Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-loop-distribution.c @ 158:494b0b89df80 default tip
...
| author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
|---|---|
| date | Mon, 25 May 2020 18:13:55 +0900 |
| parents | 1830386684a0 |
| children |
| rev | line source |
|---|---|
| 0 | 1 /* Loop distribution. |
| 145 | 2 Copyright (C) 2006-2020 Free Software Foundation, Inc. |
| 0 | 3 Contributed by Georges-Andre Silber <Georges-Andre.Silber@ensmp.fr> |
| 4 and Sebastian Pop <sebastian.pop@amd.com>. | |
| 5 | |
| 6 This file is part of GCC. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
7 |
| 0 | 8 GCC is free software; you can redistribute it and/or modify it |
| 9 under the terms of the GNU General Public License as published by the | |
| 10 Free Software Foundation; either version 3, or (at your option) any | |
| 11 later version. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
12 |
| 0 | 13 GCC is distributed in the hope that it will be useful, but WITHOUT |
| 14 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 16 for more details. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
17 |
| 0 | 18 You should have received a copy of the GNU General Public License |
| 19 along with GCC; see the file COPYING3. If not see | |
| 20 <http://www.gnu.org/licenses/>. */ | |
| 21 | |
| 22 /* This pass performs loop distribution: for example, the loop | |
| 23 | |
| 24 |DO I = 2, N | |
| 25 | A(I) = B(I) + C | |
| 26 | D(I) = A(I-1)*E | |
| 27 |ENDDO | |
| 28 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
29 is transformed to |
| 0 | 30 |
| 31 |DOALL I = 2, N | |
| 32 | A(I) = B(I) + C | |
| 33 |ENDDO | |
| 34 | | |
| 35 |DOALL I = 2, N | |
| 36 | D(I) = A(I-1)*E | |
| 37 |ENDDO | |
| 38 | |
| 111 | 39 Loop distribution is the dual of loop fusion. It separates statements |
| 40 of a loop (or loop nest) into multiple loops (or loop nests) with the | |
| 41 same loop header. The major goal is to separate statements which may | |
| 42 be vectorized from those that can't. This pass implements distribution | |
| 43 in the following steps: | |
| 44 | |
| 45 1) Seed partitions with specific type statements. For now we support | |
| 46 two types seed statements: statement defining variable used outside | |
| 47 of loop; statement storing to memory. | |
| 48 2) Build reduced dependence graph (RDG) for loop to be distributed. | |
| 49 The vertices (RDG:V) model all statements in the loop and the edges | |
| 50 (RDG:E) model flow and control dependencies between statements. | |
| 51 3) Apart from RDG, compute data dependencies between memory references. | |
| 52 4) Starting from seed statement, build up partition by adding depended | |
| 53 statements according to RDG's dependence information. Partition is | |
| 54 classified as parallel type if it can be executed paralleled; or as | |
| 55 sequential type if it can't. Parallel type partition is further | |
| 56 classified as different builtin kinds if it can be implemented as | |
| 57 builtin function calls. | |
| 58 5) Build partition dependence graph (PG) based on data dependencies. | |
| 59 The vertices (PG:V) model all partitions and the edges (PG:E) model | |
| 60 all data dependencies between every partitions pair. In general, | |
| 61 data dependence is either compilation time known or unknown. In C | |
| 62 family languages, there exists quite amount compilation time unknown | |
| 63 dependencies because of possible alias relation of data references. | |
| 64 We categorize PG's edge to two types: "true" edge that represents | |
| 65 compilation time known data dependencies; "alias" edge for all other | |
| 66 data dependencies. | |
| 67 6) Traverse subgraph of PG as if all "alias" edges don't exist. Merge | |
| 68 partitions in each strong connected component (SCC) correspondingly. | |
| 69 Build new PG for merged partitions. | |
| 70 7) Traverse PG again and this time with both "true" and "alias" edges | |
| 71 included. We try to break SCCs by removing some edges. Because | |
| 72 SCCs by "true" edges are all fused in step 6), we can break SCCs | |
| 73 by removing some "alias" edges. It's NP-hard to choose optimal | |
| 74 edge set, fortunately simple approximation is good enough for us | |
| 75 given the small problem scale. | |
| 76 8) Collect all data dependencies of the removed "alias" edges. Create | |
| 77 runtime alias checks for collected data dependencies. | |
| 78 9) Version loop under the condition of runtime alias checks. Given | |
| 79 loop distribution generally introduces additional overhead, it is | |
| 80 only useful if vectorization is achieved in distributed loop. We | |
| 81 version loop with internal function call IFN_LOOP_DIST_ALIAS. If | |
| 82 no distributed loop can be vectorized, we simply remove distributed | |
| 83 loops and recover to the original one. | |
| 84 | |
| 85 TODO: | |
| 86 1) We only distribute innermost two-level loop nest now. We should | |
| 87 extend it for arbitrary loop nests in the future. | |
| 88 2) We only fuse partitions in SCC now. A better fusion algorithm is | |
| 89 desired to minimize loop overhead, maximize parallelism and maximize | |
| 90 data reuse. */ | |
| 0 | 91 |
| 92 #include "config.h" | |
| 93 #include "system.h" | |
| 94 #include "coretypes.h" | |
| 111 | 95 #include "backend.h" |
| 96 #include "tree.h" | |
| 97 #include "gimple.h" | |
| 98 #include "cfghooks.h" | |
| 0 | 99 #include "tree-pass.h" |
| 111 | 100 #include "ssa.h" |
| 101 #include "gimple-pretty-print.h" | |
| 102 #include "fold-const.h" | |
| 103 #include "cfganal.h" | |
| 104 #include "gimple-iterator.h" | |
| 105 #include "gimplify-me.h" | |
| 106 #include "stor-layout.h" | |
| 107 #include "tree-cfg.h" | |
| 108 #include "tree-ssa-loop-manip.h" | |
| 109 #include "tree-ssa-loop-ivopts.h" | |
| 110 #include "tree-ssa-loop.h" | |
| 111 #include "tree-into-ssa.h" | |
| 112 #include "tree-ssa.h" | |
| 113 #include "cfgloop.h" | |
| 114 #include "tree-scalar-evolution.h" | |
| 115 #include "tree-vectorizer.h" | |
| 145 | 116 #include "tree-eh.h" |
| 117 #include "gimple-fold.h" | |
| 111 | 118 |
| 119 | |
| 120 #define MAX_DATAREFS_NUM \ | |
| 145 | 121 ((unsigned) param_loop_max_datarefs_for_datadeps) |
| 111 | 122 |
| 123 /* Threshold controlling number of distributed partitions. Given it may | |
| 124 be unnecessary if a memory stream cost model is invented in the future, | |
| 125 we define it as a temporary macro, rather than a parameter. */ | |
| 126 #define NUM_PARTITION_THRESHOLD (4) | |
| 127 | |
| 128 /* Hashtable helpers. */ | |
| 129 | |
| 130 struct ddr_hasher : nofree_ptr_hash <struct data_dependence_relation> | |
| 131 { | |
| 132 static inline hashval_t hash (const data_dependence_relation *); | |
| 133 static inline bool equal (const data_dependence_relation *, | |
| 134 const data_dependence_relation *); | |
| 135 }; | |
| 136 | |
| 137 /* Hash function for data dependence. */ | |
| 138 | |
| 139 inline hashval_t | |
| 140 ddr_hasher::hash (const data_dependence_relation *ddr) | |
| 141 { | |
| 142 inchash::hash h; | |
| 143 h.add_ptr (DDR_A (ddr)); | |
| 144 h.add_ptr (DDR_B (ddr)); | |
| 145 return h.end (); | |
| 146 } | |
| 147 | |
| 148 /* Hash table equality function for data dependence. */ | |
| 149 | |
| 150 inline bool | |
| 151 ddr_hasher::equal (const data_dependence_relation *ddr1, | |
| 152 const data_dependence_relation *ddr2) | |
| 153 { | |
| 154 return (DDR_A (ddr1) == DDR_A (ddr2) && DDR_B (ddr1) == DDR_B (ddr2)); | |
| 155 } | |
| 156 | |
| 145 | 157 |
| 158 | |
| 111 | 159 #define DR_INDEX(dr) ((uintptr_t) (dr)->aux) |
| 160 | |
| 161 /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ | |
| 162 struct rdg_vertex | |
| 163 { | |
| 164 /* The statement represented by this vertex. */ | |
| 165 gimple *stmt; | |
| 166 | |
| 167 /* Vector of data-references in this statement. */ | |
| 168 vec<data_reference_p> datarefs; | |
| 169 | |
| 170 /* True when the statement contains a write to memory. */ | |
| 171 bool has_mem_write; | |
| 172 | |
| 173 /* True when the statement contains a read from memory. */ | |
| 174 bool has_mem_reads; | |
| 175 }; | |
| 176 | |
| 177 #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
| 178 #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
| 179 #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
| 180 #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
| 181 #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
| 182 #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
| 183 #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
| 184 #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
| 185 | |
| 186 /* Data dependence type. */ | |
| 187 | |
| 188 enum rdg_dep_type | |
| 189 { | |
| 190 /* Read After Write (RAW). */ | |
| 191 flow_dd = 'f', | |
| 192 | |
| 193 /* Control dependence (execute conditional on). */ | |
| 194 control_dd = 'c' | |
| 195 }; | |
| 196 | |
| 197 /* Dependence information attached to an edge of the RDG. */ | |
| 198 | |
| 199 struct rdg_edge | |
| 200 { | |
| 201 /* Type of the dependence. */ | |
| 202 enum rdg_dep_type type; | |
| 203 }; | |
| 204 | |
| 205 #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
| 206 | |
| 145 | 207 /* Kind of distributed loop. */ |
| 208 enum partition_kind { | |
| 209 PKIND_NORMAL, | |
| 210 /* Partial memset stands for a paritition can be distributed into a loop | |
| 211 of memset calls, rather than a single memset call. It's handled just | |
| 212 like a normal parition, i.e, distributed as separate loop, no memset | |
| 213 call is generated. | |
| 214 | |
| 215 Note: This is a hacking fix trying to distribute ZERO-ing stmt in a | |
| 216 loop nest as deep as possible. As a result, parloop achieves better | |
| 217 parallelization by parallelizing deeper loop nest. This hack should | |
| 218 be unnecessary and removed once distributed memset can be understood | |
| 219 and analyzed in data reference analysis. See PR82604 for more. */ | |
| 220 PKIND_PARTIAL_MEMSET, | |
| 221 PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE | |
| 222 }; | |
| 223 | |
| 224 /* Type of distributed loop. */ | |
| 225 enum partition_type { | |
| 226 /* The distributed loop can be executed parallelly. */ | |
| 227 PTYPE_PARALLEL = 0, | |
| 228 /* The distributed loop has to be executed sequentially. */ | |
| 229 PTYPE_SEQUENTIAL | |
| 230 }; | |
| 231 | |
| 232 /* Builtin info for loop distribution. */ | |
| 233 struct builtin_info | |
| 234 { | |
| 235 /* data-references a kind != PKIND_NORMAL partition is about. */ | |
| 236 data_reference_p dst_dr; | |
| 237 data_reference_p src_dr; | |
| 238 /* Base address and size of memory objects operated by the builtin. Note | |
| 239 both dest and source memory objects must have the same size. */ | |
| 240 tree dst_base; | |
| 241 tree src_base; | |
| 242 tree size; | |
| 243 /* Base and offset part of dst_base after stripping constant offset. This | |
| 244 is only used in memset builtin distribution for now. */ | |
| 245 tree dst_base_base; | |
| 246 unsigned HOST_WIDE_INT dst_base_offset; | |
| 247 }; | |
| 248 | |
| 249 /* Partition for loop distribution. */ | |
| 250 struct partition | |
| 251 { | |
| 252 /* Statements of the partition. */ | |
| 253 bitmap stmts; | |
| 254 /* True if the partition defines variable which is used outside of loop. */ | |
| 255 bool reduction_p; | |
| 256 location_t loc; | |
| 257 enum partition_kind kind; | |
| 258 enum partition_type type; | |
| 259 /* Data references in the partition. */ | |
| 260 bitmap datarefs; | |
| 261 /* Information of builtin parition. */ | |
| 262 struct builtin_info *builtin; | |
| 263 }; | |
| 264 | |
| 265 /* Partitions are fused because of different reasons. */ | |
| 266 enum fuse_type | |
| 267 { | |
| 268 FUSE_NON_BUILTIN = 0, | |
| 269 FUSE_REDUCTION = 1, | |
| 270 FUSE_SHARE_REF = 2, | |
| 271 FUSE_SAME_SCC = 3, | |
| 272 FUSE_FINALIZE = 4 | |
| 273 }; | |
| 274 | |
| 275 /* Description on different fusing reason. */ | |
| 276 static const char *fuse_message[] = { | |
| 277 "they are non-builtins", | |
| 278 "they have reductions", | |
| 279 "they have shared memory refs", | |
| 280 "they are in the same dependence scc", | |
| 281 "there is no point to distribute loop"}; | |
| 282 | |
| 283 | |
| 111 | 284 /* Dump vertex I in RDG to FILE. */ |
| 285 | |
| 286 static void | |
| 287 dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
| 288 { | |
| 289 struct vertex *v = &(rdg->vertices[i]); | |
| 290 struct graph_edge *e; | |
| 291 | |
| 292 fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
| 293 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
| 294 RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
| 295 | |
| 296 if (v->pred) | |
| 297 for (e = v->pred; e; e = e->pred_next) | |
| 298 fprintf (file, " %d", e->src); | |
| 299 | |
| 300 fprintf (file, ") (out:"); | |
| 301 | |
| 302 if (v->succ) | |
| 303 for (e = v->succ; e; e = e->succ_next) | |
| 304 fprintf (file, " %d", e->dest); | |
| 305 | |
| 306 fprintf (file, ")\n"); | |
| 307 print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
| 308 fprintf (file, ")\n"); | |
| 309 } | |
| 310 | |
| 311 /* Call dump_rdg_vertex on stderr. */ | |
| 312 | |
| 313 DEBUG_FUNCTION void | |
| 314 debug_rdg_vertex (struct graph *rdg, int i) | |
| 315 { | |
| 316 dump_rdg_vertex (stderr, rdg, i); | |
| 317 } | |
| 318 | |
| 319 /* Dump the reduced dependence graph RDG to FILE. */ | |
| 320 | |
| 321 static void | |
| 322 dump_rdg (FILE *file, struct graph *rdg) | |
| 323 { | |
| 324 fprintf (file, "(rdg\n"); | |
| 325 for (int i = 0; i < rdg->n_vertices; i++) | |
| 326 dump_rdg_vertex (file, rdg, i); | |
| 327 fprintf (file, ")\n"); | |
| 328 } | |
| 329 | |
| 330 /* Call dump_rdg on stderr. */ | |
| 331 | |
| 332 DEBUG_FUNCTION void | |
| 333 debug_rdg (struct graph *rdg) | |
| 334 { | |
| 335 dump_rdg (stderr, rdg); | |
| 336 } | |
| 337 | |
| 338 static void | |
| 339 dot_rdg_1 (FILE *file, struct graph *rdg) | |
| 340 { | |
| 341 int i; | |
| 342 pretty_printer buffer; | |
| 343 pp_needs_newline (&buffer) = false; | |
| 344 buffer.buffer->stream = file; | |
| 345 | |
| 346 fprintf (file, "digraph RDG {\n"); | |
| 347 | |
| 348 for (i = 0; i < rdg->n_vertices; i++) | |
| 349 { | |
| 350 struct vertex *v = &(rdg->vertices[i]); | |
| 351 struct graph_edge *e; | |
| 352 | |
| 353 fprintf (file, "%d [label=\"[%d] ", i, i); | |
| 354 pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
| 355 pp_flush (&buffer); | |
| 356 fprintf (file, "\"]\n"); | |
| 357 | |
| 358 /* Highlight reads from memory. */ | |
| 359 if (RDG_MEM_READS_STMT (rdg, i)) | |
| 360 fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
| 361 | |
| 362 /* Highlight stores to memory. */ | |
| 363 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
| 364 fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
| 365 | |
| 366 if (v->succ) | |
| 367 for (e = v->succ; e; e = e->succ_next) | |
| 368 switch (RDGE_TYPE (e)) | |
| 369 { | |
| 370 case flow_dd: | |
| 371 /* These are the most common dependences: don't print these. */ | |
| 372 fprintf (file, "%d -> %d \n", i, e->dest); | |
| 373 break; | |
| 374 | |
| 375 case control_dd: | |
| 376 fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
| 377 break; | |
| 378 | |
| 379 default: | |
| 380 gcc_unreachable (); | |
| 381 } | |
| 382 } | |
| 383 | |
| 384 fprintf (file, "}\n\n"); | |
| 385 } | |
| 386 | |
| 387 /* Display the Reduced Dependence Graph using dotty. */ | |
| 388 | |
| 389 DEBUG_FUNCTION void | |
| 390 dot_rdg (struct graph *rdg) | |
| 391 { | |
| 392 /* When debugging, you may want to enable the following code. */ | |
| 393 #ifdef HAVE_POPEN | |
| 394 FILE *file = popen ("dot -Tx11", "w"); | |
| 395 if (!file) | |
| 396 return; | |
| 397 dot_rdg_1 (file, rdg); | |
| 398 fflush (file); | |
| 399 close (fileno (file)); | |
| 400 pclose (file); | |
| 401 #else | |
| 402 dot_rdg_1 (stderr, rdg); | |
| 403 #endif | |
| 404 } | |
| 405 | |
| 406 /* Returns the index of STMT in RDG. */ | |
| 407 | |
| 408 static int | |
| 409 rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) | |
| 410 { | |
| 411 int index = gimple_uid (stmt); | |
| 412 gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
| 413 return index; | |
| 414 } | |
| 415 | |
| 416 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is | |
| 417 the index of DEF in RDG. */ | |
| 418 | |
| 419 static void | |
