1 /* Functions to determine/estimate number of iterations of a loop. 2 Copyright (C) 2004-2015 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 "tm.h" 24 #include "hash-set.h" 25 #include "machmode.h" 26 #include "vec.h" 27 #include "double-int.h" 28 #include "input.h" 29 #include "alias.h" 30 #include "symtab.h" 31 #include "wide-int.h" 32 #include "inchash.h" 33 #include "tree.h" 34 #include "fold-const.h" 35 #include "calls.h" 36 #include "hashtab.h" 37 #include "hard-reg-set.h" 38 #include "function.h" 39 #include "rtl.h" 40 #include "flags.h" 41 #include "statistics.h" 42 #include "real.h" 43 #include "fixed-value.h" 44 #include "insn-config.h" 45 #include "expmed.h" 46 #include "dojump.h" 47 #include "explow.h" 48 #include "emit-rtl.h" 49 #include "varasm.h" 50 #include "stmt.h" 51 #include "expr.h" 52 #include "tm_p.h" 53 #include "predict.h" 54 #include "dominance.h" 55 #include "cfg.h" 56 #include "basic-block.h" 57 #include "gimple-pretty-print.h" 58 #include "intl.h" 59 #include "tree-ssa-alias.h" 60 #include "internal-fn.h" 61 #include "gimple-expr.h" 62 #include "is-a.h" 63 #include "gimple.h" 64 #include "gimplify.h" 65 #include "gimple-iterator.h" 66 #include "gimple-ssa.h" 67 #include "tree-cfg.h" 68 #include "tree-phinodes.h" 69 #include "ssa-iterators.h" 70 #include "tree-ssa-loop-ivopts.h" 71 #include "tree-ssa-loop-niter.h" 72 #include "tree-ssa-loop.h" 73 #include "dumpfile.h" 74 #include "cfgloop.h" 75 #include "tree-chrec.h" 76 #include "tree-scalar-evolution.h" 77 #include "tree-data-ref.h" 78 #include "params.h" 79 #include "diagnostic-core.h" 80 #include "tree-inline.h" 81 #include "tree-pass.h" 82 #include "stringpool.h" 83 #include "tree-ssanames.h" 84 #include "wide-int-print.h" 85 86 87 #define SWAP(X, Y) do { affine_iv *tmp = (X); (X) = (Y); (Y) = tmp; } while (0) 88 89 /* The maximum number of dominator BBs we search for conditions 90 of loop header copies we use for simplifying a conditional 91 expression. */ 92 #define MAX_DOMINATORS_TO_WALK 8 93 94 /* 95 96 Analysis of number of iterations of an affine exit test. 97 98 */ 99 100 /* Bounds on some value, BELOW <= X <= UP. */ 101 102 typedef struct 103 { 104 mpz_t below, up; 105 } bounds; 106 107 108 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */ 109 110 static void 111 split_to_var_and_offset (tree expr, tree *var, mpz_t offset) 112 { 113 tree type = TREE_TYPE (expr); 114 tree op0, op1; 115 bool negate = false; 116 117 *var = expr; 118 mpz_set_ui (offset, 0); 119 120 switch (TREE_CODE (expr)) 121 { 122 case MINUS_EXPR: 123 negate = true; 124 /* Fallthru. */ 125 126 case PLUS_EXPR: 127 case POINTER_PLUS_EXPR: 128 op0 = TREE_OPERAND (expr, 0); 129 op1 = TREE_OPERAND (expr, 1); 130 131 if (TREE_CODE (op1) != INTEGER_CST) 132 break; 133 134 *var = op0; 135 /* Always sign extend the offset. */ 136 wi::to_mpz (op1, offset, SIGNED); 137 if (negate) 138 mpz_neg (offset, offset); 139 break; 140 141 case INTEGER_CST: 142 *var = build_int_cst_type (type, 0); 143 wi::to_mpz (expr, offset, TYPE_SIGN (type)); 144 break; 145 146 default: 147 break; 148 } 149 } 150 151 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF 152 in TYPE to MIN and MAX. */ 153 154 static void 155 determine_value_range (struct loop *loop, tree type, tree var, mpz_t off, 156 mpz_t min, mpz_t max) 157 { 158 wide_int minv, maxv; 159 enum value_range_type rtype = VR_VARYING; 160 161 /* If the expression is a constant, we know its value exactly. */ 162 if (integer_zerop (var)) 163 { 164 mpz_set (min, off); 165 mpz_set (max, off); 166 return; 167 } 168 169 get_type_static_bounds (type, min, max); 170 171 /* See if we have some range info from VRP. */ 172 if (TREE_CODE (var) == SSA_NAME && INTEGRAL_TYPE_P (type)) 173 { 174 edge e = loop_preheader_edge (loop); 175 signop sgn = TYPE_SIGN (type); 176 gphi_iterator gsi; 177 178 /* Either for VAR itself... */ 179 rtype = get_range_info (var, &minv, &maxv); 180 /* Or for PHI results in loop->header where VAR is used as 181 PHI argument from the loop preheader edge. */ 182 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) 183 { 184 gphi *phi = gsi.phi (); 185 wide_int minc, maxc; 186 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var 187 && (get_range_info (gimple_phi_result (phi), &minc, &maxc) 188 == VR_RANGE)) 189 { 190 if (rtype != VR_RANGE) 191 { 192 rtype = VR_RANGE; 193 minv = minc; 194 maxv = maxc; 195 } 196 else 197 { 198 minv = wi::max (minv, minc, sgn); 199 maxv = wi::min (maxv, maxc, sgn); 200 /* If the PHI result range are inconsistent with 201 the VAR range, give up on looking at the PHI 202 results. This can happen if VR_UNDEFINED is 203 involved. */ 204 if (wi::gt_p (minv, maxv, sgn)) 205 { 206 rtype = get_range_info (var, &minv, &maxv); 207 break; 208 } 209 } 210 } 211 } 212 if (rtype == VR_RANGE) 213 { 214 mpz_t minm, maxm; 215 gcc_assert (wi::le_p (minv, maxv, sgn)); 216 mpz_init (minm); 217 mpz_init (maxm); 218 wi::to_mpz (minv, minm, sgn); 219 wi::to_mpz (maxv, maxm, sgn); 220 mpz_add (minm, minm, off); 221 mpz_add (maxm, maxm, off); 222 /* If the computation may not wrap or off is zero, then this 223 is always fine. If off is negative and minv + off isn't 224 smaller than type's minimum, or off is positive and 225 maxv + off isn't bigger than type's maximum, use the more 226 precise range too. */ 227 if (nowrap_type_p (type) 228 || mpz_sgn (off) == 0 229 || (mpz_sgn (off) < 0 && mpz_cmp (minm, min) >= 0) 230 || (mpz_sgn (off) > 0 && mpz_cmp (maxm, max) <= 0)) 231 { 232 mpz_set (min, minm); 233 mpz_set (max, maxm); 234 mpz_clear (minm); 235 mpz_clear (maxm); 236 return; 237 } 238 mpz_clear (minm); 239 mpz_clear (maxm); 240 } 241 } 242 243 /* If the computation may wrap, we know nothing about the value, except for 244 the range of the type. */ 245 if (!nowrap_type_p (type)) 246 return; 247 248 /* Since the addition of OFF does not wrap, if OFF is positive, then we may 249 add it to MIN, otherwise to MAX. */ 250 if (mpz_sgn (off) < 0) 251 mpz_add (max, max, off); 252 else 253 mpz_add (min, min, off); 254 } 255 256 /* Stores the bounds on the difference of the values of the expressions 257 (var + X) and (var + Y), computed in TYPE, to BNDS. */ 258 259 static void 260 bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y, 261 bounds *bnds) 262 { 263 int rel = mpz_cmp (x, y); 264 bool may_wrap = !nowrap_type_p (type); 265 mpz_t m; 266 267 /* If X == Y, then the expressions are always equal. 268 If X > Y, there are the following possibilities: 269 a) neither of var + X and var + Y overflow or underflow, or both of 270 them do. Then their difference is X - Y. 271 b) var + X overflows, and var + Y does not. Then the values of the 272 expressions are var + X - M and var + Y, where M is the range of 273 the type, and their difference is X - Y - M. 274 c) var + Y underflows and var + X does not. Their difference again 275 is M - X + Y. 276 Therefore, if the arithmetics in type does not overflow, then the 277 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y) 278 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or 279 (X - Y, X - Y + M). */ 280 281 if (rel == 0) 282 { 283 mpz_set_ui (bnds->below, 0); 284 mpz_set_ui (bnds->up, 0); 285 return; 286 } 287 288 mpz_init (m); 289 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), m, UNSIGNED); 290 mpz_add_ui (m, m, 1); 291 mpz_sub (bnds->up, x, y); 292 mpz_set (bnds->below, bnds->up); 293 294 if (may_wrap) 295 { 296 if (rel > 0) 297 mpz_sub (bnds->below, bnds->below, m); 298 else 299 mpz_add (bnds->up, bnds->up, m); 300 } 301 302 mpz_clear (m); 303 } 304 305 /* From condition C0 CMP C1 derives information regarding the 306 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE, 307 and stores it to BNDS. */ 308 309 static void 310 refine_bounds_using_guard (tree type, tree varx, mpz_t offx, 311 tree vary, mpz_t offy, 312 tree c0, enum tree_code cmp, tree c1, 313 bounds *bnds) 314 { 315 tree varc0, varc1, tmp, ctype; 316 mpz_t offc0, offc1, loffx, loffy, bnd; 317 bool lbound = false; 318 bool no_wrap = nowrap_type_p (type); 319 bool x_ok, y_ok; 320 321 switch (cmp) 322 { 323 case LT_EXPR: 324 case LE_EXPR: 325 case GT_EXPR: 326 case GE_EXPR: 327 STRIP_SIGN_NOPS (c0); 328 STRIP_SIGN_NOPS (c1); 329 ctype = TREE_TYPE (c0); 330 if (!useless_type_conversion_p (ctype, type)) 331 return; 332 333 break; 334 335 case EQ_EXPR: 336 /* We could derive quite precise information from EQ_EXPR, however, such 337 a guard is unlikely to appear, so we do not bother with handling 338 it. */ 339 return; 340 341 case NE_EXPR: 342 /* NE_EXPR comparisons do not contain much of useful information, except for 343 special case of comparing with the bounds of the type. */ 344 if (TREE_CODE (c1) != INTEGER_CST 345 || !INTEGRAL_TYPE_P (type)) 346 return; 347 348 /* Ensure that the condition speaks about an expression in the same type 349 as X and Y. */ 350 ctype = TREE_TYPE (c0); 351 if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type)) 352 return; 353 c0 = fold_convert (type, c0); 354 c1 = fold_convert (type, c1); 355 356 if (TYPE_MIN_VALUE (type) 357 && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0)) 358 { 359 cmp = GT_EXPR; 360 break; 361 } 362 if (TYPE_MAX_VALUE (type) 363 && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0)) 364 { 365 cmp = LT_EXPR; 366 break; 367 } 368 369 return; 370 default: 371 return; 372 } 373 374 mpz_init (offc0); 375 mpz_init (offc1); 376 split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0); 377 split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1); 378 379 /* We are only interested in comparisons of expressions based on VARX and 380 VARY. TODO -- we might also be able to derive some bounds from 381 expressions containing just one of the variables. */ 382 383 if (operand_equal_p (varx, varc1, 0)) 384 { 385 tmp = varc0; varc0 = varc1; varc1 = tmp; 386 mpz_swap (offc0, offc1); 387 cmp = swap_tree_comparison (cmp); 388 } 389 390 if (!operand_equal_p (varx, varc0, 0) 391 || !operand_equal_p (vary, varc1, 0)) 392 goto end; 393 394 mpz_init_set (loffx, offx); 395 mpz_init_set (loffy, offy); 396 397 if (cmp == GT_EXPR || cmp == GE_EXPR) 398 { 399 tmp = varx; varx = vary; vary = tmp; 400 mpz_swap (offc0, offc1); 401 mpz_swap (loffx, loffy); 402 cmp = swap_tree_comparison (cmp); 403 lbound = true; 404 } 405 406 /* If there is no overflow, the condition implies that 407 408 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0). 409 410 The overflows and underflows may complicate things a bit; each 411 overflow decreases the appropriate offset by M, and underflow 412 increases it by M. The above inequality would not necessarily be 413 true if 414 415 -- VARX + OFFX underflows and VARX + OFFC0 does not, or 416 VARX + OFFC0 overflows, but VARX + OFFX does not. 417 This may only happen if OFFX < OFFC0. 418 -- VARY + OFFY overflows and VARY + OFFC1 does not, or 419 VARY + OFFC1 underflows and VARY + OFFY does not. 420 This may only happen if OFFY > OFFC1. */ 421 422 if (no_wrap) 423 { 424 x_ok = true; 425 y_ok = true; 426 } 427 else 428 { 429 x_ok = (integer_zerop (varx) 430 || mpz_cmp (loffx, offc0) >= 0); 431 y_ok = (integer_zerop (vary) 432 || mpz_cmp (loffy, offc1) <= 0); 433 } 434 435 if (x_ok && y_ok) 436 { 437 mpz_init (bnd); 438 mpz_sub (bnd, loffx, loffy); 439 mpz_add (bnd, bnd, offc1); 440 mpz_sub (bnd, bnd, offc0); 441 442 if (cmp == LT_EXPR) 443 mpz_sub_ui (bnd, bnd, 1); 444 445 if (lbound) 446 { 447 mpz_neg (bnd, bnd); 448 if (mpz_cmp (bnds->below, bnd) < 0) 449 mpz_set (bnds->below, bnd); 450 } 451 else 452 { 453 if (mpz_cmp (bnd, bnds->up) < 0) 454 mpz_set (bnds->up, bnd); 455 } 456 mpz_clear (bnd); 457 } 458 459 mpz_clear (loffx); 460 mpz_clear (loffy); 461 end: 462 mpz_clear (offc0); 463 mpz_clear (offc1); 464 } 465 466 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS. 467 The subtraction is considered to be performed in arbitrary precision, 468 without overflows. 469 470 We do not attempt to be too clever regarding the value ranges of X and 471 Y; most of the time, they are just integers or ssa names offsetted by 472 integer. However, we try to use the information contained in the 473 comparisons before the loop (usually created by loop header copying). */ 474 475 static void 476 bound_difference (struct loop *loop, tree x, tree y, bounds *bnds) 477 { 478 tree type = TREE_TYPE (x); 479 tree varx, vary; 480 mpz_t offx, offy; 481 mpz_t minx, maxx, miny, maxy; 482 int cnt = 0; 483 edge e; 484 basic_block bb; 485 tree c0, c1; 486 gimple cond; 487 enum tree_code cmp; 488 489 /* Get rid of unnecessary casts, but preserve the value of 490 the expressions. */ 491 STRIP_SIGN_NOPS (x); 492 STRIP_SIGN_NOPS (y); 493 494 mpz_init (bnds->below); 495 mpz_init (bnds->up); 496 mpz_init (offx); 497 mpz_init (offy); 498 split_to_var_and_offset (x, &varx, offx); 499 split_to_var_and_offset (y, &vary, offy); 500 501 if (!integer_zerop (varx) 502 && operand_equal_p (varx, vary, 0)) 503 { 504 /* Special case VARX == VARY -- we just need to compare the 505 offsets. The matters are a bit more complicated in the 506 case addition of offsets may wrap. */ 507 bound_difference_of_offsetted_base (type, offx, offy, bnds); 508 } 509 else 510 { 511 /* Otherwise, use the value ranges to determine the initial 512 estimates on below and up. */ 513 mpz_init (minx); 514 mpz_init (maxx); 515 mpz_init (miny); 516 mpz_init (maxy); 517 determine_value_range (loop, type, varx, offx, minx, maxx); 518 determine_value_range (loop, type, vary, offy, miny, maxy); 519 520 mpz_sub (bnds->below, minx, maxy); 521 mpz_sub (bnds->up, maxx, miny); 522 mpz_clear (minx); 523 mpz_clear (maxx); 524 mpz_clear (miny); 525 mpz_clear (maxy); 526 } 527 528 /* If both X and Y are constants, we cannot get any more precise. */ 529 if (integer_zerop (varx) && integer_zerop (vary)) 530 goto end; 531 532 /* Now walk the dominators of the loop header and use the entry 533 guards to refine the estimates. */ 534 for (bb = loop->header; 535 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; 536 bb = get_immediate_dominator (CDI_DOMINATORS, bb)) 537 { 538 if (!single_pred_p (bb)) 539 continue; 540 e = single_pred_edge (bb); 541 542 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) 543 continue; 544 545 cond = last_stmt (e->src); 546 c0 = gimple_cond_lhs (cond); 547 cmp = gimple_cond_code (cond); 548 c1 = gimple_cond_rhs (cond); 549 550 if (e->flags & EDGE_FALSE_VALUE) 551 cmp = invert_tree_comparison (cmp, false); 552 553 refine_bounds_using_guard (type, varx, offx, vary, offy, 554 c0, cmp, c1, bnds); 555 ++cnt; 556 } 557 558 end: 559 mpz_clear (offx); 560 mpz_clear (offy); 561 } 562 563 /* Update the bounds in BNDS that restrict the value of X to the bounds 564 that restrict the value of X + DELTA. X can be obtained as a 565 difference of two values in TYPE. */ 566 567 static void 568 bounds_add (bounds *bnds, const widest_int &delta, tree type) 569 { 570 mpz_t mdelta, max; 571 572 mpz_init (mdelta); 573 wi::to_mpz (delta, mdelta, SIGNED); 574 575 mpz_init (max); 576 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED); 577 578 mpz_add (bnds->up, bnds->up, mdelta); 