1/* Optimization of PHI nodes by converting them into straightline code. 2 Copyright (C) 2004-2015 Free Software Foundation, Inc. 3 4This file is part of GCC. 5 6GCC is free software; you can redistribute it and/or modify it 7under the terms of the GNU General Public License as published by the 8Free Software Foundation; either version 3, or (at your option) any 9later version. 10 11GCC is distributed in the hope that it will be useful, but WITHOUT 12ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14for more details. 15 16You should have received a copy of the GNU General Public License 17along 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 "hash-table.h" 24#include "tm.h" 25#include "hash-set.h" 26#include "machmode.h" 27#include "vec.h" 28#include "double-int.h" 29#include "input.h" 30#include "alias.h" 31#include "symtab.h" 32#include "wide-int.h" 33#include "inchash.h" 34#include "tree.h" 35#include "fold-const.h" 36#include "stor-layout.h" 37#include "flags.h" 38#include "tm_p.h" 39#include "predict.h" 40#include "hard-reg-set.h" 41#include "function.h" 42#include "dominance.h" 43#include "cfg.h" 44#include "cfganal.h" 45#include "basic-block.h" 46#include "tree-ssa-alias.h" 47#include "internal-fn.h" 48#include "gimple-expr.h" 49#include "is-a.h" 50#include "gimple.h" 51#include "gimplify.h" 52#include "gimple-iterator.h" 53#include "gimplify-me.h" 54#include "gimple-ssa.h" 55#include "tree-cfg.h" 56#include "tree-phinodes.h" 57#include "ssa-iterators.h" 58#include "stringpool.h" 59#include "tree-ssanames.h" 60#include "hashtab.h" 61#include "rtl.h" 62#include "statistics.h" 63#include "real.h" 64#include "fixed-value.h" 65#include "insn-config.h" 66#include "expmed.h" 67#include "dojump.h" 68#include "explow.h" 69#include "calls.h" 70#include "emit-rtl.h" 71#include "varasm.h" 72#include "stmt.h" 73#include "expr.h" 74#include "tree-dfa.h" 75#include "tree-pass.h" 76#include "langhooks.h" 77#include "domwalk.h" 78#include "cfgloop.h" 79#include "tree-data-ref.h" 80#include "gimple-pretty-print.h" 81#include "insn-codes.h" 82#include "optabs.h" 83#include "tree-scalar-evolution.h" 84#include "tree-inline.h" 85 86#ifndef HAVE_conditional_move 87#define HAVE_conditional_move (0) 88#endif 89 90static unsigned int tree_ssa_phiopt_worker (bool, bool); 91static bool conditional_replacement (basic_block, basic_block, 92 edge, edge, gphi *, tree, tree); 93static int value_replacement (basic_block, basic_block, 94 edge, edge, gimple, tree, tree); 95static bool minmax_replacement (basic_block, basic_block, 96 edge, edge, gimple, tree, tree); 97static bool abs_replacement (basic_block, basic_block, 98 edge, edge, gimple, tree, tree); 99static bool cond_store_replacement (basic_block, basic_block, edge, edge, 100 hash_set<tree> *); 101static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block); 102static hash_set<tree> * get_non_trapping (); 103static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree); 104static void hoist_adjacent_loads (basic_block, basic_block, 105 basic_block, basic_block); 106static bool gate_hoist_loads (void); 107 108/* This pass tries to transform conditional stores into unconditional 109 ones, enabling further simplifications with the simpler then and else 110 blocks. In particular it replaces this: 111 112 bb0: 113 if (cond) goto bb2; else goto bb1; 114 bb1: 115 *p = RHS; 116 bb2: 117 118 with 119 120 bb0: 121 if (cond) goto bb1; else goto bb2; 122 bb1: 123 condtmp' = *p; 124 bb2: 125 condtmp = PHI <RHS, condtmp'> 126 *p = condtmp; 127 128 This transformation can only be done under several constraints, 129 documented below. It also replaces: 130 131 bb0: 132 if (cond) goto bb2; else goto bb1; 133 bb1: 134 *p = RHS1; 135 goto bb3; 136 bb2: 137 *p = RHS2; 138 bb3: 139 140 with 141 142 bb0: 143 if (cond) goto bb3; else goto bb1; 144 bb1: 145 bb3: 146 condtmp = PHI <RHS1, RHS2> 147 *p = condtmp; */ 148 149static unsigned int 150tree_ssa_cs_elim (void) 151{ 152 unsigned todo; 153 /* ??? We are not interested in loop related info, but the following 154 will create it, ICEing as we didn't init loops with pre-headers. 155 An interfacing issue of find_data_references_in_bb. */ 156 loop_optimizer_init (LOOPS_NORMAL); 157 scev_initialize (); 158 todo = tree_ssa_phiopt_worker (true, false); 159 scev_finalize (); 160 loop_optimizer_finalize (); 161 return todo; 162} 163 164/* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */ 165 166static gphi * 167single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1) 168{ 169 gimple_stmt_iterator i; 170 gphi *phi = NULL; 171 if (gimple_seq_singleton_p (seq)) 172 return as_a <gphi *> (gsi_stmt (gsi_start (seq))); 173 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i)) 174 { 175 gphi *p = as_a <gphi *> (gsi_stmt (i)); 176 /* If the PHI arguments are equal then we can skip this PHI. */ 177 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx), 178 gimple_phi_arg_def (p, e1->dest_idx))) 179 continue; 180 181 /* If we already have a PHI that has the two edge arguments are 182 different, then return it is not a singleton for these PHIs. */ 183 if (phi) 184 return NULL; 185 186 phi = p; 187 } 188 return phi; 189} 190 191/* The core routine of conditional store replacement and normal 192 phi optimizations. Both share much of the infrastructure in how 193 to match applicable basic block patterns. DO_STORE_ELIM is true 194 when we want to do conditional store replacement, false otherwise. 195 DO_HOIST_LOADS is true when we want to hoist adjacent loads out 196 of diamond control flow patterns, false otherwise. */ 197static unsigned int 198tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads) 199{ 200 basic_block bb; 201 basic_block *bb_order; 202 unsigned n, i; 203 bool cfgchanged = false; 204 hash_set<tree> *nontrap = 0; 205 206 if (do_store_elim) 207 /* Calculate the set of non-trapping memory accesses. */ 208 nontrap = get_non_trapping (); 209 210 /* Search every basic block for COND_EXPR we may be able to optimize. 211 212 We walk the blocks in order that guarantees that a block with 213 a single predecessor is processed before the predecessor. 214 This ensures that we collapse inner ifs before visiting the 215 outer ones, and also that we do not try to visit a removed 216 block. */ 217 bb_order = single_pred_before_succ_order (); 218 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS; 219 220 for (i = 0; i < n; i++) 221 { 222 gimple cond_stmt; 223 gphi *phi; 224 basic_block bb1, bb2; 225 edge e1, e2; 226 tree arg0, arg1; 227 228 bb = bb_order[i]; 229 230 cond_stmt = last_stmt (bb); 231 /* Check to see if the last statement is a GIMPLE_COND. */ 232 if (!cond_stmt 233 || gimple_code (cond_stmt) != GIMPLE_COND) 234 continue; 235 236 e1 = EDGE_SUCC (bb, 0); 237 bb1 = e1->dest; 238 e2 = EDGE_SUCC (bb, 1); 239 bb2 = e2->dest; 240 241 /* We cannot do the optimization on abnormal edges. */ 242 if ((e1->flags & EDGE_ABNORMAL) != 0 243 || (e2->flags & EDGE_ABNORMAL) != 0) 244 continue; 245 246 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */ 247 if (EDGE_COUNT (bb1->succs) == 0 248 || bb2 == NULL 249 || EDGE_COUNT (bb2->succs) == 0) 250 continue; 251 252 /* Find the bb which is the fall through to the other. */ 253 if (EDGE_SUCC (bb1, 0)->dest == bb2) 254 ; 255 else if (EDGE_SUCC (bb2, 0)->dest == bb1) 256 { 257 basic_block bb_tmp = bb1; 258 edge e_tmp = e1; 259 bb1 = bb2; 260 bb2 = bb_tmp; 261 e1 = e2; 262 e2 = e_tmp; 263 } 264 else if (do_store_elim 265 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) 266 { 267 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; 268 269 if (!single_succ_p (bb1) 270 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0 271 || !single_succ_p (bb2) 272 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0 273 || EDGE_COUNT (bb3->preds) != 2) 274 continue; 275 if (cond_if_else_store_replacement (bb1, bb2, bb3)) 276 cfgchanged = true; 277 continue; 278 } 279 else if (do_hoist_loads 280 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) 281 { 282 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; 283 284 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt))) 285 && single_succ_p (bb1) 286 && single_succ_p (bb2) 287 && single_pred_p (bb1) 288 && single_pred_p (bb2) 289 && EDGE_COUNT (bb->succs) == 2 290 && EDGE_COUNT (bb3->preds) == 2 291 /* If one edge or the other is dominant, a conditional move 292 is likely to perform worse than the well-predicted branch. */ 293 && !predictable_edge_p (EDGE_SUCC (bb, 0)) 294 && !predictable_edge_p (EDGE_SUCC (bb, 1))) 295 hoist_adjacent_loads (bb, bb1, bb2, bb3); 296 continue; 297 } 298 else 299 continue; 300 301 e1 = EDGE_SUCC (bb1, 0); 302 303 /* Make sure that bb1 is just a fall through. */ 304 if (!single_succ_p (bb1) 