Bernhard Rosenkraenzer | c83ebe5 | 2012-09-18 21:38:03 +0159 | [diff] [blame] | 1 | /* Loop autoparallelization. |
| 2 | Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011, 2012 |
| 3 | Free Software Foundation, Inc. |
| 4 | Contributed by Sebastian Pop <pop@cri.ensmp.fr> |
| 5 | Zdenek Dvorak <dvorakz@suse.cz> and Razya Ladelsky <razya@il.ibm.com>. |
| 6 | |
| 7 | This file is part of GCC. |
| 8 | |
| 9 | GCC is free software; you can redistribute it and/or modify it under |
| 10 | the terms of the GNU General Public License as published by the Free |
| 11 | Software Foundation; either version 3, or (at your option) any later |
| 12 | version. |
| 13 | |
| 14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
| 15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 17 | for more details. |
| 18 | |
| 19 | You should have received a copy of the GNU General Public License |
| 20 | along with GCC; see the file COPYING3. If not see |
| 21 | <http://www.gnu.org/licenses/>. */ |
| 22 | |
| 23 | #include "config.h" |
| 24 | #include "system.h" |
| 25 | #include "coretypes.h" |
| 26 | #include "tree-flow.h" |
| 27 | #include "cfgloop.h" |
| 28 | #include "tree-data-ref.h" |
| 29 | #include "tree-scalar-evolution.h" |
| 30 | #include "gimple-pretty-print.h" |
| 31 | #include "tree-pass.h" |
| 32 | #include "langhooks.h" |
| 33 | #include "tree-vectorizer.h" |
| 34 | |
| 35 | /* This pass tries to distribute iterations of loops into several threads. |
| 36 | The implementation is straightforward -- for each loop we test whether its |
| 37 | iterations are independent, and if it is the case (and some additional |
| 38 | conditions regarding profitability and correctness are satisfied), we |
| 39 | add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion |
| 40 | machinery do its job. |
| 41 | |
| 42 | The most of the complexity is in bringing the code into shape expected |
| 43 | by the omp expanders: |
| 44 | -- for GIMPLE_OMP_FOR, ensuring that the loop has only one induction |
| 45 | variable and that the exit test is at the start of the loop body |
| 46 | -- for GIMPLE_OMP_PARALLEL, replacing the references to local addressable |
| 47 | variables by accesses through pointers, and breaking up ssa chains |
| 48 | by storing the values incoming to the parallelized loop to a structure |
| 49 | passed to the new function as an argument (something similar is done |
| 50 | in omp gimplification, unfortunately only a small part of the code |
| 51 | can be shared). |
| 52 | |
| 53 | TODO: |
| 54 | -- if there are several parallelizable loops in a function, it may be |
| 55 | possible to generate the threads just once (using synchronization to |
| 56 | ensure that cross-loop dependences are obeyed). |
| 57 | -- handling of common reduction patterns for outer loops. |
| 58 | |
| 59 | More info can also be found at http://gcc.gnu.org/wiki/AutoParInGCC */ |
| 60 | /* |
| 61 | Reduction handling: |
| 62 | currently we use vect_force_simple_reduction() to detect reduction patterns. |
| 63 | The code transformation will be introduced by an example. |
| 64 | |
| 65 | |
| 66 | parloop |
| 67 | { |
| 68 | int sum=1; |
| 69 | |
| 70 | for (i = 0; i < N; i++) |
| 71 | { |
| 72 | x[i] = i + 3; |
| 73 | sum+=x[i]; |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | gimple-like code: |
| 78 | header_bb: |
| 79 | |
| 80 | # sum_29 = PHI <sum_11(5), 1(3)> |
| 81 | # i_28 = PHI <i_12(5), 0(3)> |
| 82 | D.1795_8 = i_28 + 3; |
| 83 | x[i_28] = D.1795_8; |
| 84 | sum_11 = D.1795_8 + sum_29; |
| 85 | i_12 = i_28 + 1; |
| 86 | if (N_6(D) > i_12) |
| 87 | goto header_bb; |
| 88 | |
| 89 | |
| 90 | exit_bb: |
| 91 | |
| 92 | # sum_21 = PHI <sum_11(4)> |
| 93 | printf (&"%d"[0], sum_21); |
| 94 | |
| 95 | |
| 96 | after reduction transformation (only relevant parts): |
| 97 | |
| 98 | parloop |
| 99 | { |
| 100 | |
| 101 | .... |
| 102 | |
| 103 | |
| 104 | # Storing the initial value given by the user. # |
| 105 | |
| 106 | .paral_data_store.32.sum.27 = 1; |
| 107 | |
| 108 | #pragma omp parallel num_threads(4) |
| 109 | |
| 110 | #pragma omp for schedule(static) |
| 111 | |
| 112 | # The neutral element corresponding to the particular |
| 113 | reduction's operation, e.g. 0 for PLUS_EXPR, |
| 114 | 1 for MULT_EXPR, etc. replaces the user's initial value. # |
| 115 | |
| 116 | # sum.27_29 = PHI <sum.27_11, 0> |
| 117 | |
| 118 | sum.27_11 = D.1827_8 + sum.27_29; |
| 119 | |
| 120 | GIMPLE_OMP_CONTINUE |
| 121 | |
| 122 | # Adding this reduction phi is done at create_phi_for_local_result() # |
| 123 | # sum.27_56 = PHI <sum.27_11, 0> |
| 124 | GIMPLE_OMP_RETURN |
| 125 | |
| 126 | # Creating the atomic operation is done at |
| 127 | create_call_for_reduction_1() # |
| 128 | |
| 129 | #pragma omp atomic_load |
| 130 | D.1839_59 = *&.paral_data_load.33_51->reduction.23; |
| 131 | D.1840_60 = sum.27_56 + D.1839_59; |
| 132 | #pragma omp atomic_store (D.1840_60); |
| 133 | |
| 134 | GIMPLE_OMP_RETURN |
| 135 | |
| 136 | # collecting the result after the join of the threads is done at |
| 137 | create_loads_for_reductions(). |
| 138 | The value computed by the threads is loaded from the |
| 139 | shared struct. # |
| 140 | |
| 141 | |
| 142 | .paral_data_load.33_52 = &.paral_data_store.32; |
| 143 | sum_37 = .paral_data_load.33_52->sum.27; |
| 144 | sum_43 = D.1795_41 + sum_37; |
| 145 | |
| 146 | exit bb: |
| 147 | # sum_21 = PHI <sum_43, sum_26> |
| 148 | printf (&"%d"[0], sum_21); |
| 149 | |
| 150 | ... |
| 151 | |
| 152 | } |
| 153 | |
| 154 | */ |
| 155 | |
| 156 | /* Minimal number of iterations of a loop that should be executed in each |
| 157 | thread. */ |
| 158 | #define MIN_PER_THREAD 100 |
| 159 | |
| 160 | /* Element of the hashtable, representing a |
| 161 | reduction in the current loop. */ |
| 162 | struct reduction_info |
| 163 | { |
| 164 | gimple reduc_stmt; /* reduction statement. */ |
| 165 | gimple reduc_phi; /* The phi node defining the reduction. */ |
| 166 | enum tree_code reduction_code;/* code for the reduction operation. */ |
| 167 | unsigned reduc_version; /* SSA_NAME_VERSION of original reduc_phi |
| 168 | result. */ |
| 169 | gimple keep_res; /* The PHI_RESULT of this phi is the resulting value |
| 170 | of the reduction variable when existing the loop. */ |
| 171 | tree initial_value; /* The initial value of the reduction var before entering the loop. */ |
| 172 | tree field; /* the name of the field in the parloop data structure intended for reduction. */ |
| 173 | tree init; /* reduction initialization value. */ |
| 174 | gimple new_phi; /* (helper field) Newly created phi node whose result |
| 175 | will be passed to the atomic operation. Represents |
| 176 | the local result each thread computed for the reduction |
| 177 | operation. */ |
| 178 | }; |
| 179 | |
| 180 | /* Equality and hash functions for hashtab code. */ |
| 181 | |
| 182 | static int |
| 183 | reduction_info_eq (const void *aa, const void *bb) |
| 184 | { |
| 185 | const struct reduction_info *a = (const struct reduction_info *) aa; |
| 186 | const struct reduction_info *b = (const struct reduction_info *) bb; |
| 187 | |
| 188 | return (a->reduc_phi == b->reduc_phi); |
| 189 | } |
| 190 | |
| 191 | static hashval_t |
| 192 | reduction_info_hash (const void *aa) |
| 193 | { |
| 194 | const struct reduction_info *a = (const struct reduction_info *) aa; |
| 195 | |
| 196 | return a->reduc_version; |
| 197 | } |
| 198 | |
| 199 | static struct reduction_info * |
| 200 | reduction_phi (htab_t reduction_list, gimple phi) |
| 201 | { |
| 202 | struct reduction_info tmpred, *red; |
| 203 | |
| 204 | if (htab_elements (reduction_list) == 0 || phi == NULL) |
| 205 | return NULL; |
| 206 | |
| 207 | tmpred.reduc_phi = phi; |
| 208 | tmpred.reduc_version = gimple_uid (phi); |
| 209 | red = (struct reduction_info *) htab_find (reduction_list, &tmpred); |
| 210 | |
| 211 | return red; |
| 212 | } |
| 213 | |
| 214 | /* Element of hashtable of names to copy. */ |
| 215 | |
| 216 | struct name_to_copy_elt |
| 217 | { |
| 218 | unsigned version; /* The version of the name to copy. */ |
| 219 | tree new_name; /* The new name used in the copy. */ |
| 220 | tree field; /* The field of the structure used to pass the |
| 221 | value. */ |
| 222 | }; |
| 223 | |
| 224 | /* Equality and hash functions for hashtab code. */ |
| 225 | |
| 226 | static int |
| 227 | name_to_copy_elt_eq (const void *aa, const void *bb) |
| 228 | { |
| 229 | const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; |
| 230 | const struct name_to_copy_elt *b = (const struct name_to_copy_elt *) bb; |
| 231 | |
| 232 | return a->version == b->version; |
| 233 | } |
| 234 | |
| 235 | static hashval_t |
| 236 | name_to_copy_elt_hash (const void *aa) |
| 237 | { |
| 238 | const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; |
| 239 | |
| 240 | return (hashval_t) a->version; |
| 241 | } |
| 242 | |
| 243 | /* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE |
| 244 | matrix. Rather than use floats, we simply keep a single DENOMINATOR that |
| 245 | represents the denominator for every element in the matrix. */ |
| 246 | typedef struct lambda_trans_matrix_s |
| 247 | { |
| 248 | lambda_matrix matrix; |
| 249 | int rowsize; |
| 250 | int colsize; |
| 251 | int denominator; |
| 252 | } *lambda_trans_matrix; |
| 253 | #define LTM_MATRIX(T) ((T)->matrix) |
| 254 | #define LTM_ROWSIZE(T) ((T)->rowsize) |
| 255 | #define LTM_COLSIZE(T) ((T)->colsize) |
| 256 | #define LTM_DENOMINATOR(T) ((T)->denominator) |
| 257 | |
| 258 | /* Allocate a new transformation matrix. */ |
| 259 | |
| 260 | static lambda_trans_matrix |
| 261 | lambda_trans_matrix_new (int colsize, int rowsize, |
| 262 | struct obstack * lambda_obstack) |
| 263 | { |
| 264 | lambda_trans_matrix ret; |
| 265 | |
| 266 | ret = (lambda_trans_matrix) |
| 267 | obstack_alloc (lambda_obstack, sizeof (struct lambda_trans_matrix_s)); |
| 268 | LTM_MATRIX (ret) = lambda_matrix_new (rowsize, colsize, lambda_obstack); |
| 269 | LTM_ROWSIZE (ret) = rowsize; |
| 270 | LTM_COLSIZE (ret) = colsize; |
| 271 | LTM_DENOMINATOR (ret) = 1; |
| 272 | return ret; |
| 273 | } |
| 274 | |
| 275 | /* Multiply a vector VEC by a matrix MAT. |
| 276 | MAT is an M*N matrix, and VEC is a vector with length N. The result |
| 277 | is stored in DEST which must be a vector of length M. */ |
| 278 | |
| 279 | static void |
| 280 | lambda_matrix_vector_mult (lambda_matrix matrix, int m, int n, |
| 281 | lambda_vector vec, lambda_vector dest) |
| 282 | { |
| 283 | int i, j; |
| 284 | |
| 285 | lambda_vector_clear (dest, m); |
| 286 | for (i = 0; i < m; i++) |
| 287 | for (j = 0; j < n; j++) |
| 288 | dest[i] += matrix[i][j] * vec[j]; |