579 mpz_add (bnds->below, bnds->below, mdelta); 580 581 if (mpz_cmp (bnds->up, max) > 0) 582 mpz_set (bnds->up, max); 583 584 mpz_neg (max, max); 585 if (mpz_cmp (bnds->below, max) < 0) 586 mpz_set (bnds->below, max); 587 588 mpz_clear (mdelta); 589 mpz_clear (max); 590 } 591 592 /* Update the bounds in BNDS that restrict the value of X to the bounds 593 that restrict the value of -X. */ 594 595 static void 596 bounds_negate (bounds *bnds) 597 { 598 mpz_t tmp; 599 600 mpz_init_set (tmp, bnds->up); 601 mpz_neg (bnds->up, bnds->below); 602 mpz_neg (bnds->below, tmp); 603 mpz_clear (tmp); 604 } 605 606 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */ 607 608 static tree 609 inverse (tree x, tree mask) 610 { 611 tree type = TREE_TYPE (x); 612 tree rslt; 613 unsigned ctr = tree_floor_log2 (mask); 614 615 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT) 616 { 617 unsigned HOST_WIDE_INT ix; 618 unsigned HOST_WIDE_INT imask; 619 unsigned HOST_WIDE_INT irslt = 1; 620 621 gcc_assert (cst_and_fits_in_hwi (x)); 622 gcc_assert (cst_and_fits_in_hwi (mask)); 623 624 ix = int_cst_value (x); 625 imask = int_cst_value (mask); 626 627 for (; ctr; ctr--) 628 { 629 irslt *= ix; 630 ix *= ix; 631 } 632 irslt &= imask; 633 634 rslt = build_int_cst_type (type, irslt); 635 } 636 else 637 { 638 rslt = build_int_cst (type, 1); 639 for (; ctr; ctr--) 640 { 641 rslt = int_const_binop (MULT_EXPR, rslt, x); 642 x = int_const_binop (MULT_EXPR, x, x); 643 } 644 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask); 645 } 646 647 return rslt; 648 } 649 650 /* Derives the upper bound BND on the number of executions of loop with exit 651 condition S * i <> C. If NO_OVERFLOW is true, then the control variable of 652 the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed 653 that the loop ends through this exit, i.e., the induction variable ever 654 reaches the value of C. 655 656 The value C is equal to final - base, where final and base are the final and 657 initial value of the actual induction variable in the analysed loop. BNDS 658 bounds the value of this difference when computed in signed type with 659 unbounded range, while the computation of C is performed in an unsigned 660 type with the range matching the range of the type of the induction variable. 661 In particular, BNDS.up contains an upper bound on C in the following cases: 662 -- if the iv must reach its final value without overflow, i.e., if 663 NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or 664 -- if final >= base, which we know to hold when BNDS.below >= 0. */ 665 666 static void 667 number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s, 668 bounds *bnds, bool exit_must_be_taken) 669 { 670 widest_int max; 671 mpz_t d; 672 tree type = TREE_TYPE (c); 673 bool bnds_u_valid = ((no_overflow && exit_must_be_taken) 674 || mpz_sgn (bnds->below) >= 0); 675 676 if (integer_onep (s) 677 || (TREE_CODE (c) == INTEGER_CST 678 && TREE_CODE (s) == INTEGER_CST 679 && wi::mod_trunc (c, s, TYPE_SIGN (type)) == 0) 680 || (TYPE_OVERFLOW_UNDEFINED (type) 681 && multiple_of_p (type, c, s))) 682 { 683 /* If C is an exact multiple of S, then its value will be reached before 684 the induction variable overflows (unless the loop is exited in some 685 other way before). Note that the actual induction variable in the 686 loop (which ranges from base to final instead of from 0 to C) may 687 overflow, in which case BNDS.up will not be giving a correct upper 688 bound on C; thus, BNDS_U_VALID had to be computed in advance. */ 689 no_overflow = true; 690 exit_must_be_taken = true; 691 } 692 693 /* If the induction variable can overflow, the number of iterations is at 694 most the period of the control variable (or infinite, but in that case 695 the whole # of iterations analysis will fail). */ 696 if (!no_overflow) 697 { 698 max = wi::mask <widest_int> (TYPE_PRECISION (type) - wi::ctz (s), false); 699 wi::to_mpz (max, bnd, UNSIGNED); 700 return; 701 } 702 703 /* Now we know that the induction variable does not overflow, so the loop 704 iterates at most (range of type / S) times. */ 705 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), bnd, UNSIGNED); 706 707 /* If the induction variable is guaranteed to reach the value of C before 708 overflow, ... */ 709 if (exit_must_be_taken) 710 { 711 /* ... then we can strengthen this to C / S, and possibly we can use 712 the upper bound on C given by BNDS. */ 713 if (TREE_CODE (c) == INTEGER_CST) 714 wi::to_mpz (c, bnd, UNSIGNED); 715 else if (bnds_u_valid) 716 mpz_set (bnd, bnds->up); 717 } 718 719 mpz_init (d); 720 wi::to_mpz (s, d, UNSIGNED); 721 mpz_fdiv_q (bnd, bnd, d); 722 mpz_clear (d); 723 } 724 725 /* Determines number of iterations of loop whose ending condition 726 is IV <> FINAL. TYPE is the type of the iv. The number of 727 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if 728 we know that the exit must be taken eventually, i.e., that the IV 729 ever reaches the value FINAL (we derived this earlier, and possibly set 730 NITER->assumptions to make sure this is the case). BNDS contains the 731 bounds on the difference FINAL - IV->base. */ 732 733 static bool 734 number_of_iterations_ne (tree type, affine_iv *iv, tree final, 735 struct tree_niter_desc *niter, bool exit_must_be_taken, 736 bounds *bnds) 737 { 738 tree niter_type = unsigned_type_for (type); 739 tree s, c, d, bits, assumption, tmp, bound; 740 mpz_t max; 741 742 niter->control = *iv; 743 niter->bound = final; 744 niter->cmp = NE_EXPR; 745 746 /* Rearrange the terms so that we get inequality S * i <> C, with S 747 positive. Also cast everything to the unsigned type. If IV does 748 not overflow, BNDS bounds the value of C. Also, this is the 749 case if the computation |FINAL - IV->base| does not overflow, i.e., 750 if BNDS->below in the result is nonnegative. */ 751 if (tree_int_cst_sign_bit (iv->step)) 752 { 753 s = fold_convert (niter_type, 754 fold_build1 (NEGATE_EXPR, type, iv->step)); 755 c = fold_build2 (MINUS_EXPR, niter_type, 756 fold_convert (niter_type, iv->base), 757 fold_convert (niter_type, final)); 758 bounds_negate (bnds); 759 } 760 else 761 { 762 s = fold_convert (niter_type, iv->step); 763 c = fold_build2 (MINUS_EXPR, niter_type, 764 fold_convert (niter_type, final), 765 fold_convert (niter_type, iv->base)); 766 } 767 768 mpz_init (max); 769 number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds, 770 exit_must_be_taken); 771 niter->max = widest_int::from (wi::from_mpz (niter_type, max, false), 772 TYPE_SIGN (niter_type)); 773 mpz_clear (max); 774 775 /* First the trivial cases -- when the step is 1. */ 776 if (integer_onep (s)) 777 { 778 niter->niter = c; 779 return true; 780 } 781 782 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop 783 is infinite. Otherwise, the number of iterations is 784 (inverse(s/d) * (c/d)) mod (size of mode/d). */ 785 bits = num_ending_zeros (s); 786 bound = build_low_bits_mask (niter_type, 787 (TYPE_PRECISION (niter_type) 788 - tree_to_uhwi (bits))); 789 790 d = fold_binary_to_constant (LSHIFT_EXPR, niter_type, 791 build_int_cst (niter_type, 1), bits); 792 s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits); 793 794 if (!exit_must_be_taken) 795 { 796 /* If we cannot assume that the exit is taken eventually, record the 797 assumptions for divisibility of c. */ 798 assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d); 799 assumption = fold_build2 (EQ_EXPR, boolean_type_node, 800 assumption, build_int_cst (niter_type, 0)); 801 if (!integer_nonzerop (assumption)) 802 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, 803 niter->assumptions, assumption); 804 } 805 806 c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d); 807 tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound)); 808 niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound); 809 return true; 810 } 811 812 /* Checks whether we can determine the final value of the control variable 813 of the loop with ending condition IV0 < IV1 (computed in TYPE). 814 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value 815 of the step. The assumptions necessary to ensure that the computation 816 of the final value does not overflow are recorded in NITER. If we 817 find the final value, we adjust DELTA and return TRUE. Otherwise 818 we return false. BNDS bounds the value of IV1->base - IV0->base, 819 and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is 820 true if we know that the exit must be taken eventually. */ 821 822 static bool 823 number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1, 824 struct tree_niter_desc *niter, 825 tree *delta, tree step, 826 bool exit_must_be_taken, bounds *bnds) 827 { 828 tree niter_type = TREE_TYPE (step); 829 tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step); 830 tree tmod; 831 mpz_t mmod; 832 tree assumption = boolean_true_node, bound, noloop; 833 bool ret = false, fv_comp_no_overflow; 834 tree type1 = type; 835 if (POINTER_TYPE_P (type)) 836 type1 = sizetype; 837 838 if (TREE_CODE (mod) != INTEGER_CST) 839 return false; 840 if (integer_nonzerop (mod)) 841 mod = fold_build2 (MINUS_EXPR, niter_type, step, mod); 842 tmod = fold_convert (type1, mod); 843 844 mpz_init (mmod); 845 wi::to_mpz (mod, mmod, UNSIGNED); 846 mpz_neg (mmod, mmod); 847 848 /* If the induction variable does not overflow and the exit is taken, 849 then the computation of the final value does not overflow. This is 850 also obviously the case if the new final value is equal to the 851 current one. Finally, we postulate this for pointer type variables, 852 as the code cannot rely on the object to that the pointer points being 853 placed at the end of the address space (and more pragmatically, 854 TYPE_{MIN,MAX}_VALUE is not defined for pointers). */ 855 if (integer_zerop (mod) || POINTER_TYPE_P (type)) 856 fv_comp_no_overflow = true; 857 else if (!exit_must_be_taken) 858 fv_comp_no_overflow = false; 859 else 860 fv_comp_no_overflow = 861 (iv0->no_overflow && integer_nonzerop (iv0->step)) 862 || (iv1->no_overflow && integer_nonzerop (iv1->step)); 863 864 if (integer_nonzerop (iv0->step)) 865 { 866 /* The final value of the iv is iv1->base + MOD, assuming that this 867 computation does not overflow, and that 868 iv0->base <= iv1->base + MOD. */ 869 if (!fv_comp_no_overflow) 870 { 871 bound = fold_build2 (MINUS_EXPR, type1, 872 TYPE_MAX_VALUE (type1), tmod); 873 assumption = fold_build2 (LE_EXPR, boolean_type_node, 874 iv1->base, bound); 875 if (integer_zerop (assumption)) 876 goto end; 877 } 878 if (mpz_cmp (mmod, bnds->below) < 0) 879 noloop = boolean_false_node; 880 else if (POINTER_TYPE_P (type)) 881 noloop = fold_build2 (GT_EXPR, boolean_type_node, 882 iv0->base, 883 fold_build_pointer_plus (iv1->base, tmod)); 884 else 885 noloop = fold_build2 (GT_EXPR, boolean_type_node, 886 iv0->base, 887 fold_build2 (PLUS_EXPR, type1, 888 iv1->base, tmod)); 889 } 890 else 891 { 892 /* The final value of the iv is iv0->base - MOD, assuming that this 893 computation does not overflow, and that 894 iv0->base - MOD <= iv1->base. */ 895 if (!fv_comp_no_overflow) 896 { 897 bound = fold_build2 (PLUS_EXPR, type1, 898 TYPE_MIN_VALUE (type1), tmod); 899 assumption = fold_build2 (GE_EXPR, boolean_type_node, 900 iv0->base, bound); 901 if (integer_zerop (assumption)) 902 goto end; 903 } 904 if (mpz_cmp (mmod, bnds->below) < 0) 905 noloop = boolean_false_node; 906 else if (POINTER_TYPE_P (type)) 907 noloop = fold_build2 (GT_EXPR, boolean_type_node, 908 fold_build_pointer_plus (iv0->base, 909 fold_build1 (NEGATE_EXPR, 910 type1, tmod)), 911 iv1->base); 912 else 913 noloop = fold_build2 (GT_EXPR, boolean_type_node, 914 fold_build2 (MINUS_EXPR, type1, 915 iv0->base, tmod), 916 iv1->base); 917 } 918 919 if (!integer_nonzerop (assumption)) 920 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, 921 niter->assumptions, 922 assumption); 923 if (!integer_zerop (noloop)) 924 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, 925 niter->may_be_zero, 926 noloop); 927 bounds_add (bnds, wi::to_widest (mod), type); 928 *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod); 929 930 ret = true; 931 end: 932 mpz_clear (mmod); 933 return ret; 934 } 935 936 /* Add assertions to NITER that ensure that the control variable of the loop 937 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1 938 are TYPE. Returns false if we can prove that there is an overflow, true 939 otherwise. STEP is the absolute value of the step. */ 940 941 static bool 942 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1, 943 struct tree_niter_desc *niter, tree step) 944 { 945 tree bound, d, assumption, diff; 946 tree niter_type = TREE_TYPE (step); 947 948 if (integer_nonzerop (iv0->step)) 949 { 950 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */ 951 if (iv0->no_overflow) 952 return true; 953 954 /* If iv0->base is a constant, we can determine the last value before 955 overflow precisely; otherwise we conservatively assume 956 MAX - STEP + 1. */ 957 958 if (TREE_CODE (iv0->base) == INTEGER_CST) 959 { 960 d = fold_build2 (MINUS_EXPR, niter_type, 961 fold_convert (niter_type, TYPE_MAX_VALUE (type)), 962 fold_convert (niter_type, iv0->base)); 963 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step); 964 } 965 else 966 diff = fold_build2 (MINUS_EXPR, niter_type, step, 967 build_int_cst (niter_type, 1)); 968 bound = fold_build2 (MINUS_EXPR, type, 969 TYPE_MAX_VALUE (type), fold_convert (type, diff)); 970 assumption = fold_build2 (LE_EXPR, boolean_type_node, 971 iv1->base, bound); 972 } 973 else 974 { 975 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */ 976 if (iv1->no_overflow) 977 return true; 978 979 if (TREE_CODE (iv1->base) == INTEGER_CST) 980 { 981 d = fold_build2 (MINUS_EXPR, niter_type, 982 fold_convert (niter_type, iv1->base), 983 fold_convert (niter_type, TYPE_MIN_VALUE (type))); 984 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step); 985 } 986 else 987 diff = fold_build2 (MINUS_EXPR, niter_type, step, 988 build_int_cst (niter_type, 1)); 989 bound = fold_build2 (PLUS_EXPR, type, 990 TYPE_MIN_VALUE (type), fold_convert (type, diff)); 991 assumption = fold_build2 (GE_EXPR, boolean_type_node, 992 iv0->base, bound); 993 } 994 995 if (integer_zerop (assumption)) 996 return false; 997 if (!integer_nonzerop (assumption)) 998 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, 999 niter->assumptions, assumption); 1000 1001 iv0->no_overflow = true; 1002 iv1->no_overflow = true; 1003 return true; 1004 } 1005 1006 /* Add an assumption to NITER that a loop whose ending condition 1007 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS 1008 bounds the value of IV1->base - IV0->base. */ 1009 1010 static void 1011 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1, 1012 struct tree_niter_desc *niter, bounds *bnds) 1013 { 1014 tree assumption = boolean_true_node, bound, diff; 1015 tree mbz, mbzl, mbzr, type1; 1016 bool rolls_p, no_overflow_p; 1017 widest_int dstep; 1018 mpz_t mstep, max; 1019 1020 /* We are going to compute the number of iterations as 1021 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned 1022 variant of TYPE. This formula only works if 1023 1024 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1 1025 1026 (where MAX is the maximum value of the unsigned variant of TYPE, and 1027 the computations in this formula are performed in full precision, 1028 i.e., without overflows). 