305 || (e1->flags & EDGE_FALLTHRU) == 0) 306 continue; 307 308 /* Also make sure that bb1 only have one predecessor and that it 309 is bb. */ 310 if (!single_pred_p (bb1) 311 || single_pred (bb1) != bb) 312 continue; 313 314 if (do_store_elim) 315 { 316 /* bb1 is the middle block, bb2 the join block, bb the split block, 317 e1 the fallthrough edge from bb1 to bb2. We can't do the 318 optimization if the join block has more than two predecessors. */ 319 if (EDGE_COUNT (bb2->preds) > 2) 320 continue; 321 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap)) 322 cfgchanged = true; 323 } 324 else 325 { 326 gimple_seq phis = phi_nodes (bb2); 327 gimple_stmt_iterator gsi; 328 bool candorest = true; 329 330 /* Value replacement can work with more than one PHI 331 so try that first. */ 332 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) 333 { 334 phi = as_a <gphi *> (gsi_stmt (gsi)); 335 arg0 = gimple_phi_arg_def (phi, e1->dest_idx); 336 arg1 = gimple_phi_arg_def (phi, e2->dest_idx); 337 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2) 338 { 339 candorest = false; 340 cfgchanged = true; 341 break; 342 } 343 } 344 345 if (!candorest) 346 continue; 347 348 phi = single_non_singleton_phi_for_edges (phis, e1, e2); 349 if (!phi) 350 continue; 351 352 arg0 = gimple_phi_arg_def (phi, e1->dest_idx); 353 arg1 = gimple_phi_arg_def (phi, e2->dest_idx); 354 355 /* Something is wrong if we cannot find the arguments in the PHI 356 node. */ 357 gcc_assert (arg0 != NULL && arg1 != NULL); 358 359 /* Do the replacement of conditional if it can be done. */ 360 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) 361 cfgchanged = true; 362 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) 363 cfgchanged = true; 364 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) 365 cfgchanged = true; 366 } 367 } 368 369 free (bb_order); 370 371 if (do_store_elim) 372 delete nontrap; 373 /* If the CFG has changed, we should cleanup the CFG. */ 374 if (cfgchanged && do_store_elim) 375 { 376 /* In cond-store replacement we have added some loads on edges 377 and new VOPS (as we moved the store, and created a load). */ 378 gsi_commit_edge_inserts (); 379 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals; 380 } 381 else if (cfgchanged) 382 return TODO_cleanup_cfg; 383 return 0; 384} 385 386/* Replace PHI node element whose edge is E in block BB with variable NEW. 387 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK 388 is known to have two edges, one of which must reach BB). */ 389 390static void 391replace_phi_edge_with_variable (basic_block cond_block, 392 edge e, gimple phi, tree new_tree) 393{ 394 basic_block bb = gimple_bb (phi); 395 basic_block block_to_remove; 396 gimple_stmt_iterator gsi; 397 398 /* Change the PHI argument to new. */ 399 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree); 400 401 /* Remove the empty basic block. */ 402 if (EDGE_SUCC (cond_block, 0)->dest == bb) 403 { 404 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU; 405 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); 406 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE; 407 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count; 408 409 block_to_remove = EDGE_SUCC (cond_block, 1)->dest; 410 } 411 else 412 { 413 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU; 414 EDGE_SUCC (cond_block, 1)->flags 415 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); 416 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE; 417 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count; 418 419 block_to_remove = EDGE_SUCC (cond_block, 0)->dest; 420 } 421 delete_basic_block (block_to_remove); 422 423 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */ 424 gsi = gsi_last_bb (cond_block); 425 gsi_remove (&gsi, true); 426 427 if (dump_file && (dump_flags & TDF_DETAILS)) 428 fprintf (dump_file, 429 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n", 430 cond_block->index, 431 bb->index); 432} 433 434/* The function conditional_replacement does the main work of doing the 435 conditional replacement. Return true if the replacement is done. 436 Otherwise return false. 437 BB is the basic block where the replacement is going to be done on. ARG0 438 is argument 0 from PHI. Likewise for ARG1. */ 439 440static bool 441conditional_replacement (basic_block cond_bb, basic_block middle_bb, 442 edge e0, edge e1, gphi *phi, 443 tree arg0, tree arg1) 444{ 445 tree result; 446 gimple stmt; 447 gassign *new_stmt; 448 tree cond; 449 gimple_stmt_iterator gsi; 450 edge true_edge, false_edge; 451 tree new_var, new_var2; 452 bool neg; 453 454 /* FIXME: Gimplification of complex type is too hard for now. */ 455 /* We aren't prepared to handle vectors either (and it is a question 456 if it would be worthwhile anyway). */ 457 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0)) 458 || POINTER_TYPE_P (TREE_TYPE (arg0))) 459 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1)) 460 || POINTER_TYPE_P (TREE_TYPE (arg1)))) 461 return false; 462 463 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then 464 convert it to the conditional. */ 465 if ((integer_zerop (arg0) && integer_onep (arg1)) 466 || (integer_zerop (arg1) && integer_onep (arg0))) 467 neg = false; 468 else if ((integer_zerop (arg0) && integer_all_onesp (arg1)) 469 || (integer_zerop (arg1) && integer_all_onesp (arg0))) 470 neg = true; 471 else 472 return false; 473 474 if (!empty_block_p (middle_bb)) 475 return false; 476 477 /* At this point we know we have a GIMPLE_COND with two successors. 478 One successor is BB, the other successor is an empty block which 479 falls through into BB. 480 481 There is a single PHI node at the join point (BB) and its arguments 482 are constants (0, 1) or (0, -1). 483 484 So, given the condition COND, and the two PHI arguments, we can 485 rewrite this PHI into non-branching code: 486 487 dest = (COND) or dest = COND' 488 489 We use the condition as-is if the argument associated with the 490 true edge has the value one or the argument associated with the 491 false edge as the value zero. Note that those conditions are not 492 the same since only one of the outgoing edges from the GIMPLE_COND 493 will directly reach BB and thus be associated with an argument. */ 494 495 stmt = last_stmt (cond_bb); 496 result = PHI_RESULT (phi); 497 498 /* To handle special cases like floating point comparison, it is easier and 499 less error-prone to build a tree and gimplify it on the fly though it is 500 less efficient. */ 501 cond = fold_build2_loc (gimple_location (stmt), 502 gimple_cond_code (stmt), boolean_type_node, 503 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt)); 504 505 /* We need to know which is the true edge and which is the false 506 edge so that we know when to invert the condition below. */ 507 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); 508 if ((e0 == true_edge && integer_zerop (arg0)) 509 || (e0 == false_edge && !integer_zerop (arg0)) 510 || (e1 == true_edge && integer_zerop (arg1)) 511 || (e1 == false_edge && !integer_zerop (arg1))) 512 cond = fold_build1_loc (gimple_location (stmt), 513 TRUTH_NOT_EXPR, TREE_TYPE (cond), cond); 514 515 if (neg) 516 { 517 cond = fold_convert_loc (gimple_location (stmt), 518 TREE_TYPE (result), cond); 519 cond = fold_build1_loc (gimple_location (stmt), 520 NEGATE_EXPR, TREE_TYPE (cond), cond); 521 } 522 523 /* Insert our new statements at the end of conditional block before the 524 COND_STMT. */ 525 gsi = gsi_for_stmt (stmt); 526 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true, 527 GSI_SAME_STMT); 528 529 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var))) 530 { 531 source_location locus_0, locus_1; 532 533 new_var2 = make_ssa_name (TREE_TYPE (result)); 534 new_stmt = gimple_build_assign (new_var2, CONVERT_EXPR, new_var); 535 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); 536 new_var = new_var2; 537 538 /* Set the locus to the first argument, unless is doesn't have one. */ 539 locus_0 = gimple_phi_arg_location (phi, 0); 540 locus_1 = gimple_phi_arg_location (phi, 1); 541 if (locus_0 == UNKNOWN_LOCATION) 542 locus_0 = locus_1; 543 gimple_set_location (new_stmt, locus_0); 544 } 545 546 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var); 547 reset_flow_sensitive_info_in_bb (cond_bb); 548 549 /* Note that we optimized this PHI. */ 550 return true; 551} 552 553/* Update *ARG which is defined in STMT so that it contains the 554 computed value if that seems profitable. Return true if the 555 statement is made dead by that rewriting. */ 556 557static bool 558jump_function_from_stmt (tree *arg, gimple stmt) 559{ 560 enum tree_code code = gimple_assign_rhs_code (stmt); 561 if (code == ADDR_EXPR) 562 { 563 /* For arg = &p->i transform it to p, if possible. */ 564 tree rhs1 = gimple_assign_rhs1 (stmt); 565 HOST_WIDE_INT offset; 566 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0), 567 &offset); 568 if (tem 569 && TREE_CODE (tem) == MEM_REF 570 && (mem_ref_offset (tem) + offset) == 0) 571 { 572 *arg = TREE_OPERAND (tem, 0); 573 return true; 574 } 575 } 576 /* TODO: Much like IPA-CP jump-functions we want to handle constant 577 additions symbolically here, and we'd need to update the comparison 578 code that compares the arg + cst tuples in our caller. For now the 579 code above exactly handles the VEC_BASE pattern from vec.h. */ 580 return false; 581} 582 583/* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional 584 of the form SSA_NAME NE 0. 