| 289 | } |
| 290 | |
| 291 | /* Return true if TRANS is a legal transformation matrix that respects |
| 292 | the dependence vectors in DISTS and DIRS. The conservative answer |
| 293 | is false. |
| 294 | |
| 295 | "Wolfe proves that a unimodular transformation represented by the |
| 296 | matrix T is legal when applied to a loop nest with a set of |
| 297 | lexicographically non-negative distance vectors RDG if and only if |
| 298 | for each vector d in RDG, (T.d >= 0) is lexicographically positive. |
| 299 | i.e.: if and only if it transforms the lexicographically positive |
| 300 | distance vectors to lexicographically positive vectors. Note that |
| 301 | a unimodular matrix must transform the zero vector (and only it) to |
| 302 | the zero vector." S.Muchnick. */ |
| 303 | |
| 304 | static bool |
| 305 | lambda_transform_legal_p (lambda_trans_matrix trans, |
| 306 | int nb_loops, |
| 307 | VEC (ddr_p, heap) *dependence_relations) |
| 308 | { |
| 309 | unsigned int i, j; |
| 310 | lambda_vector distres; |
| 311 | struct data_dependence_relation *ddr; |
| 312 | |
| 313 | gcc_assert (LTM_COLSIZE (trans) == nb_loops |
| 314 | && LTM_ROWSIZE (trans) == nb_loops); |
| 315 | |
| 316 | /* When there are no dependences, the transformation is correct. */ |
| 317 | if (VEC_length (ddr_p, dependence_relations) == 0) |
| 318 | return true; |
| 319 | |
| 320 | ddr = VEC_index (ddr_p, dependence_relations, 0); |
| 321 | if (ddr == NULL) |
| 322 | return true; |
| 323 | |
| 324 | /* When there is an unknown relation in the dependence_relations, we |
| 325 | know that it is no worth looking at this loop nest: give up. */ |
| 326 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
| 327 | return false; |
| 328 | |
| 329 | distres = lambda_vector_new (nb_loops); |
| 330 | |
| 331 | /* For each distance vector in the dependence graph. */ |
| 332 | FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr) |
| 333 | { |
| 334 | /* Don't care about relations for which we know that there is no |
| 335 | dependence, nor about read-read (aka. output-dependences): |
| 336 | these data accesses can happen in any order. */ |
| 337 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known |
| 338 | || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr)))) |
| 339 | continue; |
| 340 | |
| 341 | /* Conservatively answer: "this transformation is not valid". */ |
| 342 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
| 343 | return false; |
| 344 | |
| 345 | /* If the dependence could not be captured by a distance vector, |
| 346 | conservatively answer that the transform is not valid. */ |
| 347 | if (DDR_NUM_DIST_VECTS (ddr) == 0) |
| 348 | return false; |
| 349 | |
| 350 | /* Compute trans.dist_vect */ |
| 351 | for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++) |
| 352 | { |
| 353 | lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops, |
| 354 | DDR_DIST_VECT (ddr, j), distres); |
| 355 | |
| 356 | if (!lambda_vector_lexico_pos (distres, nb_loops)) |
| 357 | return false; |
| 358 | } |
| 359 | } |
| 360 | return true; |
| 361 | } |
| 362 | |
| 363 | /* Data dependency analysis. Returns true if the iterations of LOOP |
| 364 | are independent on each other (that is, if we can execute them |
| 365 | in parallel). */ |
| 366 | |
| 367 | static bool |
| 368 | loop_parallel_p (struct loop *loop, struct obstack * parloop_obstack) |
| 369 | { |
| 370 | VEC (loop_p, heap) *loop_nest; |
| 371 | VEC (ddr_p, heap) *dependence_relations; |
| 372 | VEC (data_reference_p, heap) *datarefs; |
| 373 | lambda_trans_matrix trans; |
| 374 | bool ret = false; |
| 375 | |
| 376 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 377 | { |
| 378 | fprintf (dump_file, "Considering loop %d\n", loop->num); |
| 379 | if (!loop->inner) |
| 380 | fprintf (dump_file, "loop is innermost\n"); |
| 381 | else |
| 382 | fprintf (dump_file, "loop NOT innermost\n"); |
| 383 | } |
| 384 | |
| 385 | /* Check for problems with dependences. If the loop can be reversed, |
| 386 | the iterations are independent. */ |
| 387 | datarefs = VEC_alloc (data_reference_p, heap, 10); |
| 388 | dependence_relations = VEC_alloc (ddr_p, heap, 10 * 10); |
| 389 | loop_nest = VEC_alloc (loop_p, heap, 3); |
| 390 | if (! compute_data_dependences_for_loop (loop, true, &loop_nest, &datarefs, |
| 391 | &dependence_relations)) |
| 392 | { |
| 393 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 394 | fprintf (dump_file, " FAILED: cannot analyze data dependencies\n"); |
| 395 | ret = false; |
| 396 | goto end; |
| 397 | } |
| 398 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 399 | dump_data_dependence_relations (dump_file, dependence_relations); |
| 400 | |
| 401 | trans = lambda_trans_matrix_new (1, 1, parloop_obstack); |
| 402 | LTM_MATRIX (trans)[0][0] = -1; |
| 403 | |
| 404 | if (lambda_transform_legal_p (trans, 1, dependence_relations)) |
| 405 | { |
| 406 | ret = true; |
| 407 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 408 | fprintf (dump_file, " SUCCESS: may be parallelized\n"); |
| 409 | } |
| 410 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
| 411 | fprintf (dump_file, |
| 412 | " FAILED: data dependencies exist across iterations\n"); |
| 413 | |
| 414 | end: |
| 415 | VEC_free (loop_p, heap, loop_nest); |
| 416 | free_dependence_relations (dependence_relations); |
| 417 | free_data_refs (datarefs); |
| 418 | |
| 419 | return ret; |
| 420 | } |
| 421 | |
| 422 | /* Return true when LOOP contains basic blocks marked with the |
| 423 | BB_IRREDUCIBLE_LOOP flag. */ |
| 424 | |
| 425 | static inline bool |
| 426 | loop_has_blocks_with_irreducible_flag (struct loop *loop) |
| 427 | { |
| 428 | unsigned i; |
| 429 | basic_block *bbs = get_loop_body_in_dom_order (loop); |
| 430 | bool res = true; |
| 431 | |
| 432 | for (i = 0; i < loop->num_nodes; i++) |
| 433 | if (bbs[i]->flags & BB_IRREDUCIBLE_LOOP) |
| 434 | goto end; |
| 435 | |
| 436 | res = false; |
| 437 | end: |
| 438 | free (bbs); |
| 439 | return res; |
| 440 | } |
| 441 | |
| 442 | /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name. |
| 443 | The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls |
| 444 | to their addresses that can be reused. The address of OBJ is known to |
| 445 | be invariant in the whole function. Other needed statements are placed |
| 446 | right before GSI. */ |
| 447 | |
| 448 | static tree |
| 449 | take_address_of (tree obj, tree type, edge entry, htab_t decl_address, |
| 450 | gimple_stmt_iterator *gsi) |
| 451 | { |
| 452 | int uid; |
| 453 | void **dslot; |
| 454 | struct int_tree_map ielt, *nielt; |
| 455 | tree *var_p, name, addr; |
| 456 | gimple stmt; |
| 457 | gimple_seq stmts; |
| 458 | |
| 459 | /* Since the address of OBJ is invariant, the trees may be shared. |
| 460 | Avoid rewriting unrelated parts of the code. */ |
| 461 | obj = unshare_expr (obj); |
| 462 | for (var_p = &obj; |
| 463 | handled_component_p (*var_p); |
| 464 | var_p = &TREE_OPERAND (*var_p, 0)) |
| 465 | continue; |
| 466 | |
| 467 | /* Canonicalize the access to base on a MEM_REF. */ |
| 468 | if (DECL_P (*var_p)) |
| 469 | *var_p = build_simple_mem_ref (build_fold_addr_expr (*var_p)); |
| 470 | |
| 471 | /* Assign a canonical SSA name to the address of the base decl used |
| 472 | in the address and share it for all accesses and addresses based |
| 473 | on it. */ |
| 474 | uid = DECL_UID (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); |
| 475 | ielt.uid = uid; |
| 476 | dslot = htab_find_slot_with_hash (decl_address, &ielt, uid, INSERT); |
| 477 | if (!*dslot) |
| 478 | { |
| 479 | if (gsi == NULL) |
| 480 | return NULL; |
| 481 | addr = TREE_OPERAND (*var_p, 0); |
| 482 | name = make_temp_ssa_name (TREE_TYPE (addr), NULL, |
| 483 | get_name (TREE_OPERAND |
| 484 | (TREE_OPERAND (*var_p, 0), 0))); |
| 485 | stmt = gimple_build_assign (name, addr); |
| 486 | gsi_insert_on_edge_immediate (entry, stmt); |
| 487 | |
| 488 | nielt = XNEW (struct int_tree_map); |
| 489 | nielt->uid = uid; |
| 490 | nielt->to = name; |
| 491 | *dslot = nielt; |
| 492 | } |
| 493 | else |
| 494 | name = ((struct int_tree_map *) *dslot)->to; |
| 495 | |
| 496 | /* Express the address in terms of the canonical SSA name. */ |
| 497 | TREE_OPERAND (*var_p, 0) = name; |
| 498 | if (gsi == NULL) |
| 499 | return build_fold_addr_expr_with_type (obj, type); |
| 500 | |
| 501 | name = force_gimple_operand (build_addr (obj, current_function_decl), |
| 502 | &stmts, true, NULL_TREE); |
| 503 | if (!gimple_seq_empty_p (stmts)) |
| 504 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
| 505 | |
| 506 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) |
| 507 | { |
| 508 | name = force_gimple_operand (fold_convert (type, name), &stmts, true, |
| 509 | NULL_TREE); |
| 510 | if (!gimple_seq_empty_p (stmts)) |
| 511 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
| 512 | } |
| 513 | |
| 514 | return name; |
| 515 | } |
| 516 | |
| 517 | /* Callback for htab_traverse. Create the initialization statement |
| 518 | for reduction described in SLOT, and place it at the preheader of |
| 519 | the loop described in DATA. */ |
| 520 | |
| 521 | static int |
| 522 | initialize_reductions (void **slot, void *data) |
| 523 | { |
| 524 | tree init, c; |
| 525 | tree bvar, type, arg; |
| 526 | edge e; |
| 527 | |
| 528 | struct reduction_info *const reduc = (struct reduction_info *) *slot; |
| 529 | struct loop *loop = (struct loop *) data; |
| 530 | |
| 531 | /* Create initialization in preheader: |
| 532 | reduction_variable = initialization value of reduction. */ |
| 533 | |
| 534 | /* In the phi node at the header, replace the argument coming |
| 535 | from the preheader with the reduction initialization value. */ |
| 536 | |
| 537 | /* Create a new variable to initialize the reduction. */ |
| 538 | type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); |
| 539 | bvar = create_tmp_var (type, "reduction"); |
| 540 | |
| 541 | c = build_omp_clause (gimple_location (reduc->reduc_stmt), |
| 542 | OMP_CLAUSE_REDUCTION); |
| 543 | OMP_CLAUSE_REDUCTION_CODE (c) = reduc->reduction_code; |
| 544 | OMP_CLAUSE_DECL (c) = SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)); |
| 545 | |
| 546 | init = omp_reduction_init (c, TREE_TYPE (bvar)); |
| 547 | reduc->init = init; |
| 548 | |
| 549 | /* Replace the argument representing the initialization value |
| 550 | with the initialization value for the reduction (neutral |
| 551 | element for the particular operation, e.g. 0 for PLUS_EXPR, |
| 552 | 1 for MULT_EXPR, etc). |
| 553 | Keep the old value in a new variable "reduction_initial", |
| 554 | that will be taken in consideration after the parallel |
