Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-loop-ch.c @ 127:4c56639505ff
fix function.c and add CbC-example Makefile
| author | mir3636 |
|---|---|
| date | Wed, 11 Apr 2018 18:46:58 +0900 |
| parents | 04ced10e8804 |
| children | 84e7813d76e9 |
| rev | line source |
|---|---|
| 0 | 1 /* Loop header copying on trees. |
| 111 | 2 Copyright (C) 2004-2017 Free Software Foundation, Inc. |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3 |
| 0 | 4 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
|
5 |
| 0 | 6 GCC is free software; you can redistribute it and/or modify it |
| 7 under the terms of the GNU General Public License as published by the | |
| 8 Free Software Foundation; either version 3, or (at your option) any | |
| 9 later version. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
10 |
| 0 | 11 GCC is distributed in the hope that it will be useful, but WITHOUT |
| 12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 14 for more details. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
15 |
| 0 | 16 You should have received a copy of the GNU General Public License |
| 17 along with GCC; see the file COPYING3. If not see | |
| 18 <http://www.gnu.org/licenses/>. */ | |
| 19 | |
| 20 #include "config.h" | |
| 21 #include "system.h" | |
| 22 #include "coretypes.h" | |
| 111 | 23 #include "backend.h" |
| 0 | 24 #include "tree.h" |
| 111 | 25 #include "gimple.h" |
| 26 #include "cfghooks.h" | |
| 0 | 27 #include "tree-pass.h" |
| 111 | 28 #include "gimple-ssa.h" |
| 29 #include "gimple-iterator.h" | |
| 30 #include "tree-cfg.h" | |
| 31 #include "tree-into-ssa.h" | |
| 0 | 32 #include "cfgloop.h" |
| 33 #include "tree-inline.h" | |
| 111 | 34 #include "tree-ssa-scopedtables.h" |
| 35 #include "tree-ssa-threadedge.h" | |
| 36 #include "params.h" | |
| 0 | 37 |
| 38 /* Duplicates headers of loops if they are small enough, so that the statements | |
| 39 in the loop body are always executed when the loop is entered. This | |
| 40 increases effectiveness of code motion optimizations, and reduces the need | |
| 41 for loop preconditioning. */ | |
| 42 | |
| 43 /* Check whether we should duplicate HEADER of LOOP. At most *LIMIT | |
| 44 instructions should be duplicated, limit is decreased by the actual | |
| 45 amount. */ | |
| 46 | |
| 47 static bool | |
| 48 should_duplicate_loop_header_p (basic_block header, struct loop *loop, | |
| 49 int *limit) | |
| 50 { | |
| 51 gimple_stmt_iterator bsi; | |
| 111 | 52 gimple *last; |
| 0 | 53 |
| 111 | 54 gcc_assert (!header->aux); |
| 0 | 55 |
| 56 /* Loop header copying usually increases size of the code. This used not to | |
| 57 be true, since quite often it is possible to verify that the condition is | |
| 58 satisfied in the first iteration and therefore to eliminate it. Jump | |
| 59 threading handles these cases now. */ | |
| 60 if (optimize_loop_for_size_p (loop)) | |
| 111 | 61 { |
| 62 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 63 fprintf (dump_file, | |
| 64 " Not duplicating bb %i: optimizing for size.\n", | |
| 65 header->index); | |
| 66 return false; | |
| 67 } | |
| 0 | 68 |
| 69 gcc_assert (EDGE_COUNT (header->succs) > 0); | |
| 70 if (single_succ_p (header)) | |
| 111 | 71 { |
| 72 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 73 fprintf (dump_file, | |
| 74 " Not duplicating bb %i: it is single succ.\n", | |
| 75 header->index); | |
| 76 return false; | |
| 77 } | |
| 78 | |
| 0 | 79 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest) |
| 80 && flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 1)->dest)) | |
| 111 | 81 { |