585 586 If RHS is fed by a simple EQ_EXPR comparison of two values, see if 587 the two input values of the EQ_EXPR match arg0 and arg1. 588 589 If so update *code and return TRUE. Otherwise return FALSE. */ 590 591static bool 592rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1, 593 enum tree_code *code, const_tree rhs) 594{ 595 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining 596 statement. */ 597 if (TREE_CODE (rhs) == SSA_NAME) 598 { 599 gimple def1 = SSA_NAME_DEF_STMT (rhs); 600 601 /* Verify the defining statement has an EQ_EXPR on the RHS. */ 602 if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR) 603 { 604 /* Finally verify the source operands of the EQ_EXPR are equal 605 to arg0 and arg1. */ 606 tree op0 = gimple_assign_rhs1 (def1); 607 tree op1 = gimple_assign_rhs2 (def1); 608 if ((operand_equal_for_phi_arg_p (arg0, op0) 609 && operand_equal_for_phi_arg_p (arg1, op1)) 610 || (operand_equal_for_phi_arg_p (arg0, op1) 611 && operand_equal_for_phi_arg_p (arg1, op0))) 612 { 613 /* We will perform the optimization. */ 614 *code = gimple_assign_rhs_code (def1); 615 return true; 616 } 617 } 618 } 619 return false; 620} 621 622/* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND. 623 624 Also return TRUE if arg0/arg1 are equal to the source arguments of a 625 an EQ comparison feeding a BIT_AND_EXPR which feeds COND. 626 627 Return FALSE otherwise. */ 628 629static bool 630operand_equal_for_value_replacement (const_tree arg0, const_tree arg1, 631 enum tree_code *code, gimple cond) 632{ 633 gimple def; 634 tree lhs = gimple_cond_lhs (cond); 635 tree rhs = gimple_cond_rhs (cond); 636 637 if ((operand_equal_for_phi_arg_p (arg0, lhs) 638 && operand_equal_for_phi_arg_p (arg1, rhs)) 639 || (operand_equal_for_phi_arg_p (arg1, lhs) 640 && operand_equal_for_phi_arg_p (arg0, rhs))) 641 return true; 642 643 /* Now handle more complex case where we have an EQ comparison 644 which feeds a BIT_AND_EXPR which feeds COND. 645 646 First verify that COND is of the form SSA_NAME NE 0. */ 647 if (*code != NE_EXPR || !integer_zerop (rhs) 648 || TREE_CODE (lhs) != SSA_NAME) 649 return false; 650 651 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */ 652 def = SSA_NAME_DEF_STMT (lhs); 653 if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR) 654 return false; 655 656 /* Now verify arg0/arg1 correspond to the source arguments of an 657 EQ comparison feeding the BIT_AND_EXPR. */ 658 659 tree tmp = gimple_assign_rhs1 (def); 660 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp)) 661 return true; 662 663 tmp = gimple_assign_rhs2 (def); 664 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp)) 665 return true; 666 667 return false; 668} 669 670/* Returns true if ARG is a neutral element for operation CODE 671 on the RIGHT side. */ 672 673static bool 674neutral_element_p (tree_code code, tree arg, bool right) 675{ 676 switch (code) 677 { 678 case PLUS_EXPR: 679 case BIT_IOR_EXPR: 680 case BIT_XOR_EXPR: 681 return integer_zerop (arg); 682 683 case LROTATE_EXPR: 684 case RROTATE_EXPR: 685 case LSHIFT_EXPR: 686 case RSHIFT_EXPR: 687 case MINUS_EXPR: 688 case POINTER_PLUS_EXPR: 689 return right && integer_zerop (arg); 690 691 case MULT_EXPR: 692 return integer_onep (arg); 693 694 case TRUNC_DIV_EXPR: 695 case CEIL_DIV_EXPR: 696 case FLOOR_DIV_EXPR: 697 case ROUND_DIV_EXPR: 698 case EXACT_DIV_EXPR: 699 return right && integer_onep (arg); 700 701 case BIT_AND_EXPR: 702 return integer_all_onesp (arg); 703 704 default: 705 return false; 706 } 707} 708 709/* Returns true if ARG is an absorbing element for operation CODE. */ 710 711static bool 712absorbing_element_p (tree_code code, tree arg) 713{ 714 switch (code) 715 { 716 case BIT_IOR_EXPR: 717 return integer_all_onesp (arg); 718 719 case MULT_EXPR: 720 case BIT_AND_EXPR: 721 return integer_zerop (arg); 722 723 default: 724 return false; 725 } 726} 727 728/* The function value_replacement does the main work of doing the value 729 replacement. Return non-zero if the replacement is done. Otherwise return 730 0. If we remove the middle basic block, return 2. 731 BB is the basic block where the replacement is going to be done on. ARG0 732 is argument 0 from the PHI. Likewise for ARG1. */ 733 734static int 735value_replacement (basic_block cond_bb, basic_block middle_bb, 736 edge e0, edge e1, gimple phi, 737 tree arg0, tree arg1) 738{ 739 gimple_stmt_iterator gsi; 740 gimple cond; 741 edge true_edge, false_edge; 742 enum tree_code code; 743 bool emtpy_or_with_defined_p = true; 744 745 /* If the type says honor signed zeros we cannot do this 746 optimization. */ 747 if (HONOR_SIGNED_ZEROS (arg1)) 748 return 0; 749 750 /* If there is a statement in MIDDLE_BB that defines one of the PHI 751 arguments, then adjust arg0 or arg1. */ 752 gsi = gsi_start_nondebug_after_labels_bb (middle_bb); 753 while (!gsi_end_p (gsi)) 754 { 755 gimple stmt = gsi_stmt (gsi); 756 tree lhs; 757 gsi_next_nondebug (&gsi); 758 if (!is_gimple_assign (stmt)) 759 { 760 emtpy_or_with_defined_p = false; 761 continue; 762 } 763 /* Now try to adjust arg0 or arg1 according to the computation 764 in the statement. */ 765 lhs = gimple_assign_lhs (stmt); 766 if (!(lhs == arg0 767 && jump_function_from_stmt (&arg0, stmt)) 768 || (lhs == arg1 769 && jump_function_from_stmt (&arg1, stmt))) 770 emtpy_or_with_defined_p = false; 771 } 772 773 cond = last_stmt (cond_bb); 774 code = gimple_cond_code (cond); 775 776 /* This transformation is only valid for equality comparisons. */ 777 if (code != NE_EXPR && code != EQ_EXPR) 778 return 0; 779 780 /* We need to know which is the true edge and which is the false 781 edge so that we know if have abs or negative abs. */ 782 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); 783 784 /* At this point we know we have a COND_EXPR with two successors. 785 One successor is BB, the other successor is an empty block which 786 falls through into BB. 787 788 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR. 789 790 There is a single PHI node at the join point (BB) with two arguments. 791 792 We now need to verify that the two arguments in the PHI node match 793 the two arguments to the equality comparison. */ 794 795 if (operand_equal_for_value_replacement (arg0, arg1, &code, cond)) 796 { 797 edge e; 798 tree arg; 799 800 /* For NE_EXPR, we want to build an assignment result = arg where 801 arg is the PHI argument associated with the true edge. For 802 EQ_EXPR we want the PHI argument associated with the false edge. */ 803 e = (code == NE_EXPR ? true_edge : false_edge); 804 805 /* Unfortunately, E may not reach BB (it may instead have gone to 806 OTHER_BLOCK). If that is the case, then we want the single outgoing 807 edge from OTHER_BLOCK which reaches BB and represents the desired 808 path from COND_BLOCK. */ 809 if (e->dest == middle_bb) 810 e = single_succ_edge (e->dest); 811 812 /* Now we know the incoming edge to BB that has the argument for the 813 RHS of our new assignment statement. */ 814 if (e0 == e) 815 arg = arg0; 816 else 817 arg = arg1; 818 819 /* If the middle basic block was empty or is defining the 820 PHI arguments and this is a single phi where the args are different 821 for the edges e0 and e1 then we can remove the middle basic block. */ 822 if (emtpy_or_with_defined_p 823 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), 824 e0, e1) == phi) 825 { 826 replace_phi_edge_with_variable (cond_bb, e1, phi, arg); 827 /* Note that we optimized this PHI. */ 828 return 2; 829 } 830 else 831 { 832 /* Replace the PHI arguments with arg. */ 833 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg); 834 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg); 835 if (dump_file && (dump_flags & TDF_DETAILS)) 836 { 837 fprintf (dump_file, "PHI "); 838 print_generic_expr (dump_file, gimple_phi_result (phi), 0); 839 fprintf (dump_file, " reduced for COND_EXPR in block %d to ", 840 cond_bb->index); 841 print_generic_expr (dump_file, arg, 0); 842 fprintf (dump_file, ".