1 /* High-level loop manipulation functions. 2 Copyright (C) 2004-2018 Free Software Foundation, Inc. 3 4 This file is part of GCC. 5 6 GCC is free software; you can redistribute it and/or modify it 7 under the terms of the GNU General Public License as published by the 8 Free Software Foundation; either version 3, or (at your option) any 9 later version. 10 11 GCC is distributed in the hope that it will be useful, but WITHOUT 12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 for more details. 15 16 You should have received a copy of the GNU General Public License 17 along with GCC; see the file COPYING3. If not see 18 <http://www.gnu.org/licenses/>. */ 19 20 #include "config.h" 21 #include "system.h" 22 #include "coretypes.h" 23 #include "backend.h" 24 #include "tree.h" 25 #include "gimple.h" 26 #include "cfghooks.h" 27 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */ 28 #include "ssa.h" 29 #include "gimple-pretty-print.h" 30 #include "fold-const.h" 31 #include "cfganal.h" 32 #include "gimplify.h" 33 #include "gimple-iterator.h" 34 #include "gimplify-me.h" 35 #include "tree-cfg.h" 36 #include "tree-ssa-loop-ivopts.h" 37 #include "tree-ssa-loop-manip.h" 38 #include "tree-ssa-loop-niter.h" 39 #include "tree-ssa-loop.h" 40 #include "tree-into-ssa.h" 41 #include "tree-ssa.h" 42 #include "cfgloop.h" 43 #include "tree-scalar-evolution.h" 44 #include "params.h" 45 #include "tree-inline.h" 46 47 /* All bitmaps for rewriting into loop-closed SSA go on this obstack, 48 so that we can free them all at once. */ 49 static bitmap_obstack loop_renamer_obstack; 50 51 /* Creates an induction variable with value BASE + STEP * iteration in LOOP. 52 It is expected that neither BASE nor STEP are shared with other expressions 53 (unless the sharing rules allow this). Use VAR as a base var_decl for it 54 (if NULL, a new temporary will be created). The increment will occur at 55 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and 56 AFTER can be computed using standard_iv_increment_position. The ssa versions 57 of the variable before and after increment will be stored in VAR_BEFORE and 58 VAR_AFTER (unless they are NULL). */ 59 60 void 61 create_iv (tree base, tree step, tree var, struct loop *loop, 62 gimple_stmt_iterator *incr_pos, bool after, 63 tree *var_before, tree *var_after) 64 { 65 gassign *stmt; 66 gphi *phi; 67 tree initial, step1; 68 gimple_seq stmts; 69 tree vb, va; 70 enum tree_code incr_op = PLUS_EXPR; 71 edge pe = loop_preheader_edge (loop); 72 73 if (var != NULL_TREE) 74 { 75 vb = make_ssa_name (var); 76 va = make_ssa_name (var); 77 } 78 else 79 { 80 vb = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp"); 81 va = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp"); 82 } 83 if (var_before) 84 *var_before = vb; 85 if (var_after) 86 *var_after = va; 87 88 /* For easier readability of the created code, produce MINUS_EXPRs 89 when suitable. */ 90 if (TREE_CODE (step) == INTEGER_CST) 91 { 92 if (TYPE_UNSIGNED (TREE_TYPE (step))) 93 { 94 step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); 95 if (tree_int_cst_lt (step1, step)) 96 { 97 incr_op = MINUS_EXPR; 98 step = step1; 99 } 100 } 101 else 102 { 103 bool ovf; 104 105 if (!tree_expr_nonnegative_warnv_p (step, &ovf) 106 && may_negate_without_overflow_p (step)) 107 { 108 incr_op = MINUS_EXPR; 109 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); 110 } 111 } 112 } 113 if (POINTER_TYPE_P (TREE_TYPE (base))) 114 { 115 if (TREE_CODE (base) == ADDR_EXPR) 116 mark_addressable (TREE_OPERAND (base, 0)); 117 step = convert_to_ptrofftype (step); 118 if (incr_op == MINUS_EXPR) 119 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); 120 incr_op = POINTER_PLUS_EXPR; 121 } 122 /* Gimplify the step if necessary. We put the computations in front of the 123 loop (i.e. the step should be loop invariant). */ 124 step = force_gimple_operand (step, &stmts, true, NULL_TREE); 125 if (stmts) 126 gsi_insert_seq_on_edge_immediate (pe, stmts); 127 128 stmt = gimple_build_assign (va, incr_op, vb, step); 129 if (after) 130 gsi_insert_after (incr_pos, stmt, GSI_NEW_STMT); 131 else 132 gsi_insert_before (incr_pos, stmt, GSI_NEW_STMT); 133 134 initial = force_gimple_operand (base, &stmts, true, var); 135 if (stmts) 136 gsi_insert_seq_on_edge_immediate (pe, stmts); 137 138 phi = create_phi_node (vb, loop->header); 139 add_phi_arg (phi, initial, loop_preheader_edge (loop), UNKNOWN_LOCATION); 140 add_phi_arg (phi, va, loop_latch_edge (loop), UNKNOWN_LOCATION); 141 } 142 143 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of 144 both DEF_LOOP and USE_LOOP. */ 145 146 static inline struct loop * 147 find_sibling_superloop (struct loop *use_loop, struct loop *def_loop) 148 { 149 unsigned ud = loop_depth (use_loop); 150 unsigned dd = loop_depth (def_loop); 151 gcc_assert (ud > 0 && dd > 0); 152 if (ud > dd) 153 use_loop = superloop_at_depth (use_loop, dd); 154 if (ud < dd) 155 def_loop = superloop_at_depth (def_loop, ud); 156 while (loop_outer (use_loop) != loop_outer (def_loop)) 157 { 158 use_loop = loop_outer (use_loop); 159 def_loop = loop_outer (def_loop); 160 gcc_assert (use_loop && def_loop); 161 } 162 return use_loop; 163 } 164 165 /* DEF_BB is a basic block containing a DEF that needs rewriting into 166 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing 167 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in 168 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B). 169 ALL_EXITS[I] is the set of all basic blocks that exit loop I. 170 171 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB 172 or one of its loop fathers, in which DEF is live. This set is returned 173 in the bitmap LIVE_EXITS. 