Fix bootstrap/PR63632
[official-gcc.git] / gcc / tree-ssa-loop-niter.c
blob4da18556b7635c9d7b9db7d3ddc9431c9bcb4ab2
1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
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.
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.
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/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "calls.h"
26 #include "expr.h"
27 #include "tm_p.h"
28 #include "basic-block.h"
29 #include "gimple-pretty-print.h"
30 #include "intl.h"
31 #include "hash-set.h"
32 #include "tree-ssa-alias.h"
33 #include "internal-fn.h"
34 #include "gimple-expr.h"
35 #include "is-a.h"
36 #include "gimple.h"
37 #include "gimplify.h"
38 #include "gimple-iterator.h"
39 #include "gimple-ssa.h"
40 #include "tree-cfg.h"
41 #include "tree-phinodes.h"
42 #include "ssa-iterators.h"
43 #include "tree-ssa-loop-ivopts.h"
44 #include "tree-ssa-loop-niter.h"
45 #include "tree-ssa-loop.h"
46 #include "dumpfile.h"
47 #include "cfgloop.h"
48 #include "tree-chrec.h"
49 #include "tree-scalar-evolution.h"
50 #include "tree-data-ref.h"
51 #include "params.h"
52 #include "flags.h"
53 #include "diagnostic-core.h"
54 #include "tree-inline.h"
55 #include "tree-pass.h"
56 #include "stringpool.h"
57 #include "tree-ssanames.h"
58 #include "wide-int-print.h"
61 #define SWAP(X, Y) do { affine_iv *tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
63 /* The maximum number of dominator BBs we search for conditions
64 of loop header copies we use for simplifying a conditional
65 expression. */
66 #define MAX_DOMINATORS_TO_WALK 8
70 Analysis of number of iterations of an affine exit test.
74 /* Bounds on some value, BELOW <= X <= UP. */
76 typedef struct
78 mpz_t below, up;
79 } bounds;
82 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
84 static void
85 split_to_var_and_offset (tree expr, tree *var, mpz_t offset)
87 tree type = TREE_TYPE (expr);
88 tree op0, op1;
89 bool negate = false;
91 *var = expr;
92 mpz_set_ui (offset, 0);
94 switch (TREE_CODE (expr))
96 case MINUS_EXPR:
97 negate = true;
98 /* Fallthru. */
100 case PLUS_EXPR:
101 case POINTER_PLUS_EXPR:
102 op0 = TREE_OPERAND (expr, 0);
103 op1 = TREE_OPERAND (expr, 1);
105 if (TREE_CODE (op1) != INTEGER_CST)
106 break;
108 *var = op0;
109 /* Always sign extend the offset. */
110 wi::to_mpz (op1, offset, SIGNED);
111 if (negate)
112 mpz_neg (offset, offset);
113 break;
115 case INTEGER_CST:
116 *var = build_int_cst_type (type, 0);
117 wi::to_mpz (expr, offset, TYPE_SIGN (type));
118 break;
120 default:
121 break;
125 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
126 in TYPE to MIN and MAX. */
128 static void
129 determine_value_range (struct loop *loop, tree type, tree var, mpz_t off,
130 mpz_t min, mpz_t max)
132 wide_int minv, maxv;
133 enum value_range_type rtype = VR_VARYING;
135 /* If the expression is a constant, we know its value exactly. */
136 if (integer_zerop (var))
138 mpz_set (min, off);
139 mpz_set (max, off);
140 return;
143 get_type_static_bounds (type, min, max);
145 /* See if we have some range info from VRP. */
146 if (TREE_CODE (var) == SSA_NAME && INTEGRAL_TYPE_P (type))
148 edge e = loop_preheader_edge (loop);
149 signop sgn = TYPE_SIGN (type);
150 gimple_stmt_iterator gsi;
152 /* Either for VAR itself... */
153 rtype = get_range_info (var, &minv, &maxv);
154 /* Or for PHI results in loop->header where VAR is used as
155 PHI argument from the loop preheader edge. */
156 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi))
158 gimple phi = gsi_stmt (gsi);
159 wide_int minc, maxc;
160 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var
161 && (get_range_info (gimple_phi_result (phi), &minc, &maxc)
162 == VR_RANGE))
164 if (rtype != VR_RANGE)
166 rtype = VR_RANGE;
167 minv = minc;
168 maxv = maxc;
170 else
172 minv = wi::max (minv, minc, sgn);
173 maxv = wi::min (maxv, maxc, sgn);
174 /* If the PHI result range are inconsistent with
175 the VAR range, give up on looking at the PHI
176 results. This can happen if VR_UNDEFINED is
177 involved. */
178 if (wi::gt_p (minv, maxv, sgn))
180 rtype = get_range_info (var, &minv, &maxv);
181 break;
186 if (rtype == VR_RANGE)
188 mpz_t minm, maxm;
189 gcc_assert (wi::le_p (minv, maxv, sgn));
190 mpz_init (minm);
191 mpz_init (maxm);
192 wi::to_mpz (minv, minm, sgn);
193 wi::to_mpz (maxv, maxm, sgn);
194 mpz_add (minm, minm, off);
195 mpz_add (maxm, maxm, off);
196 /* If the computation may not wrap or off is zero, then this
197 is always fine. If off is negative and minv + off isn't
198 smaller than type's minimum, or off is positive and
199 maxv + off isn't bigger than type's maximum, use the more
200 precise range too. */
201 if (nowrap_type_p (type)
202 || mpz_sgn (off) == 0
203 || (mpz_sgn (off) < 0 && mpz_cmp (minm, min) >= 0)
204 || (mpz_sgn (off) > 0 && mpz_cmp (maxm, max) <= 0))
206 mpz_set (min, minm);
207 mpz_set (max, maxm);
208 mpz_clear (minm);
209 mpz_clear (maxm);
210 return;
212 mpz_clear (minm);
213 mpz_clear (maxm);
217 /* If the computation may wrap, we know nothing about the value, except for
218 the range of the type. */
219 if (!nowrap_type_p (type))
220 return;
222 /* Since the addition of OFF does not wrap, if OFF is positive, then we may
223 add it to MIN, otherwise to MAX. */
224 if (mpz_sgn (off) < 0)
225 mpz_add (max, max, off);
226 else
227 mpz_add (min, min, off);
230 /* Stores the bounds on the difference of the values of the expressions
231 (var + X) and (var + Y), computed in TYPE, to BNDS. */
233 static void
234 bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y,
235 bounds *bnds)
237 int rel = mpz_cmp (x, y);
238 bool may_wrap = !nowrap_type_p (type);
239 mpz_t m;
241 /* If X == Y, then the expressions are always equal.
242 If X > Y, there are the following possibilities:
243 a) neither of var + X and var + Y overflow or underflow, or both of
244 them do. Then their difference is X - Y.
245 b) var + X overflows, and var + Y does not. Then the values of the
246 expressions are var + X - M and var + Y, where M is the range of
247 the type, and their difference is X - Y - M.
248 c) var + Y underflows and var + X does not. Their difference again
249 is M - X + Y.
250 Therefore, if the arithmetics in type does not overflow, then the
251 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y)
252 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or
253 (X - Y, X - Y + M). */
255 if (rel == 0)
257 mpz_set_ui (bnds->below, 0);
258 mpz_set_ui (bnds->up, 0);
259 return;
262 mpz_init (m);
263 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), m, UNSIGNED);
264 mpz_add_ui (m, m, 1);
265 mpz_sub (bnds->up, x, y);
266 mpz_set (bnds->below, bnds->up);
268 if (may_wrap)
270 if (rel > 0)
271 mpz_sub (bnds->below, bnds->below, m);
272 else
273 mpz_add (bnds->up, bnds->up, m);
276 mpz_clear (m);
279 /* From condition C0 CMP C1 derives information regarding the
280 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE,
281 and stores it to BNDS. */
283 static void
284 refine_bounds_using_guard (tree type, tree varx, mpz_t offx,
285 tree vary, mpz_t offy,
286 tree c0, enum tree_code cmp, tree c1,
287 bounds *bnds)
289 tree varc0, varc1, tmp, ctype;
290 mpz_t offc0, offc1, loffx, loffy, bnd;
291 bool lbound = false;
292 bool no_wrap = nowrap_type_p (type);
293 bool x_ok, y_ok;
295 switch (cmp)
297 case LT_EXPR:
298 case LE_EXPR:
299 case GT_EXPR:
300 case GE_EXPR:
301 STRIP_SIGN_NOPS (c0);
302 STRIP_SIGN_NOPS (c1);
303 ctype = TREE_TYPE (c0);
304 if (!useless_type_conversion_p (ctype, type))
305 return;
307 break;
309 case EQ_EXPR:
310 /* We could derive quite precise information from EQ_EXPR, however, such
311 a guard is unlikely to appear, so we do not bother with handling
312 it. */
313 return;
315 case NE_EXPR:
316 /* NE_EXPR comparisons do not contain much of useful information, except for
317 special case of comparing with the bounds of the type. */
318 if (TREE_CODE (c1) != INTEGER_CST
319 || !INTEGRAL_TYPE_P (type))
320 return;
322 /* Ensure that the condition speaks about an expression in the same type
323 as X and Y. */
324 ctype = TREE_TYPE (c0);
325 if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type))
326 return;
327 c0 = fold_convert (type, c0);
328 c1 = fold_convert (type, c1);
330 if (TYPE_MIN_VALUE (type)
331 && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0))
333 cmp = GT_EXPR;
334 break;
336 if (TYPE_MAX_VALUE (type)
337 && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0))
339 cmp = LT_EXPR;
340 break;
343 return;
344 default:
345 return;
348 mpz_init (offc0);
349 mpz_init (offc1);
350 split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0);
351 split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1);
353 /* We are only interested in comparisons of expressions based on VARX and
354 VARY. TODO -- we might also be able to derive some bounds from
355 expressions containing just one of the variables. */
357 if (operand_equal_p (varx, varc1, 0))
359 tmp = varc0; varc0 = varc1; varc1 = tmp;
360 mpz_swap (offc0, offc1);
361 cmp = swap_tree_comparison (cmp);
364 if (!operand_equal_p (varx, varc0, 0)
365 || !operand_equal_p (vary, varc1, 0))
366 goto end;
368 mpz_init_set (loffx, offx);
369 mpz_init_set (loffy, offy);
371 if (cmp == GT_EXPR || cmp == GE_EXPR)
373 tmp = varx; varx = vary; vary = tmp;
374 mpz_swap (offc0, offc1);
375 mpz_swap (loffx, loffy);
376 cmp = swap_tree_comparison (cmp);
377 lbound = true;
380 /* If there is no overflow, the condition implies that
382 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
384 The overflows and underflows may complicate things a bit; each
385 overflow decreases the appropriate offset by M, and underflow
386 increases it by M. The above inequality would not necessarily be
387 true if
389 -- VARX + OFFX underflows and VARX + OFFC0 does not, or
390 VARX + OFFC0 overflows, but VARX + OFFX does not.
391 This may only happen if OFFX < OFFC0.
392 -- VARY + OFFY overflows and VARY + OFFC1 does not, or
393 VARY + OFFC1 underflows and VARY + OFFY does not.
394 This may only happen if OFFY > OFFC1. */
396 if (no_wrap)
398 x_ok = true;
399 y_ok = true;
401 else
403 x_ok = (integer_zerop (varx)
404 || mpz_cmp (loffx, offc0) >= 0);
405 y_ok = (integer_zerop (vary)
406 || mpz_cmp (loffy, offc1) <= 0);
409 if (x_ok && y_ok)
411 mpz_init (bnd);
412 mpz_sub (bnd, loffx, loffy);
413 mpz_add (bnd, bnd, offc1);
414 mpz_sub (bnd, bnd, offc0);
416 if (cmp == LT_EXPR)
417 mpz_sub_ui (bnd, bnd, 1);
419 if (lbound)
421 mpz_neg (bnd, bnd);
422 if (mpz_cmp (bnds->below, bnd) < 0)
423 mpz_set (bnds->below, bnd);
425 else
427 if (mpz_cmp (bnd, bnds->up) < 0)
428 mpz_set (bnds->up, bnd);
430 mpz_clear (bnd);
433 mpz_clear (loffx);
434 mpz_clear (loffy);
435 end:
436 mpz_clear (offc0);
437 mpz_clear (offc1);
440 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS.