\n"); 843 } 844 return 1; 845 } 846 847 } 848 849 /* Now optimize (x != 0) ? x + y : y to just y. 850 The following condition is too restrictive, there can easily be another 851 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */ 852 gimple assign = last_and_only_stmt (middle_bb); 853 if (!assign || gimple_code (assign) != GIMPLE_ASSIGN 854 || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS 855 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)) 856 && !POINTER_TYPE_P (TREE_TYPE (arg0)))) 857 return 0; 858 859 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */ 860 if (!gimple_seq_empty_p (phi_nodes (middle_bb))) 861 return 0; 862 863 /* Only transform if it removes the condition. */ 864 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1)) 865 return 0; 866 867 /* Size-wise, this is always profitable. */ 868 if (optimize_bb_for_speed_p (cond_bb) 869 /* The special case is useless if it has a low probability. */ 870 && profile_status_for_fn (cfun) != PROFILE_ABSENT 871 && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN 872 /* If assign is cheap, there is no point avoiding it. */ 873 && estimate_num_insns (assign, &eni_time_weights) 874 >= 3 * estimate_num_insns (cond, &eni_time_weights)) 875 return 0; 876 877 tree lhs = gimple_assign_lhs (assign); 878 tree rhs1 = gimple_assign_rhs1 (assign); 879 tree rhs2 = gimple_assign_rhs2 (assign); 880 enum tree_code code_def = gimple_assign_rhs_code (assign); 881 tree cond_lhs = gimple_cond_lhs (cond); 882 tree cond_rhs = gimple_cond_rhs (cond); 883 884 if (((code == NE_EXPR && e1 == false_edge) 885 || (code == EQ_EXPR && e1 == true_edge)) 886 && arg0 == lhs 887 && ((arg1 == rhs1 888 && operand_equal_for_phi_arg_p (rhs2, cond_lhs) 889 && neutral_element_p (code_def, cond_rhs, true)) 890 || (arg1 == rhs2 891 && operand_equal_for_phi_arg_p (rhs1, cond_lhs) 892 && neutral_element_p (code_def, cond_rhs, false)) 893 || (operand_equal_for_phi_arg_p (arg1, cond_rhs) 894 && (operand_equal_for_phi_arg_p (rhs2, cond_lhs) 895 || operand_equal_for_phi_arg_p (rhs1, cond_lhs)) 896 && absorbing_element_p (code_def, cond_rhs)))) 897 { 898 gsi = gsi_for_stmt (cond); 899 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))) 900 { 901 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6 902 def-stmt in: 903 if (n_5 != 0) 904 goto <bb 3>; 905 else 906 goto <bb 4>; 907 908 <bb 3>: 909 # RANGE [0, 4294967294] 910 u_6 = n_5 + 4294967295; 911 912 <bb 4>: 913 # u_3 = PHI <u_6(3), 4294967295(2)> */ 914 SSA_NAME_RANGE_INFO (lhs) = NULL; 915 SSA_NAME_ANTI_RANGE_P (lhs) = 0; 916 /* If available, we can use VR of phi result at least. */ 917 tree phires = gimple_phi_result (phi); 918 struct range_info_def *phires_range_info 919 = SSA_NAME_RANGE_INFO (phires); 920 if (phires_range_info) 921 duplicate_ssa_name_range_info (lhs, SSA_NAME_RANGE_TYPE (phires), 922 phires_range_info); 923 } 924 gimple_stmt_iterator gsi_from = gsi_for_stmt (assign); 925 gsi_move_before (&gsi_from, &gsi); 926 replace_phi_edge_with_variable (cond_bb, e1, phi, lhs); 927 return 2; 928 } 929 930 return 0; 931} 932 933/* The function minmax_replacement does the main work of doing the minmax 934 replacement. Return true if the replacement is done. Otherwise return 935 false. 936 BB is the basic block where the replacement is going to be done on. ARG0 937 is argument 0 from the PHI. Likewise for ARG1. */ 938 939static bool 940minmax_replacement (basic_block cond_bb, basic_block middle_bb, 941 edge e0, edge e1, gimple phi, 942 tree arg0, tree arg1) 943{ 944 tree result, type; 945 gcond *cond; 946 gassign *new_stmt; 947 edge true_edge, false_edge; 948 enum tree_code cmp, minmax, ass_code; 949 tree smaller, larger, arg_true, arg_false; 950 gimple_stmt_iterator gsi, gsi_from; 951 952 type = TREE_TYPE (PHI_RESULT (phi)); 953 954 /* The optimization may be unsafe due to NaNs. */ 955 if (HONOR_NANS (type)) 956 return false; 957 958 cond = as_a <gcond *> (last_stmt (cond_bb)); 959 cmp = gimple_cond_code (cond); 960 961 /* This transformation is only valid for order comparisons. Record which 962 operand is smaller/larger if the result of the comparison is true. */ 963 if (cmp == LT_EXPR || cmp == LE_EXPR) 964 { 965 smaller = gimple_cond_lhs (cond); 966 larger = gimple_cond_rhs (cond); 967 } 968 else if (cmp == GT_EXPR || cmp == GE_EXPR) 969 { 970 smaller = gimple_cond_rhs (cond); 971 larger = gimple_cond_lhs (cond); 972 } 973 else 974 return false; 975 976 /* We need to know which is the true edge and which is the false 977 edge so that we know if have abs or negative abs. */ 978 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); 979 980 /* Forward the edges over the middle basic block. */ 981 if (true_edge->dest == middle_bb) 982 true_edge = EDGE_SUCC (true_edge->dest, 0); 983 if (false_edge->dest == middle_bb) 984 false_edge = EDGE_SUCC (false_edge->dest, 0); 985 986 if (true_edge == e0) 987 { 988 gcc_assert (false_edge == e1); 989 arg_true = arg0; 990 arg_false = arg1; 991 } 992 else 993 { 994 gcc_assert (false_edge == e0); 995 gcc_assert (true_edge == e1); 996 arg_true = arg1; 997 arg_false = arg0; 998 } 999 1000 if (empty_block_p (middle_bb)) 1001 { 1002 if (operand_equal_for_phi_arg_p (arg_true, smaller) 1003 && operand_equal_for_phi_arg_p (arg_false, larger)) 1004 { 1005 /* Case 1006 1007 if (smaller < larger) 1008 rslt = smaller; 1009 else 1010 rslt = larger; */ 1011 minmax = MIN_EXPR; 1012 } 1013 else if (operand_equal_for_phi_arg_p (arg_false, smaller) 1014 && operand_equal_for_phi_arg_p (arg_true, larger)) 1015 minmax = MAX_EXPR; 1016 else 1017 return false; 1018 } 1019 else 1020 { 1021 /* Recognize the following case, assuming d <= u: 1022 1023 if (a <= u) 1024 b = MAX (a, d); 1025 x = PHI <b, u> 1026 1027 This is equivalent to 1028 1029 b = MAX (a, d); 1030 x = MIN (b, u); */ 1031 1032 gimple assign = last_and_only_stmt (middle_bb); 1033 tree lhs, op0, op1, bound; 1034 1035 if (!assign 1036 || gimple_code (assign) != GIMPLE_ASSIGN) 1037 return false; 1038 1039 lhs = gimple_assign_lhs (assign); 1040 ass_code = gimple_assign_rhs_code (assign); 1041 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR) 1042 return false; 1043 op0 = gimple_assign_rhs1 (assign); 1044 op1 = gimple_assign_rhs2 (assign); 1045 1046 if (true_edge->src == middle_bb) 1047 { 1048 /* We got here if the condition is true, i.e., SMALLER < LARGER. */ 1049 if (!operand_equal_for_phi_arg_p (lhs, arg_true)) 1050 return false; 1051 1052 if (operand_equal_for_phi_arg_p (arg_false, larger)) 1053 { 1054 /* Case 1055 1056 if (smaller < larger) 1057 { 1058 r' = MAX_EXPR (smaller, bound) 1059 } 1060 r = PHI <r', larger> --> to be turned to MIN_EXPR. */ 1061 if (ass_code != MAX_EXPR) 1062 return false; 1063 1064 minmax = MIN_EXPR; 1065 if (operand_equal_for_phi_arg_p (op0, smaller)) 1066 bound = op1; 1067 else if (operand_equal_for_phi_arg_p (op1, smaller)) 1068 bound = op0; 1069 else 1070 return false; 1071 1072 /* We need BOUND <= LARGER. */ 1073 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, 1074 bound, larger))) 1075 return false; 1076 } 1077 else if (operand_equal_for_phi_arg_p (arg_false, smaller)) 1078 { 1079 /* Case 1080 1081 if (smaller < larger) 1082 { 1083 r' = MIN_EXPR (larger, bound) 1084 } 1085 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */ 1086 if (ass_code != MIN_EXPR) 1087 return false; 1088 1089 minmax = MAX_EXPR; 1090 if (operand_equal_for_phi_arg_p (op0, larger)) 1091 bound = op1; 1092 else if (operand_equal_for_phi_arg_p (op1, larger)) 1093 bound = op0; 1094 else 1095 return false; 1096 1097 /* We need BOUND >= SMALLER. */ 1098 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, 1099 bound, smaller))) 1100 return false; 1101 } 1102 else 1103 return false; 1104 } 1105 else 1106 { 1107 /* We got here if the condition is false, i.e., SMALLER > LARGER. */ 1108 if (!operand_equal_for_phi_arg_p (lhs, arg_false)) 1109 return false; 1110 1111 if (operand_equal_for_phi_arg_p (arg_true, larger)) 1112 { 1113 /* Case 1114 1115 if (smaller > larger) 1116 { 1117 r' = MIN_EXPR (smaller, bound) 1118 } 1119 r = PHI <r', larger> --> to be turned to MAX_EXPR. */ 1120 if (ass_code != MIN_EXPR) 1121 return false; 1122 1123 minmax = MAX_EXPR; 1124 if (operand_equal_for_phi_arg_p (op0, smaller)) 1125 bound = op1; 1126 else if (operand_equal_for_phi_arg_p (op1, smaller)) 1127 bound = op0; 1128 else 1129 return false; 1130 1131 /* We need BOUND >= LARGER. */ 1132 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, 1133 bound, larger))) 1134 return false; 1135 } 1136 else if (operand_equal_for_phi_arg_p (arg_true, smaller)) 1137 { 1138 /* Case 1139 1140 if (smaller > larger) 1141 { 1142 r' = MAX_EXPR (larger, bound) 1143 } 1144 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */ 1145 if (ass_code != MAX_EXPR) 1146 return false; 1147 1148 minmax = MIN_EXPR; 1149 if (operand_equal_for_phi_arg_p (op0, larger)) 1150 bound = op1; 1151 else if (operand_equal_for_phi_arg_p (op1, larger)) 1152 bound = op0; 1153 else 1154 return false; 1155 1156 /* We need BOUND <= SMALLER. */ 1157 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, 1158 bound, smaller))) 1159 return false; 1160 } 1161 else 1162 return false; 1163 } 1164 1165 /* Move the statement from the middle block. */ 1166 gsi = gsi_last_bb (cond_bb); 1167 gsi_from = gsi_last_nondebug_bb (middle_bb); 1168 gsi_move_before (&gsi_from, &gsi); 1169 } 1170 1171 /* Emit the statement to compute min/max. */ 1172 result = duplicate_ssa_name (PHI_RESULT (phi), NULL); 1173 new_stmt = gimple_build_assign (result, minmax, arg0, arg1); 1174 gsi = gsi_last_bb (cond_bb); 1175 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); 1176 1177 replace_phi_edge_with_variable (cond_bb, e1, phi, result); 1178 reset_flow_sensitive_info_in_bb (cond_bb); 1179 1180 return true; 1181} 1182 1183/* The function absolute_replacement does the main work of doing the absolute 1184 replacement. Return true if the replacement is done. Otherwise return 1185 false. 