174 175 Instead of computing the complete livein set of the def, we use the loop 176 nesting tree as a form of poor man's structure analysis. This greatly 177 speeds up the analysis, which is important because this function may be 178 called on all SSA names that need rewriting, one at a time. */ 179 180 static void 181 compute_live_loop_exits (bitmap live_exits, bitmap use_blocks, 182 bitmap *loop_exits, basic_block def_bb) 183 { 184 unsigned i; 185 bitmap_iterator bi; 186 struct loop *def_loop = def_bb->loop_father; 187 unsigned def_loop_depth = loop_depth (def_loop); 188 bitmap def_loop_exits; 189 190 /* Normally the work list size is bounded by the number of basic 191 blocks in the largest loop. We don't know this number, but we 192 can be fairly sure that it will be relatively small. */ 193 auto_vec<basic_block> worklist (MAX (8, n_basic_blocks_for_fn (cfun) / 128)); 194 195 EXECUTE_IF_SET_IN_BITMAP (use_blocks, 0, i, bi) 196 { 197 basic_block use_bb = BASIC_BLOCK_FOR_FN (cfun, i); 198 struct loop *use_loop = use_bb->loop_father; 199 gcc_checking_assert (def_loop != use_loop 200 && ! flow_loop_nested_p (def_loop, use_loop)); 201 if (! flow_loop_nested_p (use_loop, def_loop)) 202 use_bb = find_sibling_superloop (use_loop, def_loop)->header; 203 if (bitmap_set_bit (live_exits, use_bb->index)) 204 worklist.safe_push (use_bb); 205 } 206 207 /* Iterate until the worklist is empty. */ 208 while (! worklist.is_empty ()) 209 { 210 edge e; 211 edge_iterator ei; 212 213 /* Pull a block off the worklist. */ 214 basic_block bb = worklist.pop (); 215 216 /* Make sure we have at least enough room in the work list 217 for all predecessors of this block. */ 218 worklist.reserve (EDGE_COUNT (bb->preds)); 219 220 /* For each predecessor block. */ 221 FOR_EACH_EDGE (e, ei, bb->preds) 222 { 223 basic_block pred = e->src; 224 struct loop *pred_loop = pred->loop_father; 225 unsigned pred_loop_depth = loop_depth (pred_loop); 226 bool pred_visited; 227 228 /* We should have met DEF_BB along the way. */ 229 gcc_assert (pred != ENTRY_BLOCK_PTR_FOR_FN (cfun)); 230 231 if (pred_loop_depth >= def_loop_depth) 232 { 233 if (pred_loop_depth > def_loop_depth) 234 pred_loop = superloop_at_depth (pred_loop, def_loop_depth); 235 /* If we've reached DEF_LOOP, our train ends here. */ 236 if (pred_loop == def_loop) 237 continue; 238 } 239 else if (! flow_loop_nested_p (pred_loop, def_loop)) 240 pred = find_sibling_superloop (pred_loop, def_loop)->header; 241 242 /* Add PRED to the LIVEIN set. PRED_VISITED is true if 243 we had already added PRED to LIVEIN before. */ 244 pred_visited = !bitmap_set_bit (live_exits, pred->index); 245 246 /* If we have visited PRED before, don't add it to the worklist. 247 If BB dominates PRED, then we're probably looking at a loop. 248 We're only interested in looking up in the dominance tree 249 because DEF_BB dominates all the uses. */ 250 if (pred_visited || dominated_by_p (CDI_DOMINATORS, pred, bb)) 251 continue; 252 253 worklist.quick_push (pred); 254 } 255 } 256 257 def_loop_exits = BITMAP_ALLOC (&loop_renamer_obstack); 258 for (struct loop *loop = def_loop; 259 loop != current_loops->tree_root; 260 loop = loop_outer (loop)) 261 bitmap_ior_into (def_loop_exits, loop_exits[loop->num]); 262 bitmap_and_into (live_exits, def_loop_exits); 263 BITMAP_FREE (def_loop_exits); 264 } 265 266 /* Add a loop-closing PHI for VAR in basic block EXIT. */ 267 268 static void 269 add_exit_phi (basic_block exit, tree var) 270 { 271 gphi *phi; 272 edge e; 273 edge_iterator ei; 274 275 /* Check that at least one of the edges entering the EXIT block exits 276 the loop, or a superloop of that loop, that VAR is defined in. */ 277 if (flag_checking) 278 { 279 gimple *def_stmt = SSA_NAME_DEF_STMT (var); 280 basic_block def_bb = gimple_bb (def_stmt); 281 FOR_EACH_EDGE (e, ei, exit->preds) 282 { 283 struct loop *aloop = find_common_loop (def_bb->loop_father, 284 e->src->loop_father); 285 if (!flow_bb_inside_loop_p (aloop, e->dest)) 286 break; 287 } 288 gcc_assert (e); 289 } 290 291 phi = create_phi_node (NULL_TREE, exit); 292 create_new_def_for (var, phi, gimple_phi_result_ptr (phi)); 293 FOR_EACH_EDGE (e, ei, exit->preds) 294 add_phi_arg (phi, var, e, UNKNOWN_LOCATION); 295 296 if (dump_file && (dump_flags & TDF_DETAILS)) 297 { 298 fprintf (dump_file, ";; Created LCSSA PHI: "); 299 print_gimple_stmt (dump_file, phi, 0, dump_flags); 300 } 301 } 302 303 /* Add exit phis for VAR that is used in LIVEIN. 304 Exits of the loops are stored in LOOP_EXITS. */ 305 306 static void 307 add_exit_phis_var (tree var, bitmap use_blocks, bitmap *loop_exits) 308 { 309 unsigned index; 310 bitmap_iterator bi; 311 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var)); 312 bitmap live_exits = BITMAP_ALLOC (&loop_renamer_obstack); 313 314 gcc_checking_assert (! bitmap_bit_p (use_blocks, def_bb->index)); 315 316 compute_live_loop_exits (live_exits, use_blocks, loop_exits, def_bb); 317 318 EXECUTE_IF_SET_IN_BITMAP (live_exits, 0, index, bi) 319 { 320 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun, index), var); 321 } 322 323 BITMAP_FREE (live_exits); 324 } 325 326 /* Add exit phis for the names marked in NAMES_TO_RENAME. 327 Exits of the loops are stored in EXITS. Sets of blocks where the ssa 328 names are used are stored in USE_BLOCKS. */ 329 330 static void 331 add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap *loop_exits) 332 { 333 unsigned i; 334 bitmap_iterator bi; 335 336 EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi) 337 { 338 add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits); 339 } 340 } 341 342 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */ 343 344 static void 345 get_loops_exits (bitmap *loop_exits) 346 { 347 struct loop *loop; 348 unsigned j; 349 edge e; 350 351 FOR_EACH_LOOP (loop, 0) 352 { 353 vec<edge> exit_edges = get_loop_exit_edges (loop); 354 loop_exits[loop->num] = BITMAP_ALLOC (&loop_renamer_obstack); 355 FOR_EACH_VEC_ELT (exit_edges, j, e) 356 bitmap_set_bit (loop_exits[loop->num], e->dest->index); 357 exit_edges.release (); 358 } 359 } 360 361 /* For USE in BB, if it is used outside of the loop it is defined in, 362 mark it for rewrite. Record basic block BB where it is used 363 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. 