441 The subtraction is considered to be performed in arbitrary precision,
442 without overflows.
444 We do not attempt to be too clever regarding the value ranges of X and
445 Y; most of the time, they are just integers or ssa names offsetted by
446 integer. However, we try to use the information contained in the
447 comparisons before the loop (usually created by loop header copying). */
449 static void
450 bound_difference (struct loop *loop, tree x, tree y, bounds *bnds)
452 tree type = TREE_TYPE (x);
453 tree varx, vary;
454 mpz_t offx, offy;
455 mpz_t minx, maxx, miny, maxy;
456 int cnt = 0;
457 edge e;
458 basic_block bb;
459 tree c0, c1;
460 gimple cond;
461 enum tree_code cmp;
463 /* Get rid of unnecessary casts, but preserve the value of
464 the expressions. */
465 STRIP_SIGN_NOPS (x);
466 STRIP_SIGN_NOPS (y);
468 mpz_init (bnds->below);
469 mpz_init (bnds->up);
470 mpz_init (offx);
471 mpz_init (offy);
472 split_to_var_and_offset (x, &varx, offx);
473 split_to_var_and_offset (y, &vary, offy);
475 if (!integer_zerop (varx)
476 && operand_equal_p (varx, vary, 0))
478 /* Special case VARX == VARY -- we just need to compare the
479 offsets. The matters are a bit more complicated in the
480 case addition of offsets may wrap. */
481 bound_difference_of_offsetted_base (type, offx, offy, bnds);
483 else
485 /* Otherwise, use the value ranges to determine the initial
486 estimates on below and up. */
487 mpz_init (minx);
488 mpz_init (maxx);
489 mpz_init (miny);
490 mpz_init (maxy);
491 determine_value_range (loop, type, varx, offx, minx, maxx);
492 determine_value_range (loop, type, vary, offy, miny, maxy);
494 mpz_sub (bnds->below, minx, maxy);
495 mpz_sub (bnds->up, maxx, miny);
496 mpz_clear (minx);
497 mpz_clear (maxx);
498 mpz_clear (miny);
499 mpz_clear (maxy);
502 /* If both X and Y are constants, we cannot get any more precise. */
503 if (integer_zerop (varx) && integer_zerop (vary))
504 goto end;
506 /* Now walk the dominators of the loop header and use the entry
507 guards to refine the estimates. */
508 for (bb = loop->header;
509 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK;
510 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
512 if (!single_pred_p (bb))
513 continue;
514 e = single_pred_edge (bb);
516 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
517 continue;
519 cond = last_stmt (e->src);
520 c0 = gimple_cond_lhs (cond);
521 cmp = gimple_cond_code (cond);
522 c1 = gimple_cond_rhs (cond);
524 if (e->flags & EDGE_FALSE_VALUE)
525 cmp = invert_tree_comparison (cmp, false);
527 refine_bounds_using_guard (type, varx, offx, vary, offy,
528 c0, cmp, c1, bnds);
529 ++cnt;
532 end:
533 mpz_clear (offx);
534 mpz_clear (offy);
537 /* Update the bounds in BNDS that restrict the value of X to the bounds
538 that restrict the value of X + DELTA. X can be obtained as a
539 difference of two values in TYPE. */
541 static void
542 bounds_add (bounds *bnds, const widest_int &delta, tree type)
544 mpz_t mdelta, max;
546 mpz_init (mdelta);
547 wi::to_mpz (delta, mdelta, SIGNED);
549 mpz_init (max);
550 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED);
552 mpz_add (bnds->up, bnds->up, mdelta);
553 mpz_add (bnds->below, bnds->below, mdelta);
555 if (mpz_cmp (bnds->up, max) > 0)
556 mpz_set (bnds->up, max);
558 mpz_neg (max, max);
559 if (mpz_cmp (bnds->below, max) < 0)
560 mpz_set (bnds->below, max);
562 mpz_clear (mdelta);
563 mpz_clear (max);
566 /* Update the bounds in BNDS that restrict the value of X to the bounds
567 that restrict the value of -X. */
569 static void
570 bounds_negate (bounds *bnds)
572 mpz_t tmp;
574 mpz_init_set (tmp, bnds->up);
575 mpz_neg (bnds->up, bnds->below);
576 mpz_neg (bnds->below, tmp);
577 mpz_clear (tmp);
580 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
582 static tree
583 inverse (tree x, tree mask)
585 tree type = TREE_TYPE (x);
586 tree rslt;
587 unsigned ctr = tree_floor_log2 (mask);
589 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT)
591 unsigned HOST_WIDE_INT ix;
592 unsigned HOST_WIDE_INT imask;
593 unsigned HOST_WIDE_INT irslt = 1;
595 gcc_assert (cst_and_fits_in_hwi (x));
596 gcc_assert (cst_and_fits_in_hwi (mask));
598 ix = int_cst_value (x);
599 imask = int_cst_value (mask);
601 for (; ctr; ctr--)
603 irslt *= ix;
604 ix *= ix;
606 irslt &= imask;
608 rslt = build_int_cst_type (type, irslt);
610 else
612 rslt = build_int_cst (type, 1);
613 for (; ctr; ctr--)
615 rslt = int_const_binop (MULT_EXPR, rslt, x);
616 x = int_const_binop (MULT_EXPR, x, x);
618 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask);
621 return rslt;
624 /* Derives the upper bound BND on the number of executions of loop with exit
625 condition S * i <> C. If NO_OVERFLOW is true, then the control variable of
626 the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed
627 that the loop ends through this exit, i.e., the induction variable ever
628 reaches the value of C.
630 The value C is equal to final - base, where final and base are the final and
631 initial value of the actual induction variable in the analysed loop. BNDS
632 bounds the value of this difference when computed in signed type with
633 unbounded range, while the computation of C is performed in an unsigned
634 type with the range matching the range of the type of the induction variable.
635 In particular, BNDS.up contains an upper bound on C in the following cases:
636 -- if the iv must reach its final value without overflow, i.e., if
637 NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or
638 -- if final >= base, which we know to hold when BNDS.below >= 0. */
640 static void
641 number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s,
642 bounds *bnds, bool exit_must_be_taken)
644 widest_int max;
645 mpz_t d;
646 tree type = TREE_TYPE (c);
647 bool bnds_u_valid = ((no_overflow && exit_must_be_taken)
648 || mpz_sgn (bnds->below) >= 0);
650 if (integer_onep (s)
651 || (TREE_CODE (c) == INTEGER_CST
652 && TREE_CODE (s) == INTEGER_CST
653 && wi::mod_trunc (c, s, TYPE_SIGN (type)) == 0)
654 || (TYPE_OVERFLOW_UNDEFINED (type)
655 && multiple_of_p (type, c, s)))
657 /* If C is an exact multiple of S, then its value will be reached before
658 the induction variable overflows (unless the loop is exited in some
659 other way before). Note that the actual induction variable in the
660 loop (which ranges from base to final instead of from 0 to C) may
661 overflow, in which case BNDS.up will not be giving a correct upper
662 bound on C; thus, BNDS_U_VALID had to be computed in advance. */
663 no_overflow = true;
664 exit_must_be_taken = true;
667 /* If the induction variable can overflow, the number of iterations is at
668 most the period of the control variable (or infinite, but in that case
669 the whole # of iterations analysis will fail). */
670 if (!no_overflow)
672 max = wi::mask <widest_int> (TYPE_PRECISION (type) - wi::ctz (s), false);
673 wi::to_mpz (max, bnd, UNSIGNED);
674 return;
677 /* Now we know that the induction variable does not overflow, so the loop
678 iterates at most (range of type / S) times. */
679 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), bnd, UNSIGNED);
681 /* If the induction variable is guaranteed to reach the value of C before
682 overflow, ... */
683 if (exit_must_be_taken)
685 /* ... then we can strengthen this to C / S, and possibly we can use
686 the upper bound on C given by BNDS. */
687 if (TREE_CODE (c) == INTEGER_CST)
688 wi::to_mpz (c, bnd, UNSIGNED);
689 else if (bnds_u_valid)
690 mpz_set (bnd, bnds->up);
693 mpz_init (d);
694 wi::to_mpz (s, d, UNSIGNED);
695 mpz_fdiv_q (bnd, bnd, d);
696 mpz_clear (d);
699 /* Determines number of iterations of loop whose ending condition
700 is IV <> FINAL. TYPE is the type of the iv. The number of
701 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
702 we know that the exit must be taken eventually, i.e., that the IV
703 ever reaches the value FINAL (we derived this earlier, and possibly set
704 NITER->assumptions to make sure this is the case). BNDS contains the
705 bounds on the difference FINAL - IV->base. */
707 static bool
708 number_of_iterations_ne (tree type, affine_iv *iv, tree final,
709 struct tree_niter_desc *niter, bool exit_must_be_taken,
710 bounds *bnds)
712 tree niter_type = unsigned_type_for (type);
713 tree s, c, d, bits, assumption, tmp, bound;
714 mpz_t max;
716 niter->control = *iv;
717 niter->bound = final;
718 niter->cmp = NE_EXPR;
720 /* Rearrange the terms so that we get inequality S * i <> C, with S
721 positive. Also cast everything to the unsigned type. If IV does
722 not overflow, BNDS bounds the value of C. Also, this is the
723 case if the computation |FINAL - IV->base| does not overflow, i.e.,
724 if BNDS->below in the result is nonnegative. */
725 if (tree_int_cst_sign_bit (iv->step))
727 s = fold_convert (niter_type,
728 fold_build1 (NEGATE_EXPR, type, iv->step));
729 c = fold_build2 (MINUS_EXPR, niter_type,
730 fold_convert (niter_type, iv->base),
731 fold_convert (niter_type, final));
732 bounds_negate (bnds);
734 else
736 s = fold_convert (niter_type, iv->step);
737 c = fold_build2 (MINUS_EXPR, niter_type,
738 fold_convert (niter_type, final),
739 fold_convert (niter_type, iv->base));
742 mpz_init (max);
743 number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds,
744 exit_must_be_taken);
745 niter->max = widest_int::from (wi::from_mpz (niter_type, max, false),
746 TYPE_SIGN (niter_type));
747 mpz_clear (max);
749 /* First the trivial cases -- when the step is 1. */
750 if (integer_onep (s))
752 niter->niter = c;
753 return true;
756 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
757 is infinite. Otherwise, the number of iterations is
758 (inverse(s/d) * (c/d)) mod (size of mode/d). */
759 bits = num_ending_zeros (s);
760 bound = build_low_bits_mask (niter_type,
761 (TYPE_PRECISION (niter_type)
762 - tree_to_uhwi (bits)));
764 d = fold_binary_to_constant (LSHIFT_EXPR, niter_type,
765 build_int_cst (niter_type, 1), bits);
766 s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits);
768 if (!exit_must_be_taken)
770 /* If we cannot assume that the exit is taken eventually, record the
771 assumptions for divisibility of c. */
772 assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d);
773 assumption = fold_build2 (EQ_EXPR, boolean_type_node,
774 assumption, build_int_cst (niter_type, 0));
775 if (!integer_nonzerop (assumption))
776 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
777 niter->assumptions, assumption);
780 c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d);
781 tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound));
782 niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound);
783 return true;
786 /* Checks whether we can determine the final value of the control variable
787 of the loop with ending condition IV0 < IV1 (computed in TYPE).