| 420 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
| 421 { | |
| 422 use_operand_p imm_use_p; | |
| 423 imm_use_iterator iterator; | |
| 424 | |
| 425 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
| 426 { | |
| 427 struct graph_edge *e; | |
| 428 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
| 429 | |
| 430 if (use < 0) | |
| 431 continue; | |
| 432 | |
| 433 e = add_edge (rdg, idef, use); | |
| 434 e->data = XNEW (struct rdg_edge); | |
| 435 RDGE_TYPE (e) = flow_dd; | |
| 436 } | |
| 437 } | |
| 438 | |
| 439 /* Creates an edge for the control dependences of BB to the vertex V. */ | |
| 440 | |
| 441 static void | |
| 442 create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
| 443 int v, control_dependences *cd) | |
| 444 { | |
| 445 bitmap_iterator bi; | |
| 446 unsigned edge_n; | |
| 447 EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
| 448 0, edge_n, bi) | |
| 449 { | |
| 450 basic_block cond_bb = cd->get_edge_src (edge_n); | |
| 451 gimple *stmt = last_stmt (cond_bb); | |
| 452 if (stmt && is_ctrl_stmt (stmt)) | |
| 453 { | |
| 454 struct graph_edge *e; | |
| 455 int c = rdg_vertex_for_stmt (rdg, stmt); | |
| 456 if (c < 0) | |
| 457 continue; | |
| 458 | |
| 459 e = add_edge (rdg, c, v); | |
| 460 e->data = XNEW (struct rdg_edge); | |
| 461 RDGE_TYPE (e) = control_dd; | |
| 462 } | |
| 463 } | |
| 464 } | |
| 465 | |
| 466 /* Creates the edges of the reduced dependence graph RDG. */ | |
| 0 | 467 |
| 468 static void | |
| 111 | 469 create_rdg_flow_edges (struct graph *rdg) |
| 0 | 470 { |
| 111 | 471 int i; |
| 472 def_operand_p def_p; | |
| 473 ssa_op_iter iter; | |
| 474 | |
| 475 for (i = 0; i < rdg->n_vertices; i++) | |
| 476 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
| 477 iter, SSA_OP_DEF) | |
| 478 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
| 479 } | |
| 480 | |
| 481 /* Creates the edges of the reduced dependence graph RDG. */ | |
| 482 | |
| 483 static void | |
| 484 create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) | |
| 485 { | |
| 486 int i; | |
| 487 | |
| 488 for (i = 0; i < rdg->n_vertices; i++) | |
| 0 | 489 { |
| 111 | 490 gimple *stmt = RDG_STMT (rdg, i); |
| 491 if (gimple_code (stmt) == GIMPLE_PHI) | |
| 0 | 492 { |
| 111 | 493 edge_iterator ei; |
| 494 edge e; | |
| 495 FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
| 496 if (flow_bb_inside_loop_p (loop, e->src)) | |
| 497 create_edge_for_control_dependence (rdg, e->src, i, cd); | |
| 0 | 498 } |
| 499 else | |
| 111 | 500 create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); |
| 501 } | |
| 502 } | |
| 503 | |
| 145 | 504 |
| 505 class loop_distribution | |
| 506 { | |
| 507 private: | |
| 508 /* The loop (nest) to be distributed. */ | |
| 509 vec<loop_p> loop_nest; | |
| 510 | |
| 511 /* Vector of data references in the loop to be distributed. */ | |
| 512 vec<data_reference_p> datarefs_vec; | |
| 513 | |
| 514 /* If there is nonaddressable data reference in above vector. */ | |
| 515 bool has_nonaddressable_dataref_p; | |
| 516 | |
| 517 /* Store index of data reference in aux field. */ | |
| 518 | |
| 519 /* Hash table for data dependence relation in the loop to be distributed. */ | |
| 520 hash_table<ddr_hasher> *ddrs_table; | |
| 521 | |
| 522 /* Array mapping basic block's index to its topological order. */ | |
| 523 int *bb_top_order_index; | |
| 524 /* And size of the array. */ | |
| 525 int bb_top_order_index_size; | |
| 526 | |
| 527 /* Build the vertices of the reduced dependence graph RDG. Return false | |
| 528 if that failed. */ | |
| 529 bool create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop); | |
| 530 | |
| 531 /* Initialize STMTS with all the statements of LOOP. We use topological | |
| 532 order to discover all statements. The order is important because | |
| 533 generate_loops_for_partition is using the same traversal for identifying | |
| 534 statements in loop copies. */ | |
| 535 void stmts_from_loop (class loop *loop, vec<gimple *> *stmts); | |
| 536 | |
| 537 | |
| 538 /* Build the Reduced Dependence Graph (RDG) with one vertex per statement of | |
| 539 LOOP, and one edge per flow dependence or control dependence from control | |
| 540 dependence CD. During visiting each statement, data references are also | |
| 541 collected and recorded in global data DATAREFS_VEC. */ | |
| 542 struct graph * build_rdg (class loop *loop, control_dependences *cd); | |
| 543 | |
| 544 /* Merge PARTITION into the partition DEST. RDG is the reduced dependence | |
| 545 graph and we update type for result partition if it is non-NULL. */ | |
| 546 void partition_merge_into (struct graph *rdg, | |
| 547 partition *dest, partition *partition, | |
| 548 enum fuse_type ft); | |
| 549 | |
| 550 | |
| 551 /* Return data dependence relation for data references A and B. The two | |
| 552 data references must be in lexicographic order wrto reduced dependence | |
| 553 graph RDG. We firstly try to find ddr from global ddr hash table. If | |
| 554 it doesn't exist, compute the ddr and cache it. */ | |
| 555 data_dependence_relation * get_data_dependence (struct graph *rdg, | |
| 556 data_reference_p a, | |
| 557 data_reference_p b); | |
| 558 | |
| 559 | |
| 560 /* In reduced dependence graph RDG for loop distribution, return true if | |
| 561 dependence between references DR1 and DR2 leads to a dependence cycle | |
| 562 and such dependence cycle can't be resolved by runtime alias check. */ | |
| 563 bool data_dep_in_cycle_p (struct graph *rdg, data_reference_p dr1, | |
| 564 data_reference_p dr2); | |
| 565 | |
| 566 | |
| 567 /* Given reduced dependence graph RDG, PARTITION1 and PARTITION2, update | |
| 568 PARTITION1's type after merging PARTITION2 into PARTITION1. */ | |
| 569 void update_type_for_merge (struct graph *rdg, | |
| 570 partition *partition1, partition *partition2); | |
| 571 | |
| 572 | |
| 573 /* Returns a partition with all the statements needed for computing | |
| 574 the vertex V of the RDG, also including the loop exit conditions. */ | |
| 575 partition *build_rdg_partition_for_vertex (struct graph *rdg, int v); | |
| 576 | |
| 577 /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify | |
| 578 if it forms builtin memcpy or memmove call. */ | |
| 579 void classify_builtin_ldst (loop_p loop, struct graph *rdg, partition *partition, | |
| 580 data_reference_p dst_dr, data_reference_p src_dr); | |
| 581 | |
| 582 /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. | |
| 583 For the moment we detect memset, memcpy and memmove patterns. Bitmap | |
| 584 STMT_IN_ALL_PARTITIONS contains statements belonging to all partitions. | |
| 585 Returns true if there is a reduction in all partitions and we | |
| 586 possibly did not mark PARTITION as having one for this reason. */ | |
| 587 | |
| 588 bool | |
| 589 classify_partition (loop_p loop, | |
| 590 struct graph *rdg, partition *partition, | |
| 591 bitmap stmt_in_all_partitions); | |
| 592 | |
| 593 | |
| 594 /* Returns true when PARTITION1 and PARTITION2 access the same memory | |
| 595 object in RDG. */ | |
| 596 bool share_memory_accesses (struct graph *rdg, | |
| 597 partition *partition1, partition *partition2); | |
| 598 | |
| 599 /* For each seed statement in STARTING_STMTS, this function builds | |
| 600 partition for it by adding depended statements according to RDG. | |
| 601 All partitions are recorded in PARTITIONS. */ | |
| 602 void rdg_build_partitions (struct graph *rdg, | |
| 603 vec<gimple *> starting_stmts, | |
| 604 vec<partition *> *partitions); | |
| 605 | |
| 606 /* Compute partition dependence created by the data references in DRS1 | |
| 607 and DRS2, modify and return DIR according to that. IF ALIAS_DDR is | |
| 608 not NULL, we record dependence introduced by possible alias between | |
| 609 two data references in ALIAS_DDRS; otherwise, we simply ignore such | |
| 610 dependence as if it doesn't exist at all. */ | |
| 611 int pg_add_dependence_edges (struct graph *rdg, int dir, bitmap drs1, | |
| 612 bitmap drs2, vec<ddr_p> *alias_ddrs); | |
| 613 | |
| 614 | |
| 615 /* Build and return partition dependence graph for PARTITIONS. RDG is | |
| 616 reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
| 617 is true, data dependence caused by possible alias between references | |
| 618 is ignored, as if it doesn't exist at all; otherwise all depdendences | |
| 619 are considered. */ | |
| 620 struct graph *build_partition_graph (struct graph *rdg, | |
| 621 vec<struct partition *> *partitions, | |
| 622 bool ignore_alias_p); | |
| 623 | |
| 624 /* Given reduced dependence graph RDG merge strong connected components | |
| 625 of PARTITIONS. If IGNORE_ALIAS_P is true, data dependence caused by | |
| 626 possible alias between references is ignored, as if it doesn't exist | |
| 627 at all; otherwise all depdendences are considered. */ | |
| 628 void merge_dep_scc_partitions (struct graph *rdg, vec<struct partition *> | |
| 629 *partitions, bool ignore_alias_p); | |
| 630 | |
| 631 /* This is the main function breaking strong conected components in | |
| 632 PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
| 633 relations for runtime alias check in ALIAS_DDRS. */ | |
| 634 void break_alias_scc_partitions (struct graph *rdg, vec<struct partition *> | |
| 635 *partitions, vec<ddr_p> *alias_ddrs); | |
| 636 | |
| 637 | |
| 638 /* Fuse PARTITIONS of LOOP if necessary before finalizing distribution. | |
| 639 ALIAS_DDRS contains ddrs which need runtime alias check. */ | |
| 640 void finalize_partitions (class loop *loop, vec<struct partition *> | |
| 641 *partitions, vec<ddr_p> *alias_ddrs); | |
| 642 | |
| 643 /* Distributes the code from LOOP in such a way that producer statements | |
| 644 are placed before consumer statements. Tries to separate only the | |
| 645 statements from STMTS into separate loops. Returns the number of | |
| 646 distributed loops. Set NB_CALLS to number of generated builtin calls. | |
| 647 Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
| 648 int distribute_loop (class loop *loop, vec<gimple *> stmts, | |
| 649 control_dependences *cd, int *nb_calls, bool *destroy_p, | |
| 650 bool only_patterns_p); | |
| 651 | |
| 652 /* Compute topological order for basic blocks. Topological order is | |
| 653 needed because data dependence is computed for data references in | |
| 654 lexicographical order. */ | |
| 655 void bb_top_order_init (void); | |
| 656 | |
| 657 void bb_top_order_destroy (void); | |
| 658 | |
| 659 public: | |
| 660 | |
| 661 /* Getter for bb_top_order. */ | |
| 662 | |
| 663 inline int get_bb_top_order_index_size (void) | |
| 664 { | |
| 665 return bb_top_order_index_size; | |
| 666 } | |
| 667 | |
| 668 inline int get_bb_top_order_index (int i) | |
| 669 { | |
| 670 return bb_top_order_index[i]; | |
| 671 } | |
| 672 | |
| 673 unsigned int execute (function *fun); | |
| 674 }; | |
| 675 | |
| 676 | |
| 677 /* If X has a smaller topological sort number than Y, returns -1; | |
| 678 if greater, returns 1. */ | |
| 679 static int | |
| 680 bb_top_order_cmp_r (const void *x, const void *y, void *loop) | |
| 681 { | |
| 682 loop_distribution *_loop = | |
| 683 (loop_distribution *) loop; | |
| 684 | |
| 685 basic_block bb1 = *(const basic_block *) x; | |
| 686 basic_block bb2 = *(const basic_block *) y; | |
| 687 | |
| 688 int bb_top_order_index_size = _loop->get_bb_top_order_index_size (); | |
| 689 | |
| 690 gcc_assert (bb1->index < bb_top_order_index_size | |
| 691 && bb2->index < bb_top_order_index_size); | |
| 692 gcc_assert (bb1 == bb2 | |
| 693 || _loop->get_bb_top_order_index(bb1->index) | |
| 694 != _loop->get_bb_top_order_index(bb2->index)); | |
| 695 | |
| 696 return (_loop->get_bb_top_order_index(bb1->index) - | |
| 697 _loop->get_bb_top_order_index(bb2->index)); | |
| 698 } | |
| 699 | |
| 700 bool | |
| 701 loop_distribution::create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, | |
| 702 loop_p loop) | |
| 111 | 703 { |
| 704 int i; | |
| 705 gimple *stmt; | |
| 706 | |
| 707 FOR_EACH_VEC_ELT (stmts, i, stmt) | |
| 708 { | |
| 709 struct vertex *v = &(rdg->vertices[i]); | |
| 710 | |
| 711 /* Record statement to vertex mapping. */ | |
| 712 gimple_set_uid (stmt, i); | |
| 713 | |
| 714 v->data = XNEW (struct rdg_vertex); | |
| 715 RDGV_STMT (v) = stmt; | |
| 716 RDGV_DATAREFS (v).create (0); | |
| 717 RDGV_HAS_MEM_WRITE (v) = false; | |
| 718 RDGV_HAS_MEM_READS (v) = false; | |
| 719 if (gimple_code (stmt) == GIMPLE_PHI) | |
| 720 continue; | |
| 721 | |
| 722 unsigned drp = datarefs_vec.length (); | |
| 723 if (!find_data_references_in_stmt (loop, stmt, &datarefs_vec)) | |
| 724 return false; | |
| 725 for (unsigned j = drp; j < datarefs_vec.length (); ++j) | |
| 726 { | |
| 727 data_reference_p dr = datarefs_vec[j]; | |
| 728 if (DR_IS_READ (dr)) | |
| 729 RDGV_HAS_MEM_READS (v) = true; | |
| 730 else | |
| 731 RDGV_HAS_MEM_WRITE (v) = true; | |
| 732 RDGV_DATAREFS (v).safe_push (dr); | |
| 145 | 733 has_nonaddressable_dataref_p |= may_be_nonaddressable_p (dr->ref); |
| 111 | 734 } |
| 735 } | |
| 736 return true; | |
| 737 } | |
| 738 | |
| 145 | 739 void |
| 740 loop_distribution::stmts_from_loop (class loop *loop, vec<gimple *> *stmts) | |
| 111 | 741 { |
| 742 unsigned int i; | |
| 145 | 743 basic_block *bbs = get_loop_body_in_custom_order (loop, this, bb_top_order_cmp_r); |
| 111 | 744 |
| 745 for (i = 0; i < loop->num_nodes; i++) | |
| 746 { | |
| 747 basic_block bb = bbs[i]; | |
| 748 | |
| 749 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); | |
| 750 gsi_next (&bsi)) | |
| 751 if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
| 752 stmts->safe_push (bsi.phi ()); | |
| 753 | |
| 754 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); | |
| 755 gsi_next (&bsi)) | |
| 756 { | |
| 757 gimple *stmt = gsi_stmt (bsi); | |
| 758 if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) | |
| 759 stmts->safe_push (stmt); | |
| 760 } | |
| 761 } | |
| 762 | |
| 763 free (bbs); | |
| 764 } | |
| 765 | |
| 766 /* Free the reduced dependence graph RDG. */ | |
| 767 | |
| 768 static void | |
| 769 free_rdg (struct graph *rdg) | |
| 770 { | |
| 771 int i; | |
| 772 | |
| 773 for (i = 0; i < rdg->n_vertices; i++) | |
| 774 { | |
| 775 struct vertex *v = &(rdg->vertices[i]); | |
| 776 struct graph_edge *e; | |
| 777 | |
| 778 for (e = v->succ; e; e = e->succ_next) | |
| 779 free (e->data); | |
| 780 | |
| 781 if (v->data) | |
| 782 { | |
| 783 gimple_set_uid (RDGV_STMT (v), -1); | |
| 784 (RDGV_DATAREFS (v)).release (); | |
| 785 free (v->data); | |
| 786 } | |
| 787 } | |
| 788 | |
| 789 free_graph (rdg); | |
| 790 } | |
| 791 | |
| 145 | 792 struct graph * |
| 793 loop_distribution::build_rdg (class loop *loop, control_dependences *cd) | |
| 111 | 794 { |
| 795 struct graph *rdg; | |
| 796 | |
| 797 /* Create the RDG vertices from the stmts of the loop nest. */ | |
| 798 auto_vec<gimple *, 10> stmts; | |
| 799 stmts_from_loop (loop, &stmts); | |
| 800 rdg = new_graph (stmts.length ()); | |
| 801 if (!create_rdg_vertices (rdg, stmts, loop)) | |
| 802 { | |
| 803 free_rdg (rdg); | |
| 804 return NULL; | |
| 0 | 805 } |
| 111 | 806 stmts.release (); |
| 807 | |
| 808 create_rdg_flow_edges (rdg); | |
| 809 if (cd) | |
| 810 create_rdg_cd_edges (rdg, cd, loop); | |
| 811 | |
| 812 return rdg; | |
| 813 } | |
| 814 | |
| 815 | |
| 816 /* Allocate and initialize a partition from BITMAP. */ | |
| 817 | |
| 818 static partition * | |
| 819 partition_alloc (void) | |
| 820 { | |
| 821 partition *partition = XCNEW (struct partition); | |
| 822 partition->stmts = BITMAP_ALLOC (NULL); | |
| 823 partition->reduction_p = false; | |
| 145 | 824 partition->loc = UNKNOWN_LOCATION; |
| 111 | 825 partition->kind = PKIND_NORMAL; |
| 145 | 826 partition->type = PTYPE_PARALLEL; |
| 111 | 827 partition->datarefs = BITMAP_ALLOC (NULL); |
| 828 return partition; | |
| 829 } | |
| 830 | |
| 831 /* Free PARTITION. */ | |
| 832 | |
| 833 static void | |
| 834 partition_free (partition *partition) | |
| 835 { | |
| 836 BITMAP_FREE (partition->stmts); | |
| 837 BITMAP_FREE (partition->datarefs); | |
| 838 if (partition->builtin) | |
| 839 free (partition->builtin); | |
| 840 | |
| 841 free (partition); | |
| 842 } | |
| 843 | |
| 844 /* Returns true if the partition can be generated as a builtin. */ | |
| 845 | |
| 846 static bool | |
| 847 partition_builtin_p (partition *partition) | |
| 848 { | |
| 131 | 849 return partition->kind > PKIND_PARTIAL_MEMSET; |