| 82 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 83 fprintf (dump_file, | |
| 84 " Not duplicating bb %i: both sucessors are in loop.\n", | |
| 85 loop->num); | |
| 86 return false; | |
| 87 } | |
| 0 | 88 |
| 89 /* If this is not the original loop header, we want it to have just | |
| 90 one predecessor in order to match the && pattern. */ | |
| 91 if (header != loop->header && !single_pred_p (header)) | |
| 111 | 92 { |
| 93 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 94 fprintf (dump_file, | |
| 95 " Not duplicating bb %i: it has mutiple predecestors.\n", | |
| 96 header->index); | |
| 97 return false; | |
| 98 } | |
| 0 | 99 |
| 100 last = last_stmt (header); | |
| 101 if (gimple_code (last) != GIMPLE_COND) | |
| 111 | 102 { |
| 103 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 104 fprintf (dump_file, | |
| 105 " Not duplicating bb %i: it does not end by conditional.\n", | |
| 106 header->index); | |
| 107 return false; | |
| 108 } | |
| 0 | 109 |
| 111 | 110 /* Count number of instructions and punt on calls. */ |
| 0 | 111 for (bsi = gsi_start_bb (header); !gsi_end_p (bsi); gsi_next (&bsi)) |
| 112 { | |
| 113 last = gsi_stmt (bsi); | |
| 114 | |
| 115 if (gimple_code (last) == GIMPLE_LABEL) | |
| 116 continue; | |
| 117 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
118 if (is_gimple_debug (last)) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
119 continue; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
120 |
| 111 | 121 if (gimple_code (last) == GIMPLE_CALL |
| 122 && (!gimple_inexpensive_call_p (as_a <gcall *> (last)) | |
| 123 /* IFN_LOOP_DIST_ALIAS means that inner loop is distributed | |
| 124 at current loop's header. Don't copy in this case. */ | |
| 125 || gimple_call_internal_p (last, IFN_LOOP_DIST_ALIAS))) | |
| 126 { | |
| 127 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 128 fprintf (dump_file, | |
| 129 " Not duplicating bb %i: it contains call.\n", | |
| 130 header->index); | |
| 131 return false; | |
| 132 } | |
| 0 | 133 |
| 134 *limit -= estimate_num_insns (last, &eni_size_weights); | |
| 135 if (*limit < 0) | |
| 111 | 136 { |
| 137 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 138 fprintf (dump_file, | |
| 139 " Not duplicating bb %i contains too many insns.\n", | |
| 140 header->index); | |
| 141 return false; | |
| 142 } | |
| 0 | 143 } |
| 111 | 144 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 145 fprintf (dump_file, " Will duplicate bb %i\n", header->index); | |
| 0 | 146 return true; |
| 147 } | |
| 148 | |
| 149 /* Checks whether LOOP is a do-while style loop. */ | |
| 150 | |
| 151 static bool | |
| 152 do_while_loop_p (struct loop *loop) | |
| 153 { | |
| 111 | 154 gimple *stmt = last_stmt (loop->latch); |
| 0 | 155 |
| 156 /* If the latch of the loop is not empty, it is not a do-while loop. */ | |
| 157 if (stmt | |
| 158 && gimple_code (stmt) != GIMPLE_LABEL) | |
| 111 | 159 { |
| 160 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 161 fprintf (dump_file, | |
| 162 "Loop %i is not do-while loop: latch is not empty.\n", | |
| 163 loop->num); | |
| 164 return false; | |
| 165 } | |
| 0 | 166 |
| 167 /* If the header contains just a condition, it is not a do-while loop. */ | |
| 168 stmt = last_and_only_stmt (loop->header); | |
| 169 if (stmt | |
| 170 && gimple_code (stmt) == GIMPLE_COND) | |
| 111 | 171 { |
| 172 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 173 fprintf (dump_file, | |
| 174 "Loop %i is not do-while loop: " | |
| 175 "header contains just condition.\n", loop->num); | |
| 176 return false; | |
| 177 } | |