788 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
789 of the step. The assumptions necessary to ensure that the computation
790 of the final value does not overflow are recorded in NITER. If we
791 find the final value, we adjust DELTA and return TRUE. Otherwise
792 we return false. BNDS bounds the value of IV1->base - IV0->base,
793 and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is
794 true if we know that the exit must be taken eventually. */
796 static bool
797 number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1,
798 struct tree_niter_desc *niter,
799 tree *delta, tree step,
800 bool exit_must_be_taken, bounds *bnds)
802 tree niter_type = TREE_TYPE (step);
803 tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step);
804 tree tmod;
805 mpz_t mmod;
806 tree assumption = boolean_true_node, bound, noloop;
807 bool ret = false, fv_comp_no_overflow;
808 tree type1 = type;
809 if (POINTER_TYPE_P (type))
810 type1 = sizetype;
812 if (TREE_CODE (mod) != INTEGER_CST)
813 return false;
814 if (integer_nonzerop (mod))
815 mod = fold_build2 (MINUS_EXPR, niter_type, step, mod);
816 tmod = fold_convert (type1, mod);
818 mpz_init (mmod);
819 wi::to_mpz (mod, mmod, UNSIGNED);
820 mpz_neg (mmod, mmod);
822 /* If the induction variable does not overflow and the exit is taken,
823 then the computation of the final value does not overflow. This is
824 also obviously the case if the new final value is equal to the
825 current one. Finally, we postulate this for pointer type variables,
826 as the code cannot rely on the object to that the pointer points being
827 placed at the end of the address space (and more pragmatically,
828 TYPE_{MIN,MAX}_VALUE is not defined for pointers). */
829 if (integer_zerop (mod) || POINTER_TYPE_P (type))
830 fv_comp_no_overflow = true;
831 else if (!exit_must_be_taken)
832 fv_comp_no_overflow = false;
833 else
834 fv_comp_no_overflow =
835 (iv0->no_overflow && integer_nonzerop (iv0->step))
836 || (iv1->no_overflow && integer_nonzerop (iv1->step));
838 if (integer_nonzerop (iv0->step))
840 /* The final value of the iv is iv1->base + MOD, assuming that this
841 computation does not overflow, and that
842 iv0->base <= iv1->base + MOD. */
843 if (!fv_comp_no_overflow)
845 bound = fold_build2 (MINUS_EXPR, type1,
846 TYPE_MAX_VALUE (type1), tmod);
847 assumption = fold_build2 (LE_EXPR, boolean_type_node,
848 iv1->base, bound);
849 if (integer_zerop (assumption))
850 goto end;
852 if (mpz_cmp (mmod, bnds->below) < 0)
853 noloop = boolean_false_node;
854 else if (POINTER_TYPE_P (type))
855 noloop = fold_build2 (GT_EXPR, boolean_type_node,
856 iv0->base,
857 fold_build_pointer_plus (iv1->base, tmod));
858 else
859 noloop = fold_build2 (GT_EXPR, boolean_type_node,
860 iv0->base,
861 fold_build2 (PLUS_EXPR, type1,
862 iv1->base, tmod));
864 else
866 /* The final value of the iv is iv0->base - MOD, assuming that this
867 computation does not overflow, and that
868 iv0->base - MOD <= iv1->base. */
869 if (!fv_comp_no_overflow)
871 bound = fold_build2 (PLUS_EXPR, type1,
872 TYPE_MIN_VALUE (type1), tmod);
873 assumption = fold_build2 (GE_EXPR, boolean_type_node,
874 iv0->base, bound);
875 if (integer_zerop (assumption))
876 goto end;
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 fold_build_pointer_plus (iv0->base,
883 fold_build1 (NEGATE_EXPR,
884 type1, tmod)),
885 iv1->base);
886 else
887 noloop = fold_build2 (GT_EXPR, boolean_type_node,
888 fold_build2 (MINUS_EXPR, type1,
889 iv0->base, tmod),
890 iv1->base);
893 if (!integer_nonzerop (assumption))
894 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
895 niter->assumptions,
896 assumption);
897 if (!integer_zerop (noloop))
898 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
899 niter->may_be_zero,
900 noloop);
901 bounds_add (bnds, wi::to_widest (mod), type);
902 *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod);
904 ret = true;
905 end:
906 mpz_clear (mmod);
907 return ret;
910 /* Add assertions to NITER that ensure that the control variable of the loop
911 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
912 are TYPE. Returns false if we can prove that there is an overflow, true
913 otherwise. STEP is the absolute value of the step. */
915 static bool
916 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1,
917 struct tree_niter_desc *niter, tree step)
919 tree bound, d, assumption, diff;
920 tree niter_type = TREE_TYPE (step);
922 if (integer_nonzerop (iv0->step))
924 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
925 if (iv0->no_overflow)
926 return true;
928 /* If iv0->base is a constant, we can determine the last value before
929 overflow precisely; otherwise we conservatively assume
930 MAX - STEP + 1. */
932 if (TREE_CODE (iv0->base) == INTEGER_CST)
934 d = fold_build2 (MINUS_EXPR, niter_type,
935 fold_convert (niter_type, TYPE_MAX_VALUE (type)),
936 fold_convert (niter_type, iv0->base));
937 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
939 else
940 diff = fold_build2 (MINUS_EXPR, niter_type, step,
941 build_int_cst (niter_type, 1));
942 bound = fold_build2 (MINUS_EXPR, type,
943 TYPE_MAX_VALUE (type), fold_convert (type, diff));
944 assumption = fold_build2 (LE_EXPR, boolean_type_node,
945 iv1->base, bound);
947 else
949 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
950 if (iv1->no_overflow)
951 return true;
953 if (TREE_CODE (iv1->base) == INTEGER_CST)
955 d = fold_build2 (MINUS_EXPR, niter_type,
956 fold_convert (niter_type, iv1->base),
957 fold_convert (niter_type, TYPE_MIN_VALUE (type)));
958 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
960 else
961 diff = fold_build2 (MINUS_EXPR, niter_type, step,
962 build_int_cst (niter_type, 1));
963 bound = fold_build2 (PLUS_EXPR, type,
964 TYPE_MIN_VALUE (type), fold_convert (type, diff));
965 assumption = fold_build2 (GE_EXPR, boolean_type_node,
966 iv0->base, bound);
969 if (integer_zerop (assumption))
970 return false;
971 if (!integer_nonzerop (assumption))
972 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
973 niter->assumptions, assumption);
975 iv0->no_overflow = true;
976 iv1->no_overflow = true;
977 return true;
980 /* Add an assumption to NITER that a loop whose ending condition
981 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS
982 bounds the value of IV1->base - IV0->base. */
984 static void
985 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1,
986 struct tree_niter_desc *niter, bounds *bnds)
988 tree assumption = boolean_true_node, bound, diff;
989 tree mbz, mbzl, mbzr, type1;
990 bool rolls_p, no_overflow_p;
991 widest_int dstep;
992 mpz_t mstep, max;
994 /* We are going to compute the number of iterations as
995 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned
996 variant of TYPE. This formula only works if
998 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
1000 (where MAX is the maximum value of the unsigned variant of TYPE, and
1001 the computations in this formula are performed in full precision,
1002 i.e., without overflows).
1004 Usually, for loops with exit condition iv0->base + step * i < iv1->base,
1005 we have a condition of the form iv0->base - step < iv1->base before the loop,
1006 and for loops iv0->base < iv1->base - step * i the condition
1007 iv0->base < iv1->base + step, due to loop header copying, which enable us
1008 to prove the lower bound.
1010 The upper bound is more complicated. Unless the expressions for initial
1011 and final value themselves contain enough information, we usually cannot
1012 derive it from the context. */
1014 /* First check whether the answer does not follow from the bounds we gathered
1015 before. */
1016 if (integer_nonzerop (iv0->step))
1017 dstep = wi::to_widest (iv0->step);
1018 else
1020 dstep = wi::sext (wi::to_widest (iv1->step), TYPE_PRECISION (type));
1021 dstep = -dstep;
1024 mpz_init (mstep);
1025 wi::to_mpz (dstep, mstep, UNSIGNED);
1026 mpz_neg (mstep, mstep);
1027 mpz_add_ui (mstep, mstep, 1);
1029 rolls_p = mpz_cmp (mstep, bnds->below) <= 0;
1031 mpz_init (max);
1032 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED);
1033 mpz_add (max, max, mstep);
1034 no_overflow_p = (mpz_cmp (bnds->up, max) <= 0
1035 /* For pointers, only values lying inside a single object
1036 can be compared or manipulated by pointer arithmetics.
1037 Gcc in general does not allow or handle objects larger
1038 than half of the address space, hence the upper bound
1039 is satisfied for pointers. */
1040 || POINTER_TYPE_P (type));
1041 mpz_clear (mstep);
1042 mpz_clear (max);
1044 if (rolls_p && no_overflow_p)
1045 return;
1047 type1 = type;
1048 if (POINTER_TYPE_P (type))
1049 type1 = sizetype;
1051 /* Now the hard part; we must formulate the assumption(s) as expressions, and
1052 we must be careful not to introduce overflow. */
1054 if (integer_nonzerop (iv0->step))
1056 diff = fold_build2 (MINUS_EXPR, type1,
1057 iv0->step, build_int_cst (type1, 1));
1059 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
1060 0 address never belongs to any object, we can assume this for
1061 pointers. */
1062 if (!POINTER_TYPE_P (type))
1064 bound = fold_build2 (PLUS_EXPR, type1,
1065 TYPE_MIN_VALUE (type), diff);
1066 assumption = fold_build2 (GE_EXPR, boolean_type_node,
1067 iv0->base, bound);
1070 /* And then we can compute iv0->base - diff, and compare it with
1071 iv1->base. */
1072 mbzl = fold_build2 (MINUS_EXPR, type1,
1073 fold_convert (type1, iv0->base), diff);
1074 mbzr = fold_convert (type1, iv1->base);
1076 else
1078 diff = fold_build2 (PLUS_EXPR, type1,
1079 iv1->step, build_int_cst (type1, 1));
1081 if (!POINTER_TYPE_P (type))
1083 bound = fold_build2 (PLUS_EXPR, type1,
1084 TYPE_MAX_VALUE (type), diff);
1085 assumption = fold_build2 (LE_EXPR, boolean_type_node,
1086 iv1->base, bound);
1089 mbzl = fold_convert (type1, iv0->base);
1090 mbzr = fold_build2 (MINUS_EXPR, type1,
1091 fold_convert (type1, iv1->base), diff);
1094 if (!integer_nonzerop (assumption))
1095 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1096 niter->assumptions, assumption);
1097 if (!rolls_p)
1099 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr);
1100 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
1101 niter->may_be_zero, mbz);
1105 /* Determines number of iterations of loop whose ending condition
1106 is IV0 < IV1. TYPE is the type of the iv. The number of
1107 iterations is stored to NITER. BNDS bounds the difference
1108 IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know
1109 that the exit must be taken eventually. */
1111 static bool
1112 number_of_iterations_lt (tree type, affine_iv *iv0, affine_iv *iv1,
1113 struct tree_niter_desc *niter,
1114 bool exit_must_be_taken, bounds *bnds)
1116 tree niter_type = unsigned_type_for (type);
1117 tree delta, step, s;
1118 mpz_t mstep, tmp;
1120 if (integer_nonzerop (iv0->step))
1122 niter->control = *iv0;
1123 niter->cmp = LT_EXPR;
1124 niter->bound = iv1->base;
1126 else
1128 niter->control = *iv1;
1129 niter->cmp = GT_EXPR;
1130 niter->bound = iv0->base;
1133 delta = fold_build2 (MINUS_EXPR, niter_type,
1134 fold_convert (niter_type, iv1->base),
1135 fold_convert (niter_type, iv0->base));
1137 /* First handle the special case that the step is +-1. */
1138 if ((integer_onep (iv0->step) && integer_zerop (iv1->step))
1139 || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step)))
1141 /* for (i = iv0->base; i < iv1->base; i++)
1145 for (i = iv1->base; i > iv0->base; i--).