364 Note that for USEs in phis, BB should be the src of the edge corresponding to 365 the use, rather than the bb containing the phi. */ 366 367 static void 368 find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks, 369 bitmap need_phis) 370 { 371 unsigned ver; 372 basic_block def_bb; 373 struct loop *def_loop; 374 375 if (TREE_CODE (use) != SSA_NAME) 376 return; 377 378 ver = SSA_NAME_VERSION (use); 379 def_bb = gimple_bb (SSA_NAME_DEF_STMT (use)); 380 if (!def_bb) 381 return; 382 def_loop = def_bb->loop_father; 383 384 /* If the definition is not inside a loop, it is not interesting. */ 385 if (!loop_outer (def_loop)) 386 return; 387 388 /* If the use is not outside of the loop it is defined in, it is not 389 interesting. */ 390 if (flow_bb_inside_loop_p (def_loop, bb)) 391 return; 392 393 /* If we're seeing VER for the first time, we still have to allocate 394 a bitmap for its uses. */ 395 if (bitmap_set_bit (need_phis, ver)) 396 use_blocks[ver] = BITMAP_ALLOC (&loop_renamer_obstack); 397 bitmap_set_bit (use_blocks[ver], bb->index); 398 } 399 400 /* For uses matching USE_FLAGS in STMT, mark names that are used outside of the 401 loop they are defined to rewrite. Record the set of blocks in which the ssa 402 names are used to USE_BLOCKS, and the ssa names themselves to NEED_PHIS. */ 403 404 static void 405 find_uses_to_rename_stmt (gimple *stmt, bitmap *use_blocks, bitmap need_phis, 406 int use_flags) 407 { 408 ssa_op_iter iter; 409 tree var; 410 basic_block bb = gimple_bb (stmt); 411 412 if (is_gimple_debug (stmt)) 413 return; 414 415 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows SSA_OP_VIRTUAL_USES 416 only. */ 417 if (use_flags == SSA_OP_VIRTUAL_USES) 418 { 419 tree vuse = gimple_vuse (stmt); 420 if (vuse != NULL_TREE) 421 find_uses_to_rename_use (bb, gimple_vuse (stmt), use_blocks, need_phis); 422 } 423 else 424 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, use_flags) 425 find_uses_to_rename_use (bb, var, use_blocks, need_phis); 426 } 427 428 /* Marks names matching USE_FLAGS that are used in BB and outside of the loop 429 they are defined in for rewrite. Records the set of blocks in which the ssa 430 names are used to USE_BLOCKS. Record the SSA names that will 431 need exit PHIs in NEED_PHIS. */ 432 433 static void 434 find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis, 435 int use_flags) 436 { 437 edge e; 438 edge_iterator ei; 439 bool do_virtuals = (use_flags & SSA_OP_VIRTUAL_USES) != 0; 440 bool do_nonvirtuals = (use_flags & SSA_OP_USE) != 0; 441 442 FOR_EACH_EDGE (e, ei, bb->succs) 443 for (gphi_iterator bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi); 444 gsi_next (&bsi)) 445 { 446 gphi *phi = bsi.phi (); 447 bool virtual_p = virtual_operand_p (gimple_phi_result (phi)); 448 if ((virtual_p && do_virtuals) 449 || (!virtual_p && do_nonvirtuals)) 450 find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e), 451 use_blocks, need_phis); 452 } 453 454 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); 455 gsi_next (&bsi)) 456 find_uses_to_rename_stmt (gsi_stmt (bsi), use_blocks, need_phis, 457 use_flags); 458 } 459 460 /* Marks names matching USE_FLAGS that are used outside of the loop they are 461 defined in for rewrite. Records the set of blocks in which the ssa names are 462 used to USE_BLOCKS. Record the SSA names that will need exit PHIs in 463 NEED_PHIS. If CHANGED_BBS is not NULL, scan only blocks in this set. */ 464 465 static void 466 find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis, 467 int use_flags) 468 { 469 basic_block bb; 470 unsigned index; 471 bitmap_iterator bi; 472 473 if (changed_bbs) 474 EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi) 475 { 476 bb = BASIC_BLOCK_FOR_FN (cfun, index); 477 if (bb) 478 find_uses_to_rename_bb (bb, use_blocks, need_phis, use_flags); 479 } 480 else 481 FOR_EACH_BB_FN (bb, cfun) 482 find_uses_to_rename_bb (bb, use_blocks, need_phis, use_flags); 483 } 484 485 /* Mark uses of DEF that are used outside of the loop they are defined in for 486 rewrite. Record the set of blocks in which the ssa names are used to 487 USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS. */ 488 489 static void 490 find_uses_to_rename_def (tree def, bitmap *use_blocks, bitmap need_phis) 491 { 492 gimple *use_stmt; 493 imm_use_iterator imm_iter; 494 495 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) 496 { 497 if (is_gimple_debug (use_stmt)) 498 continue; 499 500 basic_block use_bb = gimple_bb (use_stmt); 501 502 use_operand_p use_p; 503 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) 504 { 505 if (gimple_code (use_stmt) == GIMPLE_PHI) 506 { 507 edge e = gimple_phi_arg_edge (as_a <gphi *> (use_stmt), 508 PHI_ARG_INDEX_FROM_USE (use_p)); 509 use_bb = e->src; 510 } 511 find_uses_to_rename_use (use_bb, USE_FROM_PTR (use_p), use_blocks, 512 need_phis); 513 } 514 } 515 } 516 517 /* Marks names matching USE_FLAGS that are defined in LOOP and used outside of 518 it for rewrite. Records the set of blocks in which the ssa names are used to 519 USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS. */ 520 521 static void 522 find_uses_to_rename_in_loop (struct loop *loop, bitmap *use_blocks, 523 bitmap need_phis, int use_flags) 524 { 525 bool do_virtuals = (use_flags & SSA_OP_VIRTUAL_USES) != 0; 526 bool do_nonvirtuals = (use_flags & SSA_OP_USE) != 0; 527 int def_flags = ((do_virtuals ? SSA_OP_VIRTUAL_DEFS : 0) 528 | (do_nonvirtuals ? SSA_OP_DEF : 0)); 529 530 531 basic_block *bbs = get_loop_body (loop); 532 533 for (unsigned int i = 0; i < loop->num_nodes; i++) 534 { 535 basic_block bb = bbs[i]; 536 537 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); 538 gsi_next (&bsi)) 539 { 540 gphi *phi = bsi.phi (); 541 tree res = gimple_phi_result (phi); 542 bool virtual_p = virtual_operand_p (res); 543 if ((virtual_p && do_virtuals) 544 || (!virtual_p && do_nonvirtuals)) 545 find_uses_to_rename_def (res, use_blocks, need_phis); 546 } 547 548 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); 549 gsi_next (&bsi)) 550 { 551 gimple *stmt = gsi_stmt (bsi); 552 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows 553 SSA_OP_VIRTUAL_DEFS only. */ 554 if (def_flags == SSA_OP_VIRTUAL_DEFS) 555 { 556 tree vdef = gimple_vdef (stmt); 557 if (vdef != NULL) 558 find_uses_to_rename_def (vdef, use_blocks, need_phis); 559 } 560 else 561 { 562 tree var; 563 ssa_op_iter iter; 564 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, def_flags) 565 find_uses_to_rename_def (var, use_blocks, need_phis); 566 } 567 } 568 } 569 570 XDELETEVEC (bbs); 571 } 572 573 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra 574 phi nodes to ensure that no variable is used outside the loop it is 575 defined in. 576 577 This strengthening of the basic ssa form has several advantages: 578 579 1) Updating it during unrolling/peeling/versioning is trivial, since 580 we do not need to care about the uses outside of the loop. 581 The same applies to virtual operands which are also rewritten into 582 loop closed SSA form. Note that virtual operands are always live 583 until function exit. 584 2) The behavior of all uses of an induction variable is the same. 585 Without this, you need to distinguish the case when the variable 586 is used outside of the loop it is defined in, for example 587 588 for (i = 0; i < 100; i++) 589 { 590 for (j = 0; j < 100; j++) 591 { 592 k = i + j; 593 use1 (k); 594 } 595 use2 (k); 596 } 597 598 Looking from the outer loop with the normal SSA form, the first use of k 599 is not well-behaved, while the second one is an induction variable with 600 base 99 and step 1. 