1186 bb is the basic block where the replacement is going to be done on. arg0 1187 is argument 0 from the phi. Likewise for arg1. */ 1188 1189static bool 1190abs_replacement (basic_block cond_bb, basic_block middle_bb, 1191 edge e0 ATTRIBUTE_UNUSED, edge e1, 1192 gimple phi, tree arg0, tree arg1) 1193{ 1194 tree result; 1195 gassign *new_stmt; 1196 gimple cond; 1197 gimple_stmt_iterator gsi; 1198 edge true_edge, false_edge; 1199 gimple assign; 1200 edge e; 1201 tree rhs, lhs; 1202 bool negate; 1203 enum tree_code cond_code; 1204 1205 /* If the type says honor signed zeros we cannot do this 1206 optimization. */ 1207 if (HONOR_SIGNED_ZEROS (arg1)) 1208 return false; 1209 1210 /* OTHER_BLOCK must have only one executable statement which must have the 1211 form arg0 = -arg1 or arg1 = -arg0. */ 1212 1213 assign = last_and_only_stmt (middle_bb); 1214 /* If we did not find the proper negation assignment, then we can not 1215 optimize. */ 1216 if (assign == NULL) 1217 return false; 1218 1219 /* If we got here, then we have found the only executable statement 1220 in OTHER_BLOCK. If it is anything other than arg = -arg1 or 1221 arg1 = -arg0, then we can not optimize. */ 1222 if (gimple_code (assign) != GIMPLE_ASSIGN) 1223 return false; 1224 1225 lhs = gimple_assign_lhs (assign); 1226 1227 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR) 1228 return false; 1229 1230 rhs = gimple_assign_rhs1 (assign); 1231 1232 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */ 1233 if (!(lhs == arg0 && rhs == arg1) 1234 && !(lhs == arg1 && rhs == arg0)) 1235 return false; 1236 1237 cond = last_stmt (cond_bb); 1238 result = PHI_RESULT (phi); 1239 1240 /* Only relationals comparing arg[01] against zero are interesting. */ 1241 cond_code = gimple_cond_code (cond); 1242 if (cond_code != GT_EXPR && cond_code != GE_EXPR 1243 && cond_code != LT_EXPR && cond_code != LE_EXPR) 1244 return false; 1245 1246 /* Make sure the conditional is arg[01] OP y. */ 1247 if (gimple_cond_lhs (cond) != rhs) 1248 return false; 1249 1250 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond))) 1251 ? real_zerop (gimple_cond_rhs (cond)) 1252 : integer_zerop (gimple_cond_rhs (cond))) 1253 ; 1254 else 1255 return false; 1256 1257 /* We need to know which is the true edge and which is the false 1258 edge so that we know if have abs or negative abs. */ 1259 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); 1260 1261 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we 1262 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if 1263 the false edge goes to OTHER_BLOCK. */ 1264 if (cond_code == GT_EXPR || cond_code == GE_EXPR) 1265 e = true_edge; 1266 else 1267 e = false_edge; 1268 1269 if (e->dest == middle_bb) 1270 negate = true; 1271 else 1272 negate = false; 1273 1274 result = duplicate_ssa_name (result, NULL); 1275 1276 if (negate) 1277 lhs = make_ssa_name (TREE_TYPE (result)); 1278 else 1279 lhs = result; 1280 1281 /* Build the modify expression with abs expression. */ 1282 new_stmt = gimple_build_assign (lhs, ABS_EXPR, rhs); 1283 1284 gsi = gsi_last_bb (cond_bb); 1285 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); 1286 1287 if (negate) 1288 { 1289 /* Get the right GSI. We want to insert after the recently 1290 added ABS_EXPR statement (which we know is the first statement 1291 in the block. */ 1292 new_stmt = gimple_build_assign (result, NEGATE_EXPR, lhs); 1293 1294 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); 1295 } 1296 1297 replace_phi_edge_with_variable (cond_bb, e1, phi, result); 1298 reset_flow_sensitive_info_in_bb (cond_bb); 1299 1300 /* Note that we optimized this PHI. */ 1301 return true; 1302} 1303 1304/* Auxiliary functions to determine the set of memory accesses which 1305 can't trap because they are preceded by accesses to the same memory 1306 portion. We do that for MEM_REFs, so we only need to track 1307 the SSA_NAME of the pointer indirectly referenced. The algorithm 1308 simply is a walk over all instructions in dominator order. When 1309 we see an MEM_REF we determine if we've already seen a same 1310 ref anywhere up to the root of the dominator tree. If we do the 1311 current access can't trap. If we don't see any dominating access 1312 the current access might trap, but might also make later accesses 1313 non-trapping, so we remember it. We need to be careful with loads 1314 or stores, for instance a load might not trap, while a store would, 1315 so if we see a dominating read access this doesn't mean that a later 1316 write access would not trap. Hence we also need to differentiate the 1317 type of access(es) seen. 1318 1319 ??? We currently are very conservative and assume that a load might 1320 trap even if a store doesn't (write-only memory). This probably is 1321 overly conservative. */ 1322 1323/* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF 1324 through it was seen, which would constitute a no-trap region for 1325 same accesses. */ 1326struct name_to_bb 1327{ 1328 unsigned int ssa_name_ver; 1329 unsigned int phase; 1330 bool store; 1331 HOST_WIDE_INT offset, size; 1332 basic_block bb; 1333}; 1334 1335/* Hashtable helpers. */ 1336 1337struct ssa_names_hasher : typed_free_remove <name_to_bb> 1338{ 1339 typedef name_to_bb value_type; 1340 typedef name_to_bb compare_type; 1341 static inline hashval_t hash (const value_type *); 1342 static inline bool equal (const value_type *, const compare_type *); 1343}; 1344 1345/* Used for quick clearing of the hash-table when we see calls. 1346 Hash entries with phase < nt_call_phase are invalid. */ 1347static unsigned int nt_call_phase; 1348 1349/* The hash function. */ 1350 1351inline hashval_t 1352ssa_names_hasher::hash (const value_type *n) 1353{ 1354 return n->ssa_name_ver ^ (((hashval_t) n->store) << 31) 1355 ^ (n->offset << 6) ^ (n->size << 3); 1356} 1357 1358/* The equality function of *P1 and *P2. */ 1359 1360inline bool 1361ssa_names_hasher::equal (const value_type *n1, const compare_type *n2) 1362{ 1363 return n1->ssa_name_ver == n2->ssa_name_ver 1364 && n1->store == n2->store 1365 && n1->offset == n2->offset 1366 && n1->size == n2->size; 1367} 1368 1369class nontrapping_dom_walker : public dom_walker 1370{ 1371public: 1372 nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps) 1373 : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {} 1374 1375 virtual void before_dom_children (basic_block); 1376 virtual void after_dom_children (basic_block); 1377 1378private: 1379 1380 /* We see the expression EXP in basic block BB. If it's an interesting 1381 expression (an MEM_REF through an SSA_NAME) possibly insert the 1382 expression into the set NONTRAP or the hash table of seen expressions. 1383 STORE is true if this expression is on the LHS, otherwise it's on 1384 the RHS. */ 1385 void add_or_mark_expr (basic_block, tree, bool); 1386 1387 hash_set<tree> *m_nontrapping; 1388 1389 /* The hash table for remembering what we've seen. */ 1390 hash_table<ssa_names_hasher> m_seen_ssa_names; 1391}; 1392 1393/* Called by walk_dominator_tree, when entering the block BB. */ 1394void 1395nontrapping_dom_walker::before_dom_children (basic_block bb) 1396{ 1397 edge e; 1398 edge_iterator ei; 1399 gimple_stmt_iterator gsi; 1400 1401 /* If we haven't seen all our predecessors, clear the hash-table. */ 1402 FOR_EACH_EDGE (e, ei, bb->preds) 1403 if ((((size_t)e->src->aux) & 2) == 0) 1404 { 1405 nt_call_phase++; 1406 break; 1407 } 1408 1409 /* Mark this BB as being on the path to dominator root and as visited. */ 1410 bb->aux = (void*)(1 | 2); 1411 1412 /* And walk the statements in order. */ 1413 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 1414 { 1415 gimple stmt = gsi_stmt (gsi); 1416 1417 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt)) 1418 nt_call_phase++; 1419 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt)) 1420 { 1421 add_or_mark_expr (bb, gimple_assign_lhs (stmt), true); 1422 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false); 1423 } 1424 } 1425} 1426 1427/* Called by walk_dominator_tree, when basic block BB is exited. */ 1428void 1429nontrapping_dom_walker::after_dom_children (basic_block bb) 1430{ 1431 /* This BB isn't on the path to dominator root anymore. */ 1432 bb->aux = (void*)2; 1433} 1434 1435/* We see the expression EXP in basic block BB. If it's an interesting 1436 expression (an MEM_REF through an SSA_NAME) possibly insert the 1437 expression into the set NONTRAP or the hash table of seen expressions. 