| 555 | computing is done. */ |
| 556 | |
| 557 | e = loop_preheader_edge (loop); |
| 558 | arg = PHI_ARG_DEF_FROM_EDGE (reduc->reduc_phi, e); |
| 559 | /* Create new variable to hold the initial value. */ |
| 560 | |
| 561 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE |
| 562 | (reduc->reduc_phi, loop_preheader_edge (loop)), init); |
| 563 | reduc->initial_value = arg; |
| 564 | return 1; |
| 565 | } |
| 566 | |
| 567 | struct elv_data |
| 568 | { |
| 569 | struct walk_stmt_info info; |
| 570 | edge entry; |
| 571 | htab_t decl_address; |
| 572 | gimple_stmt_iterator *gsi; |
| 573 | bool changed; |
| 574 | bool reset; |
| 575 | }; |
| 576 | |
| 577 | /* Eliminates references to local variables in *TP out of the single |
| 578 | entry single exit region starting at DTA->ENTRY. |
| 579 | DECL_ADDRESS contains addresses of the references that had their |
| 580 | address taken already. If the expression is changed, CHANGED is |
| 581 | set to true. Callback for walk_tree. */ |
| 582 | |
| 583 | static tree |
| 584 | eliminate_local_variables_1 (tree *tp, int *walk_subtrees, void *data) |
| 585 | { |
| 586 | struct elv_data *const dta = (struct elv_data *) data; |
| 587 | tree t = *tp, var, addr, addr_type, type, obj; |
| 588 | |
| 589 | if (DECL_P (t)) |
| 590 | { |
| 591 | *walk_subtrees = 0; |
| 592 | |
| 593 | if (!SSA_VAR_P (t) || DECL_EXTERNAL (t)) |
| 594 | return NULL_TREE; |
| 595 | |
| 596 | type = TREE_TYPE (t); |
| 597 | addr_type = build_pointer_type (type); |
| 598 | addr = take_address_of (t, addr_type, dta->entry, dta->decl_address, |
| 599 | dta->gsi); |
| 600 | if (dta->gsi == NULL && addr == NULL_TREE) |
| 601 | { |
| 602 | dta->reset = true; |
| 603 | return NULL_TREE; |
| 604 | } |
| 605 | |
| 606 | *tp = build_simple_mem_ref (addr); |
| 607 | |
| 608 | dta->changed = true; |
| 609 | return NULL_TREE; |
| 610 | } |
| 611 | |
| 612 | if (TREE_CODE (t) == ADDR_EXPR) |
| 613 | { |
| 614 | /* ADDR_EXPR may appear in two contexts: |
| 615 | -- as a gimple operand, when the address taken is a function invariant |
| 616 | -- as gimple rhs, when the resulting address in not a function |
| 617 | invariant |
| 618 | We do not need to do anything special in the latter case (the base of |
| 619 | the memory reference whose address is taken may be replaced in the |
| 620 | DECL_P case). The former case is more complicated, as we need to |
| 621 | ensure that the new address is still a gimple operand. Thus, it |
| 622 | is not sufficient to replace just the base of the memory reference -- |
| 623 | we need to move the whole computation of the address out of the |
| 624 | loop. */ |
| 625 | if (!is_gimple_val (t)) |
| 626 | return NULL_TREE; |
| 627 | |
| 628 | *walk_subtrees = 0; |
| 629 | obj = TREE_OPERAND (t, 0); |
| 630 | var = get_base_address (obj); |
| 631 | if (!var || !SSA_VAR_P (var) || DECL_EXTERNAL (var)) |
| 632 | return NULL_TREE; |
| 633 | |
| 634 | addr_type = TREE_TYPE (t); |
| 635 | addr = take_address_of (obj, addr_type, dta->entry, dta->decl_address, |
| 636 | dta->gsi); |
| 637 | if (dta->gsi == NULL && addr == NULL_TREE) |
| 638 | { |
| 639 | dta->reset = true; |
| 640 | return NULL_TREE; |
| 641 | } |
| 642 | *tp = addr; |
| 643 | |
| 644 | dta->changed = true; |
| 645 | return NULL_TREE; |
| 646 | } |
| 647 | |
| 648 | if (!EXPR_P (t)) |
| 649 | *walk_subtrees = 0; |
| 650 | |
| 651 | return NULL_TREE; |
| 652 | } |
| 653 | |
| 654 | /* Moves the references to local variables in STMT at *GSI out of the single |
| 655 | entry single exit region starting at ENTRY. DECL_ADDRESS contains |
| 656 | addresses of the references that had their address taken |
| 657 | already. */ |
| 658 | |
| 659 | static void |
| 660 | eliminate_local_variables_stmt (edge entry, gimple_stmt_iterator *gsi, |
| 661 | htab_t decl_address) |
| 662 | { |
| 663 | struct elv_data dta; |
| 664 | gimple stmt = gsi_stmt (*gsi); |
| 665 | |
| 666 | memset (&dta.info, '\0', sizeof (dta.info)); |
| 667 | dta.entry = entry; |
| 668 | dta.decl_address = decl_address; |
| 669 | dta.changed = false; |
| 670 | dta.reset = false; |
| 671 | |
| 672 | if (gimple_debug_bind_p (stmt)) |
| 673 | { |
| 674 | dta.gsi = NULL; |
| 675 | walk_tree (gimple_debug_bind_get_value_ptr (stmt), |
| 676 | eliminate_local_variables_1, &dta.info, NULL); |
| 677 | if (dta.reset) |
| 678 | { |
| 679 | gimple_debug_bind_reset_value (stmt); |
| 680 | dta.changed = true; |
| 681 | } |
| 682 | } |
| 683 | else |
| 684 | { |
| 685 | dta.gsi = gsi; |
| 686 | walk_gimple_op (stmt, eliminate_local_variables_1, &dta.info); |
| 687 | } |
| 688 | |
| 689 | if (dta.changed) |
| 690 | update_stmt (stmt); |
| 691 | } |
| 692 | |
| 693 | /* Eliminates the references to local variables from the single entry |
| 694 | single exit region between the ENTRY and EXIT edges. |
| 695 | |
| 696 | This includes: |
| 697 | 1) Taking address of a local variable -- these are moved out of the |
| 698 | region (and temporary variable is created to hold the address if |
| 699 | necessary). |
| 700 | |
| 701 | 2) Dereferencing a local variable -- these are replaced with indirect |
| 702 | references. */ |
| 703 | |
| 704 | static void |
| 705 | eliminate_local_variables (edge entry, edge exit) |
| 706 | { |
| 707 | basic_block bb; |
| 708 | VEC (basic_block, heap) *body = VEC_alloc (basic_block, heap, 3); |
| 709 | unsigned i; |
| 710 | gimple_stmt_iterator gsi; |
| 711 | bool has_debug_stmt = false; |
| 712 | htab_t decl_address = htab_create (10, int_tree_map_hash, int_tree_map_eq, |
| 713 | free); |
| 714 | basic_block entry_bb = entry->src; |
| 715 | basic_block exit_bb = exit->dest; |
| 716 | |
| 717 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
| 718 | |
| 719 | FOR_EACH_VEC_ELT (basic_block, body, i, bb) |
| 720 | if (bb != entry_bb && bb != exit_bb) |
| 721 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 722 | if (is_gimple_debug (gsi_stmt (gsi))) |
| 723 | { |
| 724 | if (gimple_debug_bind_p (gsi_stmt (gsi))) |
| 725 | has_debug_stmt = true; |
| 726 | } |
| 727 | else |
| 728 | eliminate_local_variables_stmt (entry, &gsi, decl_address); |
| 729 | |
| 730 | if (has_debug_stmt) |
| 731 | FOR_EACH_VEC_ELT (basic_block, body, i, bb) |
| 732 | if (bb != entry_bb && bb != exit_bb) |
| 733 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 734 | if (gimple_debug_bind_p (gsi_stmt (gsi))) |
| 735 | eliminate_local_variables_stmt (entry, &gsi, decl_address); |
| 736 | |
| 737 | htab_delete (decl_address); |
| 738 | VEC_free (basic_block, heap, body); |
| 739 | } |
| 740 | |
| 741 | /* Returns true if expression EXPR is not defined between ENTRY and |
| 742 | EXIT, i.e. if all its operands are defined outside of the region. */ |
| 743 | |
| 744 | static bool |
| 745 | expr_invariant_in_region_p (edge entry, edge exit, tree expr) |
| 746 | { |
| 747 | basic_block entry_bb = entry->src; |
| 748 | basic_block exit_bb = exit->dest; |
| 749 | basic_block def_bb; |
| 750 | |
| 751 | if (is_gimple_min_invariant (expr)) |
| 752 | return true; |
| 753 | |
| 754 | if (TREE_CODE (expr) == SSA_NAME) |
| 755 | { |
| 756 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
| 757 | if (def_bb |
| 758 | && dominated_by_p (CDI_DOMINATORS, def_bb, entry_bb) |
| 759 | && !dominated_by_p (CDI_DOMINATORS, def_bb, exit_bb)) |
| 760 | return false; |
| 761 | |
| 762 | return true; |
| 763 | } |
| 764 | |
| 765 | return false; |
| 766 | } |
| 767 | |
| 768 | /* If COPY_NAME_P is true, creates and returns a duplicate of NAME. |
| 769 | The copies are stored to NAME_COPIES, if NAME was already duplicated, |
| 770 | its duplicate stored in NAME_COPIES is returned. |
| 771 | |
| 772 | Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also |
| 773 | duplicated, storing the copies in DECL_COPIES. */ |
| 774 | |
| 775 | static tree |
| 776 | separate_decls_in_region_name (tree name, |
| 777 | htab_t name_copies, htab_t decl_copies, |
| 778 | bool copy_name_p) |
| 779 | { |
| 780 | tree copy, var, var_copy; |
| 781 | unsigned idx, uid, nuid; |
| 782 | struct int_tree_map ielt, *nielt; |
| 783 | struct name_to_copy_elt elt, *nelt; |
| 784 | void **slot, **dslot; |
| 785 | |
| 786 | if (TREE_CODE (name) != SSA_NAME) |
| 787 | return name; |
| 788 | |
| 789 | idx = SSA_NAME_VERSION (name); |
| 790 | elt.version = idx; |
| 791 | slot = htab_find_slot_with_hash (name_copies, &elt, idx, |
| 792 | copy_name_p ? INSERT : NO_INSERT); |
| 793 | if (slot && *slot) |
| 794 | return ((struct name_to_copy_elt *) *slot)->new_name; |
| 795 | |
| 796 | if (copy_name_p) |
| 797 | { |
| 798 | copy = duplicate_ssa_name (name, NULL); |
| 799 | nelt = XNEW (struct name_to_copy_elt); |
| 800 | nelt->version = idx; |
| 801 | nelt->new_name = copy; |
| 802 | nelt->field = NULL_TREE; |
| 803 | *slot = nelt; |
| 804 | } |
| 805 | else |
| 806 | { |
| 807 | gcc_assert (!slot); |
| 808 | copy = name; |
| 809 | } |
| 810 | |
| 811 | var = SSA_NAME_VAR (name); |
| 812 | if (!var) |
| 813 | return copy; |
| 814 | |
| 815 | uid = DECL_UID (var); |
| 816 | ielt.uid = uid; |
| 817 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, uid, INSERT); |
| 818 | if (!*dslot) |
| 819 | { |
| 820 | var_copy = create_tmp_var (TREE_TYPE (var), get_name (var)); |
| 821 | DECL_GIMPLE_REG_P (var_copy) = DECL_GIMPLE_REG_P (var); |
| 822 | nielt = XNEW (struct int_tree_map); |
| 823 | nielt->uid = uid; |
| 824 | nielt->to = var_copy; |
| 825 | *dslot = nielt; |
| 826 | |
| 827 | /* Ensure that when we meet this decl next time, we won't duplicate |
| 828 | it again. */ |
| 829 | nuid = DECL_UID (var_copy); |
| 830 | ielt.uid = nuid; |
| 831 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, nuid, INSERT); |
| 832 | gcc_assert (!*dslot); |
| 833 | nielt = XNEW (struct int_tree_map); |
| 834 | nielt->uid = nuid; |
| 835 | nielt->to = var_copy; |
| 836 | *dslot = nielt; |
| 837 | } |
| 838 | else |
| 839 | var_copy = ((struct int_tree_map *) *dslot)->to; |
| 840 | |
| 841 | replace_ssa_name_symbol (copy, var_copy); |
| 842 | return copy; |
| 843 | } |
| 844 | |
| 845 | /* Finds the ssa names used in STMT that are defined outside the |
| 846 | region between ENTRY and EXIT and replaces such ssa names with |
| 847 | their duplicates. The duplicates are stored to NAME_COPIES. Base |
| 848 | decls of all ssa names used in STMT (including those defined in |
| 849 | LOOP) are replaced with the new temporary variables; the |
| 850 | replacement decls are stored in DECL_COPIES. */ |
| 851 | |
| 852 | static void |
| 853 | separate_decls_in_region_stmt (edge entry, edge exit, gimple stmt, |
| 854 | htab_t name_copies, htab_t decl_copies) |
| 855 | { |
| 856 | use_operand_p use; |