1147 In both cases # of iterations is iv1->base - iv0->base, assuming that
1148 iv1->base >= iv0->base.
1150 First try to derive a lower bound on the value of
1151 iv1->base - iv0->base, computed in full precision. If the difference
1152 is nonnegative, we are done, otherwise we must record the
1153 condition. */
1155 if (mpz_sgn (bnds->below) < 0)
1156 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node,
1157 iv1->base, iv0->base);
1158 niter->niter = delta;
1159 niter->max = widest_int::from (wi::from_mpz (niter_type, bnds->up, false),
1160 TYPE_SIGN (niter_type));
1161 return true;
1164 if (integer_nonzerop (iv0->step))
1165 step = fold_convert (niter_type, iv0->step);
1166 else
1167 step = fold_convert (niter_type,
1168 fold_build1 (NEGATE_EXPR, type, iv1->step));
1170 /* If we can determine the final value of the control iv exactly, we can
1171 transform the condition to != comparison. In particular, this will be
1172 the case if DELTA is constant. */
1173 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step,
1174 exit_must_be_taken, bnds))
1176 affine_iv zps;
1178 zps.base = build_int_cst (niter_type, 0);
1179 zps.step = step;
1180 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1181 zps does not overflow. */
1182 zps.no_overflow = true;
1184 return number_of_iterations_ne (type, &zps, delta, niter, true, bnds);
1187 /* Make sure that the control iv does not overflow. */
1188 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step))
1189 return false;
1191 /* We determine the number of iterations as (delta + step - 1) / step. For
1192 this to work, we must know that iv1->base >= iv0->base - step + 1,
1193 otherwise the loop does not roll. */
1194 assert_loop_rolls_lt (type, iv0, iv1, niter, bnds);
1196 s = fold_build2 (MINUS_EXPR, niter_type,
1197 step, build_int_cst (niter_type, 1));
1198 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s);
1199 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step);
1201 mpz_init (mstep);
1202 mpz_init (tmp);
1203 wi::to_mpz (step, mstep, UNSIGNED);
1204 mpz_add (tmp, bnds->up, mstep);
1205 mpz_sub_ui (tmp, tmp, 1);
1206 mpz_fdiv_q (tmp, tmp, mstep);
1207 niter->max = widest_int::from (wi::from_mpz (niter_type, tmp, false),
1208 TYPE_SIGN (niter_type));
1209 mpz_clear (mstep);
1210 mpz_clear (tmp);
1212 return true;
1215 /* Determines number of iterations of loop whose ending condition
1216 is IV0 <= IV1. TYPE is the type of the iv. The number of
1217 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
1218 we know that this condition must eventually become false (we derived this
1219 earlier, and possibly set NITER->assumptions to make sure this
1220 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1222 static bool
1223 number_of_iterations_le (tree type, affine_iv *iv0, affine_iv *iv1,
1224 struct tree_niter_desc *niter, bool exit_must_be_taken,
1225 bounds *bnds)
1227 tree assumption;
1228 tree type1 = type;
1229 if (POINTER_TYPE_P (type))
1230 type1 = sizetype;
1232 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1233 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1234 value of the type. This we must know anyway, since if it is
1235 equal to this value, the loop rolls forever. We do not check
1236 this condition for pointer type ivs, as the code cannot rely on
1237 the object to that the pointer points being placed at the end of
1238 the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is
1239 not defined for pointers). */
1241 if (!exit_must_be_taken && !POINTER_TYPE_P (type))
1243 if (integer_nonzerop (iv0->step))
1244 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1245 iv1->base, TYPE_MAX_VALUE (type));
1246 else
1247 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1248 iv0->base, TYPE_MIN_VALUE (type));
1250 if (integer_zerop (assumption))
1251 return false;
1252 if (!integer_nonzerop (assumption))
1253 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1254 niter->assumptions, assumption);
1257 if (integer_nonzerop (iv0->step))
1259 if (POINTER_TYPE_P (type))
1260 iv1->base = fold_build_pointer_plus_hwi (iv1->base, 1);
1261 else
1262 iv1->base = fold_build2 (PLUS_EXPR, type1, iv1->base,
1263 build_int_cst (type1, 1));
1265 else if (POINTER_TYPE_P (type))
1266 iv0->base = fold_build_pointer_plus_hwi (iv0->base, -1);
1267 else
1268 iv0->base = fold_build2 (MINUS_EXPR, type1,
1269 iv0->base, build_int_cst (type1, 1));
1271 bounds_add (bnds, 1, type1);
1273 return number_of_iterations_lt (type, iv0, iv1, niter, exit_must_be_taken,
1274 bnds);
1277 /* Dumps description of affine induction variable IV to FILE. */
1279 static void
1280 dump_affine_iv (FILE *file, affine_iv *iv)
1282 if (!integer_zerop (iv->step))
1283 fprintf (file, "[");
1285 print_generic_expr (dump_file, iv->base, TDF_SLIM);
1287 if (!integer_zerop (iv->step))
1289 fprintf (file, ", + , ");
1290 print_generic_expr (dump_file, iv->step, TDF_SLIM);
1291 fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : "");
1295 /* Determine the number of iterations according to condition (for staying
1296 inside loop) which compares two induction variables using comparison
1297 operator CODE. The induction variable on left side of the comparison
1298 is IV0, the right-hand side is IV1. Both induction variables must have
1299 type TYPE, which must be an integer or pointer type. The steps of the
1300 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1302 LOOP is the loop whose number of iterations we are determining.
1304 ONLY_EXIT is true if we are sure this is the only way the loop could be
1305 exited (including possibly non-returning function calls, exceptions, etc.)
1306 -- in this case we can use the information whether the control induction
1307 variables can overflow or not in a more efficient way.
1309 if EVERY_ITERATION is true, we know the test is executed on every iteration.
1311 The results (number of iterations and assumptions as described in
1312 comments at struct tree_niter_desc in tree-ssa-loop.h) are stored to NITER.
1313 Returns false if it fails to determine number of iterations, true if it
1314 was determined (possibly with some assumptions). */
1316 static bool
1317 number_of_iterations_cond (struct loop *loop,
1318 tree type, affine_iv *iv0, enum tree_code code,
1319 affine_iv *iv1, struct tree_niter_desc *niter,
1320 bool only_exit, bool every_iteration)
1322 bool exit_must_be_taken = false, ret;
1323 bounds bnds;
1325 /* If the test is not executed every iteration, wrapping may make the test
1326 to pass again.
1327 TODO: the overflow case can be still used as unreliable estimate of upper
1328 bound. But we have no API to pass it down to number of iterations code
1329 and, at present, it will not use it anyway. */
1330 if (!every_iteration
1331 && (!iv0->no_overflow || !iv1->no_overflow
1332 || code == NE_EXPR || code == EQ_EXPR))
1333 return false;
1335 /* The meaning of these assumptions is this:
1336 if !assumptions
1337 then the rest of information does not have to be valid
1338 if may_be_zero then the loop does not roll, even if
1339 niter != 0. */
1340 niter->assumptions = boolean_true_node;
1341 niter->may_be_zero = boolean_false_node;
1342 niter->niter = NULL_TREE;
1343 niter->max = 0;
1344 niter->bound = NULL_TREE;
1345 niter->cmp = ERROR_MARK;
1347 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1348 the control variable is on lhs. */
1349 if (code == GE_EXPR || code == GT_EXPR
1350 || (code == NE_EXPR && integer_zerop (iv0->step)))
1352 SWAP (iv0, iv1);
1353 code = swap_tree_comparison (code);
1356 if (POINTER_TYPE_P (type))
1358 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1359 to the same object. If they do, the control variable cannot wrap
1360 (as wrap around the bounds of memory will never return a pointer
1361 that would be guaranteed to point to the same object, even if we
1362 avoid undefined behavior by casting to size_t and back). */
1363 iv0->no_overflow = true;
1364 iv1->no_overflow = true;
1367 /* If the control induction variable does not overflow and the only exit
1368 from the loop is the one that we analyze, we know it must be taken
1369 eventually. */
1370 if (only_exit)
1372 if (!integer_zerop (iv0->step) && iv0->no_overflow)
1373 exit_must_be_taken = true;
1374 else if (!integer_zerop (iv1->step) && iv1->no_overflow)
1375 exit_must_be_taken = true;
1378 /* We can handle the case when neither of the sides of the comparison is
1379 invariant, provided that the test is NE_EXPR. This rarely occurs in
1380 practice, but it is simple enough to manage. */
1381 if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step))
1383 tree step_type = POINTER_TYPE_P (type) ? sizetype : type;
1384 if (code != NE_EXPR)
1385 return false;
1387 iv0->step = fold_binary_to_constant (MINUS_EXPR, step_type,
1388 iv0->step, iv1->step);
1389 iv0->no_overflow = false;
1390 iv1->step = build_int_cst (step_type, 0);
1391 iv1->no_overflow = true;
1394 /* If the result of the comparison is a constant, the loop is weird. More
1395 precise handling would be possible, but the situation is not common enough
1396 to waste time on it. */
1397 if (integer_zerop (iv0->step) && integer_zerop (iv1->step))
1398 return false;
1400 /* Ignore loops of while (i-- < 10) type. */
1401 if (code != NE_EXPR)
1403 if (iv0->step && tree_int_cst_sign_bit (iv0->step))
1404 return false;
1406 if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step))