1029 1030 Usually, for loops with exit condition iv0->base + step * i < iv1->base, 1031 we have a condition of the form iv0->base - step < iv1->base before the loop, 1032 and for loops iv0->base < iv1->base - step * i the condition 1033 iv0->base < iv1->base + step, due to loop header copying, which enable us 1034 to prove the lower bound. 1035 1036 The upper bound is more complicated. Unless the expressions for initial 1037 and final value themselves contain enough information, we usually cannot 1038 derive it from the context. */ 1039 1040 /* First check whether the answer does not follow from the bounds we gathered 1041 before. */ 1042 if (integer_nonzerop (iv0->step)) 1043 dstep = wi::to_widest (iv0->step); 1044 else 1045 { 1046 dstep = wi::sext (wi::to_widest (iv1->step), TYPE_PRECISION (type)); 1047 dstep = -dstep; 1048 } 1049 1050 mpz_init (mstep); 1051 wi::to_mpz (dstep, mstep, UNSIGNED); 1052 mpz_neg (mstep, mstep); 1053 mpz_add_ui (mstep, mstep, 1); 1054 1055 rolls_p = mpz_cmp (mstep, bnds->below) <= 0; 1056 1057 mpz_init (max); 1058 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED); 1059 mpz_add (max, max, mstep); 1060 no_overflow_p = (mpz_cmp (bnds->up, max) <= 0 1061 /* For pointers, only values lying inside a single object 1062 can be compared or manipulated by pointer arithmetics. 1063 Gcc in general does not allow or handle objects larger 1064 than half of the address space, hence the upper bound 1065 is satisfied for pointers. */ 1066 || POINTER_TYPE_P (type)); 1067 mpz_clear (mstep); 1068 mpz_clear (max); 1069 1070 if (rolls_p && no_overflow_p) 1071 return; 1072 1073 type1 = type; 1074 if (POINTER_TYPE_P (type)) 1075 type1 = sizetype; 1076 1077 /* Now the hard part; we must formulate the assumption(s) as expressions, and 1078 we must be careful not to introduce overflow. */ 1079 1080 if (integer_nonzerop (iv0->step)) 1081 { 1082 diff = fold_build2 (MINUS_EXPR, type1, 1083 iv0->step, build_int_cst (type1, 1)); 1084 1085 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since 1086 0 address never belongs to any object, we can assume this for 1087 pointers. */ 1088 if (!POINTER_TYPE_P (type)) 1089 { 1090 bound = fold_build2 (PLUS_EXPR, type1, 1091 TYPE_MIN_VALUE (type), diff); 1092 assumption = fold_build2 (GE_EXPR, boolean_type_node, 1093 iv0->base, bound); 1094 } 1095 1096 /* And then we can compute iv0->base - diff, and compare it with 1097 iv1->base. */ 1098 mbzl = fold_build2 (MINUS_EXPR, type1, 1099 fold_convert (type1, iv0->base), diff); 1100 mbzr = fold_convert (type1, iv1->base); 1101 } 1102 else 1103 { 1104 diff = fold_build2 (PLUS_EXPR, type1, 1105 iv1->step, build_int_cst (type1, 1)); 1106 1107 if (!POINTER_TYPE_P (type)) 1108 { 1109 bound = fold_build2 (PLUS_EXPR, type1, 1110 TYPE_MAX_VALUE (type), diff); 1111 assumption = fold_build2 (LE_EXPR, boolean_type_node, 1112 iv1->base, bound); 1113 } 1114 1115 mbzl = fold_convert (type1, iv0->base); 1116 mbzr = fold_build2 (MINUS_EXPR, type1, 1117 fold_convert (type1, iv1->base), diff); 1118 } 1119 1120 if (!integer_nonzerop (assumption)) 1121 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, 1122 niter->assumptions, assumption); 1123 if (!rolls_p) 1124 { 1125 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr); 1126 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, 1127 niter->may_be_zero, mbz); 1128 } 1129 } 1130 1131 /* Determines number of iterations of loop whose ending condition 1132 is IV0 < IV1. TYPE is the type of the iv. The number of 1133 iterations is stored to NITER. BNDS bounds the difference 1134 IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know 1135 that the exit must be taken eventually. */ 1136 1137 static bool 1138 number_of_iterations_lt (tree type, affine_iv *iv0, affine_iv *iv1, 1139 struct tree_niter_desc *niter, 1140 bool exit_must_be_taken, bounds *bnds) 1141 { 1142 tree niter_type = unsigned_type_for (type); 1143 tree delta, step, s; 1144 mpz_t mstep, tmp; 1145 1146 if (integer_nonzerop (iv0->step)) 1147 { 1148 niter->control = *iv0; 1149 niter->cmp = LT_EXPR; 1150 niter->bound = iv1->base; 1151 } 1152 else 1153 { 1154 niter->control = *iv1; 1155 niter->cmp = GT_EXPR; 1156 niter->bound = iv0->base; 1157 } 1158 1159 delta = fold_build2 (MINUS_EXPR, niter_type, 1160 fold_convert (niter_type, iv1->base), 1161 fold_convert (niter_type, iv0->base)); 1162 1163 /* First handle the special case that the step is +-1. */ 1164 if ((integer_onep (iv0->step) && integer_zerop (iv1->step)) 1165 || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step))) 1166 { 1167 /* for (i = iv0->base; i < iv1->base; i++) 1168 1169 or 1170 1171 for (i = iv1->base; i > iv0->base; i--). 1172 1173 In both cases # of iterations is iv1->base - iv0->base, assuming that 1174 iv1->base >= iv0->base. 1175 1176 First try to derive a lower bound on the value of 1177 iv1->base - iv0->base, computed in full precision. If the difference 1178 is nonnegative, we are done, otherwise we must record the 1179 condition. */ 1180 1181 if (mpz_sgn (bnds->below) < 0) 1182 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node, 1183 iv1->base, iv0->base); 1184 niter->niter = delta; 1185 niter->max = widest_int::from (wi::from_mpz (niter_type, bnds->up, false), 1186 TYPE_SIGN (niter_type)); 1187 return true; 1188 } 1189 1190 if (integer_nonzerop (iv0->step)) 1191 step = fold_convert (niter_type, iv0->step); 1192 else 1193 step = fold_convert (niter_type, 1194 fold_build1 (NEGATE_EXPR, type, iv1->step)); 1195 1196 /* If we can determine the final value of the control iv exactly, we can 1197 transform the condition to != comparison. In particular, this will be 1198 the case if DELTA is constant. */ 1199 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step, 1200 exit_must_be_taken, bnds)) 1201 { 1202 affine_iv zps; 1203 1204 zps.base = build_int_cst (niter_type, 0); 1205 zps.step = step; 1206 /* number_of_iterations_lt_to_ne will add assumptions that ensure that 1207 zps does not overflow. */ 1208 zps.no_overflow = true; 1209 1210 return number_of_iterations_ne (type, &zps, delta, niter, true, bnds); 1211 } 1212 1213 /* Make sure that the control iv does not overflow. */ 1214 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step)) 1215 return false; 1216 1217 /* We determine the number of iterations as (delta + step - 1) / step. For 1218 this to work, we must know that iv1->base >= iv0->base - step + 1, 1219 otherwise the loop does not roll. */ 1220 assert_loop_rolls_lt (type, iv0, iv1, niter, bnds); 1221 1222 s = fold_build2 (MINUS_EXPR, niter_type, 1223 step, build_int_cst (niter_type, 1)); 1224 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s); 1225 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step); 1226 1227 mpz_init (mstep); 1228 mpz_init (tmp); 1229 wi::to_mpz (step, mstep, UNSIGNED); 1230 mpz_add (tmp, bnds->up, mstep); 1231 mpz_sub_ui (tmp, tmp, 1); 1232 mpz_fdiv_q (tmp, tmp, mstep); 1233 niter->max = widest_int::from (wi::from_mpz (niter_type, tmp, false), 1234 TYPE_SIGN (niter_type)); 1235 mpz_clear (mstep); 1236 mpz_clear (tmp); 1237 1238 return true; 1239 } 1240 1241 /* Determines number of iterations of loop whose ending condition 1242 is IV0 <= IV1. TYPE is the type of the iv. The number of 1243 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if 1244 we know that this condition must eventually become false (we derived this 1245 earlier, and possibly set NITER->assumptions to make sure this 1246 is the case). BNDS bounds the difference IV1->base - IV0->base. */ 1247 1248 static bool 1249 number_of_iterations_le (tree type, affine_iv *iv0, affine_iv *iv1, 1250 struct tree_niter_desc *niter, bool exit_must_be_taken, 1251 bounds *bnds) 1252 { 1253 tree assumption; 1254 tree type1 = type; 1255 if (POINTER_TYPE_P (type)) 1256 type1 = sizetype; 1257 1258 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff 1259 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest 1260 value of the type. This we must know anyway, since if it is 1261 equal to this value, the loop rolls forever. We do not check 1262 this condition for pointer type ivs, as the code cannot rely on 1263 the object to that the pointer points being placed at the end of 1264 the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is 1265 not defined for pointers). */ 1266 1267 if (!exit_must_be_taken && !POINTER_TYPE_P (type)) 1268 { 1269 if (integer_nonzerop (iv0->step)) 1270 assumption = fold_build2 (NE_EXPR, boolean_type_node, 1271 iv1->base, TYPE_MAX_VALUE (type)); 1272 else 1273 assumption = fold_build2 (NE_EXPR, boolean_type_node, 1274 iv0->base, TYPE_MIN_VALUE (type)); 1275 1276 if (integer_zerop (assumption)) 1277 return false; 1278 if (!integer_nonzerop (assumption)) 1279 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, 1280 niter->assumptions, assumption); 1281 } 1282 1283 if (integer_nonzerop (iv0->step)) 1284 { 1285 if (POINTER_TYPE_P (type)) 1286 iv1->base = fold_build_pointer_plus_hwi (iv1->base, 1); 1287 else 1288 iv1->base = fold_build2 (PLUS_EXPR, type1, iv1->base, 1289 build_int_cst (type1, 1)); 1290 } 1291 else if (POINTER_TYPE_P (type)) 1292 iv0->base = fold_build_pointer_plus_hwi (iv0->base, -1); 1293 else 1294 iv0->base = fold_build2 (MINUS_EXPR, type1, 1295 iv0->base, build_int_cst (type1, 1)); 1296 1297 bounds_add (bnds, 1, type1); 1298 1299 return number_of_iterations_lt (type, iv0, iv1, niter, exit_must_be_taken, 1300 bnds); 1301 } 1302 1303 /* Dumps description of affine induction variable IV to FILE. */ 1304 1305 static void 1306 dump_affine_iv (FILE *file, affine_iv *iv) 1307 { 1308 if (!integer_zerop (iv->step)) 1309 fprintf (file, "["); 1310 1311 print_generic_expr (dump_file, iv->base, TDF_SLIM); 1312 1313 if (!integer_zerop (iv->step)) 1314 { 1315 fprintf (file, ", + , "); 1316 print_generic_expr (dump_file, iv->step, TDF_SLIM); 1317 fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : ""); 1318 } 1319 } 1320 1321 /* Determine the number of iterations according to condition (for staying 1322 inside loop) which compares two induction variables using comparison 1323 operator CODE. The induction variable on left side of the comparison 1324 is IV0, the right-hand side is IV1. Both induction variables must have 1325 type TYPE, which must be an integer or pointer type. The steps of the 1326 ivs must be constants (or NULL_TREE, which is interpreted as constant zero). 1327 1328 LOOP is the loop whose number of iterations we are determining. 1329 1330 ONLY_EXIT is true if we are sure this is the only way the loop could be 1331 exited (including possibly non-returning function calls, exceptions, etc.) 1332 -- in this case we can use the information whether the control induction 1333 variables can overflow or not in a more efficient way. 1334 1335 if EVERY_ITERATION is true, we know the test is executed on every iteration. 1336 1337 The results (number of iterations and assumptions as described in 1338 comments at struct tree_niter_desc in tree-ssa-loop.h) are stored to NITER. 1339 Returns false if it fails to determine number of iterations, true if it 1340 was determined (possibly with some assumptions). */ 1341 1342 static bool 1343 number_of_iterations_cond (struct loop *loop, 1344 tree type, affine_iv *iv0, enum tree_code code, 1345 affine_iv *iv1, struct tree_niter_desc *niter, 1346 bool only_exit, bool every_iteration) 1347 { 1348 bool exit_must_be_taken = false, ret; 1349 bounds bnds; 1350 1351 /* If the test is not executed every iteration, wrapping may make the test 1352 to pass again. 1353 TODO: the overflow case can be still used as unreliable estimate of upper 1354 bound. But we have no API to pass it down to number of iterations code 1355 and, at present, it will not use it anyway. */ 1356 if (!every_iteration 1357 && (!iv0->no_overflow || !iv1->no_overflow 1358 || code == NE_EXPR || code == EQ_EXPR)) 1359 return false; 1360 1361 /* The meaning of these assumptions is this: 1362 if !assumptions 1363 then the rest of information does not have to be valid 1364 if may_be_zero then the loop does not roll, even if 1365 niter != 0. */ 1366 niter->assumptions = boolean_true_node; 1367 niter->may_be_zero = boolean_false_node; 1368 niter->niter = NULL_TREE; 1369 niter->max = 0; 1370 niter->bound = NULL_TREE; 1371 niter->cmp = ERROR_MARK; 1372 1373 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that 1374 the control variable is on lhs. */ 1375 if (code == GE_EXPR || code == GT_EXPR 1376 || (code == NE_EXPR && integer_zerop (iv0->step))) 1377 { 1378 SWAP (iv0, iv1); 1379 code = swap_tree_comparison (code); 1380 } 1381 1382 if (POINTER_TYPE_P (type)) 1383 { 1384 /* Comparison of pointers is undefined unless both iv0 and iv1 point 1385 to the same object. If they do, the control variable cannot wrap 1386 (as wrap around the bounds of memory will never return a pointer 1387 that would be guaranteed to point to the same object, even if we 1388 avoid undefined behavior by casting to size_t and back). */ 1389 iv0->no_overflow = true; 1390 iv1->no_overflow = true; 1391 } 1392 1393 /* If the control induction variable does not overflow and the only exit 1394 from the loop is the one that we analyze, we know it must be taken 1395 eventually. */ 1396 if (only_exit) 1397 { 1398 if (!integer_zerop (iv0->step) && iv0->no_overflow) 1399 exit_must_be_taken = true; 1400 else if (!integer_zerop (iv1->step) && iv1->no_overflow) 1401 exit_must_be_taken = true; 1402 } 1403 1404 /* We can handle the case when neither of the sides of the comparison is 1405 invariant, provided that the test is NE_EXPR. This rarely occurs in 1406 practice, but it is simple enough to manage. */ 1407 if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step)) 1408 { 1409 tree step_type = POINTER_TYPE_P (type) ? sizetype : type; 1410 if (code != NE_EXPR) 1411 return false; 1412 1413 iv0->step = fold_binary_to_constant (MINUS_EXPR, step_type, 1414 iv0->step, iv1->step); 1415 iv0->no_overflow = false; 1416 iv1->step = build_int_cst (step_type, 0); 1417 iv1->no_overflow = true; 1418 } 1419 1420 /* If the result of the comparison is a constant, the loop is weird. More 1421 precise handling would be possible, but the situation is not common enough 1422 to waste time on it. */ 1423 if (integer_zerop (iv0->step) && integer_zerop (iv1->step)) 1424 return false; 1425 1426 /* Ignore loops of while (i-- < 10) type. */ 1427 if (code != NE_EXPR) 1428 { 1429 if (iv0->step && tree_int_cst_sign_bit (iv0->step)) 1430 return false; 1431 1432 if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step)) 1433 return false; 1434 } 1435 1436 /* If the loop exits immediately, there is nothing to do. */ 1437 tree tem = fold_binary (code, boolean_type_node, iv0->base, iv1->base); 1438 if (tem && integer_zerop (tem)) 1439 { 1440 niter->niter = build_int_cst (unsigned_type_for (type), 0); 1441 niter->max = 0; 1442 return true; 1443 } 1444 1445 /* OK, now we know we have a senseful loop. Handle several cases, depending 1446 on what comparison operator is used. */ 1447 bound_difference (loop, iv1->base, iv0->base, &bnds); 1448 1449 if (dump_file && (dump_flags & TDF_DETAILS)) 1450 { 1451 fprintf (dump_file, 1452 "Analyzing # of iterations of loop %d\n", loop->num); 1453 1454 fprintf (dump_file, " exit condition "); 1455 dump_affine_iv (dump_file, iv0); 1456 fprintf (dump_file, " %s ", 1457 code == NE_EXPR ? "!