| 857 | def_operand_p def; |
| 858 | ssa_op_iter oi; |
| 859 | tree name, copy; |
| 860 | bool copy_name_p; |
| 861 | |
| 862 | FOR_EACH_PHI_OR_STMT_DEF (def, stmt, oi, SSA_OP_DEF) |
| 863 | { |
| 864 | name = DEF_FROM_PTR (def); |
| 865 | gcc_assert (TREE_CODE (name) == SSA_NAME); |
| 866 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, |
| 867 | false); |
| 868 | gcc_assert (copy == name); |
| 869 | } |
| 870 | |
| 871 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) |
| 872 | { |
| 873 | name = USE_FROM_PTR (use); |
| 874 | if (TREE_CODE (name) != SSA_NAME) |
| 875 | continue; |
| 876 | |
| 877 | copy_name_p = expr_invariant_in_region_p (entry, exit, name); |
| 878 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, |
| 879 | copy_name_p); |
| 880 | SET_USE (use, copy); |
| 881 | } |
| 882 | } |
| 883 | |
| 884 | /* Finds the ssa names used in STMT that are defined outside the |
| 885 | region between ENTRY and EXIT and replaces such ssa names with |
| 886 | their duplicates. The duplicates are stored to NAME_COPIES. Base |
| 887 | decls of all ssa names used in STMT (including those defined in |
| 888 | LOOP) are replaced with the new temporary variables; the |
| 889 | replacement decls are stored in DECL_COPIES. */ |
| 890 | |
| 891 | static bool |
| 892 | separate_decls_in_region_debug (gimple stmt, htab_t name_copies, |
| 893 | htab_t decl_copies) |
| 894 | { |
| 895 | use_operand_p use; |
| 896 | ssa_op_iter oi; |
| 897 | tree var, name; |
| 898 | struct int_tree_map ielt; |
| 899 | struct name_to_copy_elt elt; |
| 900 | void **slot, **dslot; |
| 901 | |
| 902 | if (gimple_debug_bind_p (stmt)) |
| 903 | var = gimple_debug_bind_get_var (stmt); |
| 904 | else if (gimple_debug_source_bind_p (stmt)) |
| 905 | var = gimple_debug_source_bind_get_var (stmt); |
| 906 | else |
| 907 | return true; |
| 908 | if (TREE_CODE (var) == DEBUG_EXPR_DECL || TREE_CODE (var) == LABEL_DECL) |
| 909 | return true; |
| 910 | gcc_assert (DECL_P (var) && SSA_VAR_P (var)); |
| 911 | ielt.uid = DECL_UID (var); |
| 912 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, ielt.uid, NO_INSERT); |
| 913 | if (!dslot) |
| 914 | return true; |
| 915 | if (gimple_debug_bind_p (stmt)) |
| 916 | gimple_debug_bind_set_var (stmt, ((struct int_tree_map *) *dslot)->to); |
| 917 | else if (gimple_debug_source_bind_p (stmt)) |
| 918 | gimple_debug_source_bind_set_var (stmt, ((struct int_tree_map *) *dslot)->to); |
| 919 | |
| 920 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) |
| 921 | { |
| 922 | name = USE_FROM_PTR (use); |
| 923 | if (TREE_CODE (name) != SSA_NAME) |
| 924 | continue; |
| 925 | |
| 926 | elt.version = SSA_NAME_VERSION (name); |
| 927 | slot = htab_find_slot_with_hash (name_copies, &elt, elt.version, NO_INSERT); |
| 928 | if (!slot) |
| 929 | { |
| 930 | gimple_debug_bind_reset_value (stmt); |
| 931 | update_stmt (stmt); |
| 932 | break; |
| 933 | } |
| 934 | |
| 935 | SET_USE (use, ((struct name_to_copy_elt *) *slot)->new_name); |
| 936 | } |
| 937 | |
| 938 | return false; |
| 939 | } |
| 940 | |
| 941 | /* Callback for htab_traverse. Adds a field corresponding to the reduction |
| 942 | specified in SLOT. The type is passed in DATA. */ |
| 943 | |
| 944 | static int |
| 945 | add_field_for_reduction (void **slot, void *data) |
| 946 | { |
| 947 | |
| 948 | struct reduction_info *const red = (struct reduction_info *) *slot; |
| 949 | tree const type = (tree) data; |
| 950 | tree var = SSA_NAME_VAR (gimple_assign_lhs (red->reduc_stmt)); |
| 951 | tree field = build_decl (gimple_location (red->reduc_stmt), |
| 952 | FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); |
| 953 | |
| 954 | insert_field_into_struct (type, field); |
| 955 | |
| 956 | red->field = field; |
| 957 | |
| 958 | return 1; |
| 959 | } |
| 960 | |
| 961 | /* Callback for htab_traverse. Adds a field corresponding to a ssa name |
| 962 | described in SLOT. The type is passed in DATA. */ |
| 963 | |
| 964 | static int |
| 965 | add_field_for_name (void **slot, void *data) |
| 966 | { |
| 967 | struct name_to_copy_elt *const elt = (struct name_to_copy_elt *) *slot; |
| 968 | tree type = (tree) data; |
| 969 | tree name = ssa_name (elt->version); |
| 970 | tree field = build_decl (UNKNOWN_LOCATION, |
| 971 | FIELD_DECL, SSA_NAME_IDENTIFIER (name), |
| 972 | TREE_TYPE (name)); |
| 973 | |
| 974 | insert_field_into_struct (type, field); |
| 975 | elt->field = field; |
| 976 | |
| 977 | return 1; |
| 978 | } |
| 979 | |
| 980 | /* Callback for htab_traverse. A local result is the intermediate result |
| 981 | computed by a single |
| 982 | thread, or the initial value in case no iteration was executed. |
| 983 | This function creates a phi node reflecting these values. |
| 984 | The phi's result will be stored in NEW_PHI field of the |
| 985 | reduction's data structure. */ |
| 986 | |
| 987 | static int |
| 988 | create_phi_for_local_result (void **slot, void *data) |
| 989 | { |
| 990 | struct reduction_info *const reduc = (struct reduction_info *) *slot; |
| 991 | const struct loop *const loop = (const struct loop *) data; |
| 992 | edge e; |
| 993 | gimple new_phi; |
| 994 | basic_block store_bb; |
| 995 | tree local_res; |
| 996 | source_location locus; |
| 997 | |
| 998 | /* STORE_BB is the block where the phi |
| 999 | should be stored. It is the destination of the loop exit. |
| 1000 | (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */ |
| 1001 | store_bb = FALLTHRU_EDGE (loop->latch)->dest; |
| 1002 | |
| 1003 | /* STORE_BB has two predecessors. One coming from the loop |
| 1004 | (the reduction's result is computed at the loop), |
| 1005 | and another coming from a block preceding the loop, |
| 1006 | when no iterations |
| 1007 | are executed (the initial value should be taken). */ |
| 1008 | if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (loop->latch)) |
| 1009 | e = EDGE_PRED (store_bb, 1); |
| 1010 | else |
| 1011 | e = EDGE_PRED (store_bb, 0); |
| 1012 | local_res = copy_ssa_name (gimple_assign_lhs (reduc->reduc_stmt), NULL); |
| 1013 | locus = gimple_location (reduc->reduc_stmt); |
| 1014 | new_phi = create_phi_node (local_res, store_bb); |
| 1015 | add_phi_arg (new_phi, reduc->init, e, locus); |
| 1016 | add_phi_arg (new_phi, gimple_assign_lhs (reduc->reduc_stmt), |
| 1017 | FALLTHRU_EDGE (loop->latch), locus); |
| 1018 | reduc->new_phi = new_phi; |
| 1019 | |
| 1020 | return 1; |
| 1021 | } |
| 1022 | |
| 1023 | struct clsn_data |
| 1024 | { |
| 1025 | tree store; |
| 1026 | tree load; |
| 1027 | |
| 1028 | basic_block store_bb; |
| 1029 | basic_block load_bb; |
| 1030 | }; |
| 1031 | |
| 1032 | /* Callback for htab_traverse. Create an atomic instruction for the |
| 1033 | reduction described in SLOT. |
| 1034 | DATA annotates the place in memory the atomic operation relates to, |
| 1035 | and the basic block it needs to be generated in. */ |
| 1036 | |
| 1037 | static int |
| 1038 | create_call_for_reduction_1 (void **slot, void *data) |
| 1039 | { |
| 1040 | struct reduction_info *const reduc = (struct reduction_info *) *slot; |
| 1041 | struct clsn_data *const clsn_data = (struct clsn_data *) data; |
| 1042 | gimple_stmt_iterator gsi; |
| 1043 | tree type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); |
| 1044 | tree load_struct; |
| 1045 | basic_block bb; |
| 1046 | basic_block new_bb; |
| 1047 | edge e; |
| 1048 | tree t, addr, ref, x; |
| 1049 | tree tmp_load, name; |
| 1050 | gimple load; |
| 1051 | |
| 1052 | load_struct = build_simple_mem_ref (clsn_data->load); |
| 1053 | t = build3 (COMPONENT_REF, type, load_struct, reduc->field, NULL_TREE); |
| 1054 | |
| 1055 | addr = build_addr (t, current_function_decl); |
| 1056 | |
| 1057 | /* Create phi node. */ |
| 1058 | bb = clsn_data->load_bb; |
| 1059 | |
| 1060 | e = split_block (bb, t); |
| 1061 | new_bb = e->dest; |
| 1062 | |
| 1063 | tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr)), NULL); |
| 1064 | tmp_load = make_ssa_name (tmp_load, NULL); |
| 1065 | load = gimple_build_omp_atomic_load (tmp_load, addr); |
| 1066 | SSA_NAME_DEF_STMT (tmp_load) = load; |
| 1067 | gsi = gsi_start_bb (new_bb); |
| 1068 | gsi_insert_after (&gsi, load, GSI_NEW_STMT); |
| 1069 | |
| 1070 | e = split_block (new_bb, load); |
| 1071 | new_bb = e->dest; |
| 1072 | gsi = gsi_start_bb (new_bb); |
| 1073 | ref = tmp_load; |
| 1074 | x = fold_build2 (reduc->reduction_code, |
| 1075 | TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, |
| 1076 | PHI_RESULT (reduc->new_phi)); |
| 1077 | |
| 1078 | name = force_gimple_operand_gsi (&gsi, x, true, NULL_TREE, true, |
| 1079 | GSI_CONTINUE_LINKING); |
| 1080 | |
| 1081 | gsi_insert_after (&gsi, gimple_build_omp_atomic_store (name), GSI_NEW_STMT); |
| 1082 | return 1; |
| 1083 | } |
| 1084 | |
| 1085 | /* Create the atomic operation at the join point of the threads. |
| 1086 | REDUCTION_LIST describes the reductions in the LOOP. |
| 1087 | LD_ST_DATA describes the shared data structure where |
| 1088 | shared data is stored in and loaded from. */ |
| 1089 | static void |
| 1090 | create_call_for_reduction (struct loop *loop, htab_t reduction_list, |
| 1091 | struct clsn_data *ld_st_data) |
| 1092 | { |
| 1093 | htab_traverse (reduction_list, create_phi_for_local_result, loop); |
| 1094 | /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */ |
| 1095 | ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest; |
| 1096 | htab_traverse (reduction_list, create_call_for_reduction_1, ld_st_data); |
| 1097 | } |
| 1098 | |
| 1099 | /* Callback for htab_traverse. Loads the final reduction value at the |
| 1100 | join point of all threads, and inserts it in the right place. */ |
| 1101 | |
| 1102 | static int |
| 1103 | create_loads_for_reductions (void **slot, void *data) |
| 1104 | { |
| 1105 | struct reduction_info *const red = (struct reduction_info *) *slot; |
| 1106 | struct clsn_data *const clsn_data = (struct clsn_data *) data; |
| 1107 | gimple stmt; |
| 1108 | gimple_stmt_iterator gsi; |
| 1109 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); |
| 1110 | tree load_struct; |
| 1111 | tree name; |
| 1112 | tree x; |
| 1113 | |
| 1114 | gsi = gsi_after_labels (clsn_data->load_bb); |
| 1115 | load_struct = build_simple_mem_ref (clsn_data->load); |
| 1116 | load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, |
| 1117 | NULL_TREE); |
| 1118 | |
| 1119 | x = load_struct; |
| 1120 | name = PHI_RESULT (red->keep_res); |
| 1121 | stmt = gimple_build_assign (name, x); |
| 1122 | SSA_NAME_DEF_STMT (name) = stmt; |
| 1123 | |
| 1124 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1125 | |
| 1126 | for (gsi = gsi_start_phis (gimple_bb (red->keep_res)); |