1438 STORE is true if this expression is on the LHS, otherwise it's on 1439 the RHS. */ 1440void 1441nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store) 1442{ 1443 HOST_WIDE_INT size; 1444 1445 if (TREE_CODE (exp) == MEM_REF 1446 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME 1447 && tree_fits_shwi_p (TREE_OPERAND (exp, 1)) 1448 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0) 1449 { 1450 tree name = TREE_OPERAND (exp, 0); 1451 struct name_to_bb map; 1452 name_to_bb **slot; 1453 struct name_to_bb *n2bb; 1454 basic_block found_bb = 0; 1455 1456 /* Try to find the last seen MEM_REF through the same 1457 SSA_NAME, which can trap. */ 1458 map.ssa_name_ver = SSA_NAME_VERSION (name); 1459 map.phase = 0; 1460 map.bb = 0; 1461 map.store = store; 1462 map.offset = tree_to_shwi (TREE_OPERAND (exp, 1)); 1463 map.size = size; 1464 1465 slot = m_seen_ssa_names.find_slot (&map, INSERT); 1466 n2bb = *slot; 1467 if (n2bb && n2bb->phase >= nt_call_phase) 1468 found_bb = n2bb->bb; 1469 1470 /* If we've found a trapping MEM_REF, _and_ it dominates EXP 1471 (it's in a basic block on the path from us to the dominator root) 1472 then we can't trap. */ 1473 if (found_bb && (((size_t)found_bb->aux) & 1) == 1) 1474 { 1475 m_nontrapping->add (exp); 1476 } 1477 else 1478 { 1479 /* EXP might trap, so insert it into the hash table. */ 1480 if (n2bb) 1481 { 1482 n2bb->phase = nt_call_phase; 1483 n2bb->bb = bb; 1484 } 1485 else 1486 { 1487 n2bb = XNEW (struct name_to_bb); 1488 n2bb->ssa_name_ver = SSA_NAME_VERSION (name); 1489 n2bb->phase = nt_call_phase; 1490 n2bb->bb = bb; 1491 n2bb->store = store; 1492 n2bb->offset = map.offset; 1493 n2bb->size = size; 1494 *slot = n2bb; 1495 } 1496 } 1497 } 1498} 1499 1500/* This is the entry point of gathering non trapping memory accesses. 1501 It will do a dominator walk over the whole function, and it will 1502 make use of the bb->aux pointers. It returns a set of trees 1503 (the MEM_REFs itself) which can't trap. */ 1504static hash_set<tree> * 1505get_non_trapping (void) 1506{ 1507 nt_call_phase = 0; 1508 hash_set<tree> *nontrap = new hash_set<tree>; 1509 /* We're going to do a dominator walk, so ensure that we have 1510 dominance information. */ 1511 calculate_dominance_info (CDI_DOMINATORS); 1512 1513 nontrapping_dom_walker (CDI_DOMINATORS, nontrap) 1514 .walk (cfun->cfg->x_entry_block_ptr); 1515 1516 clear_aux_for_blocks (); 1517 return nontrap; 1518} 1519 1520/* Do the main work of conditional store replacement. We already know 1521 that the recognized pattern looks like so: 1522 1523 split: 1524 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1) 1525 MIDDLE_BB: 1526 something 1527 fallthrough (edge E0) 1528 JOIN_BB: 1529 some more 1530 1531 We check that MIDDLE_BB contains only one store, that that store 1532 doesn't trap (not via NOTRAP, but via checking if an access to the same 1533 memory location dominates us) and that the store has a "simple" RHS. */ 1534 1535static bool 1536cond_store_replacement (basic_block middle_bb, basic_block join_bb, 1537 edge e0, edge e1, hash_set<tree> *nontrap) 1538{ 1539 gimple assign = last_and_only_stmt (middle_bb); 1540 tree lhs, rhs, name, name2; 1541 gphi *newphi; 1542 gassign *new_stmt; 1543 gimple_stmt_iterator gsi; 1544 source_location locus; 1545 1546 /* Check if middle_bb contains of only one store. */ 1547 if (!assign 1548 || !gimple_assign_single_p (assign) 1549 || gimple_has_volatile_ops (assign)) 1550 return false; 1551 1552 locus = gimple_location (assign); 1553 lhs = gimple_assign_lhs (assign); 1554 rhs = gimple_assign_rhs1 (assign); 1555 if (TREE_CODE (lhs) != MEM_REF 1556 || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME 1557 || !is_gimple_reg_type (TREE_TYPE (lhs))) 1558 return false; 1559 1560 /* Prove that we can move the store down. We could also check 1561 TREE_THIS_NOTRAP here, but in that case we also could move stores, 1562 whose value is not available readily, which we want to avoid. */ 1563 if (!nontrap->contains (lhs)) 1564 return false; 1565 1566 /* Now we've checked the constraints, so do the transformation: 1567 1) Remove the single store. */ 1568 gsi = gsi_for_stmt (assign); 1569 unlink_stmt_vdef (assign); 1570 gsi_remove (&gsi, true); 1571 release_defs (assign); 1572 1573 /* 2) Insert a load from the memory of the store to the temporary 1574 on the edge which did not contain the store. */ 1575 lhs = unshare_expr (lhs); 1576 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); 1577 new_stmt = gimple_build_assign (name, lhs); 1578 gimple_set_location (new_stmt, locus); 1579 gsi_insert_on_edge (e1, new_stmt); 1580 1581 /* 3) Create a PHI node at the join block, with one argument 1582 holding the old RHS, and the other holding the temporary 1583 where we stored the old memory contents. */ 1584 name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); 1585 newphi = create_phi_node (name2, join_bb); 1586 add_phi_arg (newphi, rhs, e0, locus); 1587 add_phi_arg (newphi, name, e1, locus); 1588 1589 lhs = unshare_expr (lhs); 1590 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); 1591 1592 /* 4) Insert that PHI node. */ 1593 gsi = gsi_after_labels (join_bb); 1594 if (gsi_end_p (gsi)) 1595 { 1596 gsi = gsi_last_bb (join_bb); 1597 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); 1598 } 1599 else 1600 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); 1601 1602 return true; 1603} 1604 1605/* Do the main work of conditional store replacement. */ 1606 1607static bool 1608cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb, 1609 basic_block join_bb, gimple then_assign, 1610 gimple else_assign) 1611{ 1612 tree lhs_base, lhs, then_rhs, else_rhs, name; 1613 source_location then_locus, else_locus; 1614 gimple_stmt_iterator gsi; 1615 gphi *newphi; 1616 gassign *new_stmt; 1617 1618 if (then_assign == NULL 1619 || !gimple_assign_single_p (then_assign) 1620 || gimple_clobber_p (then_assign) 1621 || gimple_has_volatile_ops (then_assign) 1622 || else_assign == NULL 1623 || !gimple_assign_single_p (else_assign) 1624 || gimple_clobber_p (else_assign) 1625 || gimple_has_volatile_ops (else_assign)) 1626 return false; 1627 1628 lhs = gimple_assign_lhs (then_assign); 1629 if (!is_gimple_reg_type (TREE_TYPE (lhs)) 1630 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0)) 1631 return false; 1632 1633 lhs_base = get_base_address (lhs); 1634 if (lhs_base == NULL_TREE 1635 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF)) 1636 return false; 1637 1638 then_rhs = gimple_assign_rhs1 (then_assign); 1639 else_rhs = gimple_assign_rhs1 (else_assign); 1640 then_locus = gimple_location (then_assign); 1641 else_locus = gimple_location (else_assign); 1642 1643 /* Now we've checked the constraints, so do the transformation: 1644 1) Remove the stores. */ 1645 gsi = gsi_for_stmt (then_assign); 1646 unlink_stmt_vdef (then_assign); 1647 gsi_remove (&gsi, true); 1648 release_defs (then_assign); 1649 1650 gsi = gsi_for_stmt (else_assign); 1651 unlink_stmt_vdef (else_assign); 1652 gsi_remove (&gsi, true); 1653 release_defs (else_assign); 1654 1655 /* 2) Create a PHI node at the join block, with one argument 1656 holding the old RHS, and the other holding the temporary 1657 where we stored the old memory contents. */ 1658 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); 1659 newphi = create_phi_node (name, join_bb); 1660 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus); 1661 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus); 1662 1663 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); 1664 1665 /* 3) Insert that PHI node. */ 1666 gsi = gsi_after_labels (join_bb); 1667 if (gsi_end_p (gsi)) 1668 { 1669 gsi = gsi_last_bb (join_bb); 1670 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); 1671 } 1672 else 1673 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); 1674 1675 return true; 1676} 1677 1678/* Conditional store replacement. We already know 1679 that the recognized pattern looks like so: 1680 1681 split: 1682 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1) 1683 THEN_BB: 1684 ... 