601 602 If LOOP is non-null, only rewrite uses that have defs in LOOP. Otherwise, 603 if CHANGED_BBS is not NULL, we look for uses outside loops only in the 604 basic blocks in this set. 605 606 USE_FLAGS allows us to specify whether we want virtual, non-virtual or 607 both variables rewritten. 608 609 UPDATE_FLAG is used in the call to update_ssa. See 610 TODO_update_ssa* for documentation. */ 611 612 void 613 rewrite_into_loop_closed_ssa_1 (bitmap changed_bbs, unsigned update_flag, 614 int use_flags, struct loop *loop) 615 { 616 bitmap *use_blocks; 617 bitmap names_to_rename; 618 619 loops_state_set (LOOP_CLOSED_SSA); 620 if (number_of_loops (cfun) <= 1) 621 return; 622 623 /* If the pass has caused the SSA form to be out-of-date, update it 624 now. */ 625 if (update_flag != 0) 626 update_ssa (update_flag); 627 else if (flag_checking) 628 verify_ssa (true, true); 629 630 bitmap_obstack_initialize (&loop_renamer_obstack); 631 632 names_to_rename = BITMAP_ALLOC (&loop_renamer_obstack); 633 634 /* Uses of names to rename. We don't have to initialize this array, 635 because we know that we will only have entries for the SSA names 636 in NAMES_TO_RENAME. */ 637 use_blocks = XNEWVEC (bitmap, num_ssa_names); 638 639 if (loop != NULL) 640 { 641 gcc_assert (changed_bbs == NULL); 642 find_uses_to_rename_in_loop (loop, use_blocks, names_to_rename, 643 use_flags); 644 } 645 else 646 { 647 gcc_assert (loop == NULL); 648 find_uses_to_rename (changed_bbs, use_blocks, names_to_rename, use_flags); 649 } 650 651 if (!bitmap_empty_p (names_to_rename)) 652 { 653 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks 654 that are the destination of an edge exiting loop number I. */ 655 bitmap *loop_exits = XNEWVEC (bitmap, number_of_loops (cfun)); 656 get_loops_exits (loop_exits); 657 658 /* Add the PHI nodes on exits of the loops for the names we need to 659 rewrite. */ 660 add_exit_phis (names_to_rename, use_blocks, loop_exits); 661 662 free (loop_exits); 663 664 /* Fix up all the names found to be used outside their original 665 loops. */ 666 update_ssa (TODO_update_ssa); 667 } 668 669 bitmap_obstack_release (&loop_renamer_obstack); 670 free (use_blocks); 671 } 672 673 /* Rewrites the non-virtual defs and uses into a loop closed ssa form. If 674 CHANGED_BBS is not NULL, we look for uses outside loops only in the basic 675 blocks in this set. UPDATE_FLAG is used in the call to update_ssa. See 676 TODO_update_ssa* for documentation. */ 677 678 void 679 rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag) 680 { 681 rewrite_into_loop_closed_ssa_1 (changed_bbs, update_flag, SSA_OP_USE, NULL); 682 } 683 684 /* Rewrites virtual defs and uses with def in LOOP into loop closed ssa 685 form. */ 686 687 void 688 rewrite_virtuals_into_loop_closed_ssa (struct loop *loop) 689 { 690 rewrite_into_loop_closed_ssa_1 (NULL, 0, SSA_OP_VIRTUAL_USES, loop); 691 } 692 693 /* Check invariants of the loop closed ssa form for the def in DEF_BB. */ 694 695 static void 696 check_loop_closed_ssa_def (basic_block def_bb, tree def) 697 { 698 use_operand_p use_p; 699 imm_use_iterator iterator; 700 FOR_EACH_IMM_USE_FAST (use_p, iterator, def) 701 { 702 if (is_gimple_debug (USE_STMT (use_p))) 703 continue; 704 705 basic_block use_bb = gimple_bb (USE_STMT (use_p)); 706 if (is_a <gphi *> (USE_STMT (use_p))) 707 use_bb = EDGE_PRED (use_bb, PHI_ARG_INDEX_FROM_USE (use_p))->src; 708 709 gcc_assert (flow_bb_inside_loop_p (def_bb->loop_father, use_bb)); 710 } 711 } 712 713 /* Checks invariants of loop closed ssa form in BB. */ 714 715 static void 716 check_loop_closed_ssa_bb (basic_block bb) 717 { 718 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); 719 gsi_next (&bsi)) 720 { 721 gphi *phi = bsi.phi (); 722 723 if (!virtual_operand_p (PHI_RESULT (phi))) 724 check_loop_closed_ssa_def (bb, PHI_RESULT (phi)); 725 } 726 727 for (gimple_stmt_iterator bsi = gsi_start_nondebug_bb (bb); !gsi_end_p (bsi); 728 gsi_next_nondebug (&bsi)) 729 { 730 ssa_op_iter iter; 731 tree var; 732 gimple *stmt = gsi_stmt (bsi); 733 734 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_DEF) 735 check_loop_closed_ssa_def (bb, var); 736 } 737 } 738 739 /* Checks that invariants of the loop closed ssa form are preserved. 740 Call verify_ssa when VERIFY_SSA_P is true. Note all loops are checked 741 if LOOP is NULL, otherwise, only LOOP is checked. */ 742 743 DEBUG_FUNCTION void 744 verify_loop_closed_ssa (bool verify_ssa_p, struct loop *loop) 745 { 746 if (number_of_loops (cfun) <= 1) 747 return; 748 749 if (verify_ssa_p) 750 verify_ssa (false, true); 751 752 timevar_push (TV_VERIFY_LOOP_CLOSED); 753 754 if (loop == NULL) 755 { 756 basic_block bb; 757 758 FOR_EACH_BB_FN (bb, cfun) 759 if (bb->loop_father && bb->loop_father->num > 0) 760 check_loop_closed_ssa_bb (bb); 761 } 762 else 763 { 764 basic_block *bbs = get_loop_body (loop); 765 766 for (unsigned i = 0; i < loop->num_nodes; ++i) 767 check_loop_closed_ssa_bb (bbs[i]); 768 769 free (bbs); 770 } 771 772 timevar_pop (TV_VERIFY_LOOP_CLOSED); 773 } 774 775 /* Split loop exit edge EXIT. The things are a bit complicated by a need to 776 preserve the loop closed ssa form. The newly created block is returned. */ 777 778 basic_block 779 split_loop_exit_edge (edge exit) 780 { 781 basic_block dest = exit->dest; 782 basic_block bb = split_edge (exit); 783 gphi *phi, *new_phi; 784 tree new_name, name; 785 use_operand_p op_p; 786 gphi_iterator psi; 787 source_location locus; 788 789 for (psi = gsi_start_phis (dest); !gsi_end_p (psi); gsi_next (&psi)) 790 { 791 phi = psi.phi (); 792 op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb)); 793 locus = gimple_phi_arg_location_from_edge (phi, single_succ_edge (bb)); 794 795 name = USE_FROM_PTR (op_p); 796 797 /* If the argument of the PHI node is a constant, we do not need 798 to keep it inside loop. */ 799 if (TREE_CODE (name) != SSA_NAME) 800 continue; 801 802 /* Otherwise create an auxiliary phi node that will copy the value 803 of the SSA name out of the loop. */ 804 new_name = duplicate_ssa_name (name, NULL); 805 new_phi = create_phi_node (new_name, bb); 806 add_phi_arg (new_phi, name, exit, locus); 807 SET_USE (op_p, new_name); 808 } 809 810 return bb; 811 } 812 813 /* Returns the basic block in that statements should be emitted for induction 814 variables incremented at the end of the LOOP. */ 815 816 basic_block 817 ip_end_pos (struct loop *loop) 818 { 819 return loop->latch; 820 } 821 822 /* Returns the basic block in that statements should be emitted for induction 823 variables incremented just before exit condition of a LOOP. */ 824 825 basic_block 826 ip_normal_pos (struct loop *loop) 827 { 828 gimple *last; 829 basic_block bb; 830 edge exit; 831 832 if (!single_pred_p (loop->latch)) 833 return NULL; 834 835 bb = single_pred (loop->latch); 836 last = last_stmt (bb); 837 if (!last 838 || gimple_code (last) != GIMPLE_COND) 839 return NULL; 840 841 exit = EDGE_SUCC (bb, 0); 842 if (exit->dest == loop->latch) 843 exit = EDGE_SUCC (bb, 1); 844 845 if (flow_bb_inside_loop_p (loop, exit->dest)) 846 return NULL; 847 848 return bb; 849 } 850 851 /* Stores the standard position for induction variable increment in LOOP 852 (just before the exit condition if it is available and latch block is empty, 853 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if 854 the increment should be inserted after *BSI. */ 855 856 void 857 standard_iv_increment_position (struct loop *loop, gimple_stmt_iterator *bsi, 858 bool *insert_after) 859 { 860 basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop); 861 gimple *last = last_stmt (latch); 862 863 if (!bb 864 || (last && gimple_code (last) != GIMPLE_LABEL)) 865 { 866 *bsi = gsi_last_bb (latch); 867 *insert_after = true; 868 } 869 else 870 { 871 *bsi = gsi_last_bb (bb); 872 *insert_after = false; 873 } 874 } 875 876 /* Copies phi node arguments for duplicated blocks. The index of the first 877 duplicated block is FIRST_NEW_BLOCK. */ 878 879 static void 880 copy_phi_node_args (unsigned first_new_block) 881 { 882 unsigned i; 883 884 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++) 885 BASIC_BLOCK_FOR_FN (cfun, i)->flags |= BB_DUPLICATED; 886 887 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++) 888 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun, i)); 889 890 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++) 891 BASIC_BLOCK_FOR_FN (cfun, i)->flags &= ~BB_DUPLICATED; 892 } 893 894 895 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also 896 updates the PHI nodes at start of the copied region. In order to 897 achieve this, only loops whose exits all lead to the same location 898 are handled. 899 900 Notice that we do not completely update the SSA web after 901 duplication. The caller is responsible for calling update_ssa 902 after the loop has been duplicated. */ 903 904 bool 905 gimple_duplicate_loop_to_header_edge (struct loop *loop, edge e, 906 unsigned int ndupl, sbitmap wont_exit, 907 edge orig, vec<edge> *to_remove, 908 int flags) 909 { 910 unsigned first_new_block; 911 912 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES)) 913 return false; 914 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS)) 915 return false; 916 917 first_new_block = last_basic_block_for_fn (cfun); 918 if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit, 919 orig, to_remove, flags)) 920 return false; 921 922 /* Readd the removed phi args for e. */ 923 flush_pending_stmts (e); 924 925 /* Copy the phi node arguments. */ 926 copy_phi_node_args (first_new_block); 927 928 scev_reset (); 929 930 return true; 931 } 932 933 /* Returns true if we can unroll LOOP FACTOR times. Number 934 of iterations of the loop is returned in NITER. */ 935 936 bool 937 can_unroll_loop_p (struct loop *loop, unsigned factor, 938 struct tree_niter_desc *niter) 939 { 940 edge exit; 941 942 /* Check whether unrolling is possible. We only want to unroll loops 943 for that we are able to determine number of iterations. We also 944 want to split the extra iterations of the loop from its end, 945 therefore we require that the loop has precisely one 946 exit. */ 947 948 exit = single_dom_exit (loop); 949 if (!exit) 950 return false; 951 952 if (!number_of_iterations_exit (loop, exit, niter, false) 953 || niter->cmp == ERROR_MARK 954 /* Scalar evolutions analysis might have copy propagated 955 the abnormal ssa names into these expressions, hence 956 emitting the computations based on them during loop 957 unrolling might create overlapping life ranges for 958 them, and failures in out-of-ssa. */ 959 || contains_abnormal_ssa_name_p (niter->may_be_zero) 960 || contains_abnormal_ssa_name_p (niter->control.base) 961 || contains_abnormal_ssa_name_p (niter->control.step) 962 || contains_abnormal_ssa_name_p (niter->bound)) 963 return false; 964 965 /* And of course, we must be able to duplicate the loop. */ 966 if (!can_duplicate_loop_p (loop)) 967 return false; 968 969 /* The final loop should be small enough. */ 970 if (tree_num_loop_insns (loop, &eni_size_weights) * factor 971 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS)) 972 return false; 973 974 return true; 975 } 976 977 /* Determines the conditions that control execution of LOOP unrolled FACTOR 978 times. DESC is number of iterations of LOOP. ENTER_COND is set to 979 condition that must be true if the main loop can be entered. 980 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing 981 how the exit from the unrolled loop should be controlled. */ 982 983 static void 984 determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc, 985 unsigned factor, tree *enter_cond, 986 tree *exit_base, tree *exit_step, 987 enum tree_code *exit_cmp, tree *exit_bound) 988 { 989 gimple_seq stmts; 990 tree base = desc->control.base; 991 tree step = desc->control.step; 992 tree bound = desc->bound; 993 tree type = TREE_TYPE (step); 994 tree bigstep, delta; 995 tree min = lower_bound_in_type (type, type); 996 tree max = upper_bound_in_type (type, type); 997 enum tree_code cmp = desc->cmp; 998 tree cond = boolean_true_node, assum; 999 1000 /* For pointers, do the arithmetics in the type of step. */ 1001 base = fold_convert (type, base); 1002 bound = fold_convert (type, bound); 1003 1004 *enter_cond = boolean_false_node; 1005 *exit_base = NULL_TREE; 1006 *exit_step = NULL_TREE; 1007 *exit_cmp = ERROR_MARK; 1008 *exit_bound = NULL_TREE; 1009 gcc_assert (cmp != ERROR_MARK); 1010 1011 /* We only need to be correct when we answer question 1012 "Do at least FACTOR more iterations remain?" in the unrolled loop. 1013 Thus, transforming BASE + STEP * i <> BOUND to 1014 BASE + STEP * i < BOUND is ok. */ 1015 if (cmp == NE_EXPR) 1016 { 1017 if (tree_int_cst_sign_bit (step)) 1018 cmp = GT_EXPR; 1019 else 1020 cmp = LT_EXPR; 1021 } 1022 else if (cmp == LT_EXPR) 1023 { 1024 gcc_assert (!tree_int_cst_sign_bit (step)); 1025 } 1026 else if (cmp == GT_EXPR) 1027 { 1028 gcc_assert (tree_int_cst_sign_bit (step)); 1029 } 1030 else 1031 gcc_unreachable (); 1032 1033 /* The main body of the loop may be entered iff: 1034 1035 1) desc->may_be_zero is false. 