=" 1458 : code == LT_EXPR ? "<" 1459 : "<="); 1460 dump_affine_iv (dump_file, iv1); 1461 fprintf (dump_file, "\n"); 1462 1463 fprintf (dump_file, " bounds on difference of bases: "); 1464 mpz_out_str (dump_file, 10, bnds.below); 1465 fprintf (dump_file, " ... "); 1466 mpz_out_str (dump_file, 10, bnds.up); 1467 fprintf (dump_file, "\n"); 1468 } 1469 1470 switch (code) 1471 { 1472 case NE_EXPR: 1473 gcc_assert (integer_zerop (iv1->step)); 1474 ret = number_of_iterations_ne (type, iv0, iv1->base, niter, 1475 exit_must_be_taken, &bnds); 1476 break; 1477 1478 case LT_EXPR: 1479 ret = number_of_iterations_lt (type, iv0, iv1, niter, exit_must_be_taken, 1480 &bnds); 1481 break; 1482 1483 case LE_EXPR: 1484 ret = number_of_iterations_le (type, iv0, iv1, niter, exit_must_be_taken, 1485 &bnds); 1486 break; 1487 1488 default: 1489 gcc_unreachable (); 1490 } 1491 1492 mpz_clear (bnds.up); 1493 mpz_clear (bnds.below); 1494 1495 if (dump_file && (dump_flags & TDF_DETAILS)) 1496 { 1497 if (ret) 1498 { 1499 fprintf (dump_file, " result:\n"); 1500 if (!integer_nonzerop (niter->assumptions)) 1501 { 1502 fprintf (dump_file, " under assumptions "); 1503 print_generic_expr (dump_file, niter->assumptions, TDF_SLIM); 1504 fprintf (dump_file, "\n"); 1505 } 1506 1507 if (!integer_zerop (niter->may_be_zero)) 1508 { 1509 fprintf (dump_file, " zero if "); 1510 print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM); 1511 fprintf (dump_file, "\n"); 1512 } 1513 1514 fprintf (dump_file, " # of iterations "); 1515 print_generic_expr (dump_file, niter->niter, TDF_SLIM); 1516 fprintf (dump_file, ", bounded by "); 1517 print_decu (niter->max, dump_file); 1518 fprintf (dump_file, "\n"); 1519 } 1520 else 1521 fprintf (dump_file, " failed\n\n"); 1522 } 1523 return ret; 1524 } 1525 1526 /* Substitute NEW for OLD in EXPR and fold the result. */ 1527 1528 static tree 1529 simplify_replace_tree (tree expr, tree old, tree new_tree) 1530 { 1531 unsigned i, n; 1532 tree ret = NULL_TREE, e, se; 1533 1534 if (!expr) 1535 return NULL_TREE; 1536 1537 /* Do not bother to replace constants. */ 1538 if (CONSTANT_CLASS_P (old)) 1539 return expr; 1540 1541 if (expr == old 1542 || operand_equal_p (expr, old, 0)) 1543 return unshare_expr (new_tree); 1544 1545 if (!EXPR_P (expr)) 1546 return expr; 1547 1548 n = TREE_OPERAND_LENGTH (expr); 1549 for (i = 0; i < n; i++) 1550 { 1551 e = TREE_OPERAND (expr, i); 1552 se = simplify_replace_tree (e, old, new_tree); 1553 if (e == se) 1554 continue; 1555 1556 if (!ret) 1557 ret = copy_node (expr); 1558 1559 TREE_OPERAND (ret, i) = se; 1560 } 1561 1562 return (ret ? fold (ret) : expr); 1563 } 1564 1565 /* Expand definitions of ssa names in EXPR as long as they are simple 1566 enough, and return the new expression. If STOP is specified, stop 1567 expanding if EXPR equals to it. */ 1568 1569 tree 1570 expand_simple_operations (tree expr, tree stop) 1571 { 1572 unsigned i, n; 1573 tree ret = NULL_TREE, e, ee, e1; 1574 enum tree_code code; 1575 gimple stmt; 1576 1577 if (expr == NULL_TREE) 1578 return expr; 1579 1580 if (is_gimple_min_invariant (expr)) 1581 return expr; 1582 1583 code = TREE_CODE (expr); 1584 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) 1585 { 1586 n = TREE_OPERAND_LENGTH (expr); 1587 for (i = 0; i < n; i++) 1588 { 1589 e = TREE_OPERAND (expr, i); 1590 ee = expand_simple_operations (e, stop); 1591 if (e == ee) 1592 continue; 1593 1594 if (!ret) 1595 ret = copy_node (expr); 1596 1597 TREE_OPERAND (ret, i) = ee; 1598 } 1599 1600 if (!ret) 1601 return expr; 1602 1603 fold_defer_overflow_warnings (); 1604 ret = fold (ret); 1605 fold_undefer_and_ignore_overflow_warnings (); 1606 return ret; 1607 } 1608 1609 /* Stop if it's not ssa name or the one we don't want to expand. */ 1610 if (TREE_CODE (expr) != SSA_NAME || expr == stop) 1611 return expr; 1612 1613 stmt = SSA_NAME_DEF_STMT (expr); 1614 if (gimple_code (stmt) == GIMPLE_PHI) 1615 { 1616 basic_block src, dest; 1617 1618 if (gimple_phi_num_args (stmt) != 1) 1619 return expr; 1620 e = PHI_ARG_DEF (stmt, 0); 1621 1622 /* Avoid propagating through loop exit phi nodes, which 1623 could break loop-closed SSA form restrictions. */ 1624 dest = gimple_bb (stmt); 1625 src = single_pred (dest); 1626 if (TREE_CODE (e) == SSA_NAME 1627 && src->loop_father != dest->loop_father) 1628 return expr; 1629 1630 return expand_simple_operations (e, stop); 1631 } 1632 if (gimple_code (stmt) != GIMPLE_ASSIGN) 1633 return expr; 1634 1635 /* Avoid expanding to expressions that contain SSA names that need 1636 to take part in abnormal coalescing. */ 1637 ssa_op_iter iter; 1638 FOR_EACH_SSA_TREE_OPERAND (e, stmt, iter, SSA_OP_USE) 1639 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e)) 1640 return expr; 1641 1642 e = gimple_assign_rhs1 (stmt); 1643 code = gimple_assign_rhs_code (stmt); 1644 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) 1645 { 1646 if (is_gimple_min_invariant (e)) 1647 return e; 1648 1649 if (code == SSA_NAME) 1650 return expand_simple_operations (e, stop); 1651 1652 return expr; 1653 } 1654 1655 switch (code) 1656 { 1657 CASE_CONVERT: 1658 /* Casts are simple. */ 1659 ee = expand_simple_operations (e, stop); 1660 return fold_build1 (code, TREE_TYPE (expr), ee); 1661 1662 case PLUS_EXPR: 1663 case MINUS_EXPR: 1664 if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (expr)) 1665 && TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr))) 1666 return expr; 1667 /* Fallthru. */ 1668 case POINTER_PLUS_EXPR: 1669 /* And increments and decrements by a constant are simple. */ 1670 e1 = gimple_assign_rhs2 (stmt); 1671 if (!is_gimple_min_invariant (e1)) 1672 return expr; 1673 1674 ee = expand_simple_operations (e, stop); 1675 return fold_build2 (code, TREE_TYPE (expr), ee, e1); 1676 1677 default: 1678 return expr; 1679 } 1680 } 1681 1682 /* Tries to simplify EXPR using the condition COND. Returns the simplified 1683 expression (or EXPR unchanged, if no simplification was possible). */ 1684 1685 static tree 1686 tree_simplify_using_condition_1 (tree cond, tree expr) 1687 { 1688 bool changed; 1689 tree e, te, e0, e1, e2, notcond; 1690 enum tree_code code = TREE_CODE (expr); 1691 1692 if (code == INTEGER_CST) 1693 return expr; 1694 1695 if (code == TRUTH_OR_EXPR 1696 || code == TRUTH_AND_EXPR 1697 || code == COND_EXPR) 1698 { 1699 changed = false; 1700 1701 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0)); 1702 if (TREE_OPERAND (expr, 0) != e0) 1703 changed = true; 1704 1705 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1)); 1706 if (TREE_OPERAND (expr, 1) != e1) 1707 changed = true; 1708 1709 if (code == COND_EXPR) 1710 { 1711 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2)); 1712 if (TREE_OPERAND (expr, 2) != e2) 1713 changed = true; 1714 } 1715 else 1716 e2 = NULL_TREE; 1717 1718 if (changed) 1719 { 1720 if (code == COND_EXPR) 1721 expr = fold_build3 (code, boolean_type_node, e0, e1, e2); 1722 else 1723 expr = fold_build2 (code, boolean_type_node, e0, e1); 1724 } 1725 1726 return expr; 1727 } 1728 1729 /* In case COND is equality, we may be able to simplify EXPR by copy/constant 1730 propagation, and vice versa. Fold does not handle this, since it is 1731 considered too expensive. */ 1732 if (TREE_CODE (cond) == EQ_EXPR) 1733 { 1734 e0 = TREE_OPERAND (cond, 0); 1735 e1 = TREE_OPERAND (cond, 1); 1736 1737 /* We know that e0 == e1. Check whether we cannot simplify expr 1738 using this fact. */ 1739 e = simplify_replace_tree (expr, e0, e1); 1740 if (integer_zerop (e) || integer_nonzerop (e)) 1741 return e; 1742 1743 e = simplify_replace_tree (expr, e1, e0); 1744 if (integer_zerop (e) || integer_nonzerop (e)) 1745 return e; 1746 } 1747 if (TREE_CODE (expr) == EQ_EXPR) 1748 { 1749 e0 = TREE_OPERAND (expr, 0); 1750 e1 = TREE_OPERAND (expr, 1); 1751 1752 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */ 1753 e = simplify_replace_tree (cond, e0, e1); 1754 if (integer_zerop (e)) 1755 return e; 1756 e = simplify_replace_tree (cond, e1, e0); 1757 if (integer_zerop (e)) 1758 return e; 1759 } 1760 if (TREE_CODE (expr) == NE_EXPR) 1761 { 1762 e0 = TREE_OPERAND (expr, 0); 1763 e1 = TREE_OPERAND (expr, 1); 1764 1765 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */ 1766 e = simplify_replace_tree (cond, e0, e1); 1767 if (integer_zerop (e)) 1768 return boolean_true_node; 1769 e = simplify_replace_tree (cond, e1, e0); 1770 if (integer_zerop (e)) 1771 return boolean_true_node; 1772 } 1773 1774 te = expand_simple_operations (expr); 1775 1776 /* Check whether COND ==> EXPR. */ 1777 notcond = invert_truthvalue (cond); 1778 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, te); 1779 if (e && integer_nonzerop (e)) 1780 return e; 1781 1782 /* Check whether COND ==> not EXPR. */ 1783 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, te); 1784 if (e && integer_zerop (e)) 1785 return e; 1786 1787 return expr; 1788 } 1789 1790 /* Tries to simplify EXPR using the condition COND. Returns the simplified 1791 expression (or EXPR unchanged, if no simplification was possible). 1792 Wrapper around tree_simplify_using_condition_1 that ensures that chains 1793 of simple operations in definitions of ssa names in COND are expanded, 1794 so that things like casts or incrementing the value of the bound before 1795 the loop do not cause us to fail. */ 1796 1797 static tree 1798 tree_simplify_using_condition (tree cond, tree expr) 1799 { 1800 cond = expand_simple_operations (cond); 1801 1802 return tree_simplify_using_condition_1 (cond, expr); 1803 } 1804 1805 /* Tries to simplify EXPR using the conditions on entry to LOOP. 1806 Returns the simplified expression (or EXPR unchanged, if no 1807 simplification was possible).*/ 1808 1809 static tree 1810 simplify_using_initial_conditions (struct loop *loop, tree expr) 1811 { 1812 edge e; 1813 basic_block bb; 1814 gimple stmt; 1815 tree cond; 1816 int cnt = 0; 1817 1818 if (TREE_CODE (expr) == INTEGER_CST) 1819 return expr; 1820 1821 /* Limit walking the dominators to avoid quadraticness in 1822 the number of BBs times the number of loops in degenerate 1823 cases. */ 1824 for (bb = loop->header; 1825 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; 1826 bb = get_immediate_dominator (CDI_DOMINATORS, bb)) 1827 { 1828 if (!single_pred_p (bb)) 1829 continue; 1830 e = single_pred_edge (bb); 1831 1832 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) 1833 continue; 1834 1835 stmt = last_stmt (e->src); 1836 cond = fold_build2 (gimple_cond_code (stmt), 1837 boolean_type_node, 1838 gimple_cond_lhs (stmt), 1839 gimple_cond_rhs (stmt)); 1840 if (e->flags & EDGE_FALSE_VALUE) 1841 cond = invert_truthvalue (cond); 1842 expr = tree_simplify_using_condition (cond, expr); 1843 ++cnt; 1844 } 1845 1846 return expr; 1847 } 1848 1849 /* Tries to simplify EXPR using the evolutions of the loop invariants 1850 in the superloops of LOOP. Returns the simplified expression 1851 (or EXPR unchanged, if no simplification was possible). */ 1852 1853 static tree 1854 simplify_using_outer_evolutions (struct loop *loop, tree expr) 1855 { 1856 enum tree_code code = TREE_CODE (expr); 1857 bool changed; 1858 tree e, e0, e1, e2; 1859 1860 if (is_gimple_min_invariant (expr)) 1861 return expr; 1862 1863 if (code == TRUTH_OR_EXPR 1864 || code == TRUTH_AND_EXPR 1865 || code == COND_EXPR) 1866 { 1867 changed = false; 1868 1869 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0)); 1870 if (TREE_OPERAND (expr, 0) != e0) 1871 changed = true; 1872 1873 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1)); 1874 if (TREE_OPERAND (expr, 1) != e1) 1875 changed = true; 1876 1877 if (code == COND_EXPR) 1878 { 1879 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2)); 1880 if (TREE_OPERAND (expr, 2) != e2) 1881 changed = true; 1882 } 1883 else 1884 e2 = NULL_TREE; 1885 1886 if (changed) 1887 { 1888 if (code == COND_EXPR) 1889 expr = fold_build3 (code, boolean_type_node, e0, e1, e2); 1890 else 1891 expr = fold_build2 (code, boolean_type_node, e0, e1); 1892 } 1893 1894 return expr; 1895 } 1896 1897 e = instantiate_parameters (loop, expr); 1898 if (is_gimple_min_invariant (e)) 1899 return e; 1900 1901 return expr; 1902 } 1903 1904 /* Returns true if EXIT is the only possible exit from LOOP. */ 1905 1906 bool 1907 loop_only_exit_p (const struct loop *loop, const_edge exit) 1908 { 1909 basic_block *body; 1910 gimple_stmt_iterator bsi; 1911 unsigned i; 1912 gimple call; 1913 1914 if (exit != single_exit (loop)) 1915 return false; 1916 1917 body = get_loop_body (loop); 1918 for (i = 0; i < loop->num_nodes; i++) 1919 { 1920 for (bsi = gsi_start_bb (body[i]); !gsi_end_p (bsi); gsi_next (&bsi)) 1921 { 1922 call = gsi_stmt (bsi); 1923 if (gimple_code (call) != GIMPLE_CALL) 1924 continue; 1925 1926 if (gimple_has_side_effects (call)) 1927 { 1928 free (body); 1929 return false; 1930 } 1931 } 1932 } 1933 1934 free (body); 1935 return true; 1936 } 1937 1938 /* Stores description of number of iterations of LOOP derived from 1939 EXIT (an exit edge of the LOOP) in NITER. Returns true if some 1940 useful information could be derived (and fields of NITER has 1941 meaning described in comments at struct tree_niter_desc 1942 declaration), false otherwise. If WARN is true and 1943 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use 1944 potentially unsafe assumptions. 1945 When EVERY_ITERATION is true, only tests that are known to be executed 1946 every iteration are considered (i.e. only test that alone bounds the loop). 