* If the common stride of all related candidates is a known |

3856 | constant, each candidate without a phi-dependence can be |

3857 | profitably replaced. Each replaces a multiply by a single |

3858 | add, with the possibility that a feeding add also goes dead. |

3859 | A candidate with a phi-dependence is replaced only if the |

3860 | compensation code it requires is offset by the strength |

3861 | reduction savings. */ |

3862 | else if (TREE_CODE (c->stride) == INTEGER_CST) |

3863 | replace_uncond_cands_and_profitable_phis (first_dep); |

3864 | |

3865 | /* When the stride is an SSA name, it may still be profitable |

3866 | to replace some or all of the dependent candidates, depending |

3867 | on whether the introduced increments can be reused, or are |

3868 | less expensive to calculate than the replaced statements. */ |

3869 | else |

3870 | { |

3871 | machine_mode mode; |

3872 | bool speed; |

3873 | |

3874 | /* Determine whether we'll be generating pointer arithmetic |

3875 | when replacing candidates. */ |

3876 | address_arithmetic_p = (c->kind == CAND_ADD |

3877 | && POINTER_TYPE_P (c->cand_type)); |

3878 | |

3879 | /* If all candidates have already been replaced under other |

3880 | interpretations, nothing remains to be done. */ |

3881 | if (!count_candidates (c)) |

3882 | continue; |

3883 | |

3884 | /* Construct an array of increments for this candidate chain. */ |

3885 | incr_vec = XNEWVEC (incr_info, MAX_INCR_VEC_LEN); |

3886 | incr_vec_len = 0; |

3887 | record_increments (c); |

3888 | |

3889 | /* Determine which increments are profitable to replace. */ |

3890 | mode = TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c->cand_stmt))); |

3891 | speed = optimize_cands_for_speed_p (c); |

3892 | analyze_increments (first_dep, mode, speed); |

3893 | |

3894 | /* Insert initializers of the form T_0 = stride * increment |

3895 | for use in profitable replacements. */ |

3896 | insert_initializers (first_dep); |

3897 | dump_incr_vec (); |

3898 | |

3899 | /* Perform the replacements. */ |

3900 | replace_profitable_candidates (first_dep); |

3901 | free (incr_vec); |

3902 | } |

3903 | } |

3904 | |

3905 | /* For conditional candidates, we may have uncommitted insertions |

3906 | on edges to clean up. */ |

3907 | gsi_commit_edge_inserts (); |

3908 | } |

3909 | |

3910 | namespace { |

3911 | |

3912 | const pass_data pass_data_strength_reduction = |

3913 | { |

3914 | GIMPLE_PASS, /* type */ |

3915 | "slsr", /* name */ |

3916 | OPTGROUP_NONE, /* optinfo_flags */ |

3917 | TV_GIMPLE_SLSR, /* tv_id */ |

3918 | ( PROP_cfg | PROP_ssa ), /* properties_required */ |

3919 | 0, /* properties_provided */ |

3920 | 0, /* properties_destroyed */ |

3921 | 0, /* todo_flags_start */ |

3922 | 0, /* todo_flags_finish */ |

3923 | }; |

3924 | |

3925 | class pass_strength_reduction : public gimple_opt_pass |

3926 | { |

3927 | public: |

3928 | pass_strength_reduction (gcc::context *ctxt) |

3929 | : gimple_opt_pass (pass_data_strength_reduction, ctxt) |

3930 | {} |

3931 | |

3932 | /* opt_pass methods: */ |

3933 | virtual bool gate (function *) { return flag_tree_slsr; } |

3934 | virtual unsigned int execute (function *); |

3935 | |

3936 | }; // class pass_strength_reduction |

3937 | |

3938 | unsigned |

3939 | pass_strength_reduction::execute (function *fun) |

3940 | { |

3941 | /* Create the obstack where candidates will reside. */ |

3942 | gcc_obstack_init (&cand_obstack); |

3943 | |

3944 | /* Allocate the candidate vector. */ |

3945 | cand_vec.create (128); |

3946 | |

3947 | /* Allocate the mapping from statements to candidate indices. */ |

3948 | stmt_cand_map = new hash_map<gimple *, slsr_cand_t>; |

3949 | |

3950 | /* Create the obstack where candidate chains will reside. */ |

3951 | gcc_obstack_init (&chain_obstack); |

3952 | |

3953 | /* Allocate the mapping from base expressions to candidate chains. */ |

3954 | base_cand_map = new hash_table<cand_chain_hasher> (500); |

3955 | |

3956 | /* Allocate the mapping from bases to alternative bases. */ |

3957 | alt_base_map = new hash_map<tree, tree>; |

3958 | |

3959 | /* Initialize the loop optimizer. We need to detect flow across |

3960 | back edges, and this gives us dominator information as well. */ |

3961 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); |

3962 | |

3963 | /* Walk the CFG in predominator order looking for strength reduction |

3964 | candidates. */ |

3965 | find_candidates_dom_walker (CDI_DOMINATORS) |

3966 | .walk (fun->cfg->x_entry_block_ptr); |

3967 | |

3968 | if (dump_file && (dump_flags & TDF_DETAILS)) |

3969 | { |

3970 | dump_cand_vec (); |

3971 | dump_cand_chains (); |

3972 | } |

3973 | |

3974 | delete alt_base_map; |

3975 | free_affine_expand_cache (&name_expansions); |

3976 | |

3977 | /* Analyze costs and make appropriate replacements. */ |

3978 | analyze_candidates_and_replace (); |

3979 | |

3980 | loop_optimizer_finalize (); |

3981 | delete base_cand_map; |

3982 | base_cand_map = NULL; |

3983 | obstack_free (&chain_obstack, NULL); |

3984 | delete stmt_cand_map; |

3985 | cand_vec.release (); |

3986 | obstack_free (&cand_obstack, NULL); |

3987 | |

3988 | return 0; |

3989 | } |

3990 | |

3991 | } // anon namespace |

3992 | |

3993 | gimple_opt_pass * |

3994 | make_pass_strength_reduction (gcc::context *ctxt) |

3995 | { |

3996 | return new pass_strength_reduction (ctxt); |

3997 | } |

3998 |