| 111 | 850 } |
| 851 | |
| 852 /* Returns true if the partition contains a reduction. */ | |
| 853 | |
| 854 static bool | |
| 855 partition_reduction_p (partition *partition) | |
| 856 { | |
| 857 return partition->reduction_p; | |
| 858 } | |
| 859 | |
| 145 | 860 void |
| 861 loop_distribution::partition_merge_into (struct graph *rdg, | |
| 862 partition *dest, partition *partition, enum fuse_type ft) | |
| 111 | 863 { |
| 864 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 865 { | |
| 866 fprintf (dump_file, "Fuse partitions because %s:\n", fuse_message[ft]); | |
| 867 fprintf (dump_file, " Part 1: "); | |
| 868 dump_bitmap (dump_file, dest->stmts); | |
| 869 fprintf (dump_file, " Part 2: "); | |
| 870 dump_bitmap (dump_file, partition->stmts); | |
| 871 } | |
| 872 | |
| 873 dest->kind = PKIND_NORMAL; | |
| 874 if (dest->type == PTYPE_PARALLEL) | |
| 875 dest->type = partition->type; | |
| 876 | |
| 877 bitmap_ior_into (dest->stmts, partition->stmts); | |
| 878 if (partition_reduction_p (partition)) | |
| 879 dest->reduction_p = true; | |
| 880 | |
| 881 /* Further check if any data dependence prevents us from executing the | |
| 882 new partition parallelly. */ | |
| 883 if (dest->type == PTYPE_PARALLEL && rdg != NULL) | |
| 884 update_type_for_merge (rdg, dest, partition); | |
| 885 | |
| 886 bitmap_ior_into (dest->datarefs, partition->datarefs); | |
| 887 } | |
| 888 | |
| 889 | |
| 890 /* Returns true when DEF is an SSA_NAME defined in LOOP and used after | |
| 891 the LOOP. */ | |
| 892 | |
| 893 static bool | |
| 894 ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
| 895 { | |
| 896 imm_use_iterator imm_iter; | |
| 897 use_operand_p use_p; | |
| 898 | |
| 899 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
| 900 { | |
| 901 if (is_gimple_debug (USE_STMT (use_p))) | |
| 902 continue; | |
| 903 | |
| 904 basic_block use_bb = gimple_bb (USE_STMT (use_p)); | |
| 905 if (!flow_bb_inside_loop_p (loop, use_bb)) | |
| 906 return true; | |
| 907 } | |
| 908 | |
| 909 return false; | |
| 910 } | |
| 911 | |
| 912 /* Returns true when STMT defines a scalar variable used after the | |
| 913 loop LOOP. */ | |
| 914 | |
| 915 static bool | |
| 916 stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) | |
| 917 { | |
| 918 def_operand_p def_p; | |
| 919 ssa_op_iter op_iter; | |
| 920 | |
| 921 if (gimple_code (stmt) == GIMPLE_PHI) | |
| 922 return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
| 923 | |
| 924 FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) | |
| 925 if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
| 926 return true; | |
| 927 | |
| 928 return false; | |
| 0 | 929 } |
| 930 | |
| 931 /* Return a copy of LOOP placed before LOOP. */ | |
| 932 | |
| 145 | 933 static class loop * |
| 934 copy_loop_before (class loop *loop) | |
| 0 | 935 { |
| 145 | 936 class loop *res; |
| 0 | 937 edge preheader = loop_preheader_edge (loop); |
| 938 | |
| 939 initialize_original_copy_tables (); | |
| 111 | 940 res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
| 941 gcc_assert (res != NULL); | |
| 0 | 942 free_original_copy_tables (); |
| 111 | 943 delete_update_ssa (); |
| 0 | 944 |
| 945 return res; | |
| 946 } | |
| 947 | |
| 948 /* Creates an empty basic block after LOOP. */ | |
| 949 | |
| 950 static void | |
| 145 | 951 create_bb_after_loop (class loop *loop) |
| 0 | 952 { |
| 953 edge exit = single_exit (loop); | |
| 954 | |
| 955 if (!exit) | |
| 956 return; | |
| 957 | |
| 958 split_edge (exit); | |
| 959 } | |
| 960 | |
| 961 /* Generate code for PARTITION from the code in LOOP. The loop is | |
| 962 copied when COPY_P is true. All the statements not flagged in the | |
| 963 PARTITION bitmap are removed from the loop or from its copy. The | |
| 964 statements are indexed in sequence inside a basic block, and the | |
| 111 | 965 basic blocks of a loop are taken in dom order. */ |
| 966 | |
| 967 static void | |
| 145 | 968 generate_loops_for_partition (class loop *loop, partition *partition, |
| 111 | 969 bool copy_p) |
| 0 | 970 { |
| 111 | 971 unsigned i; |
| 0 | 972 basic_block *bbs; |
| 973 | |
| 974 if (copy_p) | |
| 975 { | |
| 111 | 976 int orig_loop_num = loop->orig_loop_num; |
| 0 | 977 loop = copy_loop_before (loop); |
| 111 | 978 gcc_assert (loop != NULL); |
| 979 loop->orig_loop_num = orig_loop_num; | |
| 0 | 980 create_preheader (loop, CP_SIMPLE_PREHEADERS); |
| 981 create_bb_after_loop (loop); | |
| 982 } | |
| 111 | 983 else |
| 0 | 984 { |
| 111 | 985 /* Origin number is set to the new versioned loop's num. */ |
| 986 gcc_assert (loop->orig_loop_num != loop->num); | |
| 0 | 987 } |
| 988 | |
| 111 | 989 /* Remove stmts not in the PARTITION bitmap. */ |
| 0 | 990 bbs = get_loop_body_in_dom_order (loop); |
| 991 | |
| 131 | 992 if (MAY_HAVE_DEBUG_BIND_STMTS) |
| 111 | 993 for (i = 0; i < loop->num_nodes; i++) |
| 994 { | |
| 995 basic_block bb = bbs[i]; | |
| 996 | |
| 997 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); | |
| 998 gsi_next (&bsi)) | |
| 999 { | |
| 1000 gphi *phi = bsi.phi (); | |
| 1001 if (!virtual_operand_p (gimple_phi_result (phi)) | |
| 1002 && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
| 1003 reset_debug_uses (phi); | |
| 1004 } | |
| 1005 | |
| 1006 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
| 1007 { | |
| 1008 gimple *stmt = gsi_stmt (bsi); | |
| 1009 if (gimple_code (stmt) != GIMPLE_LABEL | |
| 1010 && !is_gimple_debug (stmt) | |
| 1011 && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) | |
| 1012 reset_debug_uses (stmt); | |
| 1013 } | |
| 1014 } | |
| 1015 | |
| 0 | 1016 for (i = 0; i < loop->num_nodes; i++) |
| 1017 { | |
| 1018 basic_block bb = bbs[i]; | |
| 111 | 1019 edge inner_exit = NULL; |
| 1020 | |
| 1021 if (loop != bb->loop_father) | |
| 1022 inner_exit = single_exit (bb->loop_father); | |
| 1023 | |
| 1024 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) | |
| 1025 { | |
| 1026 gphi *phi = bsi.phi (); | |
| 1027 if (!virtual_operand_p (gimple_phi_result (phi)) | |
| 1028 && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
| 1029 remove_phi_node (&bsi, true); | |
| 1030 else | |
| 1031 gsi_next (&bsi); | |
| 1032 } | |
| 1033 | |
| 1034 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) | |
| 0 | 1035 { |
| 111 | 1036 gimple *stmt = gsi_stmt (bsi); |
| 1037 if (gimple_code (stmt) != GIMPLE_LABEL | |
| 1038 && !is_gimple_debug (stmt) | |
| 1039 && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) | |
| 0 | 1040 { |
| 111 | 1041 /* In distribution of loop nest, if bb is inner loop's exit_bb, |
| 1042 we choose its exit edge/path in order to avoid generating | |
| 1043 infinite loop. For all other cases, we choose an arbitrary | |
| 1044 path through the empty CFG part that this unnecessary | |
| 1045 control stmt controls. */ | |
| 1046 if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) | |
| 1047 { | |
| 1048 if (inner_exit && inner_exit->flags & EDGE_TRUE_VALUE) | |
| 1049 gimple_cond_make_true (cond_stmt); | |
| 1050 else | |
| 1051 gimple_cond_make_false (cond_stmt); | |
| 1052 update_stmt (stmt); | |
| 1053 } | |
| 1054 else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
| 1055 { | |
| 1056 gswitch *switch_stmt = as_a <gswitch *> (stmt); | |
| 1057 gimple_switch_set_index | |
| 1058 (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); | |
| 1059 update_stmt (stmt); | |
| 1060 } | |
| 1061 else | |
| 1062 { | |
| 1063 unlink_stmt_vdef (stmt); | |
| 1064 gsi_remove (&bsi, true); | |
| 1065 release_defs (stmt); | |
| 1066 continue; | |
| 1067 } | |
| 0 | 1068 } |
| 111 | 1069 gsi_next (&bsi); |
| 0 | 1070 } |
| 1071 } | |
| 1072 | |
| 111 | 1073 free (bbs); |
| 1074 } | |
| 1075 | |
| 1076 /* If VAL memory representation contains the same value in all bytes, | |
| 1077 return that value, otherwise return -1. | |
| 1078 E.g. for 0x24242424 return 0x24, for IEEE double | |
| 1079 747708026454360457216.0 return 0x44, etc. */ | |
| 1080 | |
| 1081 static int | |
| 1082 const_with_all_bytes_same (tree val) | |
| 1083 { | |
| 1084 unsigned char buf[64]; | |
| 1085 int i, len; | |
| 1086 | |
| 1087 if (integer_zerop (val) | |
| 1088 || (TREE_CODE (val) == CONSTRUCTOR | |
| 1089 && !TREE_CLOBBER_P (val) | |
| 1090 && CONSTRUCTOR_NELTS (val) == 0)) | |
| 1091 return 0; | |
| 1092 | |
| 1093 if (real_zerop (val)) | |
| 1094 { | |
| 1095 /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
| 1096 an all bytes same memory representation. Don't transform | |
| 1097 -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
| 1098 switch (TREE_CODE (val)) | |
| 1099 { | |
| 1100 case REAL_CST: | |
| 1101 if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
| 1102 return 0; | |
| 1103 break; | |
| 1104 case COMPLEX_CST: | |
| 1105 if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
| 1106 && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
| 1107 return 0; | |
| 1108 break; | |
| 1109 case VECTOR_CST: | |
| 131 | 1110 { |
| 1111 unsigned int count = vector_cst_encoded_nelts (val); | |
| 1112 unsigned int j; | |
| 1113 for (j = 0; j < count; ++j) | |
| 1114 if (const_with_all_bytes_same (VECTOR_CST_ENCODED_ELT (val, j))) | |
| 1115 break; | |
| 1116 if (j == count) | |
| 1117 return 0; | |
| 1118 break; | |
| 1119 } | |
| 111 | 1120 default: |
| 1121 break; | |
| 1122 } | |
| 1123 } | |
| 1124 | |
| 1125 if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) | |
| 1126 return -1; | |
| 1127 | |
| 1128 len = native_encode_expr (val, buf, sizeof (buf)); | |
| 1129 if (len == 0) | |
| 1130 return -1; | |
| 1131 for (i = 1; i < len; i++) | |
| 1132 if (buf[i] != buf[0]) | |
| 1133 return -1; | |
| 1134 return buf[0]; | |
| 1135 } | |
| 1136 | |
| 1137 /* Generate a call to memset for PARTITION in LOOP. */ | |
| 1138 | |
| 1139 static void | |
| 145 | 1140 generate_memset_builtin (class loop *loop, partition *partition) |
| 111 | 1141 { |
| 1142 gimple_stmt_iterator gsi; | |
| 1143 tree mem, fn, nb_bytes; | |
| 1144 tree val; | |
| 1145 struct builtin_info *builtin = partition->builtin; | |
| 1146 gimple *fn_call; | |
| 0 | 1147 |
| 1148 /* The new statements will be placed before LOOP. */ | |
| 111 | 1149 gsi = gsi_last_bb (loop_preheader_edge (loop)->src); |
| 1150 | |
| 145 | 1151 nb_bytes = rewrite_to_non_trapping_overflow (builtin->size); |
| 111 | 1152 nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
| 1153 false, GSI_CONTINUE_LINKING); | |
| 1154 mem = builtin->dst_base; | |
| 1155 mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, | |
| 1156 false, GSI_CONTINUE_LINKING); | |
| 1157 | |
| 1158 /* This exactly matches the pattern recognition in classify_partition. */ | |
| 1159 val = gimple_assign_rhs1 (DR_STMT (builtin->dst_dr)); | |
| 1160 /* Handle constants like 0x15151515 and similarly | |
| 1161 floating point constants etc. where all bytes are the same. */ | |
| 1162 int bytev = const_with_all_bytes_same (val); | |
| 1163 if (bytev != -1) | |
| 1164 val = build_int_cst (integer_type_node, bytev); | |
| 1165 else if (TREE_CODE (val) == INTEGER_CST) | |
| 1166 val = fold_convert (integer_type_node, val); | |
| 1167 else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
| 1168 { | |
| 1169 tree tem = make_ssa_name (integer_type_node); | |
| 1170 gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); | |
| 1171 gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); | |
| 1172 val = tem; | |
| 1173 } | |
| 1174 | |
| 1175 fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); | |
| 1176 fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); | |
| 145 | 1177 gimple_set_location (fn_call, partition->loc); |
| 111 | 1178 gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
| 145 | 1179 fold_stmt (&gsi); |
| 111 | 1180 |
| 1181 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 1182 { |
| 111 | 1183 fprintf (dump_file, "generated memset"); |
| 1184 if (bytev == 0) | |
| 1185 fprintf (dump_file, " zero\n"); | |
| 1186 else | |
| 1187 fprintf (dump_file, "\n"); | |
| 1188 } | |
| 1189 } | |
| 1190 | |
| 1191 /* Generate a call to memcpy for PARTITION in LOOP. */ | |
| 1192 | |
| 1193 static void | |
| 145 | 1194 generate_memcpy_builtin (class loop *loop, partition *partition) |
| 111 | 1195 { |
| 1196 gimple_stmt_iterator gsi; | |
| 1197 gimple *fn_call; | |
| 1198 tree dest, src, fn, nb_bytes; | |
| 1199 enum built_in_function kind; | |
| 1200 struct builtin_info *builtin = partition->builtin; | |
| 1201 | |
| 1202 /* The new statements will be placed before LOOP. */ | |
| 1203 gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
| 1204 | |
| 145 | 1205 nb_bytes = rewrite_to_non_trapping_overflow (builtin->size); |
| 111 | 1206 nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
| 1207 false, GSI_CONTINUE_LINKING); | |
| 1208 dest = builtin->dst_base; | |
| 1209 src = builtin->src_base; | |
| 1210 if (partition->kind == PKIND_MEMCPY | |
| 1211 || ! ptr_derefs_may_alias_p (dest, src)) | |
| 1212 kind = BUILT_IN_MEMCPY; | |
| 1213 else | |
| 1214 kind = BUILT_IN_MEMMOVE; | |
| 1215 | |
| 1216 dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
| 1217 false, GSI_CONTINUE_LINKING); | |
| 1218 src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
| 1219 false, GSI_CONTINUE_LINKING); | |
| 1220 fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
| 1221 fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
| 145 | 1222 gimple_set_location (fn_call, partition->loc); |
| 111 | 1223 gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
| 145 | 1224 fold_stmt (&gsi); |
| 111 | 1225 |
| 1226 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 1227 { | |
| 1228 if (kind == BUILT_IN_MEMCPY) | |
| 1229 fprintf (dump_file, "generated memcpy\n"); | |
| 1230 else | |
| 1231 fprintf (dump_file, "generated memmove\n"); | |
| 0 | 1232 } |
| 1233 } | |
| 1234 | |
| 111 | 1235 /* Remove and destroy the loop LOOP. */ |
| 1236 | |
| 1237 static void | |
| 145 | 1238 destroy_loop (class loop *loop) |
| 0 | 1239 { |
| 111 | 1240 unsigned nbbs = loop->num_nodes; |
| 1241 edge exit = single_exit (loop); | |
| 1242 basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
| 1243 basic_block *bbs; | |
| 1244 unsigned i; | |
| 1245 | |
| 1246 bbs = get_loop_body_in_dom_order (loop); | |
| 1247 | |
| 145 | 1248 gimple_stmt_iterator dst_gsi = gsi_after_labels (exit->dest); |
| 1249 bool safe_p = single_pred_p (exit->dest); | |
| 1250 for (unsigned i = 0; i < nbbs; ++i) | |
| 1251 { | |
| 1252 /* We have made sure to not leave any dangling uses of SSA | |
| 1253 names defined in the loop. With the exception of virtuals. | |
| 1254 Make sure we replace all uses of virtual defs that will remain | |
| 1255 outside of the loop with the bare symbol as delete_basic_block | |
| 1256 will release them. */ | |
| 1257 for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); | |
| 1258 gsi_next (&gsi)) | |
| 1259 { | |
| 1260 gphi *phi = gsi.phi (); | |
| 1261 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 1262 mark_virtual_phi_result_for_renaming (phi); | |
| 1263 } | |
| 1264 for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi);) | |
| 1265 { | |
| 1266 gimple *stmt = gsi_stmt (gsi); | |
| 1267 tree vdef = gimple_vdef (stmt); | |
| 1268 if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
| 1269 mark_virtual_operand_for_renaming (vdef); | |
| 1270 /* Also move and eventually reset debug stmts. We can leave | |
| 1271 constant values in place in case the stmt dominates the exit. | |
| 1272 ??? Non-constant values from the last iteration can be | |
| 1273 replaced with final values if we can compute them. */ | |
| 1274 if (gimple_debug_bind_p (stmt)) | |
| 1275 { | |
| 1276 tree val = gimple_debug_bind_get_value (stmt); | |
| 1277 gsi_move_before (&gsi, &dst_gsi); | |
| 1278 if (val | |
| 1279 && (!safe_p | |
| 1280 || !is_gimple_min_invariant (val) | |
| 1281 || !dominated_by_p (CDI_DOMINATORS, exit->src, bbs[i]))) | |
| 1282 { | |
| 1283 gimple_debug_bind_reset_value (stmt); | |
| 1284 update_stmt (stmt); | |
| 1285 } | |
| 1286 } | |
| 1287 else | |
| 1288 gsi_next (&gsi); | |
| 1289 } | |
| 1290 } | |
| 1291 | |
| 111 | 1292 redirect_edge_pred (exit, src); |
| 1293 exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
| 1294 exit->flags |= EDGE_FALLTHRU; | |
| 1295 cancel_loop_tree (loop); | |
| 1296 rescan_loop_exit (exit, false, true); | |
| 1297 | |
| 1298 i = nbbs; | |
| 1299 do | |
| 1300 { | |
| 1301 --i; | |
| 1302 delete_basic_block (bbs[i]); | |
| 1303 } | |