1407 return false;
1410 /* If the loop exits immediately, there is nothing to do. */
1411 tree tem = fold_binary (code, boolean_type_node, iv0->base, iv1->base);
1412 if (tem && integer_zerop (tem))
1414 niter->niter = build_int_cst (unsigned_type_for (type), 0);
1415 niter->max = 0;
1416 return true;
1419 /* OK, now we know we have a senseful loop. Handle several cases, depending
1420 on what comparison operator is used. */
1421 bound_difference (loop, iv1->base, iv0->base, &bnds);
1423 if (dump_file && (dump_flags & TDF_DETAILS))
1425 fprintf (dump_file,
1426 "Analyzing # of iterations of loop %d\n", loop->num);
1428 fprintf (dump_file, " exit condition ");
1429 dump_affine_iv (dump_file, iv0);
1430 fprintf (dump_file, " %s ",
1431 code == NE_EXPR ? "!="
1432 : code == LT_EXPR ? "<"
1433 : "<=");
1434 dump_affine_iv (dump_file, iv1);
1435 fprintf (dump_file, "\n");
1437 fprintf (dump_file, " bounds on difference of bases: ");
1438 mpz_out_str (dump_file, 10, bnds.below);
1439 fprintf (dump_file, " ... ");
1440 mpz_out_str (dump_file, 10, bnds.up);
1441 fprintf (dump_file, "\n");
1444 switch (code)
1446 case NE_EXPR:
1447 gcc_assert (integer_zerop (iv1->step));
1448 ret = number_of_iterations_ne (type, iv0, iv1->base, niter,
1449 exit_must_be_taken, &bnds);
1450 break;
1452 case LT_EXPR:
1453 ret = number_of_iterations_lt (type, iv0, iv1, niter, exit_must_be_taken,
1454 &bnds);
1455 break;
1457 case LE_EXPR:
1458 ret = number_of_iterations_le (type, iv0, iv1, niter, exit_must_be_taken,
1459 &bnds);
1460 break;
1462 default:
1463 gcc_unreachable ();
1466 mpz_clear (bnds.up);
1467 mpz_clear (bnds.below);
1469 if (dump_file && (dump_flags & TDF_DETAILS))
1471 if (ret)
1473 fprintf (dump_file, " result:\n");
1474 if (!integer_nonzerop (niter->assumptions))
1476 fprintf (dump_file, " under assumptions ");
1477 print_generic_expr (dump_file, niter->assumptions, TDF_SLIM);
1478 fprintf (dump_file, "\n");
1481 if (!integer_zerop (niter->may_be_zero))
1483 fprintf (dump_file, " zero if ");
1484 print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM);
1485 fprintf (dump_file, "\n");
1488 fprintf (dump_file, " # of iterations ");
1489 print_generic_expr (dump_file, niter->niter, TDF_SLIM);
1490 fprintf (dump_file, ", bounded by ");
1491 print_decu (niter->max, dump_file);
1492 fprintf (dump_file, "\n");
1494 else
1495 fprintf (dump_file, " failed\n\n");
1497 return ret;
1500 /* Substitute NEW for OLD in EXPR and fold the result. */
1502 static tree
1503 simplify_replace_tree (tree expr, tree old, tree new_tree)
1505 unsigned i, n;
1506 tree ret = NULL_TREE, e, se;
1508 if (!expr)
1509 return NULL_TREE;
1511 /* Do not bother to replace constants. */
1512 if (CONSTANT_CLASS_P (old))
1513 return expr;
1515 if (expr == old
1516 || operand_equal_p (expr, old, 0))
1517 return unshare_expr (new_tree);
1519 if (!EXPR_P (expr))
1520 return expr;
1522 n = TREE_OPERAND_LENGTH (expr);
1523 for (i = 0; i < n; i++)
1525 e = TREE_OPERAND (expr, i);
1526 se = simplify_replace_tree (e, old, new_tree);
1527 if (e == se)
1528 continue;
1530 if (!ret)
1531 ret = copy_node (expr);
1533 TREE_OPERAND (ret, i) = se;
1536 return (ret ? fold (ret) : expr);
1539 /* Expand definitions of ssa names in EXPR as long as they are simple
1540 enough, and return the new expression. */
1542 tree
1543 expand_simple_operations (tree expr)
1545 unsigned i, n;
1546 tree ret = NULL_TREE, e, ee, e1;
1547 enum tree_code code;
1548 gimple stmt;
1550 if (expr == NULL_TREE)
1551 return expr;
1553 if (is_gimple_min_invariant (expr))
1554 return expr;
1556 code = TREE_CODE (expr);
1557 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code)))
1559 n = TREE_OPERAND_LENGTH (expr);
1560 for (i = 0; i < n; i++)
1562 e = TREE_OPERAND (expr, i);
1563 ee = expand_simple_operations (e);
1564 if (e == ee)
1565 continue;
1567 if (!ret)
1568 ret = copy_node (expr);
1570 TREE_OPERAND (ret, i) = ee;
1573 if (!ret)
1574 return expr;
1576 fold_defer_overflow_warnings ();
1577 ret = fold (ret);
1578 fold_undefer_and_ignore_overflow_warnings ();
1579 return ret;
1582 if (TREE_CODE (expr) != SSA_NAME)
1583 return expr;
1585 stmt = SSA_NAME_DEF_STMT (expr);
1586 if (gimple_code (stmt) == GIMPLE_PHI)
1588 basic_block src, dest;
1590 if (gimple_phi_num_args (stmt) != 1)
1591 return expr;
1592 e = PHI_ARG_DEF (stmt, 0);
1594 /* Avoid propagating through loop exit phi nodes, which
1595 could break loop-closed SSA form restrictions. */
1596 dest = gimple_bb (stmt);
1597 src = single_pred (dest);
1598 if (TREE_CODE (e) == SSA_NAME
1599 && src->loop_father != dest->loop_father)
1600 return expr;
1602 return expand_simple_operations (e);
1604 if (gimple_code (stmt) != GIMPLE_ASSIGN)
1605 return expr;
1607 /* Avoid expanding to expressions that contain SSA names that need
1608 to take part in abnormal coalescing. */
1609 ssa_op_iter iter;
1610 FOR_EACH_SSA_TREE_OPERAND (e, stmt, iter, SSA_OP_USE)
1611 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e))
1612 return expr;
1614 e = gimple_assign_rhs1 (stmt);
1615 code = gimple_assign_rhs_code (stmt);
1616 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1618 if (is_gimple_min_invariant (e))
1619 return e;
1621 if (code == SSA_NAME)
1622 return expand_simple_operations (e);
1624 return expr;
1627 switch (code)
1629 CASE_CONVERT:
1630 /* Casts are simple. */
1631 ee = expand_simple_operations (e);
1632 return fold_build1 (code, TREE_TYPE (expr), ee);
1634 case PLUS_EXPR:
1635 case MINUS_EXPR:
1636 if (TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr)))
1637 return expr;
1638 /* Fallthru. */
1639 case POINTER_PLUS_EXPR:
1640 /* And increments and decrements by a constant are simple. */
1641 e1 = gimple_assign_rhs2 (stmt);
1642 if (!is_gimple_min_invariant (e1))
1643 return expr;
1645 ee = expand_simple_operations (e);
1646 return fold_build2 (code, TREE_TYPE (expr), ee, e1);
1648 default:
1649 return expr;
1653 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1654 expression (or EXPR unchanged, if no simplification was possible). */
1656 static tree
1657 tree_simplify_using_condition_1 (tree cond, tree expr)
1659 bool changed;
1660 tree e, te, e0, e1, e2, notcond;
1661 enum tree_code code = TREE_CODE (expr);
1663 if (code == INTEGER_CST)
1664 return expr;
1666 if (code == TRUTH_OR_EXPR
1667 || code == TRUTH_AND_EXPR
1668 || code == COND_EXPR)
1670 changed = false;
1672 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0));
1673 if (TREE_OPERAND (expr, 0) != e0)
1674 changed = true;
1676 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1));
1677 if (TREE_OPERAND (expr, 1) != e1)
1678 changed = true;
1680 if (code == COND_EXPR)
1682 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2));
1683 if (TREE_OPERAND (expr, 2) != e2)
1684 changed = true;
1686 else
1687 e2 = NULL_TREE;
1689 if (changed)
1691 if (code == COND_EXPR)
1692 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1693 else
1694 expr = fold_build2 (code, boolean_type_node, e0, e1);
1697 return expr;
1700 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
1701 propagation, and vice versa. Fold does not handle this, since it is
1702 considered too expensive. */
1703 if (TREE_CODE (cond) == EQ_EXPR)
1705 e0 = TREE_OPERAND (cond, 0);
1706 e1 = TREE_OPERAND (cond, 1);
1708 /* We know that e0 == e1. Check whether we cannot simplify expr
1709 using this fact. */
1710 e = simplify_replace_tree (expr, e0, e1);
1711 if (integer_zerop (e) || integer_nonzerop (e))
1712 return e;
1714 e = simplify_replace_tree (expr, e1, e0);
1715 if (integer_zerop (e) || integer_nonzerop (e))
1716 return e;
1718 if (TREE_CODE (expr) == EQ_EXPR)
1720 e0 = TREE_OPERAND (expr, 0);
1721 e1 = TREE_OPERAND (expr, 1);
1723 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
1724 e = simplify_replace_tree (cond, e0, e1);
1725 if (integer_zerop (e))
1726 return e;
1727 e = simplify_replace_tree (cond, e1, e0);
1728 if (integer_zerop (e))
1729 return e;
1731 if (TREE_CODE (expr) == NE_EXPR)
1733 e0 = TREE_OPERAND (expr, 0);
1734 e1 = TREE_OPERAND (expr, 1);
1736 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
1737 e = simplify_replace_tree (cond, e0, e1);
1738 if (integer_zerop (e))
1739 return boolean_true_node;
1740 e = simplify_replace_tree (cond, e1, e0);
1741 if (integer_zerop (e))
1742 return boolean_true_node;
1745 te = expand_simple_operations (expr);
1747 /* Check whether COND ==> EXPR. */
1748 notcond = invert_truthvalue (cond);
1749 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, te);
1750 if (e && integer_nonzerop (e))
1751 return e;
1753 /* Check whether COND ==> not EXPR. */
1754 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, te);
1755 if (e && integer_zerop (e))
1756 return e;
1758 return expr;
1761 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1762 expression (or EXPR unchanged, if no simplification was possible).
1763 Wrapper around tree_simplify_using_condition_1 that ensures that chains
1764 of simple operations in definitions of ssa names in COND are expanded,
1765 so that things like casts or incrementing the value of the bound before
1766 the loop do not cause us to fail. */
1768 static tree
1769 tree_simplify_using_condition (tree cond, tree expr)
1771 cond = expand_simple_operations (cond);
1773 return tree_simplify_using_condition_1 (cond, expr);
1776 /* Tries to simplify EXPR using the conditions on entry to LOOP.