1036 2) it is possible to check that there are at least FACTOR iterations 1037 of the loop, i.e., BOUND - step * FACTOR does not overflow. 1038 3) # of iterations is at least FACTOR */ 1039 1040 if (!integer_zerop (desc->may_be_zero)) 1041 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, 1042 invert_truthvalue (desc->may_be_zero), 1043 cond); 1044 1045 bigstep = fold_build2 (MULT_EXPR, type, step, 1046 build_int_cst_type (type, factor)); 1047 delta = fold_build2 (MINUS_EXPR, type, bigstep, step); 1048 if (cmp == LT_EXPR) 1049 assum = fold_build2 (GE_EXPR, boolean_type_node, 1050 bound, 1051 fold_build2 (PLUS_EXPR, type, min, delta)); 1052 else 1053 assum = fold_build2 (LE_EXPR, boolean_type_node, 1054 bound, 1055 fold_build2 (PLUS_EXPR, type, max, delta)); 1056 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond); 1057 1058 bound = fold_build2 (MINUS_EXPR, type, bound, delta); 1059 assum = fold_build2 (cmp, boolean_type_node, base, bound); 1060 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond); 1061 1062 cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE); 1063 if (stmts) 1064 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); 1065 /* cond now may be a gimple comparison, which would be OK, but also any 1066 other gimple rhs (say a && b). In this case we need to force it to 1067 operand. */ 1068 if (!is_gimple_condexpr (cond)) 1069 { 1070 cond = force_gimple_operand (cond, &stmts, true, NULL_TREE); 1071 if (stmts) 1072 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); 1073 } 1074 *enter_cond = cond; 1075 1076 base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE); 1077 if (stmts) 1078 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); 1079 bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE); 1080 if (stmts) 1081 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); 1082 1083 *exit_base = base; 1084 *exit_step = bigstep; 1085 *exit_cmp = cmp; 1086 *exit_bound = bound; 1087 } 1088 1089 /* Scales the frequencies of all basic blocks in LOOP that are strictly 1090 dominated by BB by NUM/DEN. */ 1091 1092 static void 1093 scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb, 1094 profile_count num, profile_count den) 1095 { 1096 basic_block son; 1097 1098 if (!den.nonzero_p () && !(num == profile_count::zero ())) 1099 return; 1100 1101 for (son = first_dom_son (CDI_DOMINATORS, bb); 1102 son; 1103 son = next_dom_son (CDI_DOMINATORS, son)) 1104 { 1105 if (!flow_bb_inside_loop_p (loop, son)) 1106 continue; 1107 scale_bbs_frequencies_profile_count (&son, 1, num, den); 1108 scale_dominated_blocks_in_loop (loop, son, num, den); 1109 } 1110 } 1111 1112 /* Return estimated niter for LOOP after unrolling by FACTOR times. */ 1113 1114 gcov_type 1115 niter_for_unrolled_loop (struct loop *loop, unsigned factor) 1116 { 1117 gcc_assert (factor != 0); 1118 bool profile_p = false; 1119 gcov_type est_niter = expected_loop_iterations_unbounded (loop, &profile_p); 1120 /* Note that this is really CEIL (est_niter + 1, factor) - 1, where the 1121 "+ 1" converts latch iterations to loop iterations and the "- 1" 1122 converts back. */ 1123 gcov_type new_est_niter = est_niter / factor; 1124 1125 if (est_niter == -1) 1126 return -1; 1127 1128 /* Without profile feedback, loops for which we do not know a better estimate 1129 are assumed to roll 10 times. When we unroll such loop, it appears to 1130 roll too little, and it may even seem to be cold. To avoid this, we 1131 ensure that the created loop appears to roll at least 5 times (but at 1132 most as many times as before unrolling). Don't do adjustment if profile 1133 feedback is present. */ 1134 if (new_est_niter < 5 && !profile_p) 1135 { 1136 if (est_niter < 5) 1137 new_est_niter = est_niter; 1138 else 1139 new_est_niter = 5; 1140 } 1141 1142 if (loop->any_upper_bound) 1143 { 1144 /* As above, this is really CEIL (upper_bound + 1, factor) - 1. */ 1145 widest_int bound = wi::udiv_floor (loop->nb_iterations_upper_bound, 1146 factor); 1147 if (wi::ltu_p (bound, new_est_niter)) 1148 new_est_niter = bound.to_uhwi (); 1149 } 1150 1151 return new_est_niter; 1152 } 1153 1154 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP. 1155 EXIT is the exit of the loop to that DESC corresponds. 1156 1157 If N is number of iterations of the loop and MAY_BE_ZERO is the condition 1158 under that loop exits in the first iteration even if N != 0, 1159 1160 while (1) 1161 { 1162 x = phi (init, next); 1163 1164 pre; 1165 if (st) 1166 break; 1167 post; 1168 } 1169 1170 becomes (with possibly the exit conditions formulated a bit differently, 1171 avoiding the need to create a new iv): 1172 1173 if (MAY_BE_ZERO || N < FACTOR) 1174 goto rest; 1175 1176 do 1177 { 1178 x = phi (init, next); 1179 1180 pre; 1181 post; 1182 pre; 1183 post; 1184 ... 