| 1304 while (i != 0); | |
| 1305 | |
| 1306 free (bbs); | |
| 1307 | |
| 1308 set_immediate_dominator (CDI_DOMINATORS, dest, | |
| 1309 recompute_dominator (CDI_DOMINATORS, dest)); | |
| 1310 } | |
| 1311 | |
| 1312 /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ | |
| 1313 | |
| 1314 static bool | |
| 145 | 1315 generate_code_for_partition (class loop *loop, |
| 111 | 1316 partition *partition, bool copy_p) |
| 1317 { | |
| 1318 switch (partition->kind) | |
| 1319 { | |
| 1320 case PKIND_NORMAL: | |
| 131 | 1321 case PKIND_PARTIAL_MEMSET: |
| 111 | 1322 /* Reductions all have to be in the last partition. */ |
| 1323 gcc_assert (!partition_reduction_p (partition) | |
| 1324 || !copy_p); | |
| 1325 generate_loops_for_partition (loop, partition, copy_p); | |
| 1326 return false; | |
| 1327 | |
| 1328 case PKIND_MEMSET: | |
| 1329 generate_memset_builtin (loop, partition); | |
| 1330 break; | |
| 1331 | |
| 1332 case PKIND_MEMCPY: | |
| 1333 case PKIND_MEMMOVE: | |
| 1334 generate_memcpy_builtin (loop, partition); | |
| 1335 break; | |
| 1336 | |
| 1337 default: | |
| 1338 gcc_unreachable (); | |
| 1339 } | |
| 1340 | |
| 1341 /* Common tail for partitions we turn into a call. If this was the last | |
| 1342 partition for which we generate code, we have to destroy the loop. */ | |
| 1343 if (!copy_p) | |
| 0 | 1344 return true; |
| 1345 return false; | |
| 1346 } | |
| 1347 | |
| 145 | 1348 data_dependence_relation * |
| 1349 loop_distribution::get_data_dependence (struct graph *rdg, data_reference_p a, | |
| 1350 data_reference_p b) | |
| 0 | 1351 { |
| 111 | 1352 struct data_dependence_relation ent, **slot; |
| 1353 struct data_dependence_relation *ddr; | |
| 1354 | |
| 1355 gcc_assert (DR_IS_WRITE (a) || DR_IS_WRITE (b)); | |
| 1356 gcc_assert (rdg_vertex_for_stmt (rdg, DR_STMT (a)) | |
| 1357 <= rdg_vertex_for_stmt (rdg, DR_STMT (b))); | |
| 1358 ent.a = a; | |
| 1359 ent.b = b; | |
| 1360 slot = ddrs_table->find_slot (&ent, INSERT); | |
| 1361 if (*slot == NULL) | |
| 1362 { | |
| 1363 ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
| 1364 compute_affine_dependence (ddr, loop_nest[0]); | |
| 1365 *slot = ddr; | |
| 1366 } | |
| 1367 | |
| 1368 return *slot; | |
| 0 | 1369 } |
| 1370 | |
| 145 | 1371 bool |
| 1372 loop_distribution::data_dep_in_cycle_p (struct graph *rdg, | |
| 1373 data_reference_p dr1, | |
| 1374 data_reference_p dr2) | |
| 0 | 1375 { |
| 111 | 1376 struct data_dependence_relation *ddr; |
| 1377 | |
| 1378 /* Re-shuffle data-refs to be in topological order. */ | |
| 1379 if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
| 1380 > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
| 1381 std::swap (dr1, dr2); | |
| 1382 | |
| 1383 ddr = get_data_dependence (rdg, dr1, dr2); | |
| 1384 | |
| 1385 /* In case of no data dependence. */ | |
| 1386 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
| 1387 return false; | |
| 1388 /* For unknown data dependence or known data dependence which can't be | |
| 1389 expressed in classic distance vector, we check if it can be resolved | |
| 1390 by runtime alias check. If yes, we still consider data dependence | |
| 1391 as won't introduce data dependence cycle. */ | |
| 1392 else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
| 1393 || DDR_NUM_DIST_VECTS (ddr) == 0) | |
| 1394 return !runtime_alias_check_p (ddr, NULL, true); | |
| 1395 else if (DDR_NUM_DIST_VECTS (ddr) > 1) | |
| 1396 return true; | |
| 1397 else if (DDR_REVERSED_P (ddr) | |
| 1398 || lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
| 1399 return false; | |
| 1400 | |
| 1401 return true; | |
| 0 | 1402 } |
| 1403 | |
| 145 | 1404 void |
| 1405 loop_distribution::update_type_for_merge (struct graph *rdg, | |
| 1406 partition *partition1, | |
| 1407 partition *partition2) | |
| 0 | 1408 { |
| 111 | 1409 unsigned i, j; |
| 1410 bitmap_iterator bi, bj; | |
| 1411 data_reference_p dr1, dr2; | |
| 1412 | |
| 1413 EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
| 0 | 1414 { |
| 111 | 1415 unsigned start = (partition1 == partition2) ? i + 1 : 0; |
| 1416 | |
| 1417 dr1 = datarefs_vec[i]; | |
| 1418 EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, start, j, bj) | |
| 0 | 1419 { |
| 111 | 1420 dr2 = datarefs_vec[j]; |
| 1421 if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
| 1422 continue; | |
| 1423 | |
| 1424 /* Partition can only be executed sequentially if there is any | |
| 1425 data dependence cycle. */ | |
| 1426 if (data_dep_in_cycle_p (rdg, dr1, dr2)) | |
| 0 | 1427 { |
| 111 | 1428 partition1->type = PTYPE_SEQUENTIAL; |
| 1429 return; | |
| 0 | 1430 } |
| 1431 } | |
| 1432 } | |
| 1433 } | |
| 1434 | |
| 145 | 1435 partition * |
| 1436 loop_distribution::build_rdg_partition_for_vertex (struct graph *rdg, int v) | |
| 0 | 1437 { |
| 111 | 1438 partition *partition = partition_alloc (); |
| 1439 auto_vec<int, 3> nodes; | |
| 1440 unsigned i, j; | |
| 0 | 1441 int x; |
| 111 | 1442 data_reference_p dr; |
| 1443 | |
| 1444 graphds_dfs (rdg, &v, 1, &nodes, false, NULL); | |
| 1445 | |
| 1446 FOR_EACH_VEC_ELT (nodes, i, x) | |
| 1447 { | |
| 1448 bitmap_set_bit (partition->stmts, x); | |
| 1449 | |
| 1450 for (j = 0; RDG_DATAREFS (rdg, x).iterate (j, &dr); ++j) | |
| 1451 { | |
| 1452 unsigned idx = (unsigned) DR_INDEX (dr); | |
| 1453 gcc_assert (idx < datarefs_vec.length ()); | |
| 1454 | |
| 1455 /* Partition can only be executed sequentially if there is any | |
| 1456 unknown data reference. */ | |
| 1457 if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) | |
| 1458 || !DR_INIT (dr) || !DR_STEP (dr)) | |
| 1459 partition->type = PTYPE_SEQUENTIAL; | |
| 1460 | |
| 1461 bitmap_set_bit (partition->datarefs, idx); | |
| 1462 } | |
| 1463 } | |
| 1464 | |
| 1465 if (partition->type == PTYPE_SEQUENTIAL) | |
| 1466 return partition; | |
| 1467 | |
| 1468 /* Further check if any data dependence prevents us from executing the | |
| 1469 partition parallelly. */ | |
| 1470 update_type_for_merge (rdg, partition, partition); | |
| 1471 | |
| 1472 return partition; | |
| 0 | 1473 } |
| 1474 | |
| 111 | 1475 /* Given PARTITION of LOOP and RDG, record single load/store data references |
| 1476 for builtin partition in SRC_DR/DST_DR, return false if there is no such | |
| 1477 data references. */ | |
| 1478 | |
| 1479 static bool | |
| 145 | 1480 find_single_drs (class loop *loop, struct graph *rdg, partition *partition, |
| 111 | 1481 data_reference_p *dst_dr, data_reference_p *src_dr) |
| 0 | 1482 { |
| 1483 unsigned i; | |
| 111 | 1484 data_reference_p single_ld = NULL, single_st = NULL; |
| 0 | 1485 bitmap_iterator bi; |
| 111 | 1486 |
| 1487 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
| 0 | 1488 { |
| 111 | 1489 gimple *stmt = RDG_STMT (rdg, i); |
| 1490 data_reference_p dr; | |
| 1491 | |
| 1492 if (gimple_code (stmt) == GIMPLE_PHI) | |
| 1493 continue; | |
| 1494 | |
| 1495 /* Any scalar stmts are ok. */ | |
| 1496 if (!gimple_vuse (stmt)) | |
| 1497 continue; | |
| 1498 | |
| 1499 /* Otherwise just regular loads/stores. */ | |
| 1500 if (!gimple_assign_single_p (stmt)) | |
| 1501 return false; | |
| 1502 | |
| 1503 /* But exactly one store and/or load. */ | |
| 1504 for (unsigned j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) | |
| 1505 { | |
| 1506 tree type = TREE_TYPE (DR_REF (dr)); | |
| 1507 | |
| 1508 /* The memset, memcpy and memmove library calls are only | |
| 1509 able to deal with generic address space. */ | |
| 1510 if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
| 1511 return false; | |
| 1512 | |
| 1513 if (DR_IS_READ (dr)) | |
| 1514 { | |
| 1515 if (single_ld != NULL) | |
| 1516 return false; | |
| 1517 single_ld = dr; | |
| 1518 } | |
| 1519 else | |
| 1520 { | |
| 1521 if (single_st != NULL) | |
| 1522 return false; | |
| 1523 single_st = dr; | |
| 1524 } | |
| 1525 } | |
| 0 | 1526 } |
| 1527 | |
| 111 | 1528 if (!single_st) |
| 1529 return false; | |
| 1530 | |
| 1531 /* Bail out if this is a bitfield memory reference. */ | |
| 1532 if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF | |
| 1533 && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) | |
| 1534 return false; | |
| 1535 | |
| 1536 /* Data reference must be executed exactly once per iteration of each | |
| 1537 loop in the loop nest. We only need to check dominance information | |
| 1538 against the outermost one in a perfect loop nest because a bb can't | |
| 1539 dominate outermost loop's latch without dominating inner loop's. */ | |
| 1540 basic_block bb_st = gimple_bb (DR_STMT (single_st)); | |
| 1541 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) | |
| 1542 return false; | |
| 1543 | |
| 1544 if (single_ld) | |
| 1545 { | |
| 1546 gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); | |
| 1547 /* Direct aggregate copy or via an SSA name temporary. */ | |
| 1548 if (load != store | |
| 1549 && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
| 1550 return false; | |
| 1551 | |
| 1552 /* Bail out if this is a bitfield memory reference. */ | |
| 1553 if (TREE_CODE (DR_REF (single_ld)) == COMPONENT_REF | |
| 1554 && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_ld), 1))) | |
| 1555 return false; | |
| 1556 | |
| 1557 /* Load and store must be in the same loop nest. */ | |
| 1558 basic_block bb_ld = gimple_bb (DR_STMT (single_ld)); | |
| 1559 if (bb_st->loop_father != bb_ld->loop_father) | |
| 1560 return false; | |
| 1561 | |
| 1562 /* Data reference must be executed exactly once per iteration. | |
| 1563 Same as single_st, we only need to check against the outermost | |
| 1564 loop. */ | |
| 1565 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) | |
| 1566 return false; | |
| 1567 | |
| 1568 edge e = single_exit (bb_st->loop_father); | |
| 1569 bool dom_ld = dominated_by_p (CDI_DOMINATORS, e->src, bb_ld); | |
| 1570 bool dom_st = dominated_by_p (CDI_DOMINATORS, e->src, bb_st); | |
| 1571 if (dom_ld != dom_st) | |
| 1572 return false; | |
| 1573 } | |
| 1574 | |
| 1575 *src_dr = single_ld; | |
| 1576 *dst_dr = single_st; | |
| 1577 return true; | |
| 0 | 1578 } |
| 1579 | |
| 111 | 1580 /* Given data reference DR in LOOP_NEST, this function checks the enclosing |
| 1581 loops from inner to outer to see if loop's step equals to access size at | |
| 131 | 1582 each level of loop. Return 2 if we can prove this at all level loops; |
| 1583 record access base and size in BASE and SIZE; save loop's step at each | |
| 1584 level of loop in STEPS if it is not null. For example: | |
| 111 | 1585 |
| 1586 int arr[100][100][100]; | |
| 1587 for (i = 0; i < 100; i++) ;steps[2] = 40000 | |
| 1588 for (j = 100; j > 0; j--) ;steps[1] = -400 | |
| 1589 for (k = 0; k < 100; k++) ;steps[0] = 4 | |
| 131 | 1590 arr[i][j - 1][k] = 0; ;base = &arr, size = 4000000 |
| 1591 | |
| 1592 Return 1 if we can prove the equality at the innermost loop, but not all | |
| 1593 level loops. In this case, no information is recorded. | |
| 1594 | |
| 1595 Return 0 if no equality can be proven at any level loops. */ | |
| 1596 | |
| 1597 static int | |
| 111 | 1598 compute_access_range (loop_p loop_nest, data_reference_p dr, tree *base, |
| 1599 tree *size, vec<tree> *steps = NULL) | |
| 0 | 1600 { |
| 111 | 1601 location_t loc = gimple_location (DR_STMT (dr)); |
| 1602 basic_block bb = gimple_bb (DR_STMT (dr)); | |
| 145 | 1603 class loop *loop = bb->loop_father; |
| 111 | 1604 tree ref = DR_REF (dr); |
| 1605 tree access_base = build_fold_addr_expr (ref); | |
| 1606 tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
| 131 | 1607 int res = 0; |
| 111 | 1608 |
| 1609 do { | |
| 1610 tree scev_fn = analyze_scalar_evolution (loop, access_base); | |
| 1611 if (TREE_CODE (scev_fn) != POLYNOMIAL_CHREC) | |
| 131 | 1612 return res; |
| 111 | 1613 |
| 1614 access_base = CHREC_LEFT (scev_fn); | |
| 1615 if (tree_contains_chrecs (access_base, NULL)) | |
| 131 | 1616 return res; |
| 111 | 1617 |
| 1618 tree scev_step = CHREC_RIGHT (scev_fn); | |
| 1619 /* Only support constant steps. */ | |
| 1620 if (TREE_CODE (scev_step) != INTEGER_CST) | |
| 131 | 1621 return res; |
| 111 | 1622 |
| 1623 enum ev_direction access_dir = scev_direction (scev_fn); | |
| 1624 if (access_dir == EV_DIR_UNKNOWN) | |
| 131 | 1625 return res; |
| 111 | 1626 |
| 1627 if (steps != NULL) | |
| 1628 steps->safe_push (scev_step); | |
| 1629 | |
| 1630 scev_step = fold_convert_loc (loc, sizetype, scev_step); | |
| 1631 /* Compute absolute value of scev step. */ | |
| 1632 if (access_dir == EV_DIR_DECREASES) | |
| 1633 scev_step = fold_build1_loc (loc, NEGATE_EXPR, sizetype, scev_step); | |
| 1634 | |
| 1635 /* At each level of loop, scev step must equal to access size. In other | |
| 1636 words, DR must access consecutive memory between loop iterations. */ | |
| 1637 if (!operand_equal_p (scev_step, access_size, 0)) | |
| 131 | 1638 return res; |
| 1639 | |
| 1640 /* Access stride can be computed for data reference at least for the | |
| 1641 innermost loop. */ | |
| 1642 res = 1; | |
| 111 | 1643 |
| 1644 /* Compute DR's execution times in loop. */ | |
| 1645 tree niters = number_of_latch_executions (loop); | |
| 1646 niters = fold_convert_loc (loc, sizetype, niters); | |
| 1647 if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, bb)) | |
| 1648 niters = size_binop_loc (loc, PLUS_EXPR, niters, size_one_node); | |
| 1649 | |
| 1650 /* Compute DR's overall access size in loop. */ | |
| 1651 access_size = fold_build2_loc (loc, MULT_EXPR, sizetype, | |
| 1652 niters, scev_step); | |
| 1653 /* Adjust base address in case of negative step. */ | |
| 1654 if (access_dir == EV_DIR_DECREASES) | |
| 1655 { | |
| 1656 tree adj = fold_build2_loc (loc, MINUS_EXPR, sizetype, | |
| 1657 scev_step, access_size); | |
| 1658 access_base = fold_build_pointer_plus_loc (loc, access_base, adj); | |
| 1659 } | |
| 1660 } while (loop != loop_nest && (loop = loop_outer (loop)) != NULL); | |
| 1661 | |
| 1662 *base = access_base; | |
| 1663 *size = access_size; | |
| 131 | 1664 /* Access stride can be computed for data reference at each level loop. */ |
| 1665 return 2; | |
| 0 | 1666 } |
| 1667 | |
| 111 | 1668 /* Allocate and return builtin struct. Record information like DST_DR, |
| 1669 SRC_DR, DST_BASE, SRC_BASE and SIZE in the allocated struct. */ | |
| 1670 | |
| 1671 static struct builtin_info * | |
| 1672 alloc_builtin (data_reference_p dst_dr, data_reference_p src_dr, | |
| 1673 tree dst_base, tree src_base, tree size) | |
| 1674 { | |
| 1675 struct builtin_info *builtin = XNEW (struct builtin_info); | |
| 1676 builtin->dst_dr = dst_dr; | |
| 1677 builtin->src_dr = src_dr; | |
| 1678 builtin->dst_base = dst_base; | |
| 1679 builtin->src_base = src_base; | |
| 1680 builtin->size = size; | |
| 1681 return builtin; | |
| 1682 } | |
| 1683 | |
| 1684 /* Given data reference DR in loop nest LOOP, classify if it forms builtin | |
| 1685 memset call. */ | |
| 0 | 1686 |
| 1687 static void | |
| 111 | 1688 classify_builtin_st (loop_p loop, partition *partition, data_reference_p dr) |
| 0 | 1689 { |
| 111 | 1690 gimple *stmt = DR_STMT (dr); |
| 1691 tree base, size, rhs = gimple_assign_rhs1 (stmt); | |
| 1692 | |
| 1693 if (const_with_all_bytes_same (rhs) == -1 | |
| 1694 && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
| 1695 || (TYPE_MODE (TREE_TYPE (rhs)) | |
| 1696 != TYPE_MODE (unsigned_char_type_node)))) | |
| 1697 return; | |
| 1698 | |
| 1699 if (TREE_CODE (rhs) == SSA_NAME | |
| 1700 && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
| 1701 && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
| 1702 return; | |
| 1703 | |
| 131 | 1704 int res = compute_access_range (loop, dr, &base, &size); |
| 1705 if (res == 0) | |
| 1706 return; | |
| 1707 if (res == 1) | |
| 1708 { | |
| 1709 partition->kind = PKIND_PARTIAL_MEMSET; | |
| 1710 return; | |
| 1711 } | |
| 1712 | |
| 1713 poly_uint64 base_offset; | |
| 1714 unsigned HOST_WIDE_INT const_base_offset; | |
| 1715 tree base_base = strip_offset (base, &base_offset); | |
| 1716 if (!base_offset.is_constant (&const_base_offset)) | |
| 111 | 1717 return; |
| 1718 | |
| 1719 struct builtin_info *builtin; | |
| 1720 builtin = alloc_builtin (dr, NULL, base, NULL_TREE, size); | |
| 131 | 1721 builtin->dst_base_base = base_base; |
| 1722 builtin->dst_base_offset = const_base_offset; | |
| 111 | 1723 partition->builtin = builtin; |
| 1724 partition->kind = PKIND_MEMSET; | |
| 0 | 1725 } |
| 1726 | |
| 111 | 1727 /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify |
| 1728 if it forms builtin memcpy or memmove call. */ | |
| 0 | 1729 |
| 145 | 1730 void |
| 1731 loop_distribution::classify_builtin_ldst (loop_p loop, struct graph *rdg, | |
| 1732 partition *partition, | |
| 1733 data_reference_p dst_dr, | |
| 1734 data_reference_p src_dr) | |
| 0 | 1735 { |
| 111 | 1736 tree base, size, src_base, src_size; |
| 1737 auto_vec<tree> dst_steps, src_steps; | |
| 1738 | |
| 131 | 1739 /* Compute access range of both load and store. */ |
| 1740 int res = compute_access_range (loop, dst_dr, &base, &size, &dst_steps); | |
| 1741 if (res != 2) | |
| 1742 return; | |
| 1743 res = compute_access_range (loop, src_dr, &src_base, &src_size, &src_steps); | |
| 1744 if (res != 2) | |
| 1745 return; | |
| 1746 | |
| 1747 /* They much have the same access size. */ | |
| 1748 if (!operand_equal_p (size, src_size, 0)) | |
| 111 | 1749 return; |
| 1750 | |
| 1751 /* Load and store in loop nest must access memory in the same way, i.e, | |
| 1752 their must have the same steps in each loop of the nest. */ | |
| 1753 if (dst_steps.length () != src_steps.length ()) | |
| 1754 return; | |
| 1755 for (unsigned i = 0; i < dst_steps.length (); ++i) | |
| 1756 if (!operand_equal_p (dst_steps[i], src_steps[i], 0)) | |
| 1757 return; | |
| 1758 | |
| 1759 /* Now check that if there is a dependence. */ | |
| 1760 ddr_p ddr = get_data_dependence (rdg, src_dr, dst_dr); | |
| 1761 | |
| 1762 /* Classify as memcpy if no dependence between load and store. */ | |
| 1763 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
| 1764 { | |
| 1765 partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
| 1766 partition->kind = PKIND_MEMCPY; | |
| 1767 return; | |
| 1768 } | |
| 1769 | |
| 1770 /* Can't do memmove in case of unknown dependence or dependence without | |
| 1771 classical distance vector. */ | |
| 1772 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
| 1773 || DDR_NUM_DIST_VECTS (ddr) == 0) | |
| 1774 return; | |
| 1775 | |
| 1776 unsigned i; | |
| 1777 lambda_vector dist_v; | |
| 1778 int num_lev = (DDR_LOOP_NEST (ddr)).length (); | |
| 1779 FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) | |
| 0 | 1780 { |
| 111 | 1781 unsigned dep_lev = dependence_level (dist_v, num_lev); |
| 1782 /* Can't do memmove if load depends on store. */ | |
| 1783 if (dep_lev > 0 && dist_v[dep_lev - 1] > 0 && !DDR_REVERSED_P (ddr)) | |
| 1784 return; | |
| 1785 } | |
| 1786 | |
| 1787 partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
| 1788 partition->kind = PKIND_MEMMOVE; | |
| 1789 return; | |
| 1790 } | |
| 1791 | |
| 145 | 1792 bool |
| 1793 loop_distribution::classify_partition (loop_p loop, | |
| 1794 struct graph *rdg, partition *partition, | |
| 1795 bitmap stmt_in_all_partitions) | |
| 111 | 1796 { |
| 1797 bitmap_iterator bi; | |
| 1798 unsigned i; | |
| 1799 data_reference_p single_ld = NULL, single_st = NULL; | |
| 1800 bool volatiles_p = false, has_reduction = false; | |
| 1801 | |
| 1802 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
| 1803 { | |
| 1804 gimple *stmt = RDG_STMT (rdg, i); | |
| 1805 | |
| 1806 if (gimple_has_volatile_ops (stmt)) | |
| 1807 volatiles_p = true; | |
| 1808 | |
| 1809 /* If the stmt is not included by all partitions and there is uses | |
| 1810 outside of the loop, then mark the partition as reduction. */ | |
| 1811 if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
| 0 | 1812 { |
| 111 | 1813 /* Due to limitation in the transform phase we have to fuse all |
| 1814 reduction partitions. As a result, this could cancel valid | |
| 1815 loop distribution especially for loop that induction variable | |
| 1816 is used outside of loop. To workaround this issue, we skip | |
| 1817 marking partition as reudction if the reduction stmt belongs | |
| 1818 to all partitions. In such case, reduction will be computed | |
| 1819 correctly no matter how partitions are fused/distributed. */ | |
| 1820 if (!bitmap_bit_p (stmt_in_all_partitions, i)) | |
| 145 | 1821 partition->reduction_p = true; |
| 1822 else | |
| 1823 has_reduction = true; | |
| 0 | 1824 } |
| 1825 } | |
| 1826 | |
| 145 | 1827 /* Simple workaround to prevent classifying the partition as builtin |
| 1828 if it contains any use outside of loop. For the case where all | |
| 1829 partitions have the reduction this simple workaround is delayed | |
| 1830 to only affect the last partition. */ | |
| 1831 if (partition->reduction_p) | |
| 1832 return has_reduction; | |
| 1833 | |
| 111 | 1834 /* Perform general partition disqualification for builtins. */ |
| 1835 if (volatiles_p | |
| 1836 || !flag_tree_loop_distribute_patterns) | |
| 145 | 1837 return has_reduction; |
| 111 | 1838 |
| 1839 /* Find single load/store data references for builtin partition. */ | |
| 1840 if (!find_single_drs (loop, rdg, partition, &single_st, &single_ld)) | |
| 145 | 1841 return has_reduction; |
| 1842 | |
| 1843 partition->loc = gimple_location (DR_STMT (single_st)); | |
| 111 | 1844 |
| 1845 /* Classify the builtin kind. */ | |
| 1846 if (single_ld == NULL) | |
| 1847 classify_builtin_st (loop, partition, single_st); | |
| 1848 else | |
| 1849 classify_builtin_ldst (loop, rdg, partition, single_st, single_ld); | |
| 145 | 1850 return has_reduction; |
| 0 | 1851 } |
| 1852 | |
| 145 | 1853 bool |
| 1854 loop_distribution::share_memory_accesses (struct graph *rdg, | |
| 111 | 1855 partition *partition1, partition *partition2) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1856 { |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1857 unsigned i, j; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1858 bitmap_iterator bi, bj; |
| 111 | 1859 data_reference_p dr1, dr2; |
| 1860 | |
| 1861 /* First check whether in the intersection of the two partitions are | |
| 1862 any loads or stores. Common loads are the situation that happens | |
| 1863 most often. */ | |
| 1864 EXECUTE_IF_AND_IN_BITMAP (partition1->stmts, partition2->stmts, 0, i, bi) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1865 if (RDG_MEM_WRITE_STMT (rdg, i) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1866 || RDG_MEM_READS_STMT (rdg, i)) |
| 111 | 1867 return true; |
| 1868 | |
| 1869 /* Then check whether the two partitions access the same memory object. */ | |
| 1870 EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
| 1871 { | |
| 1872 dr1 = datarefs_vec[i]; | |
| 1873 | |
| 1874 if (!DR_BASE_ADDRESS (dr1) | |
| 1875 || !DR_OFFSET (dr1) || !DR_INIT (dr1) || !DR_STEP (dr1)) | |
| 1876 continue; | |
| 1877 | |
| 1878 EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, 0, j, bj) | |
| 1879 { | |
| 1880 dr2 = datarefs_vec[j]; | |
| 1881 | |
| 1882 if (!DR_BASE_ADDRESS (dr2) | |
| 1883 || !DR_OFFSET (dr2) || !DR_INIT (dr2) || !DR_STEP (dr2)) | |
| 1884 continue; | |
| 1885 | |
| 1886 if (operand_equal_p (DR_BASE_ADDRESS (dr1), DR_BASE_ADDRESS (dr2), 0) | |
| 1887 && operand_equal_p (DR_OFFSET (dr1), DR_OFFSET (dr2), 0) | |
| 1888 && operand_equal_p (DR_INIT (dr1), DR_INIT (dr2), 0) | |
| 1889 && operand_equal_p (DR_STEP (dr1), DR_STEP (dr2), 0)) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1890 return true; |
| 111 | 1891 } |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1892 } |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1893 |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1894 return false; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1895 } |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1896 |
| 111 | 1897 /* For each seed statement in STARTING_STMTS, this function builds |
| 1898 partition for it by adding depended statements according to RDG. | |
| 1899 All partitions are recorded in PARTITIONS. */ | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1900 |
| 145 | 1901 void |
| 1902 loop_distribution::rdg_build_partitions (struct graph *rdg, | |
| 1903 vec<gimple *> starting_stmts, | |
| 1904 vec<partition *> *partitions) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1905 { |
| 111 | 1906 auto_bitmap processed; |
| 0 | 1907 int i; |
| 111 | 1908 gimple *stmt; |
| 1909 | |
| 1910 FOR_EACH_VEC_ELT (starting_stmts, i, stmt) | |
| 0 | 1911 { |
| 111 | 1912 int v = rdg_vertex_for_stmt (rdg, stmt); |
| 1913 | |
| 1914 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 1915 fprintf (dump_file, | |
| 1916 "ldist asked to generate code for vertex %d\n", v); | |
| 1917 | |
| 1918 /* If the vertex is already contained in another partition so | |
| 1919 is the partition rooted at it. */ | |
| 0 | 1920 if (bitmap_bit_p (processed, v)) |
| 1921 continue; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1922 |
| 111 | 1923 partition *partition = build_rdg_partition_for_vertex (rdg, v); |
| 1924 bitmap_ior_into (processed, partition->stmts); | |
| 1925 | |
| 1926 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 1927 { |
| 111 | 1928 fprintf (dump_file, "ldist creates useful %s partition:\n", |
| 1929 partition->type == PTYPE_PARALLEL ? "parallel" : "sequent"); | |
| 1930 bitmap_print (dump_file, partition->stmts, " ", "\n"); | |
| 0 | 1931 } |
| 111 | 1932 |
| 1933 partitions->safe_push (partition); | |
| 0 | 1934 } |
| 1935 | |
| 111 | 1936 /* All vertices should have been assigned to at least one partition now, |
| 1937 other than vertices belonging to dead code. */ | |
| 0 | 1938 } |
| 1939 | |
| 1940 /* Dump to FILE the PARTITIONS. */ | |
| 1941 | |
| 1942 static void | |
| 111 | 1943 dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
| 0 | 1944 { |
| 1945 int i; | |
| 111 | 1946 partition *partition; |
| 1947 | |
| 1948 FOR_EACH_VEC_ELT (partitions, i, partition) | |
| 1949 debug_bitmap_file (file, partition->stmts); | |
| 0 | 1950 } |
| 1951 | |
| 1952 /* Debug PARTITIONS. */ | |
| 111 | 1953 extern void debug_rdg_partitions (vec<partition *> ); |
| 0 | 1954 |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1955 DEBUG_FUNCTION void |
| 111 | 1956 debug_rdg_partitions (vec<partition *> partitions) |
| 0 | 1957 { |
| 1958 dump_rdg_partitions (stderr, partitions); | |
| 1959 } | |
| 1960 | |
| 1961 /* Returns the number of read and write operations in the RDG. */ | |
| 1962 | |
| 1963 static int | |
| 1964 number_of_rw_in_rdg (struct graph *rdg) | |
| 1965 { | |
| 1966 int i, res = 0; | |
| 1967 | |
| 1968 for (i = 0; i < rdg->n_vertices; i++) | |
| 1969 { | |
| 1970 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
| 1971 ++res; | |
| 1972 | |
| 1973 if (RDG_MEM_READS_STMT (rdg, i)) | |
| 1974 ++res; | |
| 1975 } | |
| 1976 | |
| 1977 return res; | |
| 1978 } | |
| 1979 | |
| 1980 /* Returns the number of read and write operations in a PARTITION of | |
| 1981 the RDG. */ | |
| 1982 | |
| 1983 static int | |
| 111 | 1984 number_of_rw_in_partition (struct graph *rdg, partition *partition) |
| 0 | 1985 { |
| 1986 int res = 0; | |
| 1987 unsigned i; | |
| 1988 bitmap_iterator ii; | |
| 1989 | |
| 111 | 1990 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
| 0 | 1991 { |
| 1992 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
| 1993 ++res; | |
| 1994 | |
| 1995 if (RDG_MEM_READS_STMT (rdg, i)) | |
| 1996 ++res; | |
| 1997 } | |
| 1998 | |
| 1999 return res; | |
| 2000 } | |
| 2001 | |
| 2002 /* Returns true when one of the PARTITIONS contains all the read or | |
| 2003 write operations of RDG. */ | |
| 2004 | |
| 2005 static bool | |
| 111 | 2006 partition_contains_all_rw (struct graph *rdg, |
| 2007 vec<partition *> partitions) | |
| 0 | 2008 { |
| 2009 int i; | |
| 111 | 2010 partition *partition; |
| 0 | 2011 int nrw = number_of_rw_in_rdg (rdg); |
| 2012 | |
| 111 | 2013 FOR_EACH_VEC_ELT (partitions, i, partition) |
| 0 | 2014 if (nrw == number_of_rw_in_partition (rdg, partition)) |
| 2015 return true; | |
| 2016 | |
| 2017 return false; | |
| 2018 } | |
| 2019 | |
| 145 | 2020 int |
| 2021 loop_distribution::pg_add_dependence_edges (struct graph *rdg, int dir, | |
| 111 | 2022 bitmap drs1, bitmap drs2, vec<ddr_p> *alias_ddrs) |
| 2023 { | |
| 2024 unsigned i, j; | |
| 2025 bitmap_iterator bi, bj; | |
| 2026 data_reference_p dr1, dr2, saved_dr1; | |
| 2027 | |
| 2028 /* dependence direction - 0 is no dependence, -1 is back, | |
| 2029 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
| 2030 EXECUTE_IF_SET_IN_BITMAP (drs1, 0, i, bi) | |
| 2031 { | |
| 2032 dr1 = datarefs_vec[i]; | |
| 2033 | |
| 2034 EXECUTE_IF_SET_IN_BITMAP (drs2, 0, j, bj) | |
| 2035 { | |
| 2036 int res, this_dir = 1; | |
| 2037 ddr_p ddr; | |
| 2038 | |
| 2039 dr2 = datarefs_vec[j]; | |
| 2040 | |
| 2041 /* Skip all <read, read> data dependence. */ | |
| 2042 if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
| 2043 continue; | |
| 2044 | |
| 2045 saved_dr1 = dr1; | |
| 2046 /* Re-shuffle data-refs to be in topological order. */ | |
| 2047 if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
| 2048 > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
| 2049 { | |
| 2050 std::swap (dr1, dr2); | |
| 2051 this_dir = -this_dir; | |
| 2052 } | |
| 2053 ddr = get_data_dependence (rdg, dr1, dr2); | |
| 2054 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
| 2055 { | |
| 2056 this_dir = 0; | |
| 2057 res = data_ref_compare_tree (DR_BASE_ADDRESS (dr1), | |
| 2058 DR_BASE_ADDRESS (dr2)); | |
| 2059 /* Be conservative. If data references are not well analyzed, | |
| 2060 or the two data references have the same base address and | |
| 2061 offset, add dependence and consider it alias to each other. | |
| 145 | 2062 In other words, the dependence cannot be resolved by |
| 111 | 2063 runtime alias check. */ |
| 2064 if (!DR_BASE_ADDRESS (dr1) || !DR_BASE_ADDRESS (dr2) | |
| 2065 || !DR_OFFSET (dr1) || !DR_OFFSET (dr2) | |
| 2066 || !DR_INIT (dr1) || !DR_INIT (dr2) | |
| 2067 || !DR_STEP (dr1) || !tree_fits_uhwi_p (DR_STEP (dr1)) | |
| 2068 || !DR_STEP (dr2) || !tree_fits_uhwi_p (DR_STEP (dr2)) | |
| 2069 || res == 0) | |
| 2070 this_dir = 2; | |
| 2071 /* Data dependence could be resolved by runtime alias check, | |
| 2072 record it in ALIAS_DDRS. */ | |
| 2073 else if (alias_ddrs != NULL) | |
| 2074 alias_ddrs->safe_push (ddr); | |
| 2075 /* Or simply ignore it. */ | |
| 2076 } | |
| 2077 else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) | |
| 2078 { | |
| 2079 if (DDR_REVERSED_P (ddr)) | |
| 2080 this_dir = -this_dir; | |
| 2081 | |
| 2082 /* Known dependences can still be unordered througout the | |
| 2083 iteration space, see gcc.dg/tree-ssa/ldist-16.c. */ | |
| 2084 if (DDR_NUM_DIST_VECTS (ddr) != 1) | |
| 2085 this_dir = 2; | |
| 2086 /* If the overlap is exact preserve stmt order. */ | |
| 131 | 2087 else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), |
| 2088 DDR_NB_LOOPS (ddr))) | |
| 111 | 2089 ; |
| 2090 /* Else as the distance vector is lexicographic positive swap | |
| 2091 the dependence direction. */ | |
| 2092 else | |
| 2093 this_dir = -this_dir; | |
| 2094 } | |
| 2095 else | |
| 2096 this_dir = 0; | |
| 2097 if (this_dir == 2) | |
| 2098 return 2; | |
| 2099 else if (dir == 0) | |
| 2100 dir = this_dir; | |
| 2101 else if (this_dir != 0 && dir != this_dir) | |
| 2102 return 2; | |
| 2103 /* Shuffle "back" dr1. */ | |
| 2104 dr1 = saved_dr1; | |
| 2105 } | |
| 2106 } | |
| 2107 return dir; | |
| 2108 } | |
| 2109 | |
| 2110 /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
| 0 | 2111 |
| 2112 static int | |
| 111 | 2113 pgcmp (const void *v1_, const void *v2_) |
| 2114 { | |
| 2115 const vertex *v1 = (const vertex *)v1_; | |
| 2116 const vertex *v2 = (const vertex *)v2_; | |
| 2117 return v2->post - v1->post; | |
| 2118 } | |
| 2119 | |
| 2120 /* Data attached to vertices of partition dependence graph. */ | |
| 2121 struct pg_vdata | |
| 2122 { | |
| 2123 /* ID of the corresponding partition. */ | |
| 2124 int id; | |
| 2125 /* The partition. */ | |
| 2126 struct partition *partition; | |
| 2127 }; | |
| 2128 | |
| 2129 /* Data attached to edges of partition dependence graph. */ | |
| 2130 struct pg_edata | |
| 2131 { | |
| 2132 /* If the dependence edge can be resolved by runtime alias check, | |
| 2133 this vector contains data dependence relations for runtime alias | |
| 2134 check. On the other hand, if the dependence edge is introduced | |
| 2135 because of compilation time known data dependence, this vector | |
| 2136 contains nothing. */ | |
| 2137 vec<ddr_p> alias_ddrs; | |
| 2138 }; | |
| 2139 | |
| 2140 /* Callback data for traversing edges in graph. */ | |
| 2141 struct pg_edge_callback_data | |
| 0 | 2142 { |