| 1127 | !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1128 | if (gsi_stmt (gsi) == red->keep_res) |
| 1129 | { |
| 1130 | remove_phi_node (&gsi, false); |
| 1131 | return 1; |
| 1132 | } |
| 1133 | gcc_unreachable (); |
| 1134 | } |
| 1135 | |
| 1136 | /* Load the reduction result that was stored in LD_ST_DATA. |
| 1137 | REDUCTION_LIST describes the list of reductions that the |
| 1138 | loads should be generated for. */ |
| 1139 | static void |
| 1140 | create_final_loads_for_reduction (htab_t reduction_list, |
| 1141 | struct clsn_data *ld_st_data) |
| 1142 | { |
| 1143 | gimple_stmt_iterator gsi; |
| 1144 | tree t; |
| 1145 | gimple stmt; |
| 1146 | |
| 1147 | gsi = gsi_after_labels (ld_st_data->load_bb); |
| 1148 | t = build_fold_addr_expr (ld_st_data->store); |
| 1149 | stmt = gimple_build_assign (ld_st_data->load, t); |
| 1150 | |
| 1151 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
| 1152 | SSA_NAME_DEF_STMT (ld_st_data->load) = stmt; |
| 1153 | |
| 1154 | htab_traverse (reduction_list, create_loads_for_reductions, ld_st_data); |
| 1155 | |
| 1156 | } |
| 1157 | |
| 1158 | /* Callback for htab_traverse. Store the neutral value for the |
| 1159 | particular reduction's operation, e.g. 0 for PLUS_EXPR, |
| 1160 | 1 for MULT_EXPR, etc. into the reduction field. |
| 1161 | The reduction is specified in SLOT. The store information is |
| 1162 | passed in DATA. */ |
| 1163 | |
| 1164 | static int |
| 1165 | create_stores_for_reduction (void **slot, void *data) |
| 1166 | { |
| 1167 | struct reduction_info *const red = (struct reduction_info *) *slot; |
| 1168 | struct clsn_data *const clsn_data = (struct clsn_data *) data; |
| 1169 | tree t; |
| 1170 | gimple stmt; |
| 1171 | gimple_stmt_iterator gsi; |
| 1172 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); |
| 1173 | |
| 1174 | gsi = gsi_last_bb (clsn_data->store_bb); |
| 1175 | t = build3 (COMPONENT_REF, type, clsn_data->store, red->field, NULL_TREE); |
| 1176 | stmt = gimple_build_assign (t, red->initial_value); |
| 1177 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1178 | |
| 1179 | return 1; |
| 1180 | } |
| 1181 | |
| 1182 | /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and |
| 1183 | store to a field of STORE in STORE_BB for the ssa name and its duplicate |
| 1184 | specified in SLOT. */ |
| 1185 | |
| 1186 | static int |
| 1187 | create_loads_and_stores_for_name (void **slot, void *data) |
| 1188 | { |
| 1189 | struct name_to_copy_elt *const elt = (struct name_to_copy_elt *) *slot; |
| 1190 | struct clsn_data *const clsn_data = (struct clsn_data *) data; |
| 1191 | tree t; |
| 1192 | gimple stmt; |
| 1193 | gimple_stmt_iterator gsi; |
| 1194 | tree type = TREE_TYPE (elt->new_name); |
| 1195 | tree load_struct; |
| 1196 | |
| 1197 | gsi = gsi_last_bb (clsn_data->store_bb); |
| 1198 | t = build3 (COMPONENT_REF, type, clsn_data->store, elt->field, NULL_TREE); |
| 1199 | stmt = gimple_build_assign (t, ssa_name (elt->version)); |
| 1200 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1201 | |
| 1202 | gsi = gsi_last_bb (clsn_data->load_bb); |
| 1203 | load_struct = build_simple_mem_ref (clsn_data->load); |
| 1204 | t = build3 (COMPONENT_REF, type, load_struct, elt->field, NULL_TREE); |
| 1205 | stmt = gimple_build_assign (elt->new_name, t); |
| 1206 | SSA_NAME_DEF_STMT (elt->new_name) = stmt; |
| 1207 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1208 | |
| 1209 | return 1; |
| 1210 | } |
| 1211 | |
| 1212 | /* Moves all the variables used in LOOP and defined outside of it (including |
| 1213 | the initial values of loop phi nodes, and *PER_THREAD if it is a ssa |
| 1214 | name) to a structure created for this purpose. The code |
| 1215 | |
| 1216 | while (1) |
| 1217 | { |
| 1218 | use (a); |
| 1219 | use (b); |
| 1220 | } |
| 1221 | |
| 1222 | is transformed this way: |
| 1223 | |
| 1224 | bb0: |
| 1225 | old.a = a; |
| 1226 | old.b = b; |
| 1227 | |
| 1228 | bb1: |
| 1229 | a' = new->a; |
| 1230 | b' = new->b; |
| 1231 | while (1) |
| 1232 | { |
| 1233 | use (a'); |
| 1234 | use (b'); |
| 1235 | } |
| 1236 | |
| 1237 | `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The |
| 1238 | pointer `new' is intentionally not initialized (the loop will be split to a |
| 1239 | separate function later, and `new' will be initialized from its arguments). |
| 1240 | LD_ST_DATA holds information about the shared data structure used to pass |
| 1241 | information among the threads. It is initialized here, and |
| 1242 | gen_parallel_loop will pass it to create_call_for_reduction that |
| 1243 | needs this information. REDUCTION_LIST describes the reductions |
| 1244 | in LOOP. */ |
| 1245 | |
| 1246 | static void |
| 1247 | separate_decls_in_region (edge entry, edge exit, htab_t reduction_list, |
| 1248 | tree *arg_struct, tree *new_arg_struct, |
| 1249 | struct clsn_data *ld_st_data) |
| 1250 | |
| 1251 | { |
| 1252 | basic_block bb1 = split_edge (entry); |
| 1253 | basic_block bb0 = single_pred (bb1); |
| 1254 | htab_t name_copies = htab_create (10, name_to_copy_elt_hash, |
| 1255 | name_to_copy_elt_eq, free); |
| 1256 | htab_t decl_copies = htab_create (10, int_tree_map_hash, int_tree_map_eq, |
| 1257 | free); |
| 1258 | unsigned i; |
| 1259 | tree type, type_name, nvar; |
| 1260 | gimple_stmt_iterator gsi; |
| 1261 | struct clsn_data clsn_data; |
| 1262 | VEC (basic_block, heap) *body = VEC_alloc (basic_block, heap, 3); |
| 1263 | basic_block bb; |
| 1264 | basic_block entry_bb = bb1; |
| 1265 | basic_block exit_bb = exit->dest; |
| 1266 | bool has_debug_stmt = false; |
| 1267 | |
| 1268 | entry = single_succ_edge (entry_bb); |
| 1269 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
| 1270 | |
| 1271 | FOR_EACH_VEC_ELT (basic_block, body, i, bb) |
| 1272 | { |
| 1273 | if (bb != entry_bb && bb != exit_bb) |
| 1274 | { |
| 1275 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1276 | separate_decls_in_region_stmt (entry, exit, gsi_stmt (gsi), |
| 1277 | name_copies, decl_copies); |
| 1278 | |
| 1279 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1280 | { |
| 1281 | gimple stmt = gsi_stmt (gsi); |
| 1282 | |
| 1283 | if (is_gimple_debug (stmt)) |
| 1284 | has_debug_stmt = true; |
| 1285 | else |
| 1286 | separate_decls_in_region_stmt (entry, exit, stmt, |
| 1287 | name_copies, decl_copies); |
| 1288 | } |
| 1289 | } |
| 1290 | } |
| 1291 | |
| 1292 | /* Now process debug bind stmts. We must not create decls while |
| 1293 | processing debug stmts, so we defer their processing so as to |
| 1294 | make sure we will have debug info for as many variables as |
| 1295 | possible (all of those that were dealt with in the loop above), |
| 1296 | and discard those for which we know there's nothing we can |
| 1297 | do. */ |
| 1298 | if (has_debug_stmt) |
| 1299 | FOR_EACH_VEC_ELT (basic_block, body, i, bb) |
| 1300 | if (bb != entry_bb && bb != exit_bb) |
| 1301 | { |
| 1302 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) |
| 1303 | { |
| 1304 | gimple stmt = gsi_stmt (gsi); |
| 1305 | |
| 1306 | if (is_gimple_debug (stmt)) |
| 1307 | { |
| 1308 | if (separate_decls_in_region_debug (stmt, name_copies, |
| 1309 | decl_copies)) |
| 1310 | { |
| 1311 | gsi_remove (&gsi, true); |
| 1312 | continue; |
| 1313 | } |
| 1314 | } |
| 1315 | |
| 1316 | gsi_next (&gsi); |
| 1317 | } |
| 1318 | } |
| 1319 | |
| 1320 | VEC_free (basic_block, heap, body); |
| 1321 | |
| 1322 | if (htab_elements (name_copies) == 0 && htab_elements (reduction_list) == 0) |
| 1323 | { |
| 1324 | /* It may happen that there is nothing to copy (if there are only |
| 1325 | loop carried and external variables in the loop). */ |
| 1326 | *arg_struct = NULL; |
| 1327 | *new_arg_struct = NULL; |
| 1328 | } |
| 1329 | else |
| 1330 | { |
| 1331 | /* Create the type for the structure to store the ssa names to. */ |
| 1332 | type = lang_hooks.types.make_type (RECORD_TYPE); |
| 1333 | type_name = build_decl (UNKNOWN_LOCATION, |
| 1334 | TYPE_DECL, create_tmp_var_name (".paral_data"), |
| 1335 | type); |
| 1336 | TYPE_NAME (type) = type_name; |
| 1337 | |
| 1338 | htab_traverse (name_copies, add_field_for_name, type); |
| 1339 | if (reduction_list && htab_elements (reduction_list) > 0) |
| 1340 | { |
| 1341 | /* Create the fields for reductions. */ |
| 1342 | htab_traverse (reduction_list, add_field_for_reduction, |
| 1343 | type); |
| 1344 | } |
| 1345 | layout_type (type); |
| 1346 | |
| 1347 | /* Create the loads and stores. */ |
| 1348 | *arg_struct = create_tmp_var (type, ".paral_data_store"); |
| 1349 | nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); |
| 1350 | *new_arg_struct = make_ssa_name (nvar, NULL); |
| 1351 | |
| 1352 | ld_st_data->store = *arg_struct; |
| 1353 | ld_st_data->load = *new_arg_struct; |
| 1354 | ld_st_data->store_bb = bb0; |
| 1355 | ld_st_data->load_bb = bb1; |
| 1356 | |
| 1357 | htab_traverse (name_copies, create_loads_and_stores_for_name, |
| 1358 | ld_st_data); |
| 1359 | |
| 1360 | /* Load the calculation from memory (after the join of the threads). */ |
| 1361 | |
| 1362 | if (reduction_list && htab_elements (reduction_list) > 0) |
| 1363 | { |
| 1364 | htab_traverse (reduction_list, create_stores_for_reduction, |
| 1365 | ld_st_data); |
| 1366 | clsn_data.load = make_ssa_name (nvar, NULL); |
| 1367 | clsn_data.load_bb = exit->dest; |
| 1368 | clsn_data.store = ld_st_data->store; |
| 1369 | create_final_loads_for_reduction (reduction_list, &clsn_data); |
| 1370 | } |
| 1371 | } |
| 1372 | |
| 1373 | htab_delete (decl_copies); |
| 1374 | htab_delete (name_copies); |
| 1375 | } |
| 1376 | |
| 1377 | /* Bitmap containing uids of functions created by parallelization. We cannot |
| 1378 | allocate it from the default obstack, as it must live across compilation |
| 1379 | of several functions; we make it gc allocated instead. */ |
| 1380 | |
| 1381 | static GTY(()) bitmap parallelized_functions; |
| 1382 | |
| 1383 | /* Returns true if FN was created by create_loop_fn. */ |
| 1384 | |
| 1385 | bool |
| 1386 | parallelized_function_p (tree fn) |
| 1387 | { |
| 1388 | if (!parallelized_functions || !DECL_ARTIFICIAL (fn)) |
| 1389 | return false; |
| 1390 | |
| 1391 | return bitmap_bit_p (parallelized_functions, DECL_UID (fn)); |
| 1392 | } |
| 1393 | |
| 1394 | /* Creates and returns an empty function that will receive the body of |
| 1395 | a parallelized loop. */ |
| 1396 | |
| 1397 | static tree |
| 1398 | create_loop_fn (location_t loc) |
| 1399 | { |
| 1400 | char buf[100]; |
| 1401 | char *tname; |
| 1402 | tree decl, type, name, t; |
| 1403 | struct function *act_cfun = cfun; |
| 1404 | static unsigned loopfn_num; |
| 1405 | |