1947 */ 1948 1949 bool 1950 number_of_iterations_exit (struct loop *loop, edge exit, 1951 struct tree_niter_desc *niter, 1952 bool warn, bool every_iteration) 1953 { 1954 gimple last; 1955 gcond *stmt; 1956 tree type; 1957 tree op0, op1; 1958 enum tree_code code; 1959 affine_iv iv0, iv1; 1960 bool safe; 1961 1962 safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src); 1963 1964 if (every_iteration && !safe) 1965 return false; 1966 1967 niter->assumptions = boolean_false_node; 1968 last = last_stmt (exit->src); 1969 if (!last) 1970 return false; 1971 stmt = dyn_cast <gcond *> (last); 1972 if (!stmt) 1973 return false; 1974 1975 /* We want the condition for staying inside loop. */ 1976 code = gimple_cond_code (stmt); 1977 if (exit->flags & EDGE_TRUE_VALUE) 1978 code = invert_tree_comparison (code, false); 1979 1980 switch (code) 1981 { 1982 case GT_EXPR: 1983 case GE_EXPR: 1984 case LT_EXPR: 1985 case LE_EXPR: 1986 case NE_EXPR: 1987 break; 1988 1989 default: 1990 return false; 1991 } 1992 1993 op0 = gimple_cond_lhs (stmt); 1994 op1 = gimple_cond_rhs (stmt); 1995 type = TREE_TYPE (op0); 1996 1997 if (TREE_CODE (type) != INTEGER_TYPE 1998 && !POINTER_TYPE_P (type)) 1999 return false; 2000 2001 if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, false)) 2002 return false; 2003 if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, false)) 2004 return false; 2005 2006 /* We don't want to see undefined signed overflow warnings while 2007 computing the number of iterations. */ 2008 fold_defer_overflow_warnings (); 2009 2010 iv0.base = expand_simple_operations (iv0.base); 2011 iv1.base = expand_simple_operations (iv1.base); 2012 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter, 2013 loop_only_exit_p (loop, exit), safe)) 2014 { 2015 fold_undefer_and_ignore_overflow_warnings (); 2016 return false; 2017 } 2018 2019 if (optimize >= 3) 2020 { 2021 niter->assumptions = simplify_using_outer_evolutions (loop, 2022 niter->assumptions); 2023 niter->may_be_zero = simplify_using_outer_evolutions (loop, 2024 niter->may_be_zero); 2025 niter->niter = simplify_using_outer_evolutions (loop, niter->niter); 2026 } 2027 2028 niter->assumptions 2029 = simplify_using_initial_conditions (loop, 2030 niter->assumptions); 2031 niter->may_be_zero 2032 = simplify_using_initial_conditions (loop, 2033 niter->may_be_zero); 2034 2035 fold_undefer_and_ignore_overflow_warnings (); 2036 2037 /* If NITER has simplified into a constant, update MAX. */ 2038 if (TREE_CODE (niter->niter) == INTEGER_CST) 2039 niter->max = wi::to_widest (niter->niter); 2040 2041 if (integer_onep (niter->assumptions)) 2042 return true; 2043 2044 /* With -funsafe-loop-optimizations we assume that nothing bad can happen. 2045 But if we can prove that there is overflow or some other source of weird 2046 behavior, ignore the loop even with -funsafe-loop-optimizations. */ 2047 if (integer_zerop (niter->assumptions) || !single_exit (loop)) 2048 return false; 2049 2050 if (flag_unsafe_loop_optimizations) 2051 niter->assumptions = boolean_true_node; 2052 2053 if (warn) 2054 { 2055 const char *wording; 2056 location_t loc = gimple_location (stmt); 2057 2058 /* We can provide a more specific warning if one of the operator is 2059 constant and the other advances by +1 or -1. */ 2060 if (!integer_zerop (iv1.step) 2061 ? (integer_zerop (iv0.step) 2062 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step))) 2063 : (integer_onep (iv0.step) || integer_all_onesp (iv0.step))) 2064 wording = 2065 flag_unsafe_loop_optimizations 2066 ? N_("assuming that the loop is not infinite") 2067 : N_("cannot optimize possibly infinite loops"); 2068 else 2069 wording = 2070 flag_unsafe_loop_optimizations 2071 ? N_("assuming that the loop counter does not overflow") 2072 : N_("cannot optimize loop, the loop counter may overflow"); 2073 2074 warning_at ((LOCATION_LINE (loc) > 0) ? loc : input_location, 2075 OPT_Wunsafe_loop_optimizations, "%s", gettext (wording)); 2076 } 2077 2078 return flag_unsafe_loop_optimizations; 2079 } 2080 2081 /* Try to determine the number of iterations of LOOP. If we succeed, 2082 expression giving number of iterations is returned and *EXIT is 2083 set to the edge from that the information is obtained. Otherwise 2084 chrec_dont_know is returned. */ 2085 2086 tree 2087 find_loop_niter (struct loop *loop, edge *exit) 2088 { 2089 unsigned i; 2090 vec<edge> exits = get_loop_exit_edges (loop); 2091 edge ex; 2092 tree niter = NULL_TREE, aniter; 2093 struct tree_niter_desc desc; 2094 2095 *exit = NULL; 2096 FOR_EACH_VEC_ELT (exits, i, ex) 2097 { 2098 if (!number_of_iterations_exit (loop, ex, &desc, false)) 2099 continue; 2100 2101 if (integer_nonzerop (desc.may_be_zero)) 2102 { 2103 /* We exit in the first iteration through this exit. 2104 We won't find anything better. */ 2105 niter = build_int_cst (unsigned_type_node, 0); 2106 *exit = ex; 2107 break; 2108 } 2109 2110 if (!integer_zerop (desc.may_be_zero)) 2111 continue; 2112 2113 aniter = desc.niter; 2114 2115 if (!niter) 2116 { 2117 /* Nothing recorded yet. */ 2118 niter = aniter; 2119 *exit = ex; 2120 continue; 2121 } 2122 2123 /* Prefer constants, the lower the better. */ 2124 if (TREE_CODE (aniter) != INTEGER_CST) 2125 continue; 2126 2127 if (TREE_CODE (niter) != INTEGER_CST) 2128 { 2129 niter = aniter; 2130 *exit = ex; 2131 continue; 2132 } 2133 2134 if (tree_int_cst_lt (aniter, niter)) 2135 { 2136 niter = aniter; 2137 *exit = ex; 2138 continue; 2139 } 2140 } 2141 exits.release (); 2142 2143 return niter ? niter : chrec_dont_know; 2144 } 2145 2146 /* Return true if loop is known to have bounded number of iterations. */ 2147 2148 bool 2149 finite_loop_p (struct loop *loop) 2150 { 2151 widest_int nit; 2152 int flags; 2153 2154 if (flag_unsafe_loop_optimizations) 2155 return true; 2156 flags = flags_from_decl_or_type (current_function_decl); 2157 if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE)) 2158 { 2159 if (dump_file && (dump_flags & TDF_DETAILS)) 2160 fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n", 2161 loop->num); 2162 return true; 2163 } 2164 2165 if (loop->any_upper_bound 2166 || max_loop_iterations (loop, &nit)) 2167 { 2168 if (dump_file && (dump_flags & TDF_DETAILS)) 2169 fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n", 2170 loop->num); 2171 return true; 2172 } 2173 return false; 2174 } 2175 2176 /* 2177 2178 Analysis of a number of iterations of a loop by a brute-force evaluation. 2179 2180 */ 2181 2182 /* Bound on the number of iterations we try to evaluate. */ 2183 2184 #define MAX_ITERATIONS_TO_TRACK \ 2185 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK)) 2186 2187 /* Returns the loop phi node of LOOP such that ssa name X is derived from its 2188 result by a chain of operations such that all but exactly one of their 2189 operands are constants. */ 2190 2191 static gphi * 2192 chain_of_csts_start (struct loop *loop, tree x) 2193 { 2194 gimple stmt = SSA_NAME_DEF_STMT (x); 2195 tree use; 2196 basic_block bb = gimple_bb (stmt); 2197 enum tree_code code; 2198 2199 if (!bb 2200 || !flow_bb_inside_loop_p (loop, bb)) 2201 return NULL; 2202 2203 if (gimple_code (stmt) == GIMPLE_PHI) 2204 { 2205 if (bb == loop->header) 2206 return as_a <gphi *> (stmt); 2207 2208 return NULL; 2209 } 2210 2211 if (gimple_code (stmt) != GIMPLE_ASSIGN 2212 || gimple_assign_rhs_class (stmt) == GIMPLE_TERNARY_RHS) 2213 return NULL; 2214 2215 code = gimple_assign_rhs_code (stmt); 2216 if (gimple_references_memory_p (stmt) 2217 || TREE_CODE_CLASS (code) == tcc_reference 2218 || (code == ADDR_EXPR 2219 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))) 2220 return NULL; 2221 2222 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE); 2223 if (use == NULL_TREE) 2224 return NULL; 2225 2226 return chain_of_csts_start (loop, use); 2227 } 2228 2229 /* Determines whether the expression X is derived from a result of a phi node 2230 in header of LOOP such that 2231 2232 * the derivation of X consists only from operations with constants 2233 * the initial value of the phi node is constant 2234 * the value of the phi node in the next iteration can be derived from the 2235 value in the current iteration by a chain of operations with constants. 2236 2237 If such phi node exists, it is returned, otherwise NULL is returned. */ 2238 2239 static gphi * 2240 get_base_for (struct loop *loop, tree x) 2241 { 2242 gphi *phi; 2243 tree init, next; 2244 2245 if (is_gimple_min_invariant (x)) 2246 return NULL; 2247 2248 phi = chain_of_csts_start (loop, x); 2249 if (!phi) 2250 return NULL; 2251 2252 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); 2253 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); 2254 2255 if (TREE_CODE (next) != SSA_NAME) 2256 return NULL; 2257 2258 if (!is_gimple_min_invariant (init)) 2259 return NULL; 2260 2261 if (chain_of_csts_start (loop, next) != phi) 2262 return NULL; 2263 2264 return phi; 2265 } 2266 2267 /* Given an expression X, then 2268 2269 * if X is NULL_TREE, we return the constant BASE. 2270 * otherwise X is a SSA name, whose value in the considered loop is derived 2271 by a chain of operations with constant from a result of a phi node in 2272 the header of the loop. Then we return value of X when the value of the 2273 result of this phi node is given by the constant BASE. */ 2274 2275 static tree 2276 get_val_for (tree x, tree base) 2277 { 2278 gimple stmt; 2279 2280 gcc_checking_assert (is_gimple_min_invariant (base)); 2281 2282 if (!x) 2283 return base; 2284 2285 stmt = SSA_NAME_DEF_STMT (x); 2286 if (gimple_code (stmt) == GIMPLE_PHI) 2287 return base; 2288 2289 gcc_checking_assert (is_gimple_assign (stmt)); 2290 2291 /* STMT must be either an assignment of a single SSA name or an 2292 expression involving an SSA name and a constant. Try to fold that 2293 expression using the value for the SSA name. */ 2294 if (gimple_assign_ssa_name_copy_p (stmt)) 2295 return get_val_for (gimple_assign_rhs1 (stmt), base); 2296 else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS 2297 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME) 2298 { 2299 return fold_build1 (gimple_assign_rhs_code (stmt), 2300 gimple_expr_type (stmt), 2301 get_val_for (gimple_assign_rhs1 (stmt), base)); 2302 } 2303 else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS) 2304 { 2305 tree rhs1 = gimple_assign_rhs1 (stmt); 2306 tree rhs2 = gimple_assign_rhs2 (stmt); 2307 if (TREE_CODE (rhs1) == SSA_NAME) 2308 rhs1 = get_val_for (rhs1, base); 2309 else if (TREE_CODE (rhs2) == SSA_NAME) 2310 rhs2 = get_val_for (rhs2, base); 2311 else 2312 gcc_unreachable (); 2313 return fold_build2 (gimple_assign_rhs_code (stmt), 2314 gimple_expr_type (stmt), rhs1, rhs2); 2315 } 2316 else 2317 gcc_unreachable (); 2318 } 2319 2320 2321 /* Tries to count the number of iterations of LOOP till it exits by EXIT 2322 by brute force -- i.e. by determining the value of the operands of the 2323 condition at EXIT in first few iterations of the loop (assuming that 2324 these values are constant) and determining the first one in that the 2325 condition is not satisfied. Returns the constant giving the number 2326 of the iterations of LOOP if successful, chrec_dont_know otherwise. */ 2327 2328 tree 2329 loop_niter_by_eval (struct loop *loop, edge exit) 2330 { 2331 tree acnd; 2332 tree op[2], val[2], next[2], aval[2]; 2333 gphi *phi; 2334 gimple cond; 2335 unsigned i, j; 2336 enum tree_code cmp; 2337 2338 cond = last_stmt (exit->src); 2339 if (!cond || gimple_code (cond) != GIMPLE_COND) 2340 return chrec_dont_know; 2341 2342 cmp = gimple_cond_code (cond); 2343 if (exit->flags & EDGE_TRUE_VALUE) 2344 cmp = invert_tree_comparison (cmp, false); 2345 2346 switch (cmp) 2347 { 2348 case EQ_EXPR: 2349 case NE_EXPR: 2350 case GT_EXPR: 2351 case GE_EXPR: 2352 case LT_EXPR: 2353 case LE_EXPR: 2354 op[0] = gimple_cond_lhs (cond); 2355 op[1] = gimple_cond_rhs (cond); 2356 break; 2357 2358 default: 2359 return chrec_dont_know; 2360 } 2361 2362 for (j = 0; j < 2; j++) 2363 { 2364 if (is_gimple_min_invariant (op[j])) 2365 { 2366 val[j] = op[j]; 2367 next[j] = NULL_TREE; 2368 op[j] = NULL_TREE; 2369 } 2370 else 2371 { 2372 phi = get_base_for (loop, op[j]); 2373 if (!phi) 2374 return chrec_dont_know; 2375 val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); 2376 next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); 2377 } 2378 } 2379 2380 /* Don't issue signed overflow warnings. */ 2381 fold_defer_overflow_warnings (); 2382 2383 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++) 2384 { 2385 for (j = 0; j < 2; j++) 2386 aval[j] = get_val_for (op[j], val[j]); 2387 2388 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]); 2389 if (acnd && integer_zerop (acnd)) 2390 { 2391 fold_undefer_and_ignore_overflow_warnings (); 2392 if (dump_file && (dump_flags & TDF_DETAILS)) 2393 fprintf (dump_file, 2394 "Proved that loop %d iterates %d times using brute force.\n", 2395 loop->num, i); 2396 return build_int_cst (unsigned_type_node, i); 2397 } 2398 2399 for (j = 0; j < 2; j++) 2400 { 2401 val[j] = get_val_for (next[j], val[j]); 2402 if (!is_gimple_min_invariant (val[j])) 2403 { 2404 fold_undefer_and_ignore_overflow_warnings (); 2405 return chrec_dont_know; 2406 } 2407 } 2408 } 2409 2410 fold_undefer_and_ignore_overflow_warnings (); 2411 2412 return chrec_dont_know; 2413 } 2414 2415 /* Finds the exit of the LOOP by that the loop exits after a constant 2416 number of iterations and stores the exit edge to *EXIT. The constant 2417 giving the number of iterations of LOOP is returned. The number of 2418 iterations is determined using loop_niter_by_eval (i.e. by brute force 2419 evaluation). If we are unable to find the exit for that loop_niter_by_eval 2420 determines the number of iterations, chrec_dont_know is returned. */ 2421 2422 tree 2423 find_loop_niter_by_eval (struct loop *loop, edge *exit) 2424 { 2425 unsigned i; 2426 vec<edge> exits = get_loop_exit_edges (loop); 2427 edge ex; 2428 tree niter = NULL_TREE, aniter; 2429 2430 *exit = NULL; 2431 2432 /* Loops with multiple exits are expensive to handle and less important. */ 2433 if (!flag_expensive_optimizations 2434 && exits.length () > 1) 2435 { 2436 exits.release (); 2437 return chrec_dont_know; 2438 } 2439 2440 FOR_EACH_VEC_ELT (exits, i, ex) 2441 { 2442 if (!just_once_each_iteration_p (loop, ex->src)) 2443 continue; 2444 2445 aniter = loop_niter_by_eval (loop, ex); 2446 if (chrec_contains_undetermined (aniter)) 2447 continue; 2448 2449 if (niter 2450 && !tree_int_cst_lt (aniter, niter)) 2451 continue; 2452 2453 niter = aniter; 2454 *exit = ex; 2455 } 2456 exits.release (); 2457 2458 return niter ? niter : chrec_dont_know; 2459 } 2460 2461 /* 2462 2463 Analysis of upper bounds on number of iterations of a loop. 2464 2465 */ 2466 2467 static widest_int derive_constant_upper_bound_ops (tree, tree, 2468 enum tree_code, tree); 2469 2470 /* Returns a constant upper bound on the value of the right-hand side of 2471 an assignment statement STMT. */ 2472 2473 static widest_int 2474 derive_constant_upper_bound_assign (gimple stmt) 2475 { 2476 enum tree_code code = gimple_assign_rhs_code (stmt); 2477 tree op0 = gimple_assign_rhs1 (stmt); 2478 tree op1 = gimple_assign_rhs2 (stmt); 2479 2480 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)), 2481 op0, code, op1); 2482 } 2483 2484 /* Returns a constant upper bound on the value of expression VAL. VAL 2485 is considered to be unsigned. If its type is signed, its value must 2486 be nonnegative. */ 2487 2488 static widest_int 2489 derive_constant_upper_bound (tree val) 2490 { 2491 enum tree_code code; 2492 tree op0, op1, op2; 2493 2494 extract_ops_from_tree (val, &code, &op0, &op1, &op2); 2495 return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1); 2496 } 2497 2498 /* Returns a constant upper bound on the value of expression OP0 CODE OP1, 2499 whose type is TYPE. The expression is considered to be unsigned. If 2500 its type is signed, its value must be nonnegative. */ 2501 2502 static widest_int 2503 derive_constant_upper_bound_ops (tree type, tree op0, 2504 enum tree_code code, tree op1) 2505 { 2506 tree subtype, maxt; 2507 widest_int bnd, max, cst; 2508 gimple stmt; 2509 2510 if (INTEGRAL_TYPE_P (type)) 2511 maxt = TYPE_MAX_VALUE (type); 2512 else 2513 maxt = upper_bound_in_type (type, type); 2514 2515 max = wi::to_widest (maxt); 2516 2517 switch (code) 2518 { 2519 case INTEGER_CST: 2520 return wi::to_widest (op0); 2521 2522 CASE_CONVERT: 2523 subtype = TREE_TYPE (op0); 2524 if (!TYPE_UNSIGNED (subtype) 2525 /* If TYPE is also signed, the fact that VAL is nonnegative implies 2526 that OP0 is nonnegative. */ 2527 && TYPE_UNSIGNED (type) 2528 && !tree_expr_nonnegative_p (op0)) 2529 { 2530 /* If we cannot prove that the casted expression is nonnegative, 2531 we cannot establish more useful upper bound than the precision 2532 of the type gives us. */ 2533 return max; 2534 } 2535 2536 /* We now know that op0 is an nonnegative value. Try deriving an upper 2537 bound for it. */ 2538 bnd = derive_constant_upper_bound (op0); 2539 2540 /* If the bound does not fit in TYPE, max. value of TYPE could be 2541 attained. */ 2542 if (wi::ltu_p (max, bnd)) 2543 return max; 2544 2545 return bnd; 2546 2547 case PLUS_EXPR: 2548 case POINTER_PLUS_EXPR: 2549 case MINUS_EXPR: 2550 if (TREE_CODE (op1) != INTEGER_CST 2551 || !tree_expr_nonnegative_p (op0)) 2552 return max; 2553 2554 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to 2555 choose the most logical way how to treat this constant regardless 2556 of the signedness of the type. */ 2557 cst = wi::sext (wi::to_widest (op1), TYPE_PRECISION (type)); 2558 if (code != MINUS_EXPR) 2559 cst = -cst; 2560 2561 bnd = derive_constant_upper_bound (op0); 2562 2563 if (wi::neg_p (cst)) 2564 { 2565 cst = -cst; 2566 /* Avoid CST == 0x80000... */ 2567 if (wi::neg_p (cst)) 2568 return max;; 2569 2570 /* OP0 + CST. We need to check that 2571 BND <= MAX (type) - CST. */ 2572 2573 widest_int mmax = max - cst; 2574 if (wi::leu_p (bnd, mmax)) 2575 return max; 2576 2577 return bnd + cst; 2578 } 2579 else 2580 { 2581 /* OP0 - CST, where CST >= 0. 