| 111 | 2143 /* Bitmap contains strong connected components should be merged. */ |
| 2144 bitmap sccs_to_merge; | |
| 2145 /* Array constains component information for all vertices. */ | |
| 2146 int *vertices_component; | |
| 2147 /* Vector to record all data dependence relations which are needed | |
| 2148 to break strong connected components by runtime alias checks. */ | |
| 2149 vec<ddr_p> *alias_ddrs; | |
| 2150 }; | |
| 2151 | |
| 2152 /* Initialize vertice's data for partition dependence graph PG with | |
| 2153 PARTITIONS. */ | |
| 2154 | |
| 2155 static void | |
| 2156 init_partition_graph_vertices (struct graph *pg, | |
| 2157 vec<struct partition *> *partitions) | |
| 2158 { | |
| 2159 int i; | |
| 2160 partition *partition; | |
| 2161 struct pg_vdata *data; | |
| 2162 | |
| 2163 for (i = 0; partitions->iterate (i, &partition); ++i) | |
| 2164 { | |
| 2165 data = new pg_vdata; | |
| 2166 pg->vertices[i].data = data; | |
| 2167 data->id = i; | |
| 2168 data->partition = partition; | |
| 2169 } | |
| 2170 } | |
| 2171 | |
| 2172 /* Add edge <I, J> to partition dependence graph PG. Attach vector of data | |
| 2173 dependence relations to the EDGE if DDRS isn't NULL. */ | |
| 2174 | |
| 2175 static void | |
| 2176 add_partition_graph_edge (struct graph *pg, int i, int j, vec<ddr_p> *ddrs) | |
| 2177 { | |
| 2178 struct graph_edge *e = add_edge (pg, i, j); | |
| 2179 | |
| 2180 /* If the edge is attached with data dependence relations, it means this | |
| 2181 dependence edge can be resolved by runtime alias checks. */ | |
| 2182 if (ddrs != NULL) | |
| 0 | 2183 { |
| 111 | 2184 struct pg_edata *data = new pg_edata; |
| 2185 | |
| 2186 gcc_assert (ddrs->length () > 0); | |
| 2187 e->data = data; | |
| 2188 data->alias_ddrs = vNULL; | |
| 2189 data->alias_ddrs.safe_splice (*ddrs); | |
| 2190 } | |
| 2191 } | |
| 2192 | |
| 2193 /* Callback function for graph travesal algorithm. It returns true | |
| 2194 if edge E should skipped when traversing the graph. */ | |
| 2195 | |
| 2196 static bool | |
| 2197 pg_skip_alias_edge (struct graph_edge *e) | |
| 2198 { | |
| 2199 struct pg_edata *data = (struct pg_edata *)e->data; | |
| 2200 return (data != NULL && data->alias_ddrs.length () > 0); | |
| 2201 } | |
| 2202 | |
| 2203 /* Callback function freeing data attached to edge E of graph. */ | |
| 2204 | |
| 2205 static void | |
| 2206 free_partition_graph_edata_cb (struct graph *, struct graph_edge *e, void *) | |
| 2207 { | |
| 2208 if (e->data != NULL) | |
| 2209 { | |
| 2210 struct pg_edata *data = (struct pg_edata *)e->data; | |
| 2211 data->alias_ddrs.release (); | |
| 2212 delete data; | |
| 2213 } | |
| 2214 } | |
| 2215 | |
| 2216 /* Free data attached to vertice of partition dependence graph PG. */ | |
| 2217 | |
| 2218 static void | |
| 2219 free_partition_graph_vdata (struct graph *pg) | |
| 2220 { | |
| 2221 int i; | |
| 2222 struct pg_vdata *data; | |
| 2223 | |
| 2224 for (i = 0; i < pg->n_vertices; ++i) | |
| 2225 { | |
| 2226 data = (struct pg_vdata *)pg->vertices[i].data; | |
| 2227 delete data; | |
| 2228 } | |
| 2229 } | |
| 2230 | |
| 2231 /* Build and return partition dependence graph for PARTITIONS. RDG is | |
| 2232 reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
| 2233 is true, data dependence caused by possible alias between references | |
| 2234 is ignored, as if it doesn't exist at all; otherwise all depdendences | |
| 2235 are considered. */ | |
| 2236 | |
| 145 | 2237 struct graph * |
| 2238 loop_distribution::build_partition_graph (struct graph *rdg, | |
| 2239 vec<struct partition *> *partitions, | |
| 2240 bool ignore_alias_p) | |
| 111 | 2241 { |
| 2242 int i, j; | |
| 2243 struct partition *partition1, *partition2; | |
| 2244 graph *pg = new_graph (partitions->length ()); | |
| 2245 auto_vec<ddr_p> alias_ddrs, *alias_ddrs_p; | |
| 2246 | |
| 2247 alias_ddrs_p = ignore_alias_p ? NULL : &alias_ddrs; | |
| 2248 | |
| 2249 init_partition_graph_vertices (pg, partitions); | |
| 2250 | |
| 2251 for (i = 0; partitions->iterate (i, &partition1); ++i) | |
| 2252 { | |
| 2253 for (j = i + 1; partitions->iterate (j, &partition2); ++j) | |
| 0 | 2254 { |
| 111 | 2255 /* dependence direction - 0 is no dependence, -1 is back, |
| 2256 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
| 2257 int dir = 0; | |
| 2258 | |
| 2259 /* If the first partition has reduction, add back edge; if the | |
| 2260 second partition has reduction, add forth edge. This makes | |
| 2261 sure that reduction partition will be sorted as the last one. */ | |
| 2262 if (partition_reduction_p (partition1)) | |
| 2263 dir = -1; | |
| 2264 else if (partition_reduction_p (partition2)) | |
| 2265 dir = 1; | |
| 2266 | |
| 2267 /* Cleanup the temporary vector. */ | |
| 2268 alias_ddrs.truncate (0); | |
| 2269 | |
| 2270 dir = pg_add_dependence_edges (rdg, dir, partition1->datarefs, | |
| 2271 partition2->datarefs, alias_ddrs_p); | |
| 2272 | |
| 2273 /* Add edge to partition graph if there exists dependence. There | |
| 2274 are two types of edges. One type edge is caused by compilation | |
| 145 | 2275 time known dependence, this type cannot be resolved by runtime |
| 111 | 2276 alias check. The other type can be resolved by runtime alias |
| 2277 check. */ | |
| 2278 if (dir == 1 || dir == 2 | |
| 2279 || alias_ddrs.length () > 0) | |
| 2280 { | |
| 2281 /* Attach data dependence relations to edge that can be resolved | |
| 2282 by runtime alias check. */ | |
| 2283 bool alias_edge_p = (dir != 1 && dir != 2); | |
| 2284 add_partition_graph_edge (pg, i, j, | |
| 2285 (alias_edge_p) ? &alias_ddrs : NULL); | |
| 2286 } | |
| 2287 if (dir == -1 || dir == 2 | |
| 2288 || alias_ddrs.length () > 0) | |
| 2289 { | |
| 2290 /* Attach data dependence relations to edge that can be resolved | |
| 2291 by runtime alias check. */ | |
| 2292 bool alias_edge_p = (dir != -1 && dir != 2); | |
| 2293 add_partition_graph_edge (pg, j, i, | |
| 2294 (alias_edge_p) ? &alias_ddrs : NULL); | |
| 2295 } | |
| 0 | 2296 } |
| 2297 } | |
| 111 | 2298 return pg; |
| 2299 } | |
| 2300 | |
| 2301 /* Sort partitions in PG in descending post order and store them in | |
| 2302 PARTITIONS. */ | |
| 2303 | |
| 2304 static void | |
| 2305 sort_partitions_by_post_order (struct graph *pg, | |
| 2306 vec<struct partition *> *partitions) | |
| 2307 { | |
| 2308 int i; | |
| 2309 struct pg_vdata *data; | |
| 2310 | |
| 2311 /* Now order the remaining nodes in descending postorder. */ | |
| 2312 qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); | |
| 2313 partitions->truncate (0); | |
| 2314 for (i = 0; i < pg->n_vertices; ++i) | |
| 2315 { | |
| 2316 data = (struct pg_vdata *)pg->vertices[i].data; | |
| 2317 if (data->partition) | |
| 2318 partitions->safe_push (data->partition); | |
| 2319 } | |
| 2320 } | |
| 2321 | |
| 145 | 2322 void |
| 2323 loop_distribution::merge_dep_scc_partitions (struct graph *rdg, | |
| 2324 vec<struct partition *> *partitions, | |
| 2325 bool ignore_alias_p) | |
| 111 | 2326 { |
| 2327 struct partition *partition1, *partition2; | |
| 2328 struct pg_vdata *data; | |
| 2329 graph *pg = build_partition_graph (rdg, partitions, ignore_alias_p); | |
| 2330 int i, j, num_sccs = graphds_scc (pg, NULL); | |
| 2331 | |
| 2332 /* Strong connected compoenent means dependence cycle, we cannot distribute | |
| 2333 them. So fuse them together. */ | |
| 2334 if ((unsigned) num_sccs < partitions->length ()) | |
| 2335 { | |
| 2336 for (i = 0; i < num_sccs; ++i) | |
| 2337 { | |
| 2338 for (j = 0; partitions->iterate (j, &partition1); ++j) | |
| 2339 if (pg->vertices[j].component == i) | |
| 2340 break; | |
| 2341 for (j = j + 1; partitions->iterate (j, &partition2); ++j) | |
| 2342 if (pg->vertices[j].component == i) | |
| 2343 { | |
| 2344 partition_merge_into (NULL, partition1, | |
| 2345 partition2, FUSE_SAME_SCC); | |
| 2346 partition1->type = PTYPE_SEQUENTIAL; | |
| 2347 (*partitions)[j] = NULL; | |
| 2348 partition_free (partition2); | |
| 2349 data = (struct pg_vdata *)pg->vertices[j].data; | |
| 2350 data->partition = NULL; | |
| 2351 } | |
| 2352 } | |
| 2353 } | |
| 2354 | |
| 2355 sort_partitions_by_post_order (pg, partitions); | |
| 2356 gcc_assert (partitions->length () == (unsigned)num_sccs); | |
| 2357 free_partition_graph_vdata (pg); | |
| 2358 free_graph (pg); | |
| 2359 } | |
| 2360 | |
| 2361 /* Callback function for traversing edge E in graph G. DATA is private | |
| 2362 callback data. */ | |
| 2363 | |
| 2364 static void | |
| 2365 pg_collect_alias_ddrs (struct graph *g, struct graph_edge *e, void *data) | |
| 2366 { | |
| 2367 int i, j, component; | |
| 2368 struct pg_edge_callback_data *cbdata; | |
| 2369 struct pg_edata *edata = (struct pg_edata *) e->data; | |
| 2370 | |
| 2371 /* If the edge doesn't have attached data dependence, it represents | |
| 2372 compilation time known dependences. This type dependence cannot | |
| 2373 be resolved by runtime alias check. */ | |
| 2374 if (edata == NULL || edata->alias_ddrs.length () == 0) | |
| 2375 return; | |
| 2376 | |
| 2377 cbdata = (struct pg_edge_callback_data *) data; | |
| 2378 i = e->src; | |
| 2379 j = e->dest; | |
| 2380 component = cbdata->vertices_component[i]; | |
| 2381 /* Vertices are topologically sorted according to compilation time | |
| 2382 known dependences, so we can break strong connected components | |
| 2383 by removing edges of the opposite direction, i.e, edges pointing | |
| 2384 from vertice with smaller post number to vertice with bigger post | |
| 2385 number. */ | |
| 2386 if (g->vertices[i].post < g->vertices[j].post | |
| 2387 /* We only need to remove edges connecting vertices in the same | |
| 2388 strong connected component to break it. */ | |
| 2389 && component == cbdata->vertices_component[j] | |
| 2390 /* Check if we want to break the strong connected component or not. */ | |
| 2391 && !bitmap_bit_p (cbdata->sccs_to_merge, component)) | |
| 2392 cbdata->alias_ddrs->safe_splice (edata->alias_ddrs); | |
| 2393 } | |
| 2394 | |
| 2395 /* This is the main function breaking strong conected components in | |
| 2396 PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
| 2397 relations for runtime alias check in ALIAS_DDRS. */ | |
| 145 | 2398 void |
| 2399 loop_distribution::break_alias_scc_partitions (struct graph *rdg, | |
| 2400 vec<struct partition *> *partitions, | |
| 2401 vec<ddr_p> *alias_ddrs) | |
| 111 | 2402 { |
| 2403 int i, j, k, num_sccs, num_sccs_no_alias; | |
| 2404 /* Build partition dependence graph. */ | |
| 2405 graph *pg = build_partition_graph (rdg, partitions, false); | |
| 2406 | |
| 2407 alias_ddrs->truncate (0); | |
| 2408 /* Find strong connected components in the graph, with all dependence edges | |
| 2409 considered. */ | |
| 2410 num_sccs = graphds_scc (pg, NULL); | |
| 2411 /* All SCCs now can be broken by runtime alias checks because SCCs caused by | |
| 2412 compilation time known dependences are merged before this function. */ | |
| 2413 if ((unsigned) num_sccs < partitions->length ()) | |
| 2414 { | |
| 2415 struct pg_edge_callback_data cbdata; | |
| 2416 auto_bitmap sccs_to_merge; | |
| 2417 auto_vec<enum partition_type> scc_types; | |
| 2418 struct partition *partition, *first; | |
| 2419 | |
| 2420 /* If all partitions in a SCC have the same type, we can simply merge the | |
| 2421 SCC. This loop finds out such SCCS and record them in bitmap. */ | |
| 2422 bitmap_set_range (sccs_to_merge, 0, (unsigned) num_sccs); | |
| 2423 for (i = 0; i < num_sccs; ++i) | |
| 2424 { | |
| 2425 for (j = 0; partitions->iterate (j, &first); ++j) | |
| 2426 if (pg->vertices[j].component == i) | |
| 2427 break; | |
| 131 | 2428 |
| 2429 bool same_type = true, all_builtins = partition_builtin_p (first); | |
| 111 | 2430 for (++j; partitions->iterate (j, &partition); ++j) |
| 2431 { | |
| 2432 if (pg->vertices[j].component != i) | |
| 2433 continue; | |
| 2434 | |
| 2435 if (first->type != partition->type) | |
| 2436 { | |
| 131 | 2437 same_type = false; |
| 111 | 2438 break; |
| 2439 } | |
| 131 | 2440 all_builtins &= partition_builtin_p (partition); |
| 111 | 2441 } |
| 131 | 2442 /* Merge SCC if all partitions in SCC have the same type, though the |
| 2443 result partition is sequential, because vectorizer can do better | |
| 2444 runtime alias check. One expecption is all partitions in SCC are | |
| 2445 builtins. */ | |
| 2446 if (!same_type || all_builtins) | |
| 2447 bitmap_clear_bit (sccs_to_merge, i); | |
| 111 | 2448 } |
| 2449 | |
| 2450 /* Initialize callback data for traversing. */ | |
| 2451 cbdata.sccs_to_merge = sccs_to_merge; | |
| 2452 cbdata.alias_ddrs = alias_ddrs; | |
| 2453 cbdata.vertices_component = XNEWVEC (int, pg->n_vertices); | |
| 2454 /* Record the component information which will be corrupted by next | |
| 2455 graph scc finding call. */ | |
| 2456 for (i = 0; i < pg->n_vertices; ++i) | |
| 2457 cbdata.vertices_component[i] = pg->vertices[i].component; | |
| 2458 | |
| 2459 /* Collect data dependences for runtime alias checks to break SCCs. */ | |
| 2460 if (bitmap_count_bits (sccs_to_merge) != (unsigned) num_sccs) | |
| 2461 { | |
| 2462 /* Run SCC finding algorithm again, with alias dependence edges | |
| 2463 skipped. This is to topologically sort partitions according to | |
| 2464 compilation time known dependence. Note the topological order | |
| 2465 is stored in the form of pg's post order number. */ | |
| 2466 num_sccs_no_alias = graphds_scc (pg, NULL, pg_skip_alias_edge); | |
| 2467 gcc_assert (partitions->length () == (unsigned) num_sccs_no_alias); | |
| 2468 /* With topological order, we can construct two subgraphs L and R. | |
| 2469 L contains edge <x, y> where x < y in terms of post order, while | |
| 2470 R contains edge <x, y> where x > y. Edges for compilation time | |
| 2471 known dependence all fall in R, so we break SCCs by removing all | |
| 2472 (alias) edges of in subgraph L. */ | |
| 2473 for_each_edge (pg, pg_collect_alias_ddrs, &cbdata); | |
| 2474 } | |
| 2475 | |
| 2476 /* For SCC that doesn't need to be broken, merge it. */ | |
| 2477 for (i = 0; i < num_sccs; ++i) | |
| 2478 { | |
| 2479 if (!bitmap_bit_p (sccs_to_merge, i)) | |
| 2480 continue; | |
| 2481 | |
| 2482 for (j = 0; partitions->iterate (j, &first); ++j) | |
| 2483 if (cbdata.vertices_component[j] == i) | |
| 2484 break; | |
| 2485 for (k = j + 1; partitions->iterate (k, &partition); ++k) | |
| 2486 { | |
| 2487 struct pg_vdata *data; | |
| 2488 | |
| 2489 if (cbdata.vertices_component[k] != i) | |
| 2490 continue; | |
| 2491 | |
| 2492 /* Update postorder number so that merged reduction partition is | |
| 2493 sorted after other partitions. */ | |
| 2494 if (!partition_reduction_p (first) | |
| 2495 && partition_reduction_p (partition)) | |
| 2496 { | |
| 2497 gcc_assert (pg->vertices[k].post < pg->vertices[j].post); | |
| 2498 pg->vertices[j].post = pg->vertices[k].post; | |
| 2499 } | |
| 2500 partition_merge_into (NULL, first, partition, FUSE_SAME_SCC); | |
| 2501 (*partitions)[k] = NULL; | |
| 2502 partition_free (partition); | |
| 2503 data = (struct pg_vdata *)pg->vertices[k].data; | |
| 2504 gcc_assert (data->id == k); | |
| 2505 data->partition = NULL; | |
| 2506 /* The result partition of merged SCC must be sequential. */ | |
| 2507 first->type = PTYPE_SEQUENTIAL; | |
| 2508 } | |
| 2509 } | |
| 2510 } | |
| 2511 | |
| 2512 sort_partitions_by_post_order (pg, partitions); | |
| 2513 free_partition_graph_vdata (pg); | |
| 2514 for_each_edge (pg, free_partition_graph_edata_cb, NULL); | |
| 2515 free_graph (pg); | |
| 0 | 2516 |
| 2517 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 111 | 2518 { |
| 2519 fprintf (dump_file, "Possible alias data dependence to break:\n"); | |
| 2520 dump_data_dependence_relations (dump_file, *alias_ddrs); | |
| 2521 } | |
| 2522 } | |
| 2523 | |
| 2524 /* Compute and return an expression whose value is the segment length which | |
| 2525 will be accessed by DR in NITERS iterations. */ | |
| 2526 | |
| 2527 static tree | |
| 2528 data_ref_segment_size (struct data_reference *dr, tree niters) | |
| 2529 { | |
| 131 | 2530 niters = size_binop (MINUS_EXPR, |
| 2531 fold_convert (sizetype, niters), | |
| 2532 size_one_node); | |
| 2533 return size_binop (MULT_EXPR, | |
| 2534 fold_convert (sizetype, DR_STEP (dr)), | |
| 2535 fold_convert (sizetype, niters)); | |
| 111 | 2536 } |
| 2537 | |
| 2538 /* Return true if LOOP's latch is dominated by statement for data reference | |
| 2539 DR. */ | |
| 2540 | |
| 2541 static inline bool | |