| 178 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 179 fprintf (dump_file, "Loop %i is do-while loop\n", loop->num); | |
| 0 | 180 |
| 181 return true; | |
| 182 } | |
| 183 | |
| 111 | 184 namespace { |
| 185 | |
| 186 /* Common superclass for both header-copying phases. */ | |
| 187 class ch_base : public gimple_opt_pass | |
| 188 { | |
| 189 protected: | |
| 190 ch_base (pass_data data, gcc::context *ctxt) | |
| 191 : gimple_opt_pass (data, ctxt) | |
| 192 {} | |
| 193 | |
| 194 /* Copies headers of all loops in FUN for which process_loop_p is true. */ | |
| 195 unsigned int copy_headers (function *fun); | |
| 196 | |
| 197 /* Return true to copy headers of LOOP or false to skip. */ | |
| 198 virtual bool process_loop_p (struct loop *loop) = 0; | |
| 199 }; | |
| 200 | |
| 201 const pass_data pass_data_ch = | |
| 202 { | |
| 203 GIMPLE_PASS, /* type */ | |
| 204 "ch", /* name */ | |
| 205 OPTGROUP_LOOP, /* optinfo_flags */ | |
| 206 TV_TREE_CH, /* tv_id */ | |
| 207 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 208 0, /* properties_provided */ | |
| 209 0, /* properties_destroyed */ | |
| 210 0, /* todo_flags_start */ | |
| 211 0, /* todo_flags_finish */ | |
| 212 }; | |
| 213 | |
| 214 class pass_ch : public ch_base | |
| 215 { | |
| 216 public: | |
| 217 pass_ch (gcc::context *ctxt) | |
| 218 : ch_base (pass_data_ch, ctxt) | |
| 219 {} | |
| 220 | |
| 221 /* opt_pass methods: */ | |
| 222 virtual bool gate (function *) { return flag_tree_ch != 0; } | |
| 223 | |
| 224 /* Initialize and finalize loop structures, copying headers inbetween. */ | |
| 225 virtual unsigned int execute (function *); | |
| 226 | |
| 227 opt_pass * clone () { return new pass_ch (m_ctxt); } | |
| 228 | |
| 229 protected: | |
| 230 /* ch_base method: */ | |
| 231 virtual bool process_loop_p (struct loop *loop); | |
| 232 }; // class pass_ch | |
| 233 | |
| 234 const pass_data pass_data_ch_vect = | |
| 235 { | |
| 236 GIMPLE_PASS, /* type */ | |
| 237 "ch_vect", /* name */ | |
| 238 OPTGROUP_LOOP, /* optinfo_flags */ | |
| 239 TV_TREE_CH, /* tv_id */ | |
| 240 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 241 0, /* properties_provided */ | |
| 242 0, /* properties_destroyed */ | |
| 243 0, /* todo_flags_start */ | |
| 244 0, /* todo_flags_finish */ | |
| 245 }; | |
| 246 | |
| 247 /* This is a more aggressive version of the same pass, designed to run just | |
| 248 before if-conversion and vectorization, to put more loops into the form | |
| 249 required for those phases. */ | |
| 250 class pass_ch_vect : public ch_base | |
| 251 { | |
| 252 public: | |
| 253 pass_ch_vect (gcc::context *ctxt) | |
| 254 : ch_base (pass_data_ch_vect, ctxt) | |
| 255 {} | |
| 256 | |
| 257 /* opt_pass methods: */ | |
| 258 virtual bool gate (function *fun) | |
| 259 { | |
| 260 return flag_tree_ch != 0 | |
| 261 && (flag_tree_loop_vectorize != 0 || fun->has_force_vectorize_loops); | |
| 262 } | |
| 263 | |
| 264 /* Just copy headers, no initialization/finalization of loop structures. */ | |
| 265 virtual unsigned int execute (function *); | |
| 266 | |
| 267 protected: | |
| 268 /* ch_base method: */ | |
| 269 virtual bool process_loop_p (struct loop *loop); | |
| 270 }; // class pass_ch_vect | |
| 271 | |
| 0 | 272 /* For all loops, copy the condition at the end of the loop body in front |
| 273 of the loop. This is beneficial since it increases efficiency of | |
| 274 code motion optimizations. It also saves one jump on entry to the loop. */ | |
| 275 | |
| 111 | 276 unsigned int |
| 277 ch_base::copy_headers (function *fun) | |
| 0 | 278 { |
| 279 struct loop *loop; | |