Bernhard Rosenkraenzer | 84c1139 | 2012-09-27 01:39:09 +0159 | [diff] [blame] | 1406 | loc = LOCATION_LOCUS (loc); |
Bernhard Rosenkraenzer | c83ebe5 | 2012-09-18 21:38:03 +0159 | [diff] [blame] | 1407 | snprintf (buf, 100, "%s.$loopfn", current_function_name ()); |
| 1408 | ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); |
| 1409 | clean_symbol_name (tname); |
| 1410 | name = get_identifier (tname); |
| 1411 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); |
| 1412 | |
| 1413 | decl = build_decl (loc, FUNCTION_DECL, name, type); |
| 1414 | if (!parallelized_functions) |
| 1415 | parallelized_functions = BITMAP_GGC_ALLOC (); |
| 1416 | bitmap_set_bit (parallelized_functions, DECL_UID (decl)); |
| 1417 | |
| 1418 | TREE_STATIC (decl) = 1; |
| 1419 | TREE_USED (decl) = 1; |
| 1420 | DECL_ARTIFICIAL (decl) = 1; |
| 1421 | DECL_IGNORED_P (decl) = 0; |
| 1422 | TREE_PUBLIC (decl) = 0; |
| 1423 | DECL_UNINLINABLE (decl) = 1; |
| 1424 | DECL_EXTERNAL (decl) = 0; |
| 1425 | DECL_CONTEXT (decl) = NULL_TREE; |
| 1426 | DECL_INITIAL (decl) = make_node (BLOCK); |
| 1427 | |
| 1428 | t = build_decl (loc, RESULT_DECL, NULL_TREE, void_type_node); |
| 1429 | DECL_ARTIFICIAL (t) = 1; |
| 1430 | DECL_IGNORED_P (t) = 1; |
| 1431 | DECL_RESULT (decl) = t; |
| 1432 | |
| 1433 | t = build_decl (loc, PARM_DECL, get_identifier (".paral_data_param"), |
| 1434 | ptr_type_node); |
| 1435 | DECL_ARTIFICIAL (t) = 1; |
| 1436 | DECL_ARG_TYPE (t) = ptr_type_node; |
| 1437 | DECL_CONTEXT (t) = decl; |
| 1438 | TREE_USED (t) = 1; |
| 1439 | DECL_ARGUMENTS (decl) = t; |
| 1440 | |
| 1441 | allocate_struct_function (decl, false); |
| 1442 | |
| 1443 | /* The call to allocate_struct_function clobbers CFUN, so we need to restore |
| 1444 | it. */ |
| 1445 | set_cfun (act_cfun); |
| 1446 | |
| 1447 | return decl; |
| 1448 | } |
| 1449 | |
| 1450 | /* Moves the exit condition of LOOP to the beginning of its header, and |
| 1451 | duplicates the part of the last iteration that gets disabled to the |
| 1452 | exit of the loop. NIT is the number of iterations of the loop |
| 1453 | (used to initialize the variables in the duplicated part). |
| 1454 | |
| 1455 | TODO: the common case is that latch of the loop is empty and immediately |
| 1456 | follows the loop exit. In this case, it would be better not to copy the |
| 1457 | body of the loop, but only move the entry of the loop directly before the |
| 1458 | exit check and increase the number of iterations of the loop by one. |
| 1459 | This may need some additional preconditioning in case NIT = ~0. |
| 1460 | REDUCTION_LIST describes the reductions in LOOP. */ |
| 1461 | |
| 1462 | static void |
| 1463 | transform_to_exit_first_loop (struct loop *loop, htab_t reduction_list, tree nit) |
| 1464 | { |
| 1465 | basic_block *bbs, *nbbs, ex_bb, orig_header; |
| 1466 | unsigned n; |
| 1467 | bool ok; |
| 1468 | edge exit = single_dom_exit (loop), hpred; |
| 1469 | tree control, control_name, res, t; |
| 1470 | gimple phi, nphi, cond_stmt, stmt, cond_nit; |
| 1471 | gimple_stmt_iterator gsi; |
| 1472 | tree nit_1; |
| 1473 | |
| 1474 | split_block_after_labels (loop->header); |
| 1475 | orig_header = single_succ (loop->header); |
| 1476 | hpred = single_succ_edge (loop->header); |
| 1477 | |
| 1478 | cond_stmt = last_stmt (exit->src); |
| 1479 | control = gimple_cond_lhs (cond_stmt); |
| 1480 | gcc_assert (gimple_cond_rhs (cond_stmt) == nit); |
| 1481 | |
| 1482 | /* Make sure that we have phi nodes on exit for all loop header phis |
| 1483 | (create_parallel_loop requires that). */ |
| 1484 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1485 | { |
| 1486 | phi = gsi_stmt (gsi); |
| 1487 | res = PHI_RESULT (phi); |
| 1488 | t = copy_ssa_name (res, phi); |
| 1489 | SET_PHI_RESULT (phi, t); |
| 1490 | nphi = create_phi_node (res, orig_header); |
| 1491 | add_phi_arg (nphi, t, hpred, UNKNOWN_LOCATION); |
| 1492 | |
| 1493 | if (res == control) |
| 1494 | { |
| 1495 | gimple_cond_set_lhs (cond_stmt, t); |
| 1496 | update_stmt (cond_stmt); |
| 1497 | control = t; |
| 1498 | } |
| 1499 | } |
| 1500 | |
| 1501 | bbs = get_loop_body_in_dom_order (loop); |
| 1502 | |
| 1503 | for (n = 0; bbs[n] != exit->src; n++) |
| 1504 | continue; |
| 1505 | nbbs = XNEWVEC (basic_block, n); |
| 1506 | ok = gimple_duplicate_sese_tail (single_succ_edge (loop->header), exit, |
| 1507 | bbs + 1, n, nbbs); |
| 1508 | gcc_assert (ok); |
| 1509 | free (bbs); |
| 1510 | ex_bb = nbbs[0]; |
| 1511 | free (nbbs); |
| 1512 | |
| 1513 | /* Other than reductions, the only gimple reg that should be copied |
| 1514 | out of the loop is the control variable. */ |
| 1515 | exit = single_dom_exit (loop); |
| 1516 | control_name = NULL_TREE; |
| 1517 | for (gsi = gsi_start_phis (ex_bb); !gsi_end_p (gsi); ) |
| 1518 | { |
| 1519 | phi = gsi_stmt (gsi); |
| 1520 | res = PHI_RESULT (phi); |
| 1521 | if (virtual_operand_p (res)) |
| 1522 | { |
| 1523 | gsi_next (&gsi); |
| 1524 | continue; |
| 1525 | } |
| 1526 | |
| 1527 | /* Check if it is a part of reduction. If it is, |
| 1528 | keep the phi at the reduction's keep_res field. The |
| 1529 | PHI_RESULT of this phi is the resulting value of the reduction |
| 1530 | variable when exiting the loop. */ |
| 1531 | |
| 1532 | if (htab_elements (reduction_list) > 0) |
| 1533 | { |
| 1534 | struct reduction_info *red; |
| 1535 | |
| 1536 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
| 1537 | red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); |
| 1538 | if (red) |
| 1539 | { |
| 1540 | red->keep_res = phi; |
| 1541 | gsi_next (&gsi); |
| 1542 | continue; |
| 1543 | } |
| 1544 | } |
| 1545 | gcc_assert (control_name == NULL_TREE |
| 1546 | && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); |
| 1547 | control_name = res; |
| 1548 | remove_phi_node (&gsi, false); |
| 1549 | } |
| 1550 | gcc_assert (control_name != NULL_TREE); |
| 1551 | |
| 1552 | /* Initialize the control variable to number of iterations |
| 1553 | according to the rhs of the exit condition. */ |
| 1554 | gsi = gsi_after_labels (ex_bb); |
| 1555 | cond_nit = last_stmt (exit->src); |
| 1556 | nit_1 = gimple_cond_rhs (cond_nit); |
| 1557 | nit_1 = force_gimple_operand_gsi (&gsi, |
| 1558 | fold_convert (TREE_TYPE (control_name), nit_1), |
| 1559 | false, NULL_TREE, false, GSI_SAME_STMT); |
| 1560 | stmt = gimple_build_assign (control_name, nit_1); |
| 1561 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
| 1562 | SSA_NAME_DEF_STMT (control_name) = stmt; |
| 1563 | } |
| 1564 | |
| 1565 | /* Create the parallel constructs for LOOP as described in gen_parallel_loop. |
| 1566 | LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL. |
| 1567 | NEW_DATA is the variable that should be initialized from the argument |
| 1568 | of LOOP_FN. N_THREADS is the requested number of threads. Returns the |
| 1569 | basic block containing GIMPLE_OMP_PARALLEL tree. */ |
| 1570 | |
| 1571 | static basic_block |
| 1572 | create_parallel_loop (struct loop *loop, tree loop_fn, tree data, |
| 1573 | tree new_data, unsigned n_threads, location_t loc) |
| 1574 | { |
| 1575 | gimple_stmt_iterator gsi; |
| 1576 | basic_block bb, paral_bb, for_bb, ex_bb; |
| 1577 | tree t, param; |
| 1578 | gimple stmt, for_stmt, phi, cond_stmt; |
| 1579 | tree cvar, cvar_init, initvar, cvar_next, cvar_base, type; |
| 1580 | edge exit, nexit, guard, end, e; |
| 1581 | |
| 1582 | /* Prepare the GIMPLE_OMP_PARALLEL statement. */ |
| 1583 | bb = loop_preheader_edge (loop)->src; |
| 1584 | paral_bb = single_pred (bb); |
| 1585 | gsi = gsi_last_bb (paral_bb); |
| 1586 | |
| 1587 | t = build_omp_clause (loc, OMP_CLAUSE_NUM_THREADS); |
| 1588 | OMP_CLAUSE_NUM_THREADS_EXPR (t) |
| 1589 | = build_int_cst (integer_type_node, n_threads); |
| 1590 | stmt = gimple_build_omp_parallel (NULL, t, loop_fn, data); |
| 1591 | gimple_set_location (stmt, loc); |
| 1592 | |
| 1593 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1594 | |
| 1595 | /* Initialize NEW_DATA. */ |
| 1596 | if (data) |
| 1597 | { |
| 1598 | gsi = gsi_after_labels (bb); |
| 1599 | |
| 1600 | param = make_ssa_name (DECL_ARGUMENTS (loop_fn), NULL); |
| 1601 | stmt = gimple_build_assign (param, build_fold_addr_expr (data)); |
| 1602 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
| 1603 | SSA_NAME_DEF_STMT (param) = stmt; |
| 1604 | |
| 1605 | stmt = gimple_build_assign (new_data, |
| 1606 | fold_convert (TREE_TYPE (new_data), param)); |
| 1607 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
| 1608 | SSA_NAME_DEF_STMT (new_data) = stmt; |
| 1609 | } |
| 1610 | |
| 1611 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */ |
| 1612 | bb = split_loop_exit_edge (single_dom_exit (loop)); |
| 1613 | gsi = gsi_last_bb (bb); |
| 1614 | stmt = gimple_build_omp_return (false); |
| 1615 | gimple_set_location (stmt, loc); |
| 1616 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1617 | |
| 1618 | /* Extract data for GIMPLE_OMP_FOR. */ |
| 1619 | gcc_assert (loop->header == single_dom_exit (loop)->src); |
| 1620 | cond_stmt = last_stmt (loop->header); |
| 1621 | |
| 1622 | cvar = gimple_cond_lhs (cond_stmt); |
| 1623 | cvar_base = SSA_NAME_VAR (cvar); |
| 1624 | phi = SSA_NAME_DEF_STMT (cvar); |
| 1625 | cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
| 1626 | initvar = copy_ssa_name (cvar, NULL); |
| 1627 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), |
| 1628 | initvar); |
| 1629 | cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); |
| 1630 | |
| 1631 | gsi = gsi_last_nondebug_bb (loop->latch); |
| 1632 | gcc_assert (gsi_stmt (gsi) == SSA_NAME_DEF_STMT (cvar_next)); |
| 1633 | gsi_remove (&gsi, true); |
| 1634 | |
| 1635 | /* Prepare cfg. */ |
| 1636 | for_bb = split_edge (loop_preheader_edge (loop)); |
| 1637 | ex_bb = split_loop_exit_edge (single_dom_exit (loop)); |
| 1638 | extract_true_false_edges_from_block (loop->header, &nexit, &exit); |
| 1639 | gcc_assert (exit == single_dom_exit (loop)); |
| 1640 | |
| 1641 | guard = make_edge (for_bb, ex_bb, 0); |
| 1642 | single_succ_edge (loop->latch)->flags = 0; |
| 1643 | end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU); |
| 1644 | for (gsi = gsi_start_phis (ex_bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1645 | { |
| 1646 | source_location locus; |
| 1647 | tree def; |
| 1648 | phi = gsi_stmt (gsi); |
| 1649 | stmt = SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, exit)); |
| 1650 | |
| 1651 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_preheader_edge (loop)); |
| 1652 | locus = gimple_phi_arg_location_from_edge (stmt, |
| 1653 | loop_preheader_edge (loop)); |
| 1654 | add_phi_arg (phi, def, guard, locus); |
| 1655 | |
| 1656 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (loop)); |