| 145 | 2542 latch_dominated_by_data_ref (class loop *loop, data_reference *dr) |
| 111 | 2543 { |
| 2544 return dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, | |
| 2545 gimple_bb (DR_STMT (dr))); | |
| 2546 } | |
| 2547 | |
| 2548 /* Compute alias check pairs and store them in COMP_ALIAS_PAIRS for LOOP's | |
| 2549 data dependence relations ALIAS_DDRS. */ | |
| 2550 | |
| 2551 static void | |
| 145 | 2552 compute_alias_check_pairs (class loop *loop, vec<ddr_p> *alias_ddrs, |
| 111 | 2553 vec<dr_with_seg_len_pair_t> *comp_alias_pairs) |
| 2554 { | |
| 2555 unsigned int i; | |
| 2556 unsigned HOST_WIDE_INT factor = 1; | |
| 2557 tree niters_plus_one, niters = number_of_latch_executions (loop); | |
| 2558 | |
| 2559 gcc_assert (niters != NULL_TREE && niters != chrec_dont_know); | |
| 2560 niters = fold_convert (sizetype, niters); | |
| 2561 niters_plus_one = size_binop (PLUS_EXPR, niters, size_one_node); | |
| 2562 | |
| 2563 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2564 fprintf (dump_file, "Creating alias check pairs:\n"); | |
| 2565 | |
| 2566 /* Iterate all data dependence relations and compute alias check pairs. */ | |
| 2567 for (i = 0; i < alias_ddrs->length (); i++) | |
| 2568 { | |
| 2569 ddr_p ddr = (*alias_ddrs)[i]; | |
| 2570 struct data_reference *dr_a = DDR_A (ddr); | |
| 2571 struct data_reference *dr_b = DDR_B (ddr); | |
| 2572 tree seg_length_a, seg_length_b; | |
| 2573 | |
| 2574 if (latch_dominated_by_data_ref (loop, dr_a)) | |
| 2575 seg_length_a = data_ref_segment_size (dr_a, niters_plus_one); | |
| 2576 else | |
| 2577 seg_length_a = data_ref_segment_size (dr_a, niters); | |
| 2578 | |
| 2579 if (latch_dominated_by_data_ref (loop, dr_b)) | |
| 2580 seg_length_b = data_ref_segment_size (dr_b, niters_plus_one); | |
| 2581 else | |
| 2582 seg_length_b = data_ref_segment_size (dr_b, niters); | |
| 2583 | |
| 131 | 2584 unsigned HOST_WIDE_INT access_size_a |
| 2585 = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_a)))); | |
| 2586 unsigned HOST_WIDE_INT access_size_b | |
| 2587 = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_b)))); | |
| 2588 unsigned int align_a = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_a))); | |
| 2589 unsigned int align_b = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_b))); | |
| 2590 | |
| 111 | 2591 dr_with_seg_len_pair_t dr_with_seg_len_pair |
| 131 | 2592 (dr_with_seg_len (dr_a, seg_length_a, access_size_a, align_a), |
| 145 | 2593 dr_with_seg_len (dr_b, seg_length_b, access_size_b, align_b), |
| 2594 /* ??? Would WELL_ORDERED be safe? */ | |
| 2595 dr_with_seg_len_pair_t::REORDERED); | |
| 111 | 2596 |
| 2597 comp_alias_pairs->safe_push (dr_with_seg_len_pair); | |
| 2598 } | |
| 2599 | |
| 2600 if (tree_fits_uhwi_p (niters)) | |
| 2601 factor = tree_to_uhwi (niters); | |
| 2602 | |
| 2603 /* Prune alias check pairs. */ | |
| 2604 prune_runtime_alias_test_list (comp_alias_pairs, factor); | |
| 2605 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2606 fprintf (dump_file, | |
| 2607 "Improved number of alias checks from %d to %d\n", | |
| 2608 alias_ddrs->length (), comp_alias_pairs->length ()); | |
| 2609 } | |
| 2610 | |
| 2611 /* Given data dependence relations in ALIAS_DDRS, generate runtime alias | |
| 2612 checks and version LOOP under condition of these runtime alias checks. */ | |
| 2613 | |
| 2614 static void | |
| 131 | 2615 version_loop_by_alias_check (vec<struct partition *> *partitions, |
| 145 | 2616 class loop *loop, vec<ddr_p> *alias_ddrs) |
| 111 | 2617 { |
| 2618 profile_probability prob; | |
| 2619 basic_block cond_bb; | |
| 145 | 2620 class loop *nloop; |
| 111 | 2621 tree lhs, arg0, cond_expr = NULL_TREE; |
| 2622 gimple_seq cond_stmts = NULL; | |
| 2623 gimple *call_stmt = NULL; | |
| 2624 auto_vec<dr_with_seg_len_pair_t> comp_alias_pairs; | |
| 2625 | |
| 2626 /* Generate code for runtime alias checks if necessary. */ | |
| 2627 gcc_assert (alias_ddrs->length () > 0); | |
| 2628 | |
| 2629 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2630 fprintf (dump_file, | |
| 2631 "Version loop <%d> with runtime alias check\n", loop->num); | |
| 2632 | |
| 2633 compute_alias_check_pairs (loop, alias_ddrs, &comp_alias_pairs); | |
| 2634 create_runtime_alias_checks (loop, &comp_alias_pairs, &cond_expr); | |
| 2635 cond_expr = force_gimple_operand_1 (cond_expr, &cond_stmts, | |
| 2636 is_gimple_val, NULL_TREE); | |
| 2637 | |
| 2638 /* Depend on vectorizer to fold IFN_LOOP_DIST_ALIAS. */ | |
| 131 | 2639 bool cancelable_p = flag_tree_loop_vectorize; |
| 2640 if (cancelable_p) | |
| 111 | 2641 { |
| 131 | 2642 unsigned i = 0; |
| 2643 struct partition *partition; | |
| 2644 for (; partitions->iterate (i, &partition); ++i) | |
| 2645 if (!partition_builtin_p (partition)) | |
| 2646 break; | |
| 2647 | |
| 2648 /* If all partitions are builtins, distributing it would be profitable and | |
| 2649 we don't want to cancel the runtime alias checks. */ | |
| 2650 if (i == partitions->length ()) | |
| 2651 cancelable_p = false; | |
| 2652 } | |
| 2653 | |
| 2654 /* Generate internal function call for loop distribution alias check if the | |
| 2655 runtime alias check should be cancelable. */ | |
| 2656 if (cancelable_p) | |
| 2657 { | |
| 111 | 2658 call_stmt = gimple_build_call_internal (IFN_LOOP_DIST_ALIAS, |
| 2659 2, NULL_TREE, cond_expr); | |
| 2660 lhs = make_ssa_name (boolean_type_node); | |
| 2661 gimple_call_set_lhs (call_stmt, lhs); | |
| 2662 } | |
| 2663 else | |
| 2664 lhs = cond_expr; | |
| 2665 | |
| 2666 prob = profile_probability::guessed_always ().apply_scale (9, 10); | |
| 2667 initialize_original_copy_tables (); | |
| 2668 nloop = loop_version (loop, lhs, &cond_bb, prob, prob.invert (), | |
| 2669 prob, prob.invert (), true); | |
| 2670 free_original_copy_tables (); | |
| 2671 /* Record the original loop number in newly generated loops. In case of | |
| 2672 distribution, the original loop will be distributed and the new loop | |
| 2673 is kept. */ | |
| 2674 loop->orig_loop_num = nloop->num; | |
| 2675 nloop->orig_loop_num = nloop->num; | |
| 2676 nloop->dont_vectorize = true; | |
| 2677 nloop->force_vectorize = false; | |
| 2678 | |
| 2679 if (call_stmt) | |
| 2680 { | |
| 2681 /* Record new loop's num in IFN_LOOP_DIST_ALIAS because the original | |
| 2682 loop could be destroyed. */ | |
| 2683 arg0 = build_int_cst (integer_type_node, loop->orig_loop_num); | |
| 2684 gimple_call_set_arg (call_stmt, 0, arg0); | |
| 2685 gimple_seq_add_stmt_without_update (&cond_stmts, call_stmt); | |
| 2686 } | |
| 2687 | |
| 2688 if (cond_stmts) | |
| 2689 { | |
| 2690 gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb); | |
| 2691 gsi_insert_seq_before (&cond_gsi, cond_stmts, GSI_SAME_STMT); | |
| 2692 } | |
| 2693 update_ssa (TODO_update_ssa); | |
| 0 | 2694 } |
| 2695 | |
| 111 | 2696 /* Return true if loop versioning is needed to distrubute PARTITIONS. |
| 2697 ALIAS_DDRS are data dependence relations for runtime alias check. */ | |
| 2698 | |
| 2699 static inline bool | |
| 2700 version_for_distribution_p (vec<struct partition *> *partitions, | |
| 2701 vec<ddr_p> *alias_ddrs) | |
| 2702 { | |
| 2703 /* No need to version loop if we have only one partition. */ | |
| 2704 if (partitions->length () == 1) | |
| 2705 return false; | |
| 2706 | |
| 2707 /* Need to version loop if runtime alias check is necessary. */ | |
| 2708 return (alias_ddrs->length () > 0); | |
| 2709 } | |
| 2710 | |
| 2711 /* Compare base offset of builtin mem* partitions P1 and P2. */ | |
| 2712 | |
| 131 | 2713 static int |
| 2714 offset_cmp (const void *vp1, const void *vp2) | |
| 111 | 2715 { |
| 131 | 2716 struct partition *p1 = *(struct partition *const *) vp1; |
| 2717 struct partition *p2 = *(struct partition *const *) vp2; | |
| 2718 unsigned HOST_WIDE_INT o1 = p1->builtin->dst_base_offset; | |
| 2719 unsigned HOST_WIDE_INT o2 = p2->builtin->dst_base_offset; | |
| 2720 return (o2 < o1) - (o1 < o2); | |
| 111 | 2721 } |
| 2722 | |
| 2723 /* Fuse adjacent memset builtin PARTITIONS if possible. This is a special | |
| 2724 case optimization transforming below code: | |
| 2725 | |
| 2726 __builtin_memset (&obj, 0, 100); | |
| 2727 _1 = &obj + 100; | |
| 2728 __builtin_memset (_1, 0, 200); | |
| 2729 _2 = &obj + 300; | |
| 2730 __builtin_memset (_2, 0, 100); | |
| 2731 | |
| 2732 into: | |
| 2733 | |
| 2734 __builtin_memset (&obj, 0, 400); | |
| 2735 | |
| 2736 Note we don't have dependence information between different partitions | |
| 2737 at this point, as a result, we can't handle nonadjacent memset builtin | |
| 2738 partitions since dependence might be broken. */ | |
| 2739 | |
| 2740 static void | |
| 2741 fuse_memset_builtins (vec<struct partition *> *partitions) | |
| 2742 { | |
| 2743 unsigned i, j; | |
| 2744 struct partition *part1, *part2; | |
| 131 | 2745 tree rhs1, rhs2; |
| 111 | 2746 |
| 2747 for (i = 0; partitions->iterate (i, &part1);) | |
| 2748 { | |
| 2749 if (part1->kind != PKIND_MEMSET) | |
| 2750 { | |
| 2751 i++; | |
| 2752 continue; | |
| 2753 } | |
| 2754 | |
| 2755 /* Find sub-array of memset builtins of the same base. Index range | |
| 2756 of the sub-array is [i, j) with "j > i". */ | |
| 2757 for (j = i + 1; partitions->iterate (j, &part2); ++j) | |
| 2758 { | |
| 2759 if (part2->kind != PKIND_MEMSET | |
| 2760 || !operand_equal_p (part1->builtin->dst_base_base, | |
| 2761 part2->builtin->dst_base_base, 0)) | |
| 2762 break; | |
| 131 | 2763 |
| 2764 /* Memset calls setting different values can't be merged. */ | |
| 2765 rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); | |
| 2766 rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
| 2767 if (!operand_equal_p (rhs1, rhs2, 0)) | |
| 2768 break; | |
| 111 | 2769 } |
| 2770 | |
| 2771 /* Stable sort is required in order to avoid breaking dependence. */ | |
| 131 | 2772 gcc_stablesort (&(*partitions)[i], j - i, sizeof (*partitions)[i], |
| 2773 offset_cmp); | |
| 111 | 2774 /* Continue with next partition. */ |
| 2775 i = j; | |
| 2776 } | |
| 2777 | |
| 2778 /* Merge all consecutive memset builtin partitions. */ | |
| 2779 for (i = 0; i < partitions->length () - 1;) | |
| 2780 { | |
| 2781 part1 = (*partitions)[i]; | |
| 2782 if (part1->kind != PKIND_MEMSET) | |
| 2783 { | |
| 2784 i++; | |
| 2785 continue; | |
| 2786 } | |
| 2787 | |
| 2788 part2 = (*partitions)[i + 1]; | |
| 2789 /* Only merge memset partitions of the same base and with constant | |
| 2790 access sizes. */ | |
| 2791 if (part2->kind != PKIND_MEMSET | |
| 2792 || TREE_CODE (part1->builtin->size) != INTEGER_CST | |
| 2793 || TREE_CODE (part2->builtin->size) != INTEGER_CST | |
| 2794 || !operand_equal_p (part1->builtin->dst_base_base, | |
| 2795 part2->builtin->dst_base_base, 0)) | |
| 2796 { | |
| 2797 i++; | |
| 2798 continue; | |
| 2799 } | |
| 131 | 2800 rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); |
| 2801 rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
| 111 | 2802 int bytev1 = const_with_all_bytes_same (rhs1); |
| 2803 int bytev2 = const_with_all_bytes_same (rhs2); | |
| 2804 /* Only merge memset partitions of the same value. */ | |
| 2805 if (bytev1 != bytev2 || bytev1 == -1) | |
| 2806 { | |
| 2807 i++; | |
| 2808 continue; | |
| 2809 } | |
| 2810 wide_int end1 = wi::add (part1->builtin->dst_base_offset, | |
| 2811 wi::to_wide (part1->builtin->size)); | |
| 2812 /* Only merge adjacent memset partitions. */ | |
| 2813 if (wi::ne_p (end1, part2->builtin->dst_base_offset)) | |
| 2814 { | |
| 2815 i++; | |
| 2816 continue; | |
| 2817 } | |
| 2818 /* Merge partitions[i] and partitions[i+1]. */ | |
| 2819 part1->builtin->size = fold_build2 (PLUS_EXPR, sizetype, | |
| 2820 part1->builtin->size, | |
| 2821 part2->builtin->size); | |
| 2822 partition_free (part2); | |
| 2823 partitions->ordered_remove (i + 1); | |
| 2824 } | |
| 2825 } | |
| 2826 | |
| 145 | 2827 void |
| 2828 loop_distribution::finalize_partitions (class loop *loop, | |
| 2829 vec<struct partition *> *partitions, | |
| 2830 vec<ddr_p> *alias_ddrs) | |
| 111 | 2831 { |
| 2832 unsigned i; | |
| 2833 struct partition *partition, *a; | |
| 2834 | |
| 2835 if (partitions->length () == 1 | |
| 2836 || alias_ddrs->length () > 0) | |
| 2837 return; | |
| 2838 | |
| 131 | 2839 unsigned num_builtin = 0, num_normal = 0, num_partial_memset = 0; |
| 111 | 2840 bool same_type_p = true; |
| 2841 enum partition_type type = ((*partitions)[0])->type; | |
| 2842 for (i = 0; partitions->iterate (i, &partition); ++i) | |
| 2843 { | |
| 2844 same_type_p &= (type == partition->type); | |
| 131 | 2845 if (partition_builtin_p (partition)) |
| 2846 { | |
| 2847 num_builtin++; | |
| 2848 continue; | |
| 2849 } | |
| 2850 num_normal++; | |
| 2851 if (partition->kind == PKIND_PARTIAL_MEMSET) | |
| 2852 num_partial_memset++; | |
| 111 | 2853 } |
| 2854 | |
| 2855 /* Don't distribute current loop into too many loops given we don't have | |
| 2856 memory stream cost model. Be even more conservative in case of loop | |
| 2857 nest distribution. */ | |
| 131 | 2858 if ((same_type_p && num_builtin == 0 |
| 2859 && (loop->inner == NULL || num_normal != 2 || num_partial_memset != 1)) | |
| 111 | 2860 || (loop->inner != NULL |
| 2861 && i >= NUM_PARTITION_THRESHOLD && num_normal > 1) | |
| 2862 || (loop->inner == NULL | |
| 2863 && i >= NUM_PARTITION_THRESHOLD && num_normal > num_builtin)) | |
| 2864 { | |
| 2865 a = (*partitions)[0]; | |
| 2866 for (i = 1; partitions->iterate (i, &partition); ++i) | |
| 2867 { | |
| 2868 partition_merge_into (NULL, a, partition, FUSE_FINALIZE); | |
| 2869 partition_free (partition); | |
| 2870 } | |
| 2871 partitions->truncate (1); | |
| 2872 } | |
| 2873 | |
| 2874 /* Fuse memset builtins if possible. */ | |
| 2875 if (partitions->length () > 1) | |
| 2876 fuse_memset_builtins (partitions); | |
| 2877 } | |
| 2878 | |
| 2879 /* Distributes the code from LOOP in such a way that producer statements | |
| 2880 are placed before consumer statements. Tries to separate only the | |
| 2881 statements from STMTS into separate loops. Returns the number of | |
| 2882 distributed loops. Set NB_CALLS to number of generated builtin calls. | |
| 2883 Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
| 0 | 2884 |
| 145 | 2885 int |
| 2886 loop_distribution::distribute_loop (class loop *loop, vec<gimple *> stmts, | |
| 2887 control_dependences *cd, int *nb_calls, bool *destroy_p, | |
| 2888 bool only_patterns_p) | |
| 0 | 2889 { |
| 111 | 2890 ddrs_table = new hash_table<ddr_hasher> (389); |
| 0 | 2891 struct graph *rdg; |
| 111 | 2892 partition *partition; |
| 2893 int i, nbp; | |
| 2894 | |
| 2895 *destroy_p = false; | |
| 2896 *nb_calls = 0; | |
| 2897 loop_nest.create (0); | |
| 2898 if (!find_loop_nest (loop, &loop_nest)) | |
| 0 | 2899 { |
| 111 | 2900 loop_nest.release (); |
| 2901 delete ddrs_table; | |
| 2902 return 0; | |
| 0 | 2903 } |
| 2904 | |
| 111 | 2905 datarefs_vec.create (20); |
| 145 | 2906 has_nonaddressable_dataref_p = false; |
| 111 | 2907 rdg = build_rdg (loop, cd); |
| 0 | 2908 if (!rdg) |
| 2909 { | |
| 2910 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2911 fprintf (dump_file, | |
| 111 | 2912 "Loop %d not distributed: failed to build the RDG.\n", |
| 0 | 2913 loop->num); |
| 2914 | |
| 111 | 2915 loop_nest.release (); |
| 2916 free_data_refs (datarefs_vec); | |
| 2917 delete ddrs_table; | |
| 2918 return 0; | |
| 0 | 2919 } |
| 2920 | |
| 111 | 2921 if (datarefs_vec.length () > MAX_DATAREFS_NUM) |
| 2922 { | |
| 2923 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2924 fprintf (dump_file, | |
| 2925 "Loop %d not distributed: too many memory references.\n", | |
| 2926 loop->num); | |
| 2927 | |
| 2928 free_rdg (rdg); | |
| 2929 loop_nest.release (); | |
| 2930 free_data_refs (datarefs_vec); | |
| 2931 delete ddrs_table; | |
| 2932 return 0; | |
| 2933 } | |
| 2934 | |
| 2935 data_reference_p dref; | |
| 2936 for (i = 0; datarefs_vec.iterate (i, &dref); ++i) | |
| 2937 dref->aux = (void *) (uintptr_t) i; | |
| 0 | 2938 |
| 2939 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2940 dump_rdg (dump_file, rdg); | |
| 2941 | |
| 111 | 2942 auto_vec<struct partition *, 3> partitions; |
| 2943 rdg_build_partitions (rdg, stmts, &partitions); | |
| 2944 | |
| 2945 auto_vec<ddr_p> alias_ddrs; | |
| 2946 | |
| 2947 auto_bitmap stmt_in_all_partitions; | |
| 2948 bitmap_copy (stmt_in_all_partitions, partitions[0]->stmts); | |
| 2949 for (i = 1; partitions.iterate (i, &partition); ++i) | |