| 280 basic_block header; | |
| 281 edge exit, entry; | |
| 282 basic_block *bbs, *copied_bbs; | |
| 283 unsigned n_bbs; | |
| 284 unsigned bbs_size; | |
| 111 | 285 bool changed = false; |
| 0 | 286 |
| 111 | 287 if (number_of_loops (fun) <= 1) |
| 0 | 288 return 0; |
| 289 | |
| 111 | 290 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun)); |
| 291 copied_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun)); | |
| 292 bbs_size = n_basic_blocks_for_fn (fun); | |
| 0 | 293 |
| 111 | 294 FOR_EACH_LOOP (loop, 0) |
| 0 | 295 { |
| 111 | 296 int initial_limit = PARAM_VALUE (PARAM_MAX_LOOP_HEADER_INSNS); |
| 297 int remaining_limit = initial_limit; | |
| 298 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 299 fprintf (dump_file, | |
| 300 "Analyzing loop %i\n", loop->num); | |
| 0 | 301 |
| 302 header = loop->header; | |
| 303 | |
| 304 /* If the loop is already a do-while style one (either because it was | |
| 305 written as such, or because jump threading transformed it into one), | |
| 306 we might be in fact peeling the first iteration of the loop. This | |
| 307 in general is not a good idea. */ | |
| 111 | 308 if (!process_loop_p (loop)) |
| 0 | 309 continue; |
| 310 | |
| 311 /* Iterate the header copying up to limit; this takes care of the cases | |
| 312 like while (a && b) {...}, where we want to have both of the conditions | |
| 313 copied. TODO -- handle while (a || b) - like cases, by not requiring | |
| 314 the header to have just a single successor and copying up to | |
| 315 postdominator. */ | |
| 316 | |
| 317 exit = NULL; | |
| 318 n_bbs = 0; | |
| 111 | 319 while (should_duplicate_loop_header_p (header, loop, &remaining_limit)) |
| 0 | 320 { |
| 321 /* Find a successor of header that is inside a loop; i.e. the new | |
| 322 header after the condition is copied. */ | |
| 323 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest)) | |
| 324 exit = EDGE_SUCC (header, 0); | |
| 325 else | |
| 326 exit = EDGE_SUCC (header, 1); | |
| 327 bbs[n_bbs++] = header; | |
| 328 gcc_assert (bbs_size > n_bbs); | |
| 329 header = exit->dest; | |
| 330 } | |
| 331 | |
| 332 if (!exit) | |
| 333 continue; | |
| 334 | |
| 335 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 336 fprintf (dump_file, | |
| 111 | 337 "Duplicating header of the loop %d up to edge %d->%d," |
| 338 " %i insns.\n", | |
| 339 loop->num, exit->src->index, exit->dest->index, | |
| 340 initial_limit - remaining_limit); | |
| 0 | 341 |
| 342 /* Ensure that the header will have just the latch as a predecessor | |
| 343 inside the loop. */ | |
| 344 if (!single_pred_p (exit->dest)) | |
| 345 exit = single_pred_edge (split_edge (exit)); | |
| 346 | |
| 347 entry = loop_preheader_edge (loop); | |
| 348 | |
| 111 | 349 propagate_threaded_block_debug_into (exit->dest, entry->dest); |
| 350 if (!gimple_duplicate_sese_region (entry, exit, bbs, n_bbs, copied_bbs, | |
| 351 true)) | |
| 0 | 352 { |
| 353 fprintf (dump_file, "Duplication failed.\n"); | |
| 354 continue; | |
| 355 } | |
| 356 | |
| 357 /* If the loop has the form "for (i = j; i < j + 10; i++)" then | |
| 358 this copying can introduce a case where we rely on undefined | |
| 359 signed overflow to eliminate the preheader condition, because | |
| 360 we assume that "j < j + 10" is true. We don't want to warn | |
| 361 about that case for -Wstrict-overflow, because in general we | |
| 362 don't warn about overflow involving loops. Prevent the | |
| 363 warning by setting the no_warning flag in the condition. */ | |
| 364 if (warn_strict_overflow > 0) | |
| 365 { | |
| 366 unsigned int i; | |
| 367 | |