2582 2583 If TYPE is signed, we have already verified that OP0 >= 0, and we 2584 know that the result is nonnegative. This implies that 2585 VAL <= BND - CST. 2586 2587 If TYPE is unsigned, we must additionally know that OP0 >= CST, 2588 otherwise the operation underflows. 2589 */ 2590 2591 /* This should only happen if the type is unsigned; however, for 2592 buggy programs that use overflowing signed arithmetics even with 2593 -fno-wrapv, this condition may also be true for signed values. */ 2594 if (wi::ltu_p (bnd, cst)) 2595 return max; 2596 2597 if (TYPE_UNSIGNED (type)) 2598 { 2599 tree tem = fold_binary (GE_EXPR, boolean_type_node, op0, 2600 wide_int_to_tree (type, cst)); 2601 if (!tem || integer_nonzerop (tem)) 2602 return max; 2603 } 2604 2605 bnd -= cst; 2606 } 2607 2608 return bnd; 2609 2610 case FLOOR_DIV_EXPR: 2611 case EXACT_DIV_EXPR: 2612 if (TREE_CODE (op1) != INTEGER_CST 2613 || tree_int_cst_sign_bit (op1)) 2614 return max; 2615 2616 bnd = derive_constant_upper_bound (op0); 2617 return wi::udiv_floor (bnd, wi::to_widest (op1)); 2618 2619 case BIT_AND_EXPR: 2620 if (TREE_CODE (op1) != INTEGER_CST 2621 || tree_int_cst_sign_bit (op1)) 2622 return max; 2623 return wi::to_widest (op1); 2624 2625 case SSA_NAME: 2626 stmt = SSA_NAME_DEF_STMT (op0); 2627 if (gimple_code (stmt) != GIMPLE_ASSIGN 2628 || gimple_assign_lhs (stmt) != op0) 2629 return max; 2630 return derive_constant_upper_bound_assign (stmt); 2631 2632 default: 2633 return max; 2634 } 2635 } 2636 2637 /* Emit a -Waggressive-loop-optimizations warning if needed. */ 2638 2639 static void 2640 do_warn_aggressive_loop_optimizations (struct loop *loop, 2641 widest_int i_bound, gimple stmt) 2642 { 2643 /* Don't warn if the loop doesn't have known constant bound. */ 2644 if (!loop->nb_iterations 2645 || TREE_CODE (loop->nb_iterations) != INTEGER_CST 2646 || !warn_aggressive_loop_optimizations 2647 /* To avoid warning multiple times for the same loop, 2648 only start warning when we preserve loops. */ 2649 || (cfun->curr_properties & PROP_loops) == 0 2650 /* Only warn once per loop. */ 2651 || loop->warned_aggressive_loop_optimizations 2652 /* Only warn if undefined behavior gives us lower estimate than the 2653 known constant bound. */ 2654 || wi::cmpu (i_bound, wi::to_widest (loop->nb_iterations)) >= 0 2655 /* And undefined behavior happens unconditionally. */ 2656 || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt))) 2657 return; 2658 2659 edge e = single_exit (loop); 2660 if (e == NULL) 2661 return; 2662 2663 gimple estmt = last_stmt (e->src); 2664 if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations, 2665 "iteration %E invokes undefined behavior", 2666 wide_int_to_tree (TREE_TYPE (loop->nb_iterations), 2667 i_bound))) 2668 inform (gimple_location (estmt), "containing loop"); 2669 loop->warned_aggressive_loop_optimizations = true; 2670 } 2671 2672 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT 2673 is true if the loop is exited immediately after STMT, and this exit 2674 is taken at last when the STMT is executed BOUND + 1 times. 2675 REALISTIC is true if BOUND is expected to be close to the real number 2676 of iterations. UPPER is true if we are sure the loop iterates at most 2677 BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */ 2678 2679 static void 2680 record_estimate (struct loop *loop, tree bound, const widest_int &i_bound, 2681 gimple at_stmt, bool is_exit, bool realistic, bool upper) 2682 { 2683 widest_int delta; 2684 2685 if (dump_file && (dump_flags & TDF_DETAILS)) 2686 { 2687 fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : ""); 2688 print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM); 2689 fprintf (dump_file, " is %sexecuted at most ", 2690 upper ? "" : "probably "); 2691 print_generic_expr (dump_file, bound, TDF_SLIM); 2692 fprintf (dump_file, " (bounded by "); 2693 print_decu (i_bound, dump_file); 2694 fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num); 2695 } 2696 2697 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the 2698 real number of iterations. */ 2699 if (TREE_CODE (bound) != INTEGER_CST) 2700 realistic = false; 2701 else 2702 gcc_checking_assert (i_bound == wi::to_widest (bound)); 2703 if (!upper && !realistic) 2704 return; 2705 2706 /* If we have a guaranteed upper bound, record it in the appropriate 2707 list, unless this is an !is_exit bound (i.e. undefined behavior in 2708 at_stmt) in a loop with known constant number of iterations. */ 2709 if (upper 2710 && (is_exit 2711 || loop->nb_iterations == NULL_TREE 2712 || TREE_CODE (loop->nb_iterations) != INTEGER_CST)) 2713 { 2714 struct nb_iter_bound *elt = ggc_alloc<nb_iter_bound> (); 2715 2716 elt->bound = i_bound; 2717 elt->stmt = at_stmt; 2718 elt->is_exit = is_exit; 2719 elt->next = loop->bounds; 2720 loop->bounds = elt; 2721 } 2722 2723 /* If statement is executed on every path to the loop latch, we can directly 2724 infer the upper bound on the # of iterations of the loop. */ 2725 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt))) 2726 return; 2727 2728 /* Update the number of iteration estimates according to the bound. 2729 If at_stmt is an exit then the loop latch is executed at most BOUND times, 2730 otherwise it can be executed BOUND + 1 times. We will lower the estimate 2731 later if such statement must be executed on last iteration */ 2732 if (is_exit) 2733 delta = 0; 2734 else 2735 delta = 1; 2736 widest_int new_i_bound = i_bound + delta; 2737 2738 /* If an overflow occurred, ignore the result. */ 2739 if (wi::ltu_p (new_i_bound, delta)) 2740 return; 2741 2742 if (upper && !is_exit) 2743 do_warn_aggressive_loop_optimizations (loop, new_i_bound, at_stmt); 2744 record_niter_bound (loop, new_i_bound, realistic, upper); 2745 } 2746 2747 /* Record the estimate on number of iterations of LOOP based on the fact that 2748 the induction variable BASE + STEP * i evaluated in STMT does not wrap and 2749 its values belong to the range <LOW, HIGH>. REALISTIC is true if the 2750 estimated number of iterations is expected to be close to the real one. 2751 UPPER is true if we are sure the induction variable does not wrap. */ 2752 2753 static void 2754 record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple stmt, 2755 tree low, tree high, bool realistic, bool upper) 2756 { 2757 tree niter_bound, extreme, delta; 2758 tree type = TREE_TYPE (base), unsigned_type; 2759 tree orig_base = base; 2760 2761 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step)) 2762 return; 2763 2764 if (dump_file && (dump_flags & TDF_DETAILS)) 2765 { 2766 fprintf (dump_file, "Induction variable ("); 2767 print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM); 2768 fprintf (dump_file, ") "); 2769 print_generic_expr (dump_file, base, TDF_SLIM); 2770 fprintf (dump_file, " + "); 2771 print_generic_expr (dump_file, step, TDF_SLIM); 2772 fprintf (dump_file, " * iteration does not wrap in statement "); 2773 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); 2774 fprintf (dump_file, " in loop %d.\n", loop->num); 2775 } 2776 2777 unsigned_type = unsigned_type_for (type); 2778 base = fold_convert (unsigned_type, base); 2779 step = fold_convert (unsigned_type, step); 2780 2781 if (tree_int_cst_sign_bit (step)) 2782 { 2783 wide_int min, max; 2784 extreme = fold_convert (unsigned_type, low); 2785 if (TREE_CODE (orig_base) == SSA_NAME 2786 && TREE_CODE (high) == INTEGER_CST 2787 && INTEGRAL_TYPE_P (TREE_TYPE (orig_base)) 2788 && get_range_info (orig_base, &min, &max) == VR_RANGE 2789 && wi::gts_p (high, max)) 2790 base = wide_int_to_tree (unsigned_type, max); 2791 else if (TREE_CODE (base) != INTEGER_CST 2792 && dominated_by_p (CDI_DOMINATORS, 2793 loop->latch, gimple_bb (stmt))) 2794 base = fold_convert (unsigned_type, high); 2795 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme); 2796 step = fold_build1 (NEGATE_EXPR, unsigned_type, step); 2797 } 2798 else 2799 { 2800 wide_int min, max; 2801 extreme = fold_convert (unsigned_type, high); 2802 if (TREE_CODE (orig_base) == SSA_NAME 2803 && TREE_CODE (low) == INTEGER_CST 2804 && INTEGRAL_TYPE_P (TREE_TYPE (orig_base)) 2805 && get_range_info (orig_base, &min, &max) == VR_RANGE 2806 && wi::gts_p (min, low)) 2807 base = wide_int_to_tree (unsigned_type, min); 2808 else if (TREE_CODE (base) != INTEGER_CST 2809 && dominated_by_p (CDI_DOMINATORS, 2810 loop->latch, gimple_bb (stmt))) 2811 base = fold_convert (unsigned_type, low); 2812 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base); 2813 } 2814 2815 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value 2816 would get out of the range. */ 2817 niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step); 2818 widest_int max = derive_constant_upper_bound (niter_bound); 2819 record_estimate (loop, niter_bound, max, stmt, false, realistic, upper); 2820 } 2821 2822 /* Determine information about number of iterations a LOOP from the index 2823 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is 2824 guaranteed to be executed in every iteration of LOOP. Callback for 2825 for_each_index. */ 2826 2827 struct ilb_data 2828 { 2829 struct loop *loop; 2830 gimple stmt; 2831 }; 2832 2833 static bool 2834 idx_infer_loop_bounds (tree base, tree *idx, void *dta) 2835 { 2836 struct ilb_data *data = (struct ilb_data *) dta; 2837 tree ev, init, step; 2838 tree low, high, type, next; 2839 bool sign, upper = true, at_end = false; 2840 struct loop *loop = data->loop; 2841 bool reliable = true; 2842 2843 if (TREE_CODE (base) != ARRAY_REF) 2844 return true; 2845 2846 /* For arrays at the end of the structure, we are not guaranteed that they 2847 do not really extend over their declared size. However, for arrays of 2848 size greater than one, this is unlikely to be intended. */ 2849 if (array_at_struct_end_p (base)) 2850 { 2851 at_end = true; 2852 upper = false; 2853 } 2854 2855 struct loop *dloop = loop_containing_stmt (data->stmt); 2856 if (!dloop) 2857 return true; 2858 2859 ev = analyze_scalar_evolution (dloop, *idx); 2860 ev = instantiate_parameters (loop, ev); 2861 init = initial_condition (ev); 2862 step = evolution_part_in_loop_num (ev, loop->num); 2863 2864 if (!init 2865 || !step 2866 || TREE_CODE (step) != INTEGER_CST 2867 || integer_zerop (step) 2868 || tree_contains_chrecs (init, NULL) 2869 || chrec_contains_symbols_defined_in_loop (init, loop->num)) 2870 return true; 2871 2872 low = array_ref_low_bound (base); 2873 high = array_ref_up_bound (base); 2874 2875 /* The case of nonconstant bounds could be handled, but it would be 2876 complicated. */ 2877 if (TREE_CODE (low) != INTEGER_CST 2878 || !high 2879 || TREE_CODE (high) != INTEGER_CST) 2880 return true; 2881 sign = tree_int_cst_sign_bit (step); 2882 type = TREE_TYPE (step); 2883 2884 /* The array of length 1 at the end of a structure most likely extends 2885 beyond its bounds. */ 2886 if (at_end 2887 && operand_equal_p (low, high, 0)) 2888 return true; 2889 2890 /* In case the relevant bound of the array does not fit in type, or 2891 it does, but bound + step (in type) still belongs into the range of the 2892 array, the index may wrap and still stay within the range of the array 2893 (consider e.g. if the array is indexed by the full range of 2894 unsigned char). 2895 2896 To make things simpler, we require both bounds to fit into type, although 2897 there are cases where this would not be strictly necessary. */ 2898 if (!int_fits_type_p (high, type) 2899 || !int_fits_type_p (low, type)) 2900 return true; 2901 low = fold_convert (type, low); 2902 high = fold_convert (type, high); 2903 2904 if (sign) 2905 next = fold_binary (PLUS_EXPR, type, low, step); 2906 else 2907 next = fold_binary (PLUS_EXPR, type, high, step); 2908 2909 if (tree_int_cst_compare (low, next) <= 0 2910 && tree_int_cst_compare (next, high) <= 0) 2911 return true; 2912 2913 /* If access is not executed on every iteration, we must ensure that overlow may 2914 not make the access valid later. */ 2915 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt)) 2916 && scev_probably_wraps_p (initial_condition_in_loop_num (ev, loop->num), 2917 step, data->stmt, loop, true)) 2918 reliable = false; 2919 2920 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, reliable, upper); 2921 return true; 2922 } 2923 2924 /* Determine information about number of iterations a LOOP from the bounds 2925 of arrays in the data reference REF accessed in STMT. RELIABLE is true if 2926 STMT is guaranteed to be executed in every iteration of LOOP.