1777 Returns the simplified expression (or EXPR unchanged, if no
1778 simplification was possible).*/
1780 static tree
1781 simplify_using_initial_conditions (struct loop *loop, tree expr)
1783 edge e;
1784 basic_block bb;
1785 gimple stmt;
1786 tree cond;
1787 int cnt = 0;
1789 if (TREE_CODE (expr) == INTEGER_CST)
1790 return expr;
1792 /* Limit walking the dominators to avoid quadraticness in
1793 the number of BBs times the number of loops in degenerate
1794 cases. */
1795 for (bb = loop->header;
1796 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK;
1797 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
1799 if (!single_pred_p (bb))
1800 continue;
1801 e = single_pred_edge (bb);
1803 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
1804 continue;
1806 stmt = last_stmt (e->src);
1807 cond = fold_build2 (gimple_cond_code (stmt),
1808 boolean_type_node,
1809 gimple_cond_lhs (stmt),
1810 gimple_cond_rhs (stmt));
1811 if (e->flags & EDGE_FALSE_VALUE)
1812 cond = invert_truthvalue (cond);
1813 expr = tree_simplify_using_condition (cond, expr);
1814 ++cnt;
1817 return expr;
1820 /* Tries to simplify EXPR using the evolutions of the loop invariants
1821 in the superloops of LOOP. Returns the simplified expression
1822 (or EXPR unchanged, if no simplification was possible). */
1824 static tree
1825 simplify_using_outer_evolutions (struct loop *loop, tree expr)
1827 enum tree_code code = TREE_CODE (expr);
1828 bool changed;
1829 tree e, e0, e1, e2;
1831 if (is_gimple_min_invariant (expr))
1832 return expr;
1834 if (code == TRUTH_OR_EXPR
1835 || code == TRUTH_AND_EXPR
1836 || code == COND_EXPR)
1838 changed = false;
1840 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0));
1841 if (TREE_OPERAND (expr, 0) != e0)
1842 changed = true;
1844 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1));
1845 if (TREE_OPERAND (expr, 1) != e1)
1846 changed = true;
1848 if (code == COND_EXPR)
1850 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2));
1851 if (TREE_OPERAND (expr, 2) != e2)
1852 changed = true;
1854 else
1855 e2 = NULL_TREE;
1857 if (changed)
1859 if (code == COND_EXPR)
1860 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1861 else
1862 expr = fold_build2 (code, boolean_type_node, e0, e1);
1865 return expr;
1868 e = instantiate_parameters (loop, expr);
1869 if (is_gimple_min_invariant (e))
1870 return e;
1872 return expr;
1875 /* Returns true if EXIT is the only possible exit from LOOP. */
1877 bool
1878 loop_only_exit_p (const struct loop *loop, const_edge exit)
1880 basic_block *body;
1881 gimple_stmt_iterator bsi;
1882 unsigned i;
1883 gimple call;
1885 if (exit != single_exit (loop))
1886 return false;
1888 body = get_loop_body (loop);
1889 for (i = 0; i < loop->num_nodes; i++)
1891 for (bsi = gsi_start_bb (body[i]); !gsi_end_p (bsi); gsi_next (&bsi))
1893 call = gsi_stmt (bsi);
1894 if (gimple_code (call) != GIMPLE_CALL)
1895 continue;
1897 if (gimple_has_side_effects (call))
1899 free (body);
1900 return false;
1905 free (body);
1906 return true;
1909 /* Stores description of number of iterations of LOOP derived from
1910 EXIT (an exit edge of the LOOP) in NITER. Returns true if some
1911 useful information could be derived (and fields of NITER has
1912 meaning described in comments at struct tree_niter_desc
1913 declaration), false otherwise. If WARN is true and
1914 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use
1915 potentially unsafe assumptions.
1916 When EVERY_ITERATION is true, only tests that are known to be executed
1917 every iteration are considered (i.e. only test that alone bounds the loop).
1920 bool
1921 number_of_iterations_exit (struct loop *loop, edge exit,
1922 struct tree_niter_desc *niter,
1923 bool warn, bool every_iteration)
1925 gimple stmt;
1926 tree type;
1927 tree op0, op1;
1928 enum tree_code code;
1929 affine_iv iv0, iv1;
1930 bool safe;
1932 safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src);
1934 if (every_iteration && !safe)
1935 return false;
1937 niter->assumptions = boolean_false_node;
1938 stmt = last_stmt (exit->src);
1939 if (!stmt || gimple_code (stmt) != GIMPLE_COND)
1940 return false;
1942 /* We want the condition for staying inside loop. */
1943 code = gimple_cond_code (stmt);
1944 if (exit->flags & EDGE_TRUE_VALUE)
1945 code = invert_tree_comparison (code, false);
1947 switch (code)
1949 case GT_EXPR:
1950 case GE_EXPR:
1951 case LT_EXPR:
1952 case LE_EXPR:
1953 case NE_EXPR:
1954 break;
1956 default:
1957 return false;
1960 op0 = gimple_cond_lhs (stmt);
1961 op1 = gimple_cond_rhs (stmt);
1962 type = TREE_TYPE (op0);
1964 if (TREE_CODE (type) != INTEGER_TYPE
1965 && !POINTER_TYPE_P (type))
1966 return false;
1968 if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, false))
1969 return false;
1970 if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, false))
1971 return false;
1973 /* We don't want to see undefined signed overflow warnings while
1974 computing the number of iterations. */
1975 fold_defer_overflow_warnings ();
1977 iv0.base = expand_simple_operations (iv0.base);
1978 iv1.base = expand_simple_operations (iv1.base);
1979 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter,
1980 loop_only_exit_p (loop, exit), safe))
1982 fold_undefer_and_ignore_overflow_warnings ();
1983 return false;
1986 if (optimize >= 3)
1988 niter->assumptions = simplify_using_outer_evolutions (loop,
1989 niter->assumptions);
1990 niter->may_be_zero = simplify_using_outer_evolutions (loop,
1991 niter->may_be_zero);
1992 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
1995 niter->assumptions
1996 = simplify_using_initial_conditions (loop,
1997 niter->assumptions);
1998 niter->may_be_zero
1999 = simplify_using_initial_conditions (loop,
2000 niter->may_be_zero);
2002 fold_undefer_and_ignore_overflow_warnings ();
2004 /* If NITER has simplified into a constant, update MAX. */
2005 if (TREE_CODE (niter->niter) == INTEGER_CST)
2006 niter->max = wi::to_widest (niter->niter);
2008 if (integer_onep (niter->assumptions))
2009 return true;
2011 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
2012 But if we can prove that there is overflow or some other source of weird
2013 behavior, ignore the loop even with -funsafe-loop-optimizations. */
2014 if (integer_zerop (niter->assumptions) || !single_exit (loop))
2015 return false;
2017 if (flag_unsafe_loop_optimizations)
2018 niter->assumptions = boolean_true_node;
2020 if (warn)
2022 const char *wording;
2023 location_t loc = gimple_location (stmt);
2025 /* We can provide a more specific warning if one of the operator is
2026 constant and the other advances by +1 or -1. */
2027 if (!integer_zerop (iv1.step)
2028 ? (integer_zerop (iv0.step)
2029 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step)))
2030 : (integer_onep (iv0.step) || integer_all_onesp (iv0.step)))
2031 wording =
2032 flag_unsafe_loop_optimizations
2033 ? N_("assuming that the loop is not infinite")
2034 : N_("cannot optimize possibly infinite loops");
2035 else
2036 wording =
2037 flag_unsafe_loop_optimizations
2038 ? N_("assuming that the loop counter does not overflow")
2039 : N_("cannot optimize loop, the loop counter may overflow");
2041 warning_at ((LOCATION_LINE (loc) > 0) ? loc : input_location,
2042 OPT_Wunsafe_loop_optimizations, "%s", gettext (wording));
2045 return flag_unsafe_loop_optimizations;
2048 /* Try to determine the number of iterations of LOOP. If we succeed,
2049 expression giving number of iterations is returned and *EXIT is
2050 set to the edge from that the information is obtained. Otherwise
2051 chrec_dont_know is returned. */
2053 tree
2054 find_loop_niter (struct loop *loop, edge *exit)
2056 unsigned i;
2057 vec<edge> exits = get_loop_exit_edges (loop);
2058 edge ex;
2059 tree niter = NULL_TREE, aniter;
2060 struct tree_niter_desc desc;
2062 *exit = NULL;
2063 FOR_EACH_VEC_ELT (exits, i, ex)
2065 if (!number_of_iterations_exit (loop, ex, &desc, false))
2066 continue;
2068 if (integer_nonzerop (desc.may_be_zero))
2070 /* We exit in the first iteration through this exit.
2071 We won't find anything better. */
2072 niter = build_int_cst (unsigned_type_node, 0);
2073 *exit = ex;
2074 break;
2077 if (!integer_zerop (desc.may_be_zero))
2078 continue;
2080 aniter = desc.niter;
2082 if (!niter)
2084 /* Nothing recorded yet. */
2085 niter = aniter;
2086 *exit = ex;
2087 continue;
2090 /* Prefer constants, the lower the better. */
2091 if (TREE_CODE (aniter) != INTEGER_CST)
2092 continue;
2094 if (TREE_CODE (niter) != INTEGER_CST)
2096 niter = aniter;
2097 *exit = ex;
2098 continue;
2101 if (tree_int_cst_lt (aniter, niter))
2103 niter = aniter;
2104 *exit = ex;
2105 continue;
2108 exits.release ();
2110 return niter ? niter : chrec_dont_know;
2113 /* Return true if loop is known to have bounded number of iterations. */
2115 bool
2116 finite_loop_p (struct loop *loop)
2118 widest_int nit;
2119 int flags;
2121 if (flag_unsafe_loop_optimizations)
2122 return true;
2123 flags = flags_from_decl_or_type (current_function_decl);
2124 if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE))
2126 if (dump_file && (dump_flags & TDF_DETAILS))
2127 fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n",
2128 loop->num);
2129 return true;
2132 if (loop->any_upper_bound
2133 || max_loop_iterations (loop, &nit))
2135 if (dump_file && (dump_flags & TDF_DETAILS))
2136 fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n",
2137 loop->num);
2138 return true;
2140 return false;
2145 Analysis of a number of iterations of a loop by a brute-force evaluation.
2149 /* Bound on the number of iterations we try to evaluate. */
2151 #define MAX_ITERATIONS_TO_TRACK \
2152 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
2154 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
2155 result by a chain of operations such that all but exactly one of their
2156 operands are constants. */
2158 static gimple
2159 chain_of_csts_start (struct loop *loop, tree x)
2161 gimple stmt = SSA_NAME_DEF_STMT (x);
2162 tree use;
2163 basic_block bb = gimple_bb (stmt);
2164 enum tree_code code;
2166 if (!bb
2167 || !flow_bb_inside_loop_p (loop, bb))
2168 return NULL;
2170 if (gimple_code (stmt) == GIMPLE_PHI)
2172 if (bb == loop->header)
2173 return stmt;
2175 return NULL;
2178 if (gimple_code (stmt) != GIMPLE_ASSIGN
2179 || gimple_assign_rhs_class (stmt) == GIMPLE_TERNARY_RHS)
2180 return NULL;
2182 code = gimple_assign_rhs_code (stmt);
2183 if (gimple_references_memory_p (stmt)
2184 || TREE_CODE_CLASS (code) == tcc_reference
2185 || (code == ADDR_EXPR
2186 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
2187 return NULL;
2189 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
2190 if (use == NULL_TREE)
2191 return NULL;
2193 return chain_of_csts_start (loop, use);
2196 /* Determines whether the expression X is derived from a result of a phi node
2197 in header of LOOP such that
2199 * the derivation of X consists only from operations with constants
2200 * the initial value of the phi node is constant
2201 * the value of the phi node in the next iteration can be derived from the
2202 value in the current iteration by a chain of operations with constants.
2204 If such phi node exists, it is returned, otherwise NULL is returned. */
2206 static gimple
2207 get_base_for (struct loop *loop, tree x)
2209 gimple phi;
2210 tree init, next;
2212 if (is_gimple_min_invariant (x))
2213 return NULL;
2215 phi = chain_of_csts_start (loop, x);
2216 if (!phi)
2217 return NULL;
2219 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2220 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2222 if (TREE_CODE (next) != SSA_NAME)
2223 return NULL;
2225 if (!is_gimple_min_invariant (init))
2226 return NULL;
2228 if (chain_of_csts_start (loop, next) != phi)
2229 return NULL;
2231 return phi;
2234 /* Given an expression X, then
2236 * if X is NULL_TREE, we return the constant BASE.