| 1657 | locus = gimple_phi_arg_location_from_edge (stmt, loop_latch_edge (loop)); |
| 1658 | add_phi_arg (phi, def, end, locus); |
| 1659 | } |
| 1660 | e = redirect_edge_and_branch (exit, nexit->dest); |
| 1661 | PENDING_STMT (e) = NULL; |
| 1662 | |
| 1663 | /* Emit GIMPLE_OMP_FOR. */ |
| 1664 | gimple_cond_set_lhs (cond_stmt, cvar_base); |
| 1665 | type = TREE_TYPE (cvar); |
| 1666 | t = build_omp_clause (loc, OMP_CLAUSE_SCHEDULE); |
| 1667 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; |
| 1668 | |
| 1669 | for_stmt = gimple_build_omp_for (NULL, t, 1, NULL); |
| 1670 | gimple_set_location (for_stmt, loc); |
| 1671 | gimple_omp_for_set_index (for_stmt, 0, initvar); |
| 1672 | gimple_omp_for_set_initial (for_stmt, 0, cvar_init); |
| 1673 | gimple_omp_for_set_final (for_stmt, 0, gimple_cond_rhs (cond_stmt)); |
| 1674 | gimple_omp_for_set_cond (for_stmt, 0, gimple_cond_code (cond_stmt)); |
| 1675 | gimple_omp_for_set_incr (for_stmt, 0, build2 (PLUS_EXPR, type, |
| 1676 | cvar_base, |
| 1677 | build_int_cst (type, 1))); |
| 1678 | |
| 1679 | gsi = gsi_last_bb (for_bb); |
| 1680 | gsi_insert_after (&gsi, for_stmt, GSI_NEW_STMT); |
| 1681 | SSA_NAME_DEF_STMT (initvar) = for_stmt; |
| 1682 | |
| 1683 | /* Emit GIMPLE_OMP_CONTINUE. */ |
| 1684 | gsi = gsi_last_bb (loop->latch); |
| 1685 | stmt = gimple_build_omp_continue (cvar_next, cvar); |
| 1686 | gimple_set_location (stmt, loc); |
| 1687 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1688 | SSA_NAME_DEF_STMT (cvar_next) = stmt; |
| 1689 | |
| 1690 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */ |
| 1691 | gsi = gsi_last_bb (ex_bb); |
| 1692 | stmt = gimple_build_omp_return (true); |
| 1693 | gimple_set_location (stmt, loc); |
| 1694 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
| 1695 | |
| 1696 | /* After the above dom info is hosed. Re-compute it. */ |
| 1697 | free_dominance_info (CDI_DOMINATORS); |
| 1698 | calculate_dominance_info (CDI_DOMINATORS); |
| 1699 | |
| 1700 | return paral_bb; |
| 1701 | } |
| 1702 | |
| 1703 | /* Generates code to execute the iterations of LOOP in N_THREADS |
| 1704 | threads in parallel. |
| 1705 | |
| 1706 | NITER describes number of iterations of LOOP. |
| 1707 | REDUCTION_LIST describes the reductions existent in the LOOP. */ |
| 1708 | |
| 1709 | static void |
| 1710 | gen_parallel_loop (struct loop *loop, htab_t reduction_list, |
| 1711 | unsigned n_threads, struct tree_niter_desc *niter) |
| 1712 | { |
| 1713 | loop_iterator li; |
| 1714 | tree many_iterations_cond, type, nit; |
| 1715 | tree arg_struct, new_arg_struct; |
| 1716 | gimple_seq stmts; |
| 1717 | basic_block parallel_head; |
| 1718 | edge entry, exit; |
| 1719 | struct clsn_data clsn_data; |
| 1720 | unsigned prob; |
| 1721 | location_t loc; |
| 1722 | gimple cond_stmt; |
| 1723 | unsigned int m_p_thread=2; |
| 1724 | |
| 1725 | /* From |
| 1726 | |
| 1727 | --------------------------------------------------------------------- |
| 1728 | loop |
| 1729 | { |
| 1730 | IV = phi (INIT, IV + STEP) |
| 1731 | BODY1; |
| 1732 | if (COND) |
| 1733 | break; |
| 1734 | BODY2; |
| 1735 | } |
| 1736 | --------------------------------------------------------------------- |
| 1737 | |
| 1738 | with # of iterations NITER (possibly with MAY_BE_ZERO assumption), |
| 1739 | we generate the following code: |
| 1740 | |
| 1741 | --------------------------------------------------------------------- |
| 1742 | |
| 1743 | if (MAY_BE_ZERO |
| 1744 | || NITER < MIN_PER_THREAD * N_THREADS) |
| 1745 | goto original; |
| 1746 | |
| 1747 | BODY1; |
| 1748 | store all local loop-invariant variables used in body of the loop to DATA. |
| 1749 | GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); |
| 1750 | load the variables from DATA. |
| 1751 | GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) |
| 1752 | BODY2; |
| 1753 | BODY1; |
| 1754 | GIMPLE_OMP_CONTINUE; |
| 1755 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR |
| 1756 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL |
| 1757 | goto end; |
| 1758 | |
| 1759 | original: |
| 1760 | loop |
| 1761 | { |
| 1762 | IV = phi (INIT, IV + STEP) |
| 1763 | BODY1; |
| 1764 | if (COND) |
| 1765 | break; |
| 1766 | BODY2; |
| 1767 | } |
| 1768 | |
| 1769 | end: |
| 1770 | |
| 1771 | */ |
| 1772 | |
| 1773 | /* Create two versions of the loop -- in the old one, we know that the |
| 1774 | number of iterations is large enough, and we will transform it into the |
| 1775 | loop that will be split to loop_fn, the new one will be used for the |
| 1776 | remaining iterations. */ |
| 1777 | |
| 1778 | /* We should compute a better number-of-iterations value for outer loops. |
| 1779 | That is, if we have |
| 1780 | |
| 1781 | for (i = 0; i < n; ++i) |
| 1782 | for (j = 0; j < m; ++j) |
| 1783 | ... |
| 1784 | |
| 1785 | we should compute nit = n * m, not nit = n. |
| 1786 | Also may_be_zero handling would need to be adjusted. */ |
| 1787 | |
| 1788 | type = TREE_TYPE (niter->niter); |
| 1789 | nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, |
| 1790 | NULL_TREE); |
| 1791 | if (stmts) |
| 1792 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
| 1793 | |
| 1794 | if (loop->inner) |
| 1795 | m_p_thread=2; |
| 1796 | else |
| 1797 | m_p_thread=MIN_PER_THREAD; |
| 1798 | |
| 1799 | many_iterations_cond = |
| 1800 | fold_build2 (GE_EXPR, boolean_type_node, |
| 1801 | nit, build_int_cst (type, m_p_thread * n_threads)); |
| 1802 | |
| 1803 | many_iterations_cond |
| 1804 | = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
| 1805 | invert_truthvalue (unshare_expr (niter->may_be_zero)), |
| 1806 | many_iterations_cond); |
| 1807 | many_iterations_cond |
| 1808 | = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); |
| 1809 | if (stmts) |
| 1810 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
| 1811 | if (!is_gimple_condexpr (many_iterations_cond)) |
| 1812 | { |
| 1813 | many_iterations_cond |
| 1814 | = force_gimple_operand (many_iterations_cond, &stmts, |
| 1815 | true, NULL_TREE); |
| 1816 | if (stmts) |
| 1817 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
| 1818 | } |
| 1819 | |
| 1820 | initialize_original_copy_tables (); |
| 1821 | |
| 1822 | /* We assume that the loop usually iterates a lot. */ |
| 1823 | prob = 4 * REG_BR_PROB_BASE / 5; |
| 1824 | loop_version (loop, many_iterations_cond, NULL, |
| 1825 | prob, prob, REG_BR_PROB_BASE - prob, true); |
| 1826 | update_ssa (TODO_update_ssa); |
| 1827 | free_original_copy_tables (); |
| 1828 | |
| 1829 | /* Base all the induction variables in LOOP on a single control one. */ |
| 1830 | canonicalize_loop_ivs (loop, &nit, true); |
| 1831 | |
| 1832 | /* Ensure that the exit condition is the first statement in the loop. */ |
| 1833 | transform_to_exit_first_loop (loop, reduction_list, nit); |
| 1834 | |
| 1835 | /* Generate initializations for reductions. */ |
| 1836 | if (htab_elements (reduction_list) > 0) |
| 1837 | htab_traverse (reduction_list, initialize_reductions, loop); |
| 1838 | |
| 1839 | /* Eliminate the references to local variables from the loop. */ |
| 1840 | gcc_assert (single_exit (loop)); |
| 1841 | entry = loop_preheader_edge (loop); |
| 1842 | exit = single_dom_exit (loop); |
| 1843 | |
| 1844 | eliminate_local_variables (entry, exit); |
| 1845 | /* In the old loop, move all variables non-local to the loop to a structure |
| 1846 | and back, and create separate decls for the variables used in loop. */ |
| 1847 | separate_decls_in_region (entry, exit, reduction_list, &arg_struct, |
| 1848 | &new_arg_struct, &clsn_data); |
| 1849 | |
| 1850 | /* Create the parallel constructs. */ |
| 1851 | loc = UNKNOWN_LOCATION; |
| 1852 | cond_stmt = last_stmt (loop->header); |
| 1853 | if (cond_stmt) |
| 1854 | loc = gimple_location (cond_stmt); |
| 1855 | parallel_head = create_parallel_loop (loop, create_loop_fn (loc), arg_struct, |
| 1856 | new_arg_struct, n_threads, loc); |
| 1857 | if (htab_elements (reduction_list) > 0) |
| 1858 | create_call_for_reduction (loop, reduction_list, &clsn_data); |
| 1859 | |
| 1860 | scev_reset (); |
| 1861 | |
| 1862 | /* Cancel the loop (it is simpler to do it here rather than to teach the |
| 1863 | expander to do it). */ |
| 1864 | cancel_loop_tree (loop); |
| 1865 | |
| 1866 | /* Free loop bound estimations that could contain references to |
| 1867 | removed statements. */ |
| 1868 | FOR_EACH_LOOP (li, loop, 0) |
| 1869 | free_numbers_of_iterations_estimates_loop (loop); |
| 1870 | |
| 1871 | /* Expand the parallel constructs. We do it directly here instead of running |
| 1872 | a separate expand_omp pass, since it is more efficient, and less likely to |
| 1873 | cause troubles with further analyses not being able to deal with the |
| 1874 | OMP trees. */ |
| 1875 | |
| 1876 | omp_expand_local (parallel_head); |
| 1877 | } |
| 1878 | |
| 1879 | /* Returns true when LOOP contains vector phi nodes. */ |
| 1880 | |
| 1881 | static bool |
| 1882 | loop_has_vector_phi_nodes (struct loop *loop ATTRIBUTE_UNUSED) |
| 1883 | { |
| 1884 | unsigned i; |
| 1885 | basic_block *bbs = get_loop_body_in_dom_order (loop); |
| 1886 | gimple_stmt_iterator gsi; |
| 1887 | bool res = true; |
| 1888 | |
| 1889 | for (i = 0; i < loop->num_nodes; i++) |
| 1890 | for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1891 | if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi_stmt (gsi)))) == VECTOR_TYPE) |
| 1892 | goto end; |
| 1893 | |
| 1894 | res = false; |
| 1895 | end: |
| 1896 | free (bbs); |
| 1897 | return res; |
| 1898 | } |
| 1899 | |
| 1900 | /* Create a reduction_info struct, initialize it with REDUC_STMT |
| 1901 | and PHI, insert it to the REDUCTION_LIST. */ |
| 1902 | |
| 1903 | static void |
| 1904 | build_new_reduction (htab_t reduction_list, gimple reduc_stmt, gimple phi) |
| 1905 | { |
| 1906 | PTR *slot; |
| 1907 | struct reduction_info *new_reduction; |
| 1908 | |
| 1909 | gcc_assert (reduc_stmt); |
| 1910 | |
| 1911 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 1912 | { |
| 1913 | fprintf (dump_file, |
| 1914 | "Detected reduction. reduction stmt is: \n"); |
| 1915 | print_gimple_stmt (dump_file, reduc_stmt, 0, 0); |
| 1916 | fprintf (dump_file, "\n"); |
| 1917 | } |
| 1918 | |
| 1919 | new_reduction = XCNEW (struct reduction_info); |
| 1920 | |
| 1921 | new_reduction->reduc_stmt = reduc_stmt; |
| 1922 | new_reduction->reduc_phi = phi; |
| 1923 | new_reduction->reduc_version = SSA_NAME_VERSION (gimple_phi_result (phi)); |
| 1924 | new_reduction->reduction_code = gimple_assign_rhs_code (reduc_stmt); |
| 1925 | slot = htab_find_slot (reduction_list, new_reduction, INSERT); |
| 1926 | *slot = new_reduction; |
| 1927 | } |