| 368 for (i = 0; i < n_bbs; ++i) | |
| 369 { | |
| 370 gimple_stmt_iterator bsi; | |
| 371 | |
| 372 for (bsi = gsi_start_bb (copied_bbs[i]); | |
| 373 !gsi_end_p (bsi); | |
| 374 gsi_next (&bsi)) | |
| 375 { | |
| 111 | 376 gimple *stmt = gsi_stmt (bsi); |
| 0 | 377 if (gimple_code (stmt) == GIMPLE_COND) |
| 378 gimple_set_no_warning (stmt, true); | |
| 379 else if (is_gimple_assign (stmt)) | |
| 380 { | |
| 381 enum tree_code rhs_code = gimple_assign_rhs_code (stmt); | |
| 382 if (TREE_CODE_CLASS (rhs_code) == tcc_comparison) | |
| 383 gimple_set_no_warning (stmt, true); | |
| 384 } | |
| 385 } | |
| 386 } | |
| 387 } | |
| 388 | |
| 389 /* Ensure that the latch and the preheader is simple (we know that they | |
| 390 are not now, since there was the loop exit condition. */ | |
| 391 split_edge (loop_preheader_edge (loop)); | |
| 392 split_edge (loop_latch_edge (loop)); | |
| 111 | 393 |
| 394 changed = true; | |
| 0 | 395 } |
| 396 | |
| 111 | 397 if (changed) |
| 398 update_ssa (TODO_update_ssa); | |
| 0 | 399 free (bbs); |
| 400 free (copied_bbs); | |
| 401 | |
| 111 | 402 return changed ? TODO_cleanup_cfg : 0; |
| 403 } | |
| 404 | |
| 405 /* Initialize the loop structures we need, and finalize after. */ | |
| 406 | |
| 407 unsigned int | |
| 408 pass_ch::execute (function *fun) | |
| 409 { | |
| 410 loop_optimizer_init (LOOPS_HAVE_PREHEADERS | |
| 411 | LOOPS_HAVE_SIMPLE_LATCHES); | |
| 412 | |
| 413 unsigned int res = copy_headers (fun); | |
| 414 | |
| 0 | 415 loop_optimizer_finalize (); |
| 111 | 416 return res; |
| 0 | 417 } |
| 418 | |
| 111 | 419 /* Assume an earlier phase has already initialized all the loop structures that |
| 420 we need here (and perhaps others too), and that these will be finalized by | |
| 421 a later phase. */ | |
| 422 | |
| 423 unsigned int | |
| 424 pass_ch_vect::execute (function *fun) | |
| 0 | 425 { |
| 111 | 426 return copy_headers (fun); |
| 427 } | |
| 428 | |
| 429 /* Apply header copying according to a very simple test of do-while shape. */ | |
| 430 | |
| 431 bool | |
| 432 pass_ch::process_loop_p (struct loop *loop) | |
| 433 { | |
| 434 return !do_while_loop_p (loop); | |
| 0 | 435 } |
| 436 | |
| 111 | 437 /* Apply header-copying to loops where we might enable vectorization. */ |
| 438 | |
| 439 bool | |
| 440 pass_ch_vect::process_loop_p (struct loop *loop) | |
| 0 | 441 { |
| 111 | 442 if (!flag_tree_loop_vectorize && !loop->force_vectorize) |
| 443 return false; | |
| 444 | |
| 445 if (loop->dont_vectorize) | |
| 446 return false; | |
| 447 | |
| 448 if (!do_while_loop_p (loop)) | |
| 449 return true; | |
| 450 | |
| 451 /* The vectorizer won't handle anything with multiple exits, so skip. */ | |
| 452 edge exit = single_exit (loop); | |
| 453 if (!exit) | |
| 454 return false; | |
| 455 | |
| 456 /* Copy headers iff there looks to be code in the loop after the exit block, | |
| 457 i.e. the exit block has an edge to another block (besides the latch, | |
| 458 which should be empty). */ | |
| 459 edge_iterator ei; | |
| 460 edge e; | |
| 461 FOR_EACH_EDGE (e, ei, exit->src->succs) | |
| 462 if (!loop_exit_edge_p (loop, e) | |
| 463 && e->dest != loop->header | |
| 464 && e->dest != loop->latch) | |
| 465 return true; | |
| 466 | |
| 467 return false; | |
| 468 } | |
| 469 | |
| 470 } // anon namespace | |
| 471 | |
| 472 gimple_opt_pass * | |
| 473 make_pass_ch_vect (gcc::context *ctxt) | |
| 474 { | |
| 475 return new pass_ch_vect (ctxt); | |
| 476 } | |
| 477 | |
| 478 gimple_opt_pass * | |
| 479 make_pass_ch (gcc::context *ctxt) | |
| 480 { | |
| 481 return new pass_ch (ctxt); | |
| 482 } |