2237 * otherwise X is a SSA name, whose value in the considered loop is derived
2238 by a chain of operations with constant from a result of a phi node in
2239 the header of the loop. Then we return value of X when the value of the
2240 result of this phi node is given by the constant BASE. */
2242 static tree
2243 get_val_for (tree x, tree base)
2245 gimple stmt;
2247 gcc_checking_assert (is_gimple_min_invariant (base));
2249 if (!x)
2250 return base;
2252 stmt = SSA_NAME_DEF_STMT (x);
2253 if (gimple_code (stmt) == GIMPLE_PHI)
2254 return base;
2256 gcc_checking_assert (is_gimple_assign (stmt));
2258 /* STMT must be either an assignment of a single SSA name or an
2259 expression involving an SSA name and a constant. Try to fold that
2260 expression using the value for the SSA name. */
2261 if (gimple_assign_ssa_name_copy_p (stmt))
2262 return get_val_for (gimple_assign_rhs1 (stmt), base);
2263 else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS
2264 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
2266 return fold_build1 (gimple_assign_rhs_code (stmt),
2267 gimple_expr_type (stmt),
2268 get_val_for (gimple_assign_rhs1 (stmt), base));
2270 else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS)
2272 tree rhs1 = gimple_assign_rhs1 (stmt);
2273 tree rhs2 = gimple_assign_rhs2 (stmt);
2274 if (TREE_CODE (rhs1) == SSA_NAME)
2275 rhs1 = get_val_for (rhs1, base);
2276 else if (TREE_CODE (rhs2) == SSA_NAME)
2277 rhs2 = get_val_for (rhs2, base);
2278 else
2279 gcc_unreachable ();
2280 return fold_build2 (gimple_assign_rhs_code (stmt),
2281 gimple_expr_type (stmt), rhs1, rhs2);
2283 else
2284 gcc_unreachable ();
2288 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2289 by brute force -- i.e. by determining the value of the operands of the
2290 condition at EXIT in first few iterations of the loop (assuming that
2291 these values are constant) and determining the first one in that the
2292 condition is not satisfied. Returns the constant giving the number
2293 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2295 tree
2296 loop_niter_by_eval (struct loop *loop, edge exit)
2298 tree acnd;
2299 tree op[2], val[2], next[2], aval[2];
2300 gimple phi, cond;
2301 unsigned i, j;
2302 enum tree_code cmp;
2304 cond = last_stmt (exit->src);
2305 if (!cond || gimple_code (cond) != GIMPLE_COND)
2306 return chrec_dont_know;
2308 cmp = gimple_cond_code (cond);
2309 if (exit->flags & EDGE_TRUE_VALUE)
2310 cmp = invert_tree_comparison (cmp, false);
2312 switch (cmp)
2314 case EQ_EXPR:
2315 case NE_EXPR:
2316 case GT_EXPR:
2317 case GE_EXPR:
2318 case LT_EXPR:
2319 case LE_EXPR:
2320 op[0] = gimple_cond_lhs (cond);
2321 op[1] = gimple_cond_rhs (cond);
2322 break;
2324 default:
2325 return chrec_dont_know;
2328 for (j = 0; j < 2; j++)
2330 if (is_gimple_min_invariant (op[j]))
2332 val[j] = op[j];
2333 next[j] = NULL_TREE;
2334 op[j] = NULL_TREE;
2336 else
2338 phi = get_base_for (loop, op[j]);
2339 if (!phi)
2340 return chrec_dont_know;
2341 val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2342 next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2346 /* Don't issue signed overflow warnings. */
2347 fold_defer_overflow_warnings ();
2349 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
2351 for (j = 0; j < 2; j++)
2352 aval[j] = get_val_for (op[j], val[j]);
2354 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
2355 if (acnd && integer_zerop (acnd))
2357 fold_undefer_and_ignore_overflow_warnings ();
2358 if (dump_file && (dump_flags & TDF_DETAILS))
2359 fprintf (dump_file,
2360 "Proved that loop %d iterates %d times using brute force.\n",
2361 loop->num, i);
2362 return build_int_cst (unsigned_type_node, i);
2365 for (j = 0; j < 2; j++)
2367 val[j] = get_val_for (next[j], val[j]);
2368 if (!is_gimple_min_invariant (val[j]))
2370 fold_undefer_and_ignore_overflow_warnings ();
2371 return chrec_dont_know;
2376 fold_undefer_and_ignore_overflow_warnings ();
2378 return chrec_dont_know;
2381 /* Finds the exit of the LOOP by that the loop exits after a constant
2382 number of iterations and stores the exit edge to *EXIT. The constant
2383 giving the number of iterations of LOOP is returned. The number of
2384 iterations is determined using loop_niter_by_eval (i.e. by brute force
2385 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2386 determines the number of iterations, chrec_dont_know is returned. */
2388 tree
2389 find_loop_niter_by_eval (struct loop *loop, edge *exit)
2391 unsigned i;
2392 vec<edge> exits = get_loop_exit_edges (loop);
2393 edge ex;
2394 tree niter = NULL_TREE, aniter;
2396 *exit = NULL;
2398 /* Loops with multiple exits are expensive to handle and less important. */
2399 if (!flag_expensive_optimizations
2400 && exits.length () > 1)
2402 exits.release ();
2403 return chrec_dont_know;
2406 FOR_EACH_VEC_ELT (exits, i, ex)
2408 if (!just_once_each_iteration_p (loop, ex->src))
2409 continue;
2411 aniter = loop_niter_by_eval (loop, ex);
2412 if (chrec_contains_undetermined (aniter))
2413 continue;
2415 if (niter
2416 && !tree_int_cst_lt (aniter, niter))
2417 continue;
2419 niter = aniter;
2420 *exit = ex;
2422 exits.release ();
2424 return niter ? niter : chrec_dont_know;
2429 Analysis of upper bounds on number of iterations of a loop.
2433 static widest_int derive_constant_upper_bound_ops (tree, tree,
2434 enum tree_code, tree);
2436 /* Returns a constant upper bound on the value of the right-hand side of
2437 an assignment statement STMT. */
2439 static widest_int
2440 derive_constant_upper_bound_assign (gimple stmt)
2442 enum tree_code code = gimple_assign_rhs_code (stmt);
2443 tree op0 = gimple_assign_rhs1 (stmt);
2444 tree op1 = gimple_assign_rhs2 (stmt);
2446 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)),
2447 op0, code, op1);
2450 /* Returns a constant upper bound on the value of expression VAL. VAL
2451 is considered to be unsigned. If its type is signed, its value must
2452 be nonnegative. */
2454 static widest_int
2455 derive_constant_upper_bound (tree val)
2457 enum tree_code code;
2458 tree op0, op1;
2460 extract_ops_from_tree (val, &code, &op0, &op1);
2461 return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1);
2464 /* Returns a constant upper bound on the value of expression OP0 CODE OP1,
2465 whose type is TYPE. The expression is considered to be unsigned. If
2466 its type is signed, its value must be nonnegative. */
2468 static widest_int
2469 derive_constant_upper_bound_ops (tree type, tree op0,
2470 enum tree_code code, tree op1)
2472 tree subtype, maxt;
2473 widest_int bnd, max, mmax, cst;
2474 gimple stmt;
2476 if (INTEGRAL_TYPE_P (type))
2477 maxt = TYPE_MAX_VALUE (type);
2478 else
2479 maxt = upper_bound_in_type (type, type);
2481 max = wi::to_widest (maxt);
2483 switch (code)
2485 case INTEGER_CST:
2486 return wi::to_widest (op0);
2488 CASE_CONVERT:
2489 subtype = TREE_TYPE (op0);
2490 if (!TYPE_UNSIGNED (subtype)
2491 /* If TYPE is also signed, the fact that VAL is nonnegative implies
2492 that OP0 is nonnegative. */
2493 && TYPE_UNSIGNED (type)
2494 && !tree_expr_nonnegative_p (op0))
2496 /* If we cannot prove that the casted expression is nonnegative,
2497 we cannot establish more useful upper bound than the precision
2498 of the type gives us. */
2499 return max;
2502 /* We now know that op0 is an nonnegative value. Try deriving an upper
2503 bound for it. */
2504 bnd = derive_constant_upper_bound (op0);
2506 /* If the bound does not fit in TYPE, max. value of TYPE could be
2507 attained. */
2508 if (wi::ltu_p (max, bnd))
2509 return max;
2511 return bnd;
2513 case PLUS_EXPR:
2514 case POINTER_PLUS_EXPR:
2515 case MINUS_EXPR:
2516 if (TREE_CODE (op1) != INTEGER_CST
2517 || !tree_expr_nonnegative_p (op0))
2518 return max;
2520 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
2521 choose the most logical way how to treat this constant regardless
2522 of the signedness of the type. */
2523 cst = wi::sext (wi::to_widest (op1), TYPE_PRECISION (type));
2524 if (code != MINUS_EXPR)
2525 cst = -cst;
2527 bnd = derive_constant_upper_bound (op0);
2529 if (wi::neg_p (cst))
2531 cst = -cst;
2532 /* Avoid CST == 0x80000... */
2533 if (wi::neg_p (cst))
2534 return max;;
2536 /* OP0 + CST. We need to check that
2537 BND <= MAX (type) - CST. */
2539 mmax -= cst;
2540 if (wi::ltu_p (bnd, max))
2541 return max;
2543 return bnd + cst;
2545 else
2547 /* OP0 - CST, where CST >= 0.
2549 If TYPE is signed, we have already verified that OP0 >= 0, and we
2550 know that the result is nonnegative. This implies that
2551 VAL <= BND - CST.
2553 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2554 otherwise the operation underflows.
2557 /* This should only happen if the type is unsigned; however, for
2558 buggy programs that use overflowing signed arithmetics even with
2559 -fno-wrapv, this condition may also be true for signed values. */
2560 if (wi::ltu_p (bnd, cst))
2561 return max;
2563 if (TYPE_UNSIGNED (type))
2565 tree tem = fold_binary (GE_EXPR, boolean_type_node, op0,
2566 wide_int_to_tree (type, cst));
2567 if (!tem || integer_nonzerop (tem))
2568 return max;
2571 bnd -= cst;
2574 return bnd;
2576 case FLOOR_DIV_EXPR:
2577 case EXACT_DIV_EXPR:
2578 if (TREE_CODE (op1) != INTEGER_CST
2579 || tree_int_cst_sign_bit (op1))
2580 return max;
2582 bnd = derive_constant_upper_bound (op0);
2583 return wi::udiv_floor (bnd, wi::to_widest (op1));
2585 case BIT_AND_EXPR:
2586 if (TREE_CODE (op1) != INTEGER_CST
2587 || tree_int_cst_sign_bit (op1))
2588 return max;
2589 return wi::to_widest (op1);
2591 case SSA_NAME:
2592 stmt = SSA_NAME_DEF_STMT (op0);
2593 if (gimple_code (stmt) != GIMPLE_ASSIGN
2594 || gimple_assign_lhs (stmt) != op0)
2595 return max;
2596 return derive_constant_upper_bound_assign (stmt);
2598 default:
2599 return max;
2603 /* Emit a -Waggressive-loop-optimizations warning if needed. */
2605 static void
2606 do_warn_aggressive_loop_optimizations (struct loop *loop,
2607 widest_int i_bound, gimple stmt)
2609 /* Don't warn if the loop doesn't have known constant bound. */
2610 if (!loop->nb_iterations
2611 || TREE_CODE (loop->nb_iterations) != INTEGER_CST
2612 || !warn_aggressive_loop_optimizations
2613 /* To avoid warning multiple times for the same loop,
2614 only start warning when we preserve loops. */
2615 || (cfun->curr_properties & PROP_loops) == 0
2616 /* Only warn once per loop. */
2617 || loop->warned_aggressive_loop_optimizations
2618 /* Only warn if undefined behavior gives us lower estimate than the
2619 known constant bound. */
2620 || wi::cmpu (i_bound, wi::to_widest (loop->nb_iterations)) >= 0
2621 /* And undefined behavior happens unconditionally. */
2622 || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt)))
2623 return;
2625 edge e = single_exit (loop);
2626 if (e == NULL)
2627 return;
2629 gimple estmt = last_stmt (e->src);
2630 if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations,
2631 "iteration %E invokes undefined behavior",
2632 wide_int_to_tree (TREE_TYPE (loop->nb_iterations),
2633 i_bound)))
2634 inform (gimple_location (estmt), "containing loop");
2635 loop->warned_aggressive_loop_optimizations = true;
2638 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
2639 is true if the loop is exited immediately after STMT, and this exit
2640 is taken at last when the STMT is executed BOUND + 1 times.