| 1928 | |
| 1929 | /* Callback for htab_traverse. Sets gimple_uid of reduc_phi stmts. */ |
| 1930 | |
| 1931 | static int |
| 1932 | set_reduc_phi_uids (void **slot, void *data ATTRIBUTE_UNUSED) |
| 1933 | { |
| 1934 | struct reduction_info *const red = (struct reduction_info *) *slot; |
| 1935 | gimple_set_uid (red->reduc_phi, red->reduc_version); |
| 1936 | return 1; |
| 1937 | } |
| 1938 | |
| 1939 | /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */ |
| 1940 | |
| 1941 | static void |
| 1942 | gather_scalar_reductions (loop_p loop, htab_t reduction_list) |
| 1943 | { |
| 1944 | gimple_stmt_iterator gsi; |
| 1945 | loop_vec_info simple_loop_info; |
| 1946 | |
Bernhard Rosenkraenzer | c83ebe5 | 2012-09-18 21:38:03 +0159 | [diff] [blame] | 1947 | simple_loop_info = vect_analyze_loop_form (loop); |
| 1948 | |
| 1949 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 1950 | { |
| 1951 | gimple phi = gsi_stmt (gsi); |
| 1952 | affine_iv iv; |
| 1953 | tree res = PHI_RESULT (phi); |
| 1954 | bool double_reduc; |
| 1955 | |
| 1956 | if (virtual_operand_p (res)) |
| 1957 | continue; |
| 1958 | |
| 1959 | if (!simple_iv (loop, loop, res, &iv, true) |
| 1960 | && simple_loop_info) |
| 1961 | { |
| 1962 | gimple reduc_stmt = vect_force_simple_reduction (simple_loop_info, |
| 1963 | phi, true, |
| 1964 | &double_reduc); |
| 1965 | if (reduc_stmt && !double_reduc) |
| 1966 | build_new_reduction (reduction_list, reduc_stmt, phi); |
| 1967 | } |
| 1968 | } |
| 1969 | destroy_loop_vec_info (simple_loop_info, true); |
| 1970 | |
| 1971 | /* As gimple_uid is used by the vectorizer in between vect_analyze_loop_form |
| 1972 | and destroy_loop_vec_info, we can set gimple_uid of reduc_phi stmts |
| 1973 | only now. */ |
| 1974 | htab_traverse (reduction_list, set_reduc_phi_uids, NULL); |
| 1975 | } |
| 1976 | |
| 1977 | /* Try to initialize NITER for code generation part. */ |
| 1978 | |
| 1979 | static bool |
| 1980 | try_get_loop_niter (loop_p loop, struct tree_niter_desc *niter) |
| 1981 | { |
| 1982 | edge exit = single_dom_exit (loop); |
| 1983 | |
| 1984 | gcc_assert (exit); |
| 1985 | |
| 1986 | /* We need to know # of iterations, and there should be no uses of values |
| 1987 | defined inside loop outside of it, unless the values are invariants of |
| 1988 | the loop. */ |
| 1989 | if (!number_of_iterations_exit (loop, exit, niter, false)) |
| 1990 | { |
| 1991 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 1992 | fprintf (dump_file, " FAILED: number of iterations not known\n"); |
| 1993 | return false; |
| 1994 | } |
| 1995 | |
| 1996 | return true; |
| 1997 | } |
| 1998 | |
| 1999 | /* Try to initialize REDUCTION_LIST for code generation part. |
| 2000 | REDUCTION_LIST describes the reductions. */ |
| 2001 | |
| 2002 | static bool |
| 2003 | try_create_reduction_list (loop_p loop, htab_t reduction_list) |
| 2004 | { |
| 2005 | edge exit = single_dom_exit (loop); |
| 2006 | gimple_stmt_iterator gsi; |
| 2007 | |
| 2008 | gcc_assert (exit); |
| 2009 | |
| 2010 | gather_scalar_reductions (loop, reduction_list); |
| 2011 | |
| 2012 | |
| 2013 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 2014 | { |
| 2015 | gimple phi = gsi_stmt (gsi); |
| 2016 | struct reduction_info *red; |
| 2017 | imm_use_iterator imm_iter; |
| 2018 | use_operand_p use_p; |
| 2019 | gimple reduc_phi; |
| 2020 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
| 2021 | |
| 2022 | if (!virtual_operand_p (val)) |
| 2023 | { |
| 2024 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2025 | { |
| 2026 | fprintf (dump_file, "phi is "); |
| 2027 | print_gimple_stmt (dump_file, phi, 0, 0); |
| 2028 | fprintf (dump_file, "arg of phi to exit: value "); |
| 2029 | print_generic_expr (dump_file, val, 0); |
| 2030 | fprintf (dump_file, " used outside loop\n"); |
| 2031 | fprintf (dump_file, |
| 2032 | " checking if it a part of reduction pattern: \n"); |
| 2033 | } |
| 2034 | if (htab_elements (reduction_list) == 0) |
| 2035 | { |
| 2036 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2037 | fprintf (dump_file, |
| 2038 | " FAILED: it is not a part of reduction.\n"); |
| 2039 | return false; |
| 2040 | } |
| 2041 | reduc_phi = NULL; |
| 2042 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) |
| 2043 | { |
| 2044 | if (!gimple_debug_bind_p (USE_STMT (use_p)) |
| 2045 | && flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) |
| 2046 | { |
| 2047 | reduc_phi = USE_STMT (use_p); |
| 2048 | break; |
| 2049 | } |
| 2050 | } |
| 2051 | red = reduction_phi (reduction_list, reduc_phi); |
| 2052 | if (red == NULL) |
| 2053 | { |
| 2054 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2055 | fprintf (dump_file, |
| 2056 | " FAILED: it is not a part of reduction.\n"); |
| 2057 | return false; |
| 2058 | } |
| 2059 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2060 | { |
| 2061 | fprintf (dump_file, "reduction phi is "); |
| 2062 | print_gimple_stmt (dump_file, red->reduc_phi, 0, 0); |
| 2063 | fprintf (dump_file, "reduction stmt is "); |
| 2064 | print_gimple_stmt (dump_file, red->reduc_stmt, 0, 0); |
| 2065 | } |
| 2066 | } |
| 2067 | } |
| 2068 | |
| 2069 | /* The iterations of the loop may communicate only through bivs whose |
| 2070 | iteration space can be distributed efficiently. */ |
| 2071 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 2072 | { |
| 2073 | gimple phi = gsi_stmt (gsi); |
| 2074 | tree def = PHI_RESULT (phi); |
| 2075 | affine_iv iv; |
| 2076 | |
| 2077 | if (!virtual_operand_p (def) && !simple_iv (loop, loop, def, &iv, true)) |
| 2078 | { |
| 2079 | struct reduction_info *red; |
| 2080 | |
| 2081 | red = reduction_phi (reduction_list, phi); |
| 2082 | if (red == NULL) |
| 2083 | { |
| 2084 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2085 | fprintf (dump_file, |
| 2086 | " FAILED: scalar dependency between iterations\n"); |
| 2087 | return false; |
| 2088 | } |
| 2089 | } |
| 2090 | } |
| 2091 | |
| 2092 | |
| 2093 | return true; |
| 2094 | } |
| 2095 | |
| 2096 | /* Detect parallel loops and generate parallel code using libgomp |
| 2097 | primitives. Returns true if some loop was parallelized, false |
| 2098 | otherwise. */ |
| 2099 | |
| 2100 | bool |
| 2101 | parallelize_loops (void) |
| 2102 | { |
| 2103 | unsigned n_threads = flag_tree_parallelize_loops; |
| 2104 | bool changed = false; |
| 2105 | struct loop *loop; |
| 2106 | struct tree_niter_desc niter_desc; |
| 2107 | loop_iterator li; |
| 2108 | htab_t reduction_list; |
| 2109 | struct obstack parloop_obstack; |
| 2110 | HOST_WIDE_INT estimated; |
| 2111 | LOC loop_loc; |
| 2112 | |
| 2113 | /* Do not parallelize loops in the functions created by parallelization. */ |
| 2114 | if (parallelized_function_p (cfun->decl)) |
| 2115 | return false; |
| 2116 | if (cfun->has_nonlocal_label) |
| 2117 | return false; |
| 2118 | |
| 2119 | gcc_obstack_init (&parloop_obstack); |
| 2120 | reduction_list = htab_create (10, reduction_info_hash, |
| 2121 | reduction_info_eq, free); |
| 2122 | init_stmt_vec_info_vec (); |
| 2123 | |
| 2124 | FOR_EACH_LOOP (li, loop, 0) |
| 2125 | { |
| 2126 | htab_empty (reduction_list); |
| 2127 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2128 | { |
| 2129 | fprintf (dump_file, "Trying loop %d as candidate\n",loop->num); |
| 2130 | if (loop->inner) |
| 2131 | fprintf (dump_file, "loop %d is not innermost\n",loop->num); |
| 2132 | else |
| 2133 | fprintf (dump_file, "loop %d is innermost\n",loop->num); |
| 2134 | } |
| 2135 | |
| 2136 | /* If we use autopar in graphite pass, we use its marked dependency |
| 2137 | checking results. */ |
| 2138 | if (flag_loop_parallelize_all && !loop->can_be_parallel) |
| 2139 | { |
| 2140 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2141 | fprintf (dump_file, "loop is not parallel according to graphite\n"); |
| 2142 | continue; |
| 2143 | } |
| 2144 | |
| 2145 | if (!single_dom_exit (loop)) |
| 2146 | { |
| 2147 | |
| 2148 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2149 | fprintf (dump_file, "loop is !single_dom_exit\n"); |
| 2150 | |
| 2151 | continue; |
| 2152 | } |
| 2153 | |
| 2154 | if (/* And of course, the loop must be parallelizable. */ |
| 2155 | !can_duplicate_loop_p (loop) |
| 2156 | || loop_has_blocks_with_irreducible_flag (loop) |
| 2157 | || (loop_preheader_edge (loop)->src->flags & BB_IRREDUCIBLE_LOOP) |
| 2158 | /* FIXME: the check for vector phi nodes could be removed. */ |
| 2159 | || loop_has_vector_phi_nodes (loop)) |
| 2160 | continue; |
| 2161 | |
| 2162 | estimated = estimated_stmt_executions_int (loop); |
| 2163 | if (estimated == -1) |
| 2164 | estimated = max_stmt_executions_int (loop); |
| 2165 | /* FIXME: Bypass this check as graphite doesn't update the |
| 2166 | count and frequency correctly now. */ |
| 2167 | if (!flag_loop_parallelize_all |
| 2168 | && ((estimated != -1 |
| 2169 | && estimated <= (HOST_WIDE_INT) n_threads * MIN_PER_THREAD) |
| 2170 | /* Do not bother with loops in cold areas. */ |
| 2171 | || optimize_loop_nest_for_size_p (loop))) |
| 2172 | continue; |
| 2173 | |
| 2174 | if (!try_get_loop_niter (loop, &niter_desc)) |
| 2175 | continue; |
| 2176 | |
| 2177 | if (!try_create_reduction_list (loop, reduction_list)) |
| 2178 | continue; |
| 2179 | |
| 2180 | if (!flag_loop_parallelize_all |
| 2181 | && !loop_parallel_p (loop, &parloop_obstack)) |
| 2182 | continue; |
| 2183 | |
| 2184 | changed = true; |
| 2185 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2186 | { |
| 2187 | if (loop->inner) |
| 2188 | fprintf (dump_file, "parallelizing outer loop %d\n",loop->header->index); |
| 2189 | else |
| 2190 | fprintf (dump_file, "parallelizing inner loop %d\n",loop->header->index); |
| 2191 | loop_loc = find_loop_location (loop); |
| 2192 | if (loop_loc != UNKNOWN_LOC) |
| 2193 | fprintf (dump_file, "\nloop at %s:%d: ", |
| 2194 | LOC_FILE (loop_loc), LOC_LINE (loop_loc)); |
| 2195 | } |
| 2196 | gen_parallel_loop (loop, reduction_list, |
| 2197 | n_threads, &niter_desc); |
| 2198 | #ifdef ENABLE_CHECKING |
| 2199 | verify_flow_info (); |
| 2200 | verify_loop_structure (); |
| 2201 | verify_loop_closed_ssa (true); |
| 2202 | #endif |
| 2203 | } |
| 2204 | |
| 2205 | free_stmt_vec_info_vec (); |
| 2206 | htab_delete (reduction_list); |
| 2207 | obstack_free (&parloop_obstack, NULL); |
| 2208 | |
| 2209 | /* Parallelization will cause new function calls to be inserted through |
| 2210 | which local variables will escape. Reset the points-to solution |
| 2211 | for ESCAPED. */ |
| 2212 | if (changed) |
| 2213 | pt_solution_reset (&cfun->gimple_df->escaped); |
| 2214 | |
| 2215 | return changed; |
| 2216 | } |
| 2217 | |
| 2218 | #include "gt-tree-parloops.h" |