1685 X = Y; 1686 ... 1687 goto JOIN_BB; 1688 ELSE_BB: 1689 ... 1690 X = Z; 1691 ... 1692 fallthrough (edge E0) 1693 JOIN_BB: 1694 some more 1695 1696 We check that it is safe to sink the store to JOIN_BB by verifying that 1697 there are no read-after-write or write-after-write dependencies in 1698 THEN_BB and ELSE_BB. */ 1699 1700static bool 1701cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb, 1702 basic_block join_bb) 1703{ 1704 gimple then_assign = last_and_only_stmt (then_bb); 1705 gimple else_assign = last_and_only_stmt (else_bb); 1706 vec<data_reference_p> then_datarefs, else_datarefs; 1707 vec<ddr_p> then_ddrs, else_ddrs; 1708 gimple then_store, else_store; 1709 bool found, ok = false, res; 1710 struct data_dependence_relation *ddr; 1711 data_reference_p then_dr, else_dr; 1712 int i, j; 1713 tree then_lhs, else_lhs; 1714 basic_block blocks[3]; 1715 1716 if (MAX_STORES_TO_SINK == 0) 1717 return false; 1718 1719 /* Handle the case with single statement in THEN_BB and ELSE_BB. */ 1720 if (then_assign && else_assign) 1721 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, 1722 then_assign, else_assign); 1723 1724 /* Find data references. */ 1725 then_datarefs.create (1); 1726 else_datarefs.create (1); 1727 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs) 1728 == chrec_dont_know) 1729 || !then_datarefs.length () 1730 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs) 1731 == chrec_dont_know) 1732 || !else_datarefs.length ()) 1733 { 1734 free_data_refs (then_datarefs); 1735 free_data_refs (else_datarefs); 1736 return false; 1737 } 1738 1739 /* Find pairs of stores with equal LHS. */ 1740 auto_vec<gimple, 1> then_stores, else_stores; 1741 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr) 1742 { 1743 if (DR_IS_READ (then_dr)) 1744 continue; 1745 1746 then_store = DR_STMT (then_dr); 1747 then_lhs = gimple_get_lhs (then_store); 1748 if (then_lhs == NULL_TREE) 1749 continue; 1750 found = false; 1751 1752 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr) 1753 { 1754 if (DR_IS_READ (else_dr)) 1755 continue; 1756 1757 else_store = DR_STMT (else_dr); 1758 else_lhs = gimple_get_lhs (else_store); 1759 if (else_lhs == NULL_TREE) 1760 continue; 1761 1762 if (operand_equal_p (then_lhs, else_lhs, 0)) 1763 { 1764 found = true; 1765 break; 1766 } 1767 } 1768 1769 if (!found) 1770 continue; 1771 1772 then_stores.safe_push (then_store); 1773 else_stores.safe_push (else_store); 1774 } 1775 1776 /* No pairs of stores found. */ 1777 if (!then_stores.length () 1778 || then_stores.length () > (unsigned) MAX_STORES_TO_SINK) 1779 { 1780 free_data_refs (then_datarefs); 1781 free_data_refs (else_datarefs); 1782 return false; 1783 } 1784 1785 /* Compute and check data dependencies in both basic blocks. */ 1786 then_ddrs.create (1); 1787 else_ddrs.create (1); 1788 if (!compute_all_dependences (then_datarefs, &then_ddrs, 1789 vNULL, false) 1790 || !compute_all_dependences (else_datarefs, &else_ddrs, 1791 vNULL, false)) 1792 { 1793 free_dependence_relations (then_ddrs); 1794 free_dependence_relations (else_ddrs); 1795 free_data_refs (then_datarefs); 1796 free_data_refs (else_datarefs); 1797 return false; 1798 } 1799 blocks[0] = then_bb; 1800 blocks[1] = else_bb; 1801 blocks[2] = join_bb; 1802 renumber_gimple_stmt_uids_in_blocks (blocks, 3); 1803 1804 /* Check that there are no read-after-write or write-after-write dependencies 1805 in THEN_BB. */ 1806 FOR_EACH_VEC_ELT (then_ddrs, i, ddr) 1807 { 1808 struct data_reference *dra = DDR_A (ddr); 1809 struct data_reference *drb = DDR_B (ddr); 1810 1811 if (DDR_ARE_DEPENDENT (ddr) != chrec_known 1812 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) 1813 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) 1814 || (DR_IS_READ (drb) && DR_IS_WRITE (dra) 1815 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) 1816 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) 1817 { 1818 free_dependence_relations (then_ddrs); 1819 free_dependence_relations (else_ddrs); 1820 free_data_refs (then_datarefs); 1821 free_data_refs (else_datarefs); 1822 return false; 1823 } 1824 } 1825 1826 /* Check that there are no read-after-write or write-after-write dependencies 1827 in ELSE_BB. */ 1828 FOR_EACH_VEC_ELT (else_ddrs, i, ddr) 1829 { 1830 struct data_reference *dra = DDR_A (ddr); 1831 struct data_reference *drb = DDR_B (ddr); 1832 1833 if (DDR_ARE_DEPENDENT (ddr) != chrec_known 1834 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) 1835 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) 1836 || (DR_IS_READ (drb) && DR_IS_WRITE (dra) 1837 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) 1838 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) 1839 { 1840 free_dependence_relations (then_ddrs); 1841 free_dependence_relations (else_ddrs); 1842 free_data_refs (then_datarefs); 1843 free_data_refs (else_datarefs); 1844 return false; 1845 } 1846 } 1847 1848 /* Sink stores with same LHS. */ 1849 FOR_EACH_VEC_ELT (then_stores, i, then_store) 1850 { 1851 else_store = else_stores[i]; 1852 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, 1853 then_store, else_store); 1854 ok = ok || res; 1855 } 1856 1857 free_dependence_relations (then_ddrs); 1858 free_dependence_relations (else_ddrs); 1859 free_data_refs (then_datarefs); 1860 free_data_refs (else_datarefs); 1861 1862 return ok; 1863} 1864 1865/* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */ 1866 1867static bool 1868local_mem_dependence (gimple stmt, basic_block bb) 1869{ 1870 tree vuse = gimple_vuse (stmt); 1871 gimple def; 1872 1873 if (!vuse) 1874 return false; 1875 1876 def = SSA_NAME_DEF_STMT (vuse); 1877 return (def && gimple_bb (def) == bb); 1878} 1879 1880/* Given a "diamond" control-flow pattern where BB0 tests a condition, 1881 BB1 and BB2 are "then" and "else" blocks dependent on this test, 1882 and BB3 rejoins control flow following BB1 and BB2, look for 1883 opportunities to hoist loads as follows. If BB3 contains a PHI of 1884 two loads, one each occurring in BB1 and BB2, and the loads are 1885 provably of adjacent fields in the same structure, then move both 1886 loads into BB0. Of course this can only be done if there are no 1887 dependencies preventing such motion. 1888 1889 One of the hoisted loads will always be speculative, so the 1890 transformation is currently conservative: 1891 1892 - The fields must be strictly adjacent. 1893 - The two fields must occupy a single memory block that is 1894 guaranteed to not cross a page boundary. 1895 1896 The last is difficult to prove, as such memory blocks should be 1897 aligned on the minimum of the stack alignment boundary and the 1898 alignment guaranteed by heap allocation interfaces. Thus we rely 1899 on a parameter for the alignment value. 1900 1901 Provided a good value is used for the last case, the first 1902 restriction could possibly be relaxed. */ 1903 1904static void 1905hoist_adjacent_loads (basic_block bb0, basic_block bb1, 1906 basic_block bb2, basic_block bb3) 1907{ 1908 int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE); 1909 unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT); 1910 gphi_iterator gsi; 1911 1912 /* Walk the phis in bb3 looking for an opportunity. We are looking 1913 for phis of two SSA names, one each of which is defined in bb1 and 1914 bb2. */ 1915 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi)) 1916 { 1917 gphi *phi_stmt = gsi.phi (); 1918 gimple def1, def2, defswap; 1919 tree arg1, arg2, ref1, ref2, field1, field2, fieldswap; 1920 tree tree_offset1, tree_offset2, tree_size2, next; 1921 int offset1, offset2, size2; 1922 unsigned align1; 1923 gimple_stmt_iterator gsi2; 1924 basic_block bb_for_def1, bb_for_def2; 1925 1926 if (gimple_phi_num_args (phi_stmt) != 2 1927 || virtual_operand_p (gimple_phi_result (phi_stmt))) 1928 continue; 1929 1930 arg1 = gimple_phi_arg_def (phi_stmt, 0); 1931 arg2 = gimple_phi_arg_def (phi_stmt, 1); 1932 1933 if (TREE_CODE (arg1) != SSA_NAME 1934 || TREE_CODE (arg2) != SSA_NAME 1935 || SSA_NAME_IS_DEFAULT_DEF (arg1) 1936 || SSA_NAME_IS_DEFAULT_DEF (arg2)) 1937 continue; 1938 1939 def1 = SSA_NAME_DEF_STMT (arg1); 1940 def2 = SSA_NAME_DEF_STMT (arg2); 1941 1942 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2) 1943 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2)) 1944 continue; 1945 1946 /* Check the mode of the arguments to be sure a conditional move 1947 can be generated for it. */ 1948 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1))) 1949 == CODE_FOR_nothing) 1950 continue; 1951 1952 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */ 1953 if (!gimple_assign_single_p (def1) 1954 || !gimple_assign_single_p (def2) 1955 || gimple_has_volatile_ops (def1) 1956 || gimple_has_volatile_ops (def2)) 1957 continue; 1958 1959 ref1 = gimple_assign_rhs1 (def1); 1960 ref2 = gimple_assign_rhs1 (def2); 1961 1962 if (TREE_CODE (ref1) != COMPONENT_REF 1963 || TREE_CODE (ref2) != COMPONENT_REF) 1964 continue; 1965 1966 /* The zeroth operand of the two component references must be 1967 identical. It is not sufficient to compare get_base_address of 1968 the two references, because this could allow for different 1969 elements of the same array in the two trees. It is not safe to 1970 assume that the existence of one array element implies the 1971 existence of a different one. */ 1972 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0)) 1973 continue; 1974 1975 field1 = TREE_OPERAND (ref1, 1); 1976 field2 = TREE_OPERAND (ref2, 1); 1977 1978 /* Check for field adjacency, and ensure field1 comes first. */ 1979 for (next = DECL_CHAIN (field1); 1980 next && TREE_CODE (next) != FIELD_DECL; 1981 next = DECL_CHAIN (next)) 1982 ; 1983 1984 if (next != field2) 1985 { 1986 for (next = DECL_CHAIN (field2); 1987 next && TREE_CODE (next) != FIELD_DECL; 1988 next = DECL_CHAIN (next)) 1989 ; 1990 1991 if (next != field1) 1992 continue; 1993 1994 fieldswap = field1; 1995 field1 = field2; 1996 field2 = fieldswap; 1997 defswap = def1; 1998 def1 = def2; 1999 def2 = defswap; 2000 } 2001 2002 bb_for_def1 = gimple_bb (def1); 2003 bb_for_def2 = gimple_bb (def2); 2004 2005 /* Check for proper alignment of the first field. */ 2006 tree_offset1 = bit_position (field1); 2007 tree_offset2 = bit_position (field2); 2008 tree_size2 = DECL_SIZE (field2); 2009 2010 if (!tree_fits_uhwi_p (tree_offset1) 2011 || !tree_fits_uhwi_p (tree_offset2) 2012 || !tree_fits_uhwi_p (tree_size2)) 2013 continue; 2014 2015 offset1 = tree_to_uhwi (tree_offset1); 2016 offset2 = tree_to_uhwi (tree_offset2); 2017 size2 = tree_to_uhwi (tree_size2); 2018 align1 = DECL_ALIGN (field1) % param_align_bits; 2019 2020 if (offset1 % BITS_PER_UNIT != 0) 2021 continue; 2022 2023 /* For profitability, the two field references should fit within 2024 a single cache line. */ 2025 if (align1 + offset2 - offset1 + size2 > param_align_bits) 2026 continue; 2027 2028 /* The two expressions cannot be dependent upon vdefs defined 2029 in bb1/bb2. */ 2030 if (local_mem_dependence (def1, bb_for_def1) 2031 || local_mem_dependence (def2, bb_for_def2)) 2032 continue; 2033 2034 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into 2035 bb0. We hoist the first one first so that a cache miss is handled 2036 efficiently regardless of hardware cache-fill policy. */ 2037 gsi2 = gsi_for_stmt (def1); 2038 gsi_move_to_bb_end (&gsi2, bb0); 2039 gsi2 = gsi_for_stmt (def2); 2040 gsi_move_to_bb_end (&gsi2, bb0); 2041 2042 if (dump_file && (dump_flags & TDF_DETAILS)) 2043 { 2044 fprintf (dump_file, 2045 "\nHoisting adjacent loads from %d and %d into %d: \n", 2046 bb_for_def1->index, bb_for_def2->index, bb0->index); 2047 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS); 2048 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS); 2049 } 2050 } 2051} 2052 2053/* Determine whether we should attempt to hoist adjacent loads out of 2054 diamond patterns in pass_phiopt. Always hoist loads if 2055 -fhoist-adjacent-loads is specified and the target machine has 2056 both a conditional move instruction and a defined cache line size. */ 2057 2058static bool 2059gate_hoist_loads (void) 2060{ 2061 return (flag_hoist_adjacent_loads == 1 2062 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE) 2063 && HAVE_conditional_move); 2064} 2065 2066/* This pass tries to replaces an if-then-else block with an 2067 assignment. We have four kinds of transformations. Some of these 2068 transformations are also performed by the ifcvt RTL optimizer. 2069 2070 Conditional Replacement 2071 ----------------------- 2072 2073 This transformation, implemented in conditional_replacement, 2074 replaces 2075 2076 bb0: 2077 if (cond) goto bb2; else goto bb1; 2078 bb1: 2079 bb2: 2080 x = PHI <0 (bb1), 1 (bb0), ...>; 2081 2082 with 2083 2084 bb0: 2085 x' = cond; 2086 goto bb2; 2087 bb2: 2088 x = PHI <x' (bb0), ...>; 2089 2090 We remove bb1 as it becomes unreachable. This occurs often due to 2091 gimplification of conditionals. 2092 2093 Value Replacement 2094 ----------------- 2095 2096 This transformation, implemented in value_replacement, replaces 2097 2098 bb0: 2099 if (a != b) goto bb2; else goto bb1; 2100 bb1: 2101 bb2: 2102 x = PHI <a (bb1), b (bb0), ...>; 2103 2104 with 2105 2106 bb0: 2107 bb2: 2108 x = PHI <b (bb0), ...>; 2109 2110 This opportunity can sometimes occur as a result of other 2111 optimizations. 2112 2113 2114 Another case caught by value replacement looks like this: 2115 2116 bb0: 2117 t1 = a == CONST; 2118 t2 = b > c; 2119 t3 = t1 & t2; 2120 if (t3 != 0) goto bb1; else goto bb2; 2121 bb1: 2122 bb2: 2123 x = PHI (CONST, a) 2124 2125 Gets replaced with: 2126 bb0: 2127 bb2: 2128 t1 = a == CONST; 2129 t2 = b > c; 2130 t3 = t1 & t2; 2131 x = a; 2132 2133 ABS Replacement 2134 --------------- 2135 2136 This transformation, implemented in abs_replacement, replaces 2137 2138 bb0: 2139 if (a >= 0) goto bb2; else goto bb1; 2140 bb1: 2141 x = -a; 2142 bb2: 2143 x = PHI <x (bb1), a (bb0), ...>; 2144 2145 with 2146 2147 bb0: 2148 x' = ABS_EXPR< a >; 2149 bb2: 2150 x = PHI <x' (bb0), ...>; 2151 2152 MIN/MAX Replacement 2153 ------------------- 2154 2155 This transformation, minmax_replacement replaces 2156 2157 bb0: 2158 if (a <= b) goto bb2; else goto bb1; 2159 bb1: 2160 bb2: 2161 x = PHI <b (bb1), a (bb0), ...>; 2162 2163 with 2164 2165 bb0: 2166 x' = MIN_EXPR (a, b) 2167 bb2: 2168 x = PHI <x' (bb0), ...>; 2169 2170 A similar transformation is done for MAX_EXPR. 2171 2172 2173 This pass also performs a fifth transformation of a slightly different 2174 flavor. 2175 2176 Adjacent Load Hoisting 2177 ---------------------- 2178 2179 This transformation replaces 2180 2181 bb0: 2182 if (...) goto bb2; else goto bb1; 2183 bb1: 2184 x1 = (<expr>).field1; 2185 goto bb3; 2186 bb2: 2187 x2 = (<expr>).field2; 2188 bb3: 2189 # x = PHI <x1, x2>; 2190 2191 with 2192 2193 bb0: 2194 x1 = (<expr>).field1; 2195 x2 = (<expr>).field2; 2196 if (...) goto bb2; else goto bb1; 2197 bb1: 2198 goto bb3; 2199 bb2: 2200 bb3: 2201 # x = PHI <x1, x2>; 2202 2203 The purpose of this transformation is to enable generation of conditional 2204 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of 2205 the loads is speculative, the transformation is restricted to very 2206 specific cases to avoid introducing a page fault. We are looking for 2207 the common idiom: 2208 2209 if (...) 2210 x = y->left; 2211 else 2212 x = y->right; 2213 2214 where left and right are typically adjacent pointers in a tree structure. */ 2215 2216namespace { 2217 2218const pass_data pass_data_phiopt = 2219{ 2220 GIMPLE_PASS, /* type */ 2221 "phiopt", /* name */ 2222 OPTGROUP_NONE, /* optinfo_flags */ 2223 TV_TREE_PHIOPT, /* tv_id */ 2224 ( PROP_cfg | PROP_ssa ), /* properties_required */ 2225 0, /* properties_provided */ 2226 0, /* properties_destroyed */ 2227 0, /* todo_flags_start */ 2228 0, /* todo_flags_finish */ 2229}; 2230 2231class pass_phiopt : public gimple_opt_pass 2232{ 2233public: 2234 pass_phiopt (gcc::context *ctxt) 2235 : gimple_opt_pass (pass_data_phiopt, ctxt) 2236 {} 2237 2238 /* opt_pass methods: */ 2239 opt_pass * clone () { return new pass_phiopt (m_ctxt); } 2240 virtual bool gate (function *) { return flag_ssa_phiopt; } 2241 virtual unsigned int execute (function *) 2242 { 2243 return tree_ssa_phiopt_worker (false, gate_hoist_loads ()); 2244 } 2245 2246}; // class pass_phiopt 2247 2248} // anon namespace 2249 2250gimple_opt_pass * 2251make_pass_phiopt (gcc::context *ctxt) 2252{ 2253 return new pass_phiopt (ctxt); 2254} 2255 2256namespace { 2257 2258const pass_data pass_data_cselim = 2259{ 2260 GIMPLE_PASS, /* type */ 2261 "cselim", /* name */ 2262 OPTGROUP_NONE, /* optinfo_flags */ 2263 TV_TREE_PHIOPT, /* tv_id */ 2264 ( PROP_cfg | PROP_ssa ), /* properties_required */ 2265 0, /* properties_provided */ 2266 0, /* properties_destroyed */ 2267 0, /* todo_flags_start */ 2268 0, /* todo_flags_finish */ 2269}; 2270 2271class pass_cselim : public gimple_opt_pass 2272{ 2273public: 2274 pass_cselim (gcc::context *ctxt) 2275 : gimple_opt_pass (pass_data_cselim, ctxt) 2276 {} 2277 2278 /* opt_pass methods: */ 2279 virtual bool gate (function *) { return flag_tree_cselim; } 2280 virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); } 2281 2282}; // class pass_cselim 2283 2284} // anon namespace 2285 2286gimple_opt_pass * 2287make_pass_cselim (gcc::context *ctxt) 2288{ 2289 return new pass_cselim (ctxt); 2290} 2291