1185 pre; 1186 post; 1187 N -= FACTOR; 1188 1189 } while (N >= FACTOR); 1190 1191 rest: 1192 init' = phi (init, x); 1193 1194 while (1) 1195 { 1196 x = phi (init', next); 1197 1198 pre; 1199 if (st) 1200 break; 1201 post; 1202 } 1203 1204 Before the loop is unrolled, TRANSFORM is called for it (only for the 1205 unrolled loop, but not for its versioned copy). DATA is passed to 1206 TRANSFORM. */ 1207 1208 /* Probability in % that the unrolled loop is entered. Just a guess. */ 1209 #define PROB_UNROLLED_LOOP_ENTERED 90 1210 1211 void 1212 tree_transform_and_unroll_loop (struct loop *loop, unsigned factor, 1213 edge exit, struct tree_niter_desc *desc, 1214 transform_callback transform, 1215 void *data) 1216 { 1217 gcond *exit_if; 1218 tree ctr_before, ctr_after; 1219 tree enter_main_cond, exit_base, exit_step, exit_bound; 1220 enum tree_code exit_cmp; 1221 gphi *phi_old_loop, *phi_new_loop, *phi_rest; 1222 gphi_iterator psi_old_loop, psi_new_loop; 1223 tree init, next, new_init; 1224 struct loop *new_loop; 1225 basic_block rest, exit_bb; 1226 edge old_entry, new_entry, old_latch, precond_edge, new_exit; 1227 edge new_nonexit, e; 1228 gimple_stmt_iterator bsi; 1229 use_operand_p op; 1230 bool ok; 1231 unsigned i; 1232 profile_probability prob, prob_entry, scale_unrolled; 1233 profile_count freq_e, freq_h; 1234 gcov_type new_est_niter = niter_for_unrolled_loop (loop, factor); 1235 unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP; 1236 auto_vec<edge> to_remove; 1237 1238 determine_exit_conditions (loop, desc, factor, 1239 &enter_main_cond, &exit_base, &exit_step, 1240 &exit_cmp, &exit_bound); 1241 1242 /* Let us assume that the unrolled loop is quite likely to be entered. */ 1243 if (integer_nonzerop (enter_main_cond)) 1244 prob_entry = profile_probability::always (); 1245 else 1246 prob_entry = profile_probability::guessed_always () 1247 .apply_scale (PROB_UNROLLED_LOOP_ENTERED, 100); 1248 1249 /* The values for scales should keep profile consistent, and somewhat close 1250 to correct. 1251 1252 TODO: The current value of SCALE_REST makes it appear that the loop that 1253 is created by splitting the remaining iterations of the unrolled loop is 1254 executed the same number of times as the original loop, and with the same 1255 frequencies, which is obviously wrong. This does not appear to cause 1256 problems, so we do not bother with fixing it for now. To make the profile 1257 correct, we would need to change the probability of the exit edge of the 1258 loop, and recompute the distribution of frequencies in its body because 1259 of this change (scale the frequencies of blocks before and after the exit 1260 by appropriate factors). */ 1261 scale_unrolled = prob_entry; 1262 1263 new_loop = loop_version (loop, enter_main_cond, NULL, prob_entry, 1264 prob_entry.invert (), scale_unrolled, 1265 profile_probability::guessed_always (), 1266 true); 1267 gcc_assert (new_loop != NULL); 1268 update_ssa (TODO_update_ssa); 1269 1270 /* Prepare the cfg and update the phi nodes. Move the loop exit to the 1271 loop latch (and make its condition dummy, for the moment). */ 1272 rest = loop_preheader_edge (new_loop)->src; 1273 precond_edge = single_pred_edge (rest); 1274 split_edge (loop_latch_edge (loop)); 1275 exit_bb = single_pred (loop->latch); 1276 1277 /* Since the exit edge will be removed, the frequency of all the blocks 1278 in the loop that are dominated by it must be scaled by 1279 1 / (1 - exit->probability). */ 1280 if (exit->probability.initialized_p ()) 1281 scale_dominated_blocks_in_loop (loop, exit->src, 1282 /* We are scaling up here so probability 1283 does not fit. */ 1284 loop->header->count, 1285 loop->header->count 1286 - loop->header->count.apply_probability 1287 (exit->probability)); 1288 1289 bsi = gsi_last_bb (exit_bb); 1290 exit_if = gimple_build_cond (EQ_EXPR, integer_zero_node, 1291 integer_zero_node, 1292 NULL_TREE, NULL_TREE); 1293 1294 gsi_insert_after (&bsi, exit_if, GSI_NEW_STMT); 1295 new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr); 1296 rescan_loop_exit (new_exit, true, false); 1297 1298 /* Set the probability of new exit to the same of the old one. Fix 1299 the frequency of the latch block, by scaling it back by 1300 1 - exit->probability. */ 1301 new_exit->probability = exit->probability; 1302 new_nonexit = single_pred_edge (loop->latch); 1303 new_nonexit->probability = exit->probability.invert (); 1304 new_nonexit->flags = EDGE_TRUE_VALUE; 1305 if (new_nonexit->probability.initialized_p ()) 1306 scale_bbs_frequencies (&loop->latch, 1, new_nonexit->probability); 1307 1308 old_entry = loop_preheader_edge (loop); 1309 new_entry = loop_preheader_edge (new_loop); 1310 old_latch = loop_latch_edge (loop); 1311 for (psi_old_loop = gsi_start_phis (loop->header), 1312 psi_new_loop = gsi_start_phis (new_loop->header); 1313 !gsi_end_p (psi_old_loop); 1314 gsi_next (&psi_old_loop), gsi_next (&psi_new_loop)) 1315 { 1316 phi_old_loop = psi_old_loop.phi (); 1317 phi_new_loop = psi_new_loop.phi (); 1318 1319 init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry); 1320 op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry); 1321 gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op))); 1322 next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch); 1323 1324 /* Prefer using original variable as a base for the new ssa name. 1325 This is necessary for virtual ops, and useful in order to avoid 1326 losing debug info for real ops. */ 1327 if (TREE_CODE (next) == SSA_NAME 1328 && useless_type_conversion_p (TREE_TYPE (next), 1329 TREE_TYPE (init))) 1330 new_init = copy_ssa_name (next); 1331 else if (TREE_CODE (init) == SSA_NAME 1332 && useless_type_conversion_p (TREE_TYPE (init), 1333 TREE_TYPE (next))) 1334 new_init = copy_ssa_name (init); 1335 else if (useless_type_conversion_p (TREE_TYPE (next), TREE_TYPE (init))) 1336 new_init = make_temp_ssa_name (TREE_TYPE (next), NULL, "unrinittmp"); 1337 else 1338 new_init = make_temp_ssa_name (TREE_TYPE (init), NULL, "unrinittmp"); 1339 1340 phi_rest = create_phi_node (new_init, rest); 1341 1342 add_phi_arg (phi_rest, init, precond_edge, UNKNOWN_LOCATION); 1343 add_phi_arg (phi_rest, next, new_exit, UNKNOWN_LOCATION); 1344 SET_USE (op, new_init); 1345 } 1346 1347 remove_path (exit); 1348 1349 /* Transform the loop. */ 1350 if (transform) 1351 (*transform) (loop, data); 1352 1353 /* Unroll the loop and remove the exits in all iterations except for the 1354 last one. */ 1355 auto_sbitmap wont_exit (factor); 1356 bitmap_ones (wont_exit); 1357 bitmap_clear_bit (wont_exit, factor - 1); 1358 1359 ok = gimple_duplicate_loop_to_header_edge 1360 (loop, loop_latch_edge (loop), factor - 1, 1361 wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ); 1362 gcc_assert (ok); 1363 1364 FOR_EACH_VEC_ELT (to_remove, i, e) 1365 { 1366 ok = remove_path (e); 1367 gcc_assert (ok); 1368 } 1369 update_ssa (TODO_update_ssa); 1370 1371 /* Ensure that the frequencies in the loop match the new estimated 1372 number of iterations, and change the probability of the new 1373 exit edge. */ 1374 1375 freq_h = loop->header->count; 1376 freq_e = (loop_preheader_edge (loop))->count (); 1377 if (freq_h.nonzero_p ()) 1378 { 1379 /* Avoid dropping loop body profile counter to 0 because of zero count 1380 in loop's preheader. */ 1381 if (freq_h.nonzero_p () && !