| 2950 bitmap_and_into (stmt_in_all_partitions, partitions[i]->stmts); | |
| 2951 | |
| 145 | 2952 bool any_builtin = false; |
| 2953 bool reduction_in_all = false; | |
| 111 | 2954 FOR_EACH_VEC_ELT (partitions, i, partition) |
| 2955 { | |
| 145 | 2956 reduction_in_all |
| 2957 |= classify_partition (loop, rdg, partition, stmt_in_all_partitions); | |
| 111 | 2958 any_builtin |= partition_builtin_p (partition); |
| 2959 } | |
| 2960 | |
| 2961 /* If we are only distributing patterns but did not detect any, | |
| 2962 simply bail out. */ | |
| 145 | 2963 if (only_patterns_p |
| 111 | 2964 && !any_builtin) |
| 2965 { | |
| 2966 nbp = 0; | |
| 2967 goto ldist_done; | |
| 2968 } | |
| 2969 | |
| 2970 /* If we are only distributing patterns fuse all partitions that | |
| 2971 were not classified as builtins. This also avoids chopping | |
| 2972 a loop into pieces, separated by builtin calls. That is, we | |
| 2973 only want no or a single loop body remaining. */ | |
| 2974 struct partition *into; | |
| 145 | 2975 if (only_patterns_p) |
| 0 | 2976 { |
| 111 | 2977 for (i = 0; partitions.iterate (i, &into); ++i) |
| 2978 if (!partition_builtin_p (into)) | |
| 2979 break; | |
| 2980 for (++i; partitions.iterate (i, &partition); ++i) | |
| 2981 if (!partition_builtin_p (partition)) | |
| 2982 { | |
| 2983 partition_merge_into (NULL, into, partition, FUSE_NON_BUILTIN); | |
| 2984 partitions.unordered_remove (i); | |
| 2985 partition_free (partition); | |
| 2986 i--; | |
| 2987 } | |
| 2988 } | |
| 2989 | |
| 2990 /* Due to limitations in the transform phase we have to fuse all | |
| 2991 reduction partitions into the last partition so the existing | |
| 2992 loop will contain all loop-closed PHI nodes. */ | |
| 2993 for (i = 0; partitions.iterate (i, &into); ++i) | |
| 2994 if (partition_reduction_p (into)) | |
| 2995 break; | |
| 2996 for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
| 2997 if (partition_reduction_p (partition)) | |
| 2998 { | |
| 2999 partition_merge_into (rdg, into, partition, FUSE_REDUCTION); | |
| 3000 partitions.unordered_remove (i); | |
| 3001 partition_free (partition); | |
| 3002 i--; | |
| 3003 } | |
| 3004 | |
| 3005 /* Apply our simple cost model - fuse partitions with similar | |
| 3006 memory accesses. */ | |
| 3007 for (i = 0; partitions.iterate (i, &into); ++i) | |
| 3008 { | |
| 3009 bool changed = false; | |
| 131 | 3010 if (partition_builtin_p (into) || into->kind == PKIND_PARTIAL_MEMSET) |
| 111 | 3011 continue; |
| 3012 for (int j = i + 1; | |
| 3013 partitions.iterate (j, &partition); ++j) | |
| 0 | 3014 { |
| 111 | 3015 if (share_memory_accesses (rdg, into, partition)) |
| 3016 { | |
| 3017 partition_merge_into (rdg, into, partition, FUSE_SHARE_REF); | |
| 3018 partitions.unordered_remove (j); | |
| 3019 partition_free (partition); | |
| 3020 j--; | |
| 3021 changed = true; | |
| 3022 } | |
| 3023 } | |
| 3024 /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar | |
| 3025 accesses when 1 and 2 have similar accesses but not 0 and 1 | |
| 3026 then in the next iteration we will fail to consider merging | |
| 3027 1 into 0,2. So try again if we did any merging into 0. */ | |
| 3028 if (changed) | |
| 3029 i--; | |
| 3030 } | |
| 3031 | |
| 145 | 3032 /* Put a non-builtin partition last if we need to preserve a reduction. |
| 3033 ??? This is a workaround that makes sort_partitions_by_post_order do | |
| 3034 the correct thing while in reality it should sort each component | |
| 3035 separately and then put the component with a reduction or a non-builtin | |
| 3036 last. */ | |
| 3037 if (reduction_in_all | |
| 3038 && partition_builtin_p (partitions.last())) | |
| 3039 FOR_EACH_VEC_ELT (partitions, i, partition) | |
| 3040 if (!partition_builtin_p (partition)) | |
| 3041 { | |
| 3042 partitions.unordered_remove (i); | |
| 3043 partitions.quick_push (partition); | |
| 3044 break; | |
| 3045 } | |
| 3046 | |
| 111 | 3047 /* Build the partition dependency graph and fuse partitions in strong |
| 3048 connected component. */ | |
| 3049 if (partitions.length () > 1) | |
| 3050 { | |
| 3051 /* Don't support loop nest distribution under runtime alias check | |
| 145 | 3052 since it's not likely to enable many vectorization opportunities. |
| 3053 Also if loop has any data reference which may be not addressable | |
| 3054 since alias check needs to take, compare address of the object. */ | |
| 3055 if (loop->inner || has_nonaddressable_dataref_p) | |
| 111 | 3056 merge_dep_scc_partitions (rdg, &partitions, false); |
| 3057 else | |
| 3058 { | |
| 3059 merge_dep_scc_partitions (rdg, &partitions, true); | |
| 3060 if (partitions.length () > 1) | |
| 3061 break_alias_scc_partitions (rdg, &partitions, &alias_ddrs); | |
| 0 | 3062 } |
| 3063 } | |
| 3064 | |
| 111 | 3065 finalize_partitions (loop, &partitions, &alias_ddrs); |
| 3066 | |
| 145 | 3067 /* If there is a reduction in all partitions make sure the last one |
| 3068 is not classified for builtin code generation. */ | |
| 3069 if (reduction_in_all) | |
| 3070 { | |
| 3071 partition = partitions.last (); | |
| 3072 if (only_patterns_p | |
| 3073 && partition_builtin_p (partition) | |
| 3074 && !partition_builtin_p (partitions[0])) | |
| 3075 { | |
| 3076 nbp = 0; | |
| 3077 goto ldist_done; | |
| 3078 } | |
| 3079 partition->kind = PKIND_NORMAL; | |
| 3080 } | |
| 3081 | |
| 111 | 3082 nbp = partitions.length (); |
| 3083 if (nbp == 0 | |
| 3084 || (nbp == 1 && !partition_builtin_p (partitions[0])) | |
| 3085 || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
| 3086 { | |
| 3087 nbp = 0; | |
| 3088 goto ldist_done; | |
| 3089 } | |
| 3090 | |
| 3091 if (version_for_distribution_p (&partitions, &alias_ddrs)) | |
| 131 | 3092 version_loop_by_alias_check (&partitions, loop, &alias_ddrs); |
| 111 | 3093 |
| 3094 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 3095 { | |
| 3096 fprintf (dump_file, | |
| 3097 "distribute loop <%d> into partitions:\n", loop->num); | |
| 3098 dump_rdg_partitions (dump_file, partitions); | |
| 3099 } | |
| 3100 | |
| 3101 FOR_EACH_VEC_ELT (partitions, i, partition) | |
| 3102 { | |
| 3103 if (partition_builtin_p (partition)) | |
| 3104 (*nb_calls)++; | |
| 3105 *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); | |
| 3106 } | |
| 3107 | |
| 3108 ldist_done: | |
| 3109 loop_nest.release (); | |
| 3110 free_data_refs (datarefs_vec); | |
| 3111 for (hash_table<ddr_hasher>::iterator iter = ddrs_table->begin (); | |
| 3112 iter != ddrs_table->end (); ++iter) | |
| 3113 { | |
| 3114 free_dependence_relation (*iter); | |
| 3115 *iter = NULL; | |
| 3116 } | |
| 3117 delete ddrs_table; | |
| 3118 | |
| 3119 FOR_EACH_VEC_ELT (partitions, i, partition) | |
| 3120 partition_free (partition); | |
| 3121 | |
| 0 | 3122 free_rdg (rdg); |
| 111 | 3123 return nbp - *nb_calls; |
| 0 | 3124 } |
| 3125 | |
| 145 | 3126 |
| 3127 void loop_distribution::bb_top_order_init (void) | |
| 3128 { | |
| 3129 int rpo_num; | |
| 3130 int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
| 3131 | |
| 3132 bb_top_order_index = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
| 3133 bb_top_order_index_size = last_basic_block_for_fn (cfun); | |
| 3134 rpo_num = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, true); | |
| 3135 for (int i = 0; i < rpo_num; i++) | |
| 3136 bb_top_order_index[rpo[i]] = i; | |
| 3137 | |
| 3138 free (rpo); | |
| 3139 } | |
| 3140 | |
| 3141 void loop_distribution::bb_top_order_destroy () | |
| 3142 { | |
| 3143 free (bb_top_order_index); | |
| 3144 bb_top_order_index = NULL; | |
| 3145 bb_top_order_index_size = 0; | |
| 3146 } | |
| 3147 | |
| 3148 | |
| 3149 /* Given LOOP, this function records seed statements for distribution in | |
| 3150 WORK_LIST. Return false if there is nothing for distribution. */ | |
| 3151 | |
| 3152 static bool | |
| 3153 find_seed_stmts_for_distribution (class loop *loop, vec<gimple *> *work_list) | |
| 3154 { | |
| 3155 basic_block *bbs = get_loop_body_in_dom_order (loop); | |
| 3156 | |
| 3157 /* Initialize the worklist with stmts we seed the partitions with. */ | |
| 3158 for (unsigned i = 0; i < loop->num_nodes; ++i) | |
| 3159 { | |
| 3160 for (gphi_iterator gsi = gsi_start_phis (bbs[i]); | |
| 3161 !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 3162 { | |
| 3163 gphi *phi = gsi.phi (); | |
| 3164 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 3165 continue; | |
| 3166 /* Distribute stmts which have defs that are used outside of | |
| 3167 the loop. */ | |
| 3168 if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) | |
| 3169 continue; | |
| 3170 work_list->safe_push (phi); | |
| 3171 } | |
| 3172 for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); | |
| 3173 !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 3174 { | |
| 3175 gimple *stmt = gsi_stmt (gsi); | |
| 3176 | |
| 3177 /* Ignore clobbers, they do not have true side effects. */ | |
| 3178 if (gimple_clobber_p (stmt)) | |
| 3179 continue; | |
| 3180 | |
| 3181 /* If there is a stmt with side-effects bail out - we | |
| 3182 cannot and should not distribute this loop. */ | |
| 3183 if (gimple_has_side_effects (stmt)) | |
| 3184 { | |
| 3185 free (bbs); | |
| 3186 return false; | |
| 3187 } | |
| 3188 | |
| 3189 /* Distribute stmts which have defs that are used outside of | |
| 3190 the loop. */ | |
| 3191 if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
| 3192 ; | |
| 3193 /* Otherwise only distribute stores for now. */ | |
| 3194 else if (!gimple_vdef (stmt)) | |
| 3195 continue; | |
| 3196 | |
| 3197 work_list->safe_push (stmt); | |
| 3198 } | |
| 3199 } | |
| 3200 free (bbs); | |
| 3201 return work_list->length () > 0; | |
| 3202 } | |
| 3203 | |
| 3204 /* Given innermost LOOP, return the outermost enclosing loop that forms a | |
| 3205 perfect loop nest. */ | |
| 3206 | |
| 3207 static class loop * | |
| 3208 prepare_perfect_loop_nest (class loop *loop) | |
| 3209 { | |
| 3210 class loop *outer = loop_outer (loop); | |
| 3211 tree niters = number_of_latch_executions (loop); | |
| 3212 | |
| 3213 /* TODO: We only support the innermost 3-level loop nest distribution | |
| 3214 because of compilation time issue for now. This should be relaxed | |
| 3215 in the future. Note we only allow 3-level loop nest distribution | |
| 3216 when parallelizing loops. */ | |
| 3217 while ((loop->inner == NULL | |
| 3218 || (loop->inner->inner == NULL && flag_tree_parallelize_loops > 1)) | |
| 3219 && loop_outer (outer) | |
| 3220 && outer->inner == loop && loop->next == NULL | |
| 3221 && single_exit (outer) | |
| 3222 && !chrec_contains_symbols_defined_in_loop (niters, outer->num) | |
| 3223 && (niters = number_of_latch_executions (outer)) != NULL_TREE | |
| 3224 && niters != chrec_dont_know) | |
| 3225 { | |
| 3226 loop = outer; | |
| 3227 outer = loop_outer (loop); | |
| 3228 } | |
| 3229 | |
| 3230 return loop; | |
| 3231 } | |
| 3232 | |
| 3233 | |
| 3234 unsigned int | |
| 3235 loop_distribution::execute (function *fun) | |
| 3236 { | |
| 3237 class loop *loop; | |
| 3238 bool changed = false; | |
| 3239 basic_block bb; | |
| 3240 control_dependences *cd = NULL; | |
| 3241 auto_vec<loop_p> loops_to_be_destroyed; | |
| 3242 | |
| 3243 if (number_of_loops (fun) <= 1) | |
| 3244 return 0; | |
| 3245 | |
| 3246 bb_top_order_init (); | |
| 3247 | |
| 3248 FOR_ALL_BB_FN (bb, fun) | |
| 3249 { | |
| 3250 gimple_stmt_iterator gsi; | |
| 3251 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 3252 gimple_set_uid (gsi_stmt (gsi), -1); | |
| 3253 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 3254 gimple_set_uid (gsi_stmt (gsi), -1); | |
| 3255 } | |
| 3256 | |
| 3257 /* We can at the moment only distribute non-nested loops, thus restrict | |
| 3258 walking to innermost loops. */ | |
| 3259 FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) | |
| 3260 { | |
| 3261 /* Don't distribute multiple exit edges loop, or cold loop when | |
| 3262 not doing pattern detection. */ | |
| 3263 if (!single_exit (loop) | |
| 3264 || (!flag_tree_loop_distribute_patterns | |
| 3265 && !optimize_loop_for_speed_p (loop))) | |
| 3266 continue; | |
| 3267 | |
| 3268 /* Don't distribute loop if niters is unknown. */ | |
| 3269 tree niters = number_of_latch_executions (loop); | |
| 3270 if (niters == NULL_TREE || niters == chrec_dont_know) | |
| 3271 continue; | |
| 3272 | |
| 3273 /* Get the perfect loop nest for distribution. */ | |
| 3274 loop = prepare_perfect_loop_nest (loop); | |
| 3275 for (; loop; loop = loop->inner) | |
| 3276 { | |
| 3277 auto_vec<gimple *> work_list; | |
| 3278 if (!find_seed_stmts_for_distribution (loop, &work_list)) | |
| 3279 break; | |
| 3280 | |
| 3281 const char *str = loop->inner ? " nest" : ""; | |
| 3282 dump_user_location_t loc = find_loop_location (loop); | |
| 3283 if (!cd) | |
| 3284 { | |
| 3285 calculate_dominance_info (CDI_DOMINATORS); | |
| 3286 calculate_dominance_info (CDI_POST_DOMINATORS); | |
| 3287 cd = new control_dependences (); | |
| 3288 free_dominance_info (CDI_POST_DOMINATORS); | |
| 3289 } | |
| 3290 | |
| 3291 bool destroy_p; | |
| 3292 int nb_generated_loops, nb_generated_calls; | |
| 3293 nb_generated_loops | |
| 3294 = distribute_loop (loop, work_list, cd, &nb_generated_calls, | |
| 3295 &destroy_p, (!optimize_loop_for_speed_p (loop) | |
| 3296 || !flag_tree_loop_distribution)); | |
| 3297 if (destroy_p) | |
| 3298 loops_to_be_destroyed.safe_push (loop); | |
| 3299 | |
| 3300 if (nb_generated_loops + nb_generated_calls > 0) | |
| 3301 { | |
| 3302 changed = true; | |
| 3303 if (dump_enabled_p ()) | |
| 3304 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
| 3305 loc, "Loop%s %d distributed: split to %d loops " | |
| 3306 "and %d library calls.\n", str, loop->num, | |
| 3307 nb_generated_loops, nb_generated_calls); | |
| 3308 | |
| 3309 break; | |
| 3310 } | |
| 3311 | |
| 3312 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 3313 fprintf (dump_file, "Loop%s %d not distributed.\n", str, loop->num); | |
| 3314 } | |
| 3315 } | |
| 3316 | |
| 3317 if (cd) | |
| 3318 delete cd; | |
| 3319 | |
| 3320 if (bb_top_order_index != NULL) | |
| 3321 bb_top_order_destroy (); | |
| 3322 | |
| 3323 if (changed) | |
| 3324 { | |
| 3325 /* Destroy loop bodies that could not be reused. Do this late as we | |
| 3326 otherwise can end up refering to stale data in control dependences. */ | |
| 3327 unsigned i; | |
| 3328 FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
| 3329 destroy_loop (loop); | |
| 3330 | |
| 3331 /* Cached scalar evolutions now may refer to wrong or non-existing | |
| 3332 loops. */ | |
| 3333 scev_reset_htab (); | |
| 3334 mark_virtual_operands_for_renaming (fun); | |
| 3335 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); | |
| 3336 } | |
| 3337 | |
| 3338 checking_verify_loop_structure (); | |
| 3339 | |
| 3340 return changed ? TODO_cleanup_cfg : 0; | |
| 3341 } | |
| 3342 | |
| 3343 | |
| 0 | 3344 /* Distribute all loops in the current function. */ |
| 3345 | |
| 111 | 3346 namespace { |
| 3347 | |
| 3348 const pass_data pass_data_loop_distribution = | |
| 3349 { | |
| 3350 GIMPLE_PASS, /* type */ | |
| 3351 "ldist", /* name */ | |
| 3352 OPTGROUP_LOOP, /* optinfo_flags */ | |
| 3353 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ | |
| 3354 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 3355 0, /* properties_provided */ | |
| 3356 0, /* properties_destroyed */ | |
| 3357 0, /* todo_flags_start */ | |
| 3358 0, /* todo_flags_finish */ | |
| 3359 }; | |
| 3360 | |
| 3361 class pass_loop_distribution : public gimple_opt_pass | |
| 3362 { | |
| 3363 public: | |
| 3364 pass_loop_distribution (gcc::context *ctxt) | |
| 3365 : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
| 3366 {} | |
| 3367 | |
| 3368 /* opt_pass methods: */ | |
| 3369 virtual bool gate (function *) | |
| 3370 { | |
| 3371 return flag_tree_loop_distribution | |
| 3372 || flag_tree_loop_distribute_patterns; | |
| 3373 } | |
| 3374 | |
| 3375 virtual unsigned int execute (function *); | |
| 3376 | |
| 3377 }; // class pass_loop_distribution | |
| 3378 | |
| 3379 unsigned int | |
| 3380 pass_loop_distribution::execute (function *fun) | |
| 0 | 3381 { |
| 145 | 3382 return loop_distribution ().execute (fun); |
| 0 | 3383 } |
| 3384 | |
| 111 | 3385 } // anon namespace |
| 3386 | |
| 3387 gimple_opt_pass * | |
| 3388 make_pass_loop_distribution (gcc::context *ctxt) | |
| 0 | 3389 { |
| 111 | 3390 return new pass_loop_distribution (ctxt); |
| 3391 } |