*/ 2927 2928 static void 2929 infer_loop_bounds_from_ref (struct loop *loop, gimple stmt, tree ref) 2930 { 2931 struct ilb_data data; 2932 2933 data.loop = loop; 2934 data.stmt = stmt; 2935 for_each_index (&ref, idx_infer_loop_bounds, &data); 2936 } 2937 2938 /* Determine information about number of iterations of a LOOP from the way 2939 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be 2940 executed in every iteration of LOOP. */ 2941 2942 static void 2943 infer_loop_bounds_from_array (struct loop *loop, gimple stmt) 2944 { 2945 if (is_gimple_assign (stmt)) 2946 { 2947 tree op0 = gimple_assign_lhs (stmt); 2948 tree op1 = gimple_assign_rhs1 (stmt); 2949 2950 /* For each memory access, analyze its access function 2951 and record a bound on the loop iteration domain. */ 2952 if (REFERENCE_CLASS_P (op0)) 2953 infer_loop_bounds_from_ref (loop, stmt, op0); 2954 2955 if (REFERENCE_CLASS_P (op1)) 2956 infer_loop_bounds_from_ref (loop, stmt, op1); 2957 } 2958 else if (is_gimple_call (stmt)) 2959 { 2960 tree arg, lhs; 2961 unsigned i, n = gimple_call_num_args (stmt); 2962 2963 lhs = gimple_call_lhs (stmt); 2964 if (lhs && REFERENCE_CLASS_P (lhs)) 2965 infer_loop_bounds_from_ref (loop, stmt, lhs); 2966 2967 for (i = 0; i < n; i++) 2968 { 2969 arg = gimple_call_arg (stmt, i); 2970 if (REFERENCE_CLASS_P (arg)) 2971 infer_loop_bounds_from_ref (loop, stmt, arg); 2972 } 2973 } 2974 } 2975 2976 /* Determine information about number of iterations of a LOOP from the fact 2977 that pointer arithmetics in STMT does not overflow. */ 2978 2979 static void 2980 infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple stmt) 2981 { 2982 tree def, base, step, scev, type, low, high; 2983 tree var, ptr; 2984 2985 if (!is_gimple_assign (stmt) 2986 || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR) 2987 return; 2988 2989 def = gimple_assign_lhs (stmt); 2990 if (TREE_CODE (def) != SSA_NAME) 2991 return; 2992 2993 type = TREE_TYPE (def); 2994 if (!nowrap_type_p (type)) 2995 return; 2996 2997 ptr = gimple_assign_rhs1 (stmt); 2998 if (!expr_invariant_in_loop_p (loop, ptr)) 2999 return; 3000 3001 var = gimple_assign_rhs2 (stmt); 3002 if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var))) 3003 return; 3004 3005 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def)); 3006 if (chrec_contains_undetermined (scev)) 3007 return; 3008 3009 base = initial_condition_in_loop_num (scev, loop->num); 3010 step = evolution_part_in_loop_num (scev, loop->num); 3011 3012 if (!base || !step 3013 || TREE_CODE (step) != INTEGER_CST 3014 || tree_contains_chrecs (base, NULL) 3015 || chrec_contains_symbols_defined_in_loop (base, loop->num)) 3016 return; 3017 3018 low = lower_bound_in_type (type, type); 3019 high = upper_bound_in_type (type, type); 3020 3021 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot 3022 produce a NULL pointer. The contrary would mean NULL points to an object, 3023 while NULL is supposed to compare unequal with the address of all objects. 3024 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a 3025 NULL pointer since that would mean wrapping, which we assume here not to 3026 happen. So, we can exclude NULL from the valid range of pointer 3027 arithmetic. */ 3028 if (flag_delete_null_pointer_checks && int_cst_value (low) == 0) 3029 low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type))); 3030 3031 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true); 3032 } 3033 3034 /* Determine information about number of iterations of a LOOP from the fact 3035 that signed arithmetics in STMT does not overflow. */ 3036 3037 static void 3038 infer_loop_bounds_from_signedness (struct loop *loop, gimple stmt) 3039 { 3040 tree def, base, step, scev, type, low, high; 3041 3042 if (gimple_code (stmt) != GIMPLE_ASSIGN) 3043 return; 3044 3045 def = gimple_assign_lhs (stmt); 3046 3047 if (TREE_CODE (def) != SSA_NAME) 3048 return; 3049 3050 type = TREE_TYPE (def); 3051 if (!INTEGRAL_TYPE_P (type) 3052 || !TYPE_OVERFLOW_UNDEFINED (type)) 3053 return; 3054 3055 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def)); 3056 if (chrec_contains_undetermined (scev)) 3057 return; 3058 3059 base = initial_condition_in_loop_num (scev, loop->num); 3060 step = evolution_part_in_loop_num (scev, loop->num); 3061 3062 if (!base || !step 3063 || TREE_CODE (step) != INTEGER_CST 3064 || tree_contains_chrecs (base, NULL) 3065 || chrec_contains_symbols_defined_in_loop (base, loop->num)) 3066 return; 3067 3068 low = lower_bound_in_type (type, type); 3069 high = upper_bound_in_type (type, type); 3070 3071 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true); 3072 } 3073 3074 /* The following analyzers are extracting informations on the bounds 3075 of LOOP from the following undefined behaviors: 3076 3077 - data references should not access elements over the statically 3078 allocated size, 3079 3080 - signed variables should not overflow when flag_wrapv is not set. 3081 */ 3082 3083 static void 3084 infer_loop_bounds_from_undefined (struct loop *loop) 3085 { 3086 unsigned i; 3087 basic_block *bbs; 3088 gimple_stmt_iterator bsi; 3089 basic_block bb; 3090 bool reliable; 3091 3092 bbs = get_loop_body (loop); 3093 3094 for (i = 0; i < loop->num_nodes; i++) 3095 { 3096 bb = bbs[i]; 3097 3098 /* If BB is not executed in each iteration of the loop, we cannot 3099 use the operations in it to infer reliable upper bound on the 3100 # of iterations of the loop. However, we can use it as a guess. 3101 Reliable guesses come only from array bounds. */ 3102 reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb); 3103 3104 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) 3105 { 3106 gimple stmt = gsi_stmt (bsi); 3107 3108 infer_loop_bounds_from_array (loop, stmt); 3109 3110 if (reliable) 3111 { 3112 infer_loop_bounds_from_signedness (loop, stmt); 3113 infer_loop_bounds_from_pointer_arith (loop, stmt); 3114 } 3115 } 3116 3117 } 3118 3119 free (bbs); 3120 } 3121 3122 /* Compare wide ints, callback for qsort. */ 3123 3124 static int 3125 wide_int_cmp (const void *p1, const void *p2) 3126 { 3127 const widest_int *d1 = (const widest_int *) p1; 3128 const widest_int *d2 = (const widest_int *) p2; 3129 return wi::cmpu (*d1, *d2); 3130 } 3131 3132 /* Return index of BOUND in BOUNDS array sorted in increasing order. 3133 Lookup by binary search. */ 3134 3135 static int 3136 bound_index (vec<widest_int> bounds, const widest_int &bound) 3137 { 3138 unsigned int end = bounds.length (); 3139 unsigned int begin = 0; 3140 3141 /* Find a matching index by means of a binary search. */ 3142 while (begin != end) 3143 { 3144 unsigned int middle = (begin + end) / 2; 3145 widest_int index = bounds[middle]; 3146 3147 if (index == bound) 3148 return middle; 3149 else if (wi::ltu_p (index, bound)) 3150 begin = middle + 1; 3151 else 3152 end = middle; 3153 } 3154 gcc_unreachable (); 3155 } 3156 3157 /* We recorded loop bounds only for statements dominating loop latch (and thus 3158 executed each loop iteration). If there are any bounds on statements not 3159 dominating the loop latch we can improve the estimate by walking the loop 3160 body and seeing if every path from loop header to loop latch contains 3161 some bounded statement. */ 3162 3163 static void 3164 discover_iteration_bound_by_body_walk (struct loop *loop) 3165 { 3166 struct nb_iter_bound *elt; 3167 vec<widest_int> bounds = vNULL; 3168 vec<vec<basic_block> > queues = vNULL; 3169 vec<basic_block> queue = vNULL; 3170 ptrdiff_t queue_index; 3171 ptrdiff_t latch_index = 0; 3172 3173 /* Discover what bounds may interest us. */ 3174 for (elt = loop->bounds; elt; elt = elt->next) 3175 { 3176 widest_int bound = elt->bound; 3177 3178 /* Exit terminates loop at given iteration, while non-exits produce undefined 3179 effect on the next iteration. */ 3180 if (!elt->is_exit) 3181 { 3182 bound += 1; 3183 /* If an overflow occurred, ignore the result. */ 3184 if (bound == 0) 3185 continue; 3186 } 3187 3188 if (!loop->any_upper_bound 3189 || wi::ltu_p (bound, loop->nb_iterations_upper_bound)) 3190 bounds.safe_push (bound); 3191 } 3192 3193 /* Exit early if there is nothing to do. */ 3194 if (!bounds.exists ()) 3195 return; 3196 3197 if (dump_file && (dump_flags & TDF_DETAILS)) 3198 fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n"); 3199 3200 /* Sort the bounds in decreasing order. */ 3201 bounds.qsort (wide_int_cmp); 3202 3203 /* For every basic block record the lowest bound that is guaranteed to 3204 terminate the loop. */ 3205 3206 hash_map<basic_block, ptrdiff_t> bb_bounds; 3207 for (elt = loop->bounds; elt; elt = elt->next) 3208 { 3209 widest_int bound = elt->bound; 3210 if (!elt->is_exit) 3211 { 3212 bound += 1; 3213 /* If an overflow occurred, ignore the result. */ 3214 if (bound == 0) 3215 continue; 3216 } 3217 3218 if (!loop->any_upper_bound 3219 || wi::ltu_p (bound, loop->nb_iterations_upper_bound)) 3220 { 3221 ptrdiff_t index = bound_index (bounds, bound); 3222 ptrdiff_t *entry = bb_bounds.get (gimple_bb (elt->stmt)); 3223 if (!entry) 3224 bb_bounds.put (gimple_bb (elt->stmt), index); 3225 else if ((ptrdiff_t)*entry > index) 3226 *entry = index; 3227 } 3228 } 3229 3230 hash_map<basic_block, ptrdiff_t> block_priority; 3231 3232 /* Perform shortest path discovery loop->header ... loop->latch. 3233 3234 The "distance" is given by the smallest loop bound of basic block 3235 present in the path and we look for path with largest smallest bound 3236 on it. 3237 3238 To avoid the need for fibonacci heap on double ints we simply compress 3239 double ints into indexes to BOUNDS array and then represent the queue 3240 as arrays of queues for every index. 3241 Index of BOUNDS.length() means that the execution of given BB has 3242 no bounds determined. 3243 3244 VISITED is a pointer map translating basic block into smallest index 3245 it was inserted into the priority queue with. */ 3246 latch_index = -1; 3247 3248 /* Start walk in loop header with index set to infinite bound. */ 3249 queue_index = bounds.length (); 3250 queues.safe_grow_cleared (queue_index + 1); 3251 queue.safe_push (loop->header); 3252 queues[queue_index] = queue; 3253 block_priority.put (loop->header, queue_index); 3254 3255 for (; queue_index >= 0; queue_index--) 3256 { 3257 if (latch_index < queue_index) 3258 { 3259 while (queues[queue_index].length ()) 3260 { 3261 basic_block bb; 3262 ptrdiff_t bound_index = queue_index; 3263 edge e; 3264 edge_iterator ei; 3265 3266 queue = queues[queue_index]; 3267 bb = queue.pop (); 3268 3269 /* OK, we later inserted the BB with lower priority, skip it. */ 3270 if (*block_priority.get (bb) > queue_index) 3271 continue; 3272 3273 /* See if we can improve the bound. */ 3274 ptrdiff_t *entry = bb_bounds.get (bb); 3275 if (entry && *entry < bound_index) 3276 bound_index = *entry; 3277 3278 /* Insert succesors into the queue, watch for latch edge 3279 and record greatest index we saw. */ 3280 FOR_EACH_EDGE (e, ei, bb->succs) 3281 { 3282 bool insert = false; 3283 3284 if (loop_exit_edge_p (loop, e)) 3285 continue; 3286 3287 if (e == loop_latch_edge (loop) 3288 && latch_index < bound_index) 3289 latch_index = bound_index; 3290 else if (!(entry = block_priority.get (e->dest))) 3291 { 3292 insert = true; 3293 block_priority.put (e->dest, bound_index); 3294 } 3295 else if (*entry < bound_index) 3296 { 3297 insert = true; 3298 *entry = bound_index; 3299 } 3300 3301 if (insert) 3302 queues[bound_index].safe_push (e->dest); 3303 } 3304 } 3305 } 3306 queues[queue_index].release (); 3307 } 3308 3309 gcc_assert (latch_index >= 0); 3310 if ((unsigned)latch_index < bounds.length ()) 3311 { 3312 if (dump_file && (dump_flags & TDF_DETAILS)) 3313 { 3314 fprintf (dump_file, "Found better loop bound "); 3315 print_decu (bounds[latch_index], dump_file); 3316 fprintf (dump_file, "\n"); 3317 } 3318 record_niter_bound (loop, bounds[latch_index], false, true); 3319 } 3320 3321 queues.release (); 3322 bounds.release (); 3323 } 3324 3325 /* See if every path cross the loop goes through a statement that is known 3326 to not execute at the last iteration. In that case we can decrese iteration 3327 count by 1. */ 3328 3329 static void 3330 maybe_lower_iteration_bound (struct loop *loop) 3331 { 3332 hash_set<gimple> *not_executed_last_iteration = NULL; 3333 struct nb_iter_bound *elt; 3334 bool found_exit = false; 3335 vec<basic_block> queue = vNULL; 3336 bitmap visited; 3337 3338 /* Collect all statements with interesting (i.e. lower than 3339 nb_iterations_upper_bound) bound on them. 3340 3341 TODO: Due to the way record_estimate choose estimates to store, the bounds 3342 will be always nb_iterations_upper_bound-1. We can change this to record 3343 also statements not dominating the loop latch and update the walk bellow 3344 to the shortest path algorthm. */ 3345 for (elt = loop->bounds; elt; elt = elt->next) 3346 { 3347 if (!elt->is_exit 3348 && wi::ltu_p (elt->bound, loop->nb_iterations_upper_bound)) 3349 { 3350 if (!not_executed_last_iteration) 3351 not_executed_last_iteration = new hash_set<gimple>; 3352 not_executed_last_iteration->add (elt->stmt); 3353 } 3354 } 3355 if (!not_executed_last_iteration) 3356 return; 3357 3358 /* Start DFS walk in the loop header and see if we can reach the 3359 loop latch or any of the exits (including statements with side 3360 effects that may terminate the loop otherwise) without visiting 3361 any of the statements known to have undefined effect on the last 3362 iteration. */ 3363 queue.safe_push (loop->header); 3364 visited = BITMAP_ALLOC (NULL); 3365 bitmap_set_bit (visited, loop->header->index); 3366 found_exit = false; 3367 3368 do 3369 { 3370 basic_block bb = queue.pop (); 3371 gimple_stmt_iterator gsi; 3372 bool stmt_found = false; 3373 3374 /* Loop for possible exits and statements bounding the execution. */ 3375 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 3376 { 3377 gimple stmt = gsi_stmt (gsi); 3378 if (not_executed_last_iteration->contains (stmt)) 3379 { 3380 stmt_found = true; 3381 break; 3382 } 3383 if (gimple_has_side_effects (stmt)) 3384 { 3385 found_exit = true; 3386 break; 3387 } 3388 } 3389 if (found_exit) 3390 break; 3391 3392 /* If no bounding statement is found, continue the walk. */ 3393 if (!stmt_found) 3394 { 3395 edge e; 3396 edge_iterator ei; 3397 3398 FOR_EACH_EDGE (e, ei, bb->succs) 3399 { 3400 if (loop_exit_edge_p (loop, e) 3401 || e == loop_latch_edge (loop)) 3402 { 3403 found_exit = true; 3404 break; 3405 } 3406 if (bitmap_set_bit (visited, e->dest->index)) 3407 queue.safe_push (e->dest); 3408 } 3409 } 3410 } 3411 while (queue.length () && !found_exit); 3412 3413 /* If every path through the loop reach bounding statement before exit, 3414 then we know the last iteration of the loop will have undefined effect 3415 and we can decrease number of iterations. */ 3416 3417 if (!found_exit) 3418 { 3419 if (dump_file && (dump_flags & TDF_DETAILS)) 3420 fprintf (dump_file, "Reducing loop iteration estimate by 1; " 3421 "undefined statement must be executed at the last iteration.