(freq_e == profile_count::zero ())) 1382 freq_e = freq_e.force_nonzero (); 1383 scale_loop_frequencies (loop, freq_e.probability_in (freq_h)); 1384 } 1385 1386 exit_bb = single_pred (loop->latch); 1387 new_exit = find_edge (exit_bb, rest); 1388 new_exit->probability = profile_probability::always () 1389 .apply_scale (1, new_est_niter + 1); 1390 1391 rest->count += new_exit->count (); 1392 1393 new_nonexit = single_pred_edge (loop->latch); 1394 prob = new_nonexit->probability; 1395 new_nonexit->probability = new_exit->probability.invert (); 1396 prob = new_nonexit->probability / prob; 1397 if (prob.initialized_p ()) 1398 scale_bbs_frequencies (&loop->latch, 1, prob); 1399 1400 /* Finally create the new counter for number of iterations and add the new 1401 exit instruction. */ 1402 bsi = gsi_last_nondebug_bb (exit_bb); 1403 exit_if = as_a <gcond *> (gsi_stmt (bsi)); 1404 create_iv (exit_base, exit_step, NULL_TREE, loop, 1405 &bsi, false, &ctr_before, &ctr_after); 1406 gimple_cond_set_code (exit_if, exit_cmp); 1407 gimple_cond_set_lhs (exit_if, ctr_after); 1408 gimple_cond_set_rhs (exit_if, exit_bound); 1409 update_stmt (exit_if); 1410 1411 checking_verify_flow_info (); 1412 checking_verify_loop_structure (); 1413 checking_verify_loop_closed_ssa (true, loop); 1414 checking_verify_loop_closed_ssa (true, new_loop); 1415 } 1416 1417 /* Wrapper over tree_transform_and_unroll_loop for case we do not 1418 want to transform the loop before unrolling. The meaning 1419 of the arguments is the same as for tree_transform_and_unroll_loop. */ 1420 1421 void 1422 tree_unroll_loop (struct loop *loop, unsigned factor, 1423 edge exit, struct tree_niter_desc *desc) 1424 { 1425 tree_transform_and_unroll_loop (loop, factor, exit, desc, 1426 NULL, NULL); 1427 } 1428 1429 /* Rewrite the phi node at position PSI in function of the main 1430 induction variable MAIN_IV and insert the generated code at GSI. */ 1431 1432 static void 1433 rewrite_phi_with_iv (loop_p loop, 1434 gphi_iterator *psi, 1435 gimple_stmt_iterator *gsi, 1436 tree main_iv) 1437 { 1438 affine_iv iv; 1439 gassign *stmt; 1440 gphi *phi = psi->phi (); 1441 tree atype, mtype, val, res = PHI_RESULT (phi); 1442 1443 if (virtual_operand_p (res) || res == main_iv) 1444 { 1445 gsi_next (psi); 1446 return; 1447 } 1448 1449 if (!simple_iv (loop, loop, res, &iv, true)) 1450 { 1451 gsi_next (psi); 1452 return; 1453 } 1454 1455 remove_phi_node (psi, false); 1456 1457 atype = TREE_TYPE (res); 1458 mtype = POINTER_TYPE_P (atype) ? sizetype : atype; 1459 val = fold_build2 (MULT_EXPR, mtype, unshare_expr (iv.step), 1460 fold_convert (mtype, main_iv)); 1461 val = fold_build2 (POINTER_TYPE_P (atype) 1462 ? POINTER_PLUS_EXPR : PLUS_EXPR, 1463 atype, unshare_expr (iv.base), val); 1464 val = force_gimple_operand_gsi (gsi, val, false, NULL_TREE, true, 1465 GSI_SAME_STMT); 1466 stmt = gimple_build_assign (res, val); 1467 gsi_insert_before (gsi, stmt, GSI_SAME_STMT); 1468 } 1469 1470 /* Rewrite all the phi nodes of LOOP in function of the main induction 1471 variable MAIN_IV. */ 1472 1473 static void 1474 rewrite_all_phi_nodes_with_iv (loop_p loop, tree main_iv) 1475 { 1476 unsigned i; 1477 basic_block *bbs = get_loop_body_in_dom_order (loop); 1478 gphi_iterator psi; 1479 1480 for (i = 0; i < loop->num_nodes; i++) 1481 { 1482 basic_block bb = bbs[i]; 1483 gimple_stmt_iterator gsi = gsi_after_labels (bb); 1484 1485 if (bb->loop_father != loop) 1486 continue; 1487 1488 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); ) 1489 rewrite_phi_with_iv (loop, &psi, &gsi, main_iv); 1490 } 1491 1492 free (bbs); 1493 } 1494 1495 /* Bases all the induction variables in LOOP on a single induction variable 1496 (with base 0 and step 1), whose final value is compared with *NIT. When the 1497 IV type precision has to be larger than *NIT type precision, *NIT is 1498 converted to the larger type, the conversion code is inserted before the 1499 loop, and *NIT is updated to the new definition. When BUMP_IN_LATCH is true, 1500 the induction variable is incremented in the loop latch, otherwise it is 1501 incremented in the loop header. Return the induction variable that was 1502 created. */ 1503 1504 tree 1505 canonicalize_loop_ivs (struct loop *loop, tree *nit, bool bump_in_latch) 1506 { 1507 unsigned precision = TYPE_PRECISION (TREE_TYPE (*nit)); 1508 unsigned original_precision = precision; 1509 tree type, var_before; 1510 gimple_stmt_iterator gsi; 1511 gphi_iterator psi; 1512 gcond *stmt; 1513 edge exit = single_dom_exit (loop); 1514 gimple_seq stmts; 1515 bool unsigned_p = false; 1516 1517 for (psi = gsi_start_phis (loop->header); 1518 !gsi_end_p (psi); gsi_next (&psi)) 1519 { 1520 gphi *phi = psi.phi (); 1521 tree res = PHI_RESULT (phi); 1522 bool uns; 1523 1524 type = TREE_TYPE (res); 1525 if (virtual_operand_p (res) 1526 || (!INTEGRAL_TYPE_P (type) 1527 && !POINTER_TYPE_P (type)) 1528 || TYPE_PRECISION (type) < precision) 1529 continue; 1530 1531 uns = POINTER_TYPE_P (type) | TYPE_UNSIGNED (type); 1532 1533 if (TYPE_PRECISION (type) > precision) 1534 unsigned_p = uns; 1535 else 1536 unsigned_p |= uns; 1537 1538 precision = TYPE_PRECISION (type); 1539 } 1540 1541 scalar_int_mode mode = smallest_int_mode_for_size (precision); 1542 precision = GET_MODE_PRECISION (mode); 1543 type = build_nonstandard_integer_type (precision, unsigned_p); 1544 1545 if (original_precision != precision) 1546 { 1547 *nit = fold_convert (type, *nit); 1548 *nit = force_gimple_operand (*nit, &stmts, true, NULL_TREE); 1549 if (stmts) 1550 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); 1551 } 1552 1553 if (bump_in_latch) 1554 gsi = gsi_last_bb (loop->latch); 1555 else 1556 gsi = gsi_last_nondebug_bb (loop->header); 1557 create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE, 1558 loop, &gsi, bump_in_latch, &var_before, NULL); 1559 1560 rewrite_all_phi_nodes_with_iv (loop, var_before); 1561 1562 stmt = as_a <gcond *> (last_stmt (exit->src)); 1563 /* Make the loop exit if the control condition is not satisfied. */ 1564 if (exit->flags & EDGE_TRUE_VALUE) 1565 { 1566 edge te, fe; 1567 1568 extract_true_false_edges_from_block (exit->src, &te, &fe); 1569 te->flags = EDGE_FALSE_VALUE; 1570 fe->flags = EDGE_TRUE_VALUE; 1571 } 1572 gimple_cond_set_code (stmt, LT_EXPR); 1573 gimple_cond_set_lhs (stmt, var_before); 1574 gimple_cond_set_rhs (stmt, *nit); 1575 update_stmt (stmt); 1576 1577 return var_before; 1578 } 1579