\n"); 3422 record_niter_bound (loop, loop->nb_iterations_upper_bound - 1, 3423 false, true); 3424 } 3425 3426 BITMAP_FREE (visited); 3427 queue.release (); 3428 delete not_executed_last_iteration; 3429 } 3430 3431 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P 3432 is true also use estimates derived from undefined behavior. */ 3433 3434 static void 3435 estimate_numbers_of_iterations_loop (struct loop *loop) 3436 { 3437 vec<edge> exits; 3438 tree niter, type; 3439 unsigned i; 3440 struct tree_niter_desc niter_desc; 3441 edge ex; 3442 widest_int bound; 3443 edge likely_exit; 3444 3445 /* Give up if we already have tried to compute an estimation. */ 3446 if (loop->estimate_state != EST_NOT_COMPUTED) 3447 return; 3448 3449 loop->estimate_state = EST_AVAILABLE; 3450 /* Force estimate compuation but leave any existing upper bound in place. */ 3451 loop->any_estimate = false; 3452 3453 /* Ensure that loop->nb_iterations is computed if possible. If it turns out 3454 to be constant, we avoid undefined behavior implied bounds and instead 3455 diagnose those loops with -Waggressive-loop-optimizations. */ 3456 number_of_latch_executions (loop); 3457 3458 exits = get_loop_exit_edges (loop); 3459 likely_exit = single_likely_exit (loop); 3460 FOR_EACH_VEC_ELT (exits, i, ex) 3461 { 3462 if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false)) 3463 continue; 3464 3465 niter = niter_desc.niter; 3466 type = TREE_TYPE (niter); 3467 if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST) 3468 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero, 3469 build_int_cst (type, 0), 3470 niter); 3471 record_estimate (loop, niter, niter_desc.max, 3472 last_stmt (ex->src), 3473 true, ex == likely_exit, true); 3474 } 3475 exits.release (); 3476 3477 if (flag_aggressive_loop_optimizations) 3478 infer_loop_bounds_from_undefined (loop); 3479 3480 discover_iteration_bound_by_body_walk (loop); 3481 3482 maybe_lower_iteration_bound (loop); 3483 3484 /* If we have a measured profile, use it to estimate the number of 3485 iterations. */ 3486 if (loop->header->count != 0) 3487 { 3488 gcov_type nit = expected_loop_iterations_unbounded (loop) + 1; 3489 bound = gcov_type_to_wide_int (nit); 3490 record_niter_bound (loop, bound, true, false); 3491 } 3492 3493 /* If we know the exact number of iterations of this loop, try to 3494 not break code with undefined behavior by not recording smaller 3495 maximum number of iterations. */ 3496 if (loop->nb_iterations 3497 && TREE_CODE (loop->nb_iterations) == INTEGER_CST) 3498 { 3499 loop->any_upper_bound = true; 3500 loop->nb_iterations_upper_bound = wi::to_widest (loop->nb_iterations); 3501 } 3502 } 3503 3504 /* Sets NIT to the estimated number of executions of the latch of the 3505 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as 3506 large as the number of iterations. If we have no reliable estimate, 3507 the function returns false, otherwise returns true. */ 3508 3509 bool 3510 estimated_loop_iterations (struct loop *loop, widest_int *nit) 3511 { 3512 /* When SCEV information is available, try to update loop iterations 3513 estimate. Otherwise just return whatever we recorded earlier. */ 3514 if (scev_initialized_p ()) 3515 estimate_numbers_of_iterations_loop (loop); 3516 3517 return (get_estimated_loop_iterations (loop, nit)); 3518 } 3519 3520 /* Similar to estimated_loop_iterations, but returns the estimate only 3521 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate 3522 on the number of iterations of LOOP could not be derived, returns -1. */ 3523 3524 HOST_WIDE_INT 3525 estimated_loop_iterations_int (struct loop *loop) 3526 { 3527 widest_int nit; 3528 HOST_WIDE_INT hwi_nit; 3529 3530 if (!estimated_loop_iterations (loop, &nit)) 3531 return -1; 3532 3533 if (!wi::fits_shwi_p (nit)) 3534 return -1; 3535 hwi_nit = nit.to_shwi (); 3536 3537 return hwi_nit < 0 ? -1 : hwi_nit; 3538 } 3539 3540 3541 /* Sets NIT to an upper bound for the maximum number of executions of the 3542 latch of the LOOP. If we have no reliable estimate, the function returns 3543 false, otherwise returns true. */ 3544 3545 bool 3546 max_loop_iterations (struct loop *loop, widest_int *nit) 3547 { 3548 /* When SCEV information is available, try to update loop iterations 3549 estimate. Otherwise just return whatever we recorded earlier. */ 3550 if (scev_initialized_p ()) 3551 estimate_numbers_of_iterations_loop (loop); 3552 3553 return get_max_loop_iterations (loop, nit); 3554 } 3555 3556 /* Similar to max_loop_iterations, but returns the estimate only 3557 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate 3558 on the number of iterations of LOOP could not be derived, returns -1. */ 3559 3560 HOST_WIDE_INT 3561 max_loop_iterations_int (struct loop *loop) 3562 { 3563 widest_int nit; 3564 HOST_WIDE_INT hwi_nit; 3565 3566 if (!max_loop_iterations (loop, &nit)) 3567 return -1; 3568 3569 if (!wi::fits_shwi_p (nit)) 3570 return -1; 3571 hwi_nit = nit.to_shwi (); 3572 3573 return hwi_nit < 0 ? -1 : hwi_nit; 3574 } 3575 3576 /* Returns an estimate for the number of executions of statements 3577 in the LOOP. For statements before the loop exit, this exceeds 3578 the number of execution of the latch by one. */ 3579 3580 HOST_WIDE_INT 3581 estimated_stmt_executions_int (struct loop *loop) 3582 { 3583 HOST_WIDE_INT nit = estimated_loop_iterations_int (loop); 3584 HOST_WIDE_INT snit; 3585 3586 if (nit == -1) 3587 return -1; 3588 3589 snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1); 3590 3591 /* If the computation overflows, return -1. */ 3592 return snit < 0 ? -1 : snit; 3593 } 3594 3595 /* Sets NIT to the estimated maximum number of executions of the latch of the 3596 LOOP, plus one. If we have no reliable estimate, the function returns 3597 false, otherwise returns true. */ 3598 3599 bool 3600 max_stmt_executions (struct loop *loop, widest_int *nit) 3601 { 3602 widest_int nit_minus_one; 3603 3604 if (!max_loop_iterations (loop, nit)) 3605 return false; 3606 3607 nit_minus_one = *nit; 3608 3609 *nit += 1; 3610 3611 return wi::gtu_p (*nit, nit_minus_one); 3612 } 3613 3614 /* Sets NIT to the estimated number of executions of the latch of the 3615 LOOP, plus one. If we have no reliable estimate, the function returns 3616 false, otherwise returns true. */ 3617 3618 bool 3619 estimated_stmt_executions (struct loop *loop, widest_int *nit) 3620 { 3621 widest_int nit_minus_one; 3622 3623 if (!estimated_loop_iterations (loop, nit)) 3624 return false; 3625 3626 nit_minus_one = *nit; 3627 3628 *nit += 1; 3629 3630 return wi::gtu_p (*nit, nit_minus_one); 3631 } 3632 3633 /* Records estimates on numbers of iterations of loops. */ 3634 3635 void 3636 estimate_numbers_of_iterations (void) 3637 { 3638 struct loop *loop; 3639 3640 /* We don't want to issue signed overflow warnings while getting 3641 loop iteration estimates. */ 3642 fold_defer_overflow_warnings (); 3643 3644 FOR_EACH_LOOP (loop, 0) 3645 { 3646 estimate_numbers_of_iterations_loop (loop); 3647 } 3648 3649 fold_undefer_and_ignore_overflow_warnings (); 3650 } 3651 3652 /* Returns true if statement S1 dominates statement S2. */ 3653 3654 bool 3655 stmt_dominates_stmt_p (gimple s1, gimple s2) 3656 { 3657 basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2); 3658 3659 if (!bb1 3660 || s1 == s2) 3661 return true; 3662 3663 if (bb1 == bb2) 3664 { 3665 gimple_stmt_iterator bsi; 3666 3667 if (gimple_code (s2) == GIMPLE_PHI) 3668 return false; 3669 3670 if (gimple_code (s1) == GIMPLE_PHI) 3671 return true; 3672 3673 for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi)) 3674 if (gsi_stmt (bsi) == s1) 3675 return true; 3676 3677 return false; 3678 } 3679 3680 return dominated_by_p (CDI_DOMINATORS, bb2, bb1); 3681 } 3682 3683 /* Returns true when we can prove that the number of executions of 3684 STMT in the loop is at most NITER, according to the bound on 3685 the number of executions of the statement NITER_BOUND->stmt recorded in 3686 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT. 3687 3688 ??? This code can become quite a CPU hog - we can have many bounds, 3689 and large basic block forcing stmt_dominates_stmt_p to be queried 3690 many times on a large basic blocks, so the whole thing is O(n^2) 3691 for scev_probably_wraps_p invocation (that can be done n times). 3692 3693 It would make more sense (and give better answers) to remember BB 3694 bounds computed by discover_iteration_bound_by_body_walk. */ 3695 3696 static bool 3697 n_of_executions_at_most (gimple stmt, 3698 struct nb_iter_bound *niter_bound, 3699 tree niter) 3700 { 3701 widest_int bound = niter_bound->bound; 3702 tree nit_type = TREE_TYPE (niter), e; 3703 enum tree_code cmp; 3704 3705 gcc_assert (TYPE_UNSIGNED (nit_type)); 3706 3707 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that 3708 the number of iterations is small. */ 3709 if (!wi::fits_to_tree_p (bound, nit_type)) 3710 return false; 3711 3712 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1 3713 times. This means that: 3714 3715 -- if NITER_BOUND->is_exit is true, then everything after 3716 it at most NITER_BOUND->bound times. 3717 3718 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT 3719 is executed, then NITER_BOUND->stmt is executed as well in the same 3720 iteration then STMT is executed at most NITER_BOUND->bound + 1 times. 3721 3722 If we can determine that NITER_BOUND->stmt is always executed 3723 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times. 3724 We conclude that if both statements belong to the same 3725 basic block and STMT is before NITER_BOUND->stmt and there are no 3726 statements with side effects in between. */ 3727 3728 if (niter_bound->is_exit) 3729 { 3730 if (stmt == niter_bound->stmt 3731 || !stmt_dominates_stmt_p (niter_bound->stmt, stmt)) 3732 return false; 3733 cmp = GE_EXPR; 3734 } 3735 else 3736 { 3737 if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt)) 3738 { 3739 gimple_stmt_iterator bsi; 3740 if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt) 3741 || gimple_code (stmt) == GIMPLE_PHI 3742 || gimple_code (niter_bound->stmt) == GIMPLE_PHI) 3743 return false; 3744 3745 /* By stmt_dominates_stmt_p we already know that STMT appears 3746 before NITER_BOUND->STMT. Still need to test that the loop 3747 can not be terinated by a side effect in between. */ 3748 for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt; 3749 gsi_next (&bsi)) 3750 if (gimple_has_side_effects (gsi_stmt (bsi))) 3751 return false; 3752 bound += 1; 3753 if (bound == 0 3754 || !wi::fits_to_tree_p (bound, nit_type)) 3755 return false; 3756 } 3757 cmp = GT_EXPR; 3758 } 3759 3760 e = fold_binary (cmp, boolean_type_node, 3761 niter, wide_int_to_tree (nit_type, bound)); 3762 return e && integer_nonzerop (e); 3763 } 3764 3765 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */ 3766 3767 bool 3768 nowrap_type_p (tree type) 3769 { 3770 if (INTEGRAL_TYPE_P (type) 3771 && TYPE_OVERFLOW_UNDEFINED (type)) 3772 return true; 3773 3774 if (POINTER_TYPE_P (type)) 3775 return true; 3776 3777 return false; 3778 } 3779 3780 /* Return false only when the induction variable BASE + STEP * I is 3781 known to not overflow: i.e. when the number of iterations is small 3782 enough with respect to the step and initial condition in order to 3783 keep the evolution confined in TYPEs bounds. Return true when the 3784 iv is known to overflow or when the property is not computable. 3785 3786 USE_OVERFLOW_SEMANTICS is true if this function should assume that 3787 the rules for overflow of the given language apply (e.g., that signed 3788 arithmetics in C does not overflow). */ 3789 3790 bool 3791 scev_probably_wraps_p (tree base, tree step, 3792 gimple at_stmt, struct loop *loop, 3793 bool use_overflow_semantics) 3794 { 3795 tree delta, step_abs; 3796 tree unsigned_type, valid_niter; 3797 tree type = TREE_TYPE (step); 3798 tree e; 3799 widest_int niter; 3800 struct nb_iter_bound *bound; 3801 3802 /* FIXME: We really need something like 3803 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html. 3804 3805 We used to test for the following situation that frequently appears 3806 during address arithmetics: 3807 3808 D.1621_13 = (long unsigned intD.4) D.1620_12; 3809 D.1622_14 = D.1621_13 * 8; 3810 D.1623_15 = (doubleD.29 *) D.1622_14; 3811 3812 And derived that the sequence corresponding to D_14 3813 can be proved to not wrap because it is used for computing a 3814 memory access; however, this is not really the case -- for example, 3815 if D_12 = (unsigned char) [254,+,1], then D_14 has values 3816 2032, 2040, 0, 8, ..., but the code is still legal. */ 3817 3818 if (chrec_contains_undetermined (base) 3819 || chrec_contains_undetermined (step)) 3820 return true; 3821 3822 if (integer_zerop (step)) 3823 return false; 3824 3825 /* If we can use the fact that signed and pointer arithmetics does not 3826 wrap, we are done. */ 3827 if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base))) 3828 return false; 3829 3830 /* To be able to use estimates on number of iterations of the loop, 3831 we must have an upper bound on the absolute value of the step. */ 3832 if (TREE_CODE (step) != INTEGER_CST) 3833 return true; 3834 3835 /* Don't issue signed overflow warnings. */ 3836 fold_defer_overflow_warnings (); 3837 3838 /* Otherwise, compute the number of iterations before we reach the 3839 bound of the type, and verify that the loop is exited before this 3840 occurs. */ 3841 unsigned_type = unsigned_type_for (type); 3842 base = fold_convert (unsigned_type, base); 3843 3844 if (tree_int_cst_sign_bit (step)) 3845 { 3846 tree extreme = fold_convert (unsigned_type, 3847 lower_bound_in_type (type, type)); 3848 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme); 3849 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type, 3850 fold_convert (unsigned_type, step)); 3851 } 3852 else 3853 { 3854 tree extreme = fold_convert (unsigned_type, 3855 upper_bound_in_type (type, type)); 3856 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base); 3857 step_abs = fold_convert (unsigned_type, step); 3858 } 3859 3860 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs); 3861 3862 estimate_numbers_of_iterations_loop (loop); 3863 3864 if (max_loop_iterations (loop, &niter) 3865 && wi::fits_to_tree_p (niter, TREE_TYPE (valid_niter)) 3866 && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter, 3867 wide_int_to_tree (TREE_TYPE (valid_niter), 3868 niter))) != NULL 3869 && integer_nonzerop (e)) 3870 { 3871 fold_undefer_and_ignore_overflow_warnings (); 3872 return false; 3873 } 3874 if (at_stmt) 3875 for (bound = loop->bounds; bound; bound = bound->next) 3876 { 3877 if (n_of_executions_at_most (at_stmt, bound, valid_niter)) 3878 { 3879 fold_undefer_and_ignore_overflow_warnings (); 3880 return false; 3881 } 3882 } 3883 3884 fold_undefer_and_ignore_overflow_warnings (); 3885 3886 /* At this point we still don't have a proof that the iv does not 3887 overflow: give up. */ 3888 return true; 3889 } 3890 3891 /* Frees the information on upper bounds on numbers of iterations of LOOP. */ 3892 3893 void 3894 free_numbers_of_iterations_estimates_loop (struct loop *loop) 3895 { 3896 struct nb_iter_bound *bound, *next; 3897 3898 loop->nb_iterations = NULL; 3899 loop->estimate_state = EST_NOT_COMPUTED; 3900 for (bound = loop->bounds; bound; bound = next) 3901 { 3902 next = bound->next; 3903 ggc_free (bound); 3904 } 3905 3906 loop->bounds = NULL; 3907 } 3908 3909 /* Frees the information on upper bounds on numbers of iterations of loops. */ 3910 3911 void 3912 free_numbers_of_iterations_estimates (void) 3913 { 3914 struct loop *loop; 3915 3916 FOR_EACH_LOOP (loop, 0) 3917 { 3918 free_numbers_of_iterations_estimates_loop (loop); 3919 } 3920 } 3921 3922 /* Substitute value VAL for ssa name NAME inside expressions held 3923 at LOOP. */ 3924 3925 void 3926 substitute_in_loop_info (struct loop *loop, tree name, tree val) 3927 { 3928 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val); 3929 } 3930