2641 REALISTIC is true if BOUND is expected to be close to the real number
2642 of iterations. UPPER is true if we are sure the loop iterates at most
2643 BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */
2645 static void
2646 record_estimate (struct loop *loop, tree bound, const widest_int &i_bound,
2647 gimple at_stmt, bool is_exit, bool realistic, bool upper)
2649 widest_int delta;
2651 if (dump_file && (dump_flags & TDF_DETAILS))
2653 fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : "");
2654 print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM);
2655 fprintf (dump_file, " is %sexecuted at most ",
2656 upper ? "" : "probably ");
2657 print_generic_expr (dump_file, bound, TDF_SLIM);
2658 fprintf (dump_file, " (bounded by ");
2659 print_decu (i_bound, dump_file);
2660 fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num);
2663 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
2664 real number of iterations. */
2665 if (TREE_CODE (bound) != INTEGER_CST)
2666 realistic = false;
2667 else
2668 gcc_checking_assert (i_bound == wi::to_widest (bound));
2669 if (!upper && !realistic)
2670 return;
2672 /* If we have a guaranteed upper bound, record it in the appropriate
2673 list, unless this is an !is_exit bound (i.e. undefined behavior in
2674 at_stmt) in a loop with known constant number of iterations. */
2675 if (upper
2676 && (is_exit
2677 || loop->nb_iterations == NULL_TREE
2678 || TREE_CODE (loop->nb_iterations) != INTEGER_CST))
2680 struct nb_iter_bound *elt = ggc_alloc<nb_iter_bound> ();
2682 elt->bound = i_bound;
2683 elt->stmt = at_stmt;
2684 elt->is_exit = is_exit;
2685 elt->next = loop->bounds;
2686 loop->bounds = elt;
2689 /* If statement is executed on every path to the loop latch, we can directly
2690 infer the upper bound on the # of iterations of the loop. */
2691 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt)))
2692 return;
2694 /* Update the number of iteration estimates according to the bound.
2695 If at_stmt is an exit then the loop latch is executed at most BOUND times,
2696 otherwise it can be executed BOUND + 1 times. We will lower the estimate
2697 later if such statement must be executed on last iteration */
2698 if (is_exit)
2699 delta = 0;
2700 else
2701 delta = 1;
2702 widest_int new_i_bound = i_bound + delta;
2704 /* If an overflow occurred, ignore the result. */
2705 if (wi::ltu_p (new_i_bound, delta))
2706 return;
2708 if (upper && !is_exit)
2709 do_warn_aggressive_loop_optimizations (loop, new_i_bound, at_stmt);
2710 record_niter_bound (loop, new_i_bound, realistic, upper);
2713 /* Record the estimate on number of iterations of LOOP based on the fact that
2714 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
2715 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
2716 estimated number of iterations is expected to be close to the real one.
2717 UPPER is true if we are sure the induction variable does not wrap. */
2719 static void
2720 record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple stmt,
2721 tree low, tree high, bool realistic, bool upper)
2723 tree niter_bound, extreme, delta;
2724 tree type = TREE_TYPE (base), unsigned_type;
2726 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step))
2727 return;
2729 if (dump_file && (dump_flags & TDF_DETAILS))
2731 fprintf (dump_file, "Induction variable (");
2732 print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM);
2733 fprintf (dump_file, ") ");
2734 print_generic_expr (dump_file, base, TDF_SLIM);
2735 fprintf (dump_file, " + ");
2736 print_generic_expr (dump_file, step, TDF_SLIM);
2737 fprintf (dump_file, " * iteration does not wrap in statement ");
2738 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2739 fprintf (dump_file, " in loop %d.\n", loop->num);
2742 unsigned_type = unsigned_type_for (type);
2743 base = fold_convert (unsigned_type, base);
2744 step = fold_convert (unsigned_type, step);
2746 if (tree_int_cst_sign_bit (step))
2748 extreme = fold_convert (unsigned_type, low);
2749 if (TREE_CODE (base) != INTEGER_CST)
2750 base = fold_convert (unsigned_type, high);
2751 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
2752 step = fold_build1 (NEGATE_EXPR, unsigned_type, step);
2754 else
2756 extreme = fold_convert (unsigned_type, high);
2757 if (TREE_CODE (base) != INTEGER_CST)
2758 base = fold_convert (unsigned_type, low);
2759 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
2762 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
2763 would get out of the range. */
2764 niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step);
2765 widest_int max = derive_constant_upper_bound (niter_bound);
2766 record_estimate (loop, niter_bound, max, stmt, false, realistic, upper);
2769 /* Determine information about number of iterations a LOOP from the index
2770 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
2771 guaranteed to be executed in every iteration of LOOP. Callback for
2772 for_each_index. */
2774 struct ilb_data
2776 struct loop *loop;
2777 gimple stmt;
2780 static bool
2781 idx_infer_loop_bounds (tree base, tree *idx, void *dta)
2783 struct ilb_data *data = (struct ilb_data *) dta;
2784 tree ev, init, step;
2785 tree low, high, type, next;
2786 bool sign, upper = true, at_end = false;
2787 struct loop *loop = data->loop;
2788 bool reliable = true;
2790 if (TREE_CODE (base) != ARRAY_REF)
2791 return true;
2793 /* For arrays at the end of the structure, we are not guaranteed that they
2794 do not really extend over their declared size. However, for arrays of
2795 size greater than one, this is unlikely to be intended. */
2796 if (array_at_struct_end_p (base))
2798 at_end = true;
2799 upper = false;
2802 struct loop *dloop = loop_containing_stmt (data->stmt);
2803 if (!dloop)
2804 return true;
2806 ev = analyze_scalar_evolution (dloop, *idx);
2807 ev = instantiate_parameters (loop, ev);
2808 init = initial_condition (ev);
2809 step = evolution_part_in_loop_num (ev, loop->num);
2811 if (!init
2812 || !step
2813 || TREE_CODE (step) != INTEGER_CST
2814 || integer_zerop (step)
2815 || tree_contains_chrecs (init, NULL)
2816 || chrec_contains_symbols_defined_in_loop (init, loop->num))
2817 return true;
2819 low = array_ref_low_bound (base);
2820 high = array_ref_up_bound (base);
2822 /* The case of nonconstant bounds could be handled, but it would be
2823 complicated. */
2824 if (TREE_CODE (low) != INTEGER_CST
2825 || !high
2826 || TREE_CODE (high) != INTEGER_CST)
2827 return true;
2828 sign = tree_int_cst_sign_bit (step);
2829 type = TREE_TYPE (step);
2831 /* The array of length 1 at the end of a structure most likely extends
2832 beyond its bounds. */
2833 if (at_end
2834 && operand_equal_p (low, high, 0))
2835 return true;
2837 /* In case the relevant bound of the array does not fit in type, or
2838 it does, but bound + step (in type) still belongs into the range of the
2839 array, the index may wrap and still stay within the range of the array
2840 (consider e.g. if the array is indexed by the full range of
2841 unsigned char).
2843 To make things simpler, we require both bounds to fit into type, although
2844 there are cases where this would not be strictly necessary. */
2845 if (!int_fits_type_p (high, type)
2846 || !int_fits_type_p (low, type))
2847 return true;
2848 low = fold_convert (type, low);
2849 high = fold_convert (type, high);
2851 if (sign)
2852 next = fold_binary (PLUS_EXPR, type, low, step);
2853 else
2854 next = fold_binary (PLUS_EXPR, type, high, step);
2856 if (tree_int_cst_compare (low, next) <= 0
2857 && tree_int_cst_compare (next, high) <= 0)
2858 return true;
2860 /* If access is not executed on every iteration, we must ensure that overlow may
2861 not make the access valid later. */
2862 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt))
2863 && scev_probably_wraps_p (initial_condition_in_loop_num (ev, loop->num),
2864 step, data->stmt, loop, true))
2865 reliable = false;
2867 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, reliable, upper);
2868 return true;
2871 /* Determine information about number of iterations a LOOP from the bounds
2872 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
2873 STMT is guaranteed to be executed in every iteration of LOOP.*/
2875 static void
2876 infer_loop_bounds_from_ref (struct loop *loop, gimple stmt, tree ref)
2878 struct ilb_data data;
2880 data.loop = loop;
2881 data.stmt = stmt;
2882 for_each_index (&ref, idx_infer_loop_bounds, &data);
2885 /* Determine information about number of iterations of a LOOP from the way
2886 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
2887 executed in every iteration of LOOP. */
2889 static void
2890 infer_loop_bounds_from_array (struct loop *loop, gimple stmt)
2892 if (is_gimple_assign (stmt))
2894 tree op0 = gimple_assign_lhs (stmt);
2895 tree op1 = gimple_assign_rhs1 (stmt);
2897 /* For each memory access, analyze its access function
2898 and record a bound on the loop iteration domain. */
2899 if (REFERENCE_CLASS_P (op0))
2900 infer_loop_bounds_from_ref (loop, stmt, op0);
2902 if (REFERENCE_CLASS_P (op1))
2903 infer_loop_bounds_from_ref (loop, stmt, op1);
2905 else if (is_gimple_call (stmt))
2907 tree arg, lhs;
2908 unsigned i, n = gimple_call_num_args (stmt);
2910 lhs = gimple_call_lhs (stmt);
2911 if (lhs && REFERENCE_CLASS_P (lhs))
2912 infer_loop_bounds_from_ref (loop, stmt, lhs);
2914 for (i = 0; i < n; i++)
2916 arg = gimple_call_arg (stmt, i);
2917 if (REFERENCE_CLASS_P (arg))
2918 infer_loop_bounds_from_ref (loop, stmt, arg);
2923 /* Determine information about number of iterations of a LOOP from the fact
2924 that pointer arithmetics in STMT does not overflow. */
2926 static void
2927 infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple stmt)
2929 tree def, base, step, scev, type, low, high;
2930 tree var, ptr;
2932 if (!is_gimple_assign (stmt)
2933 || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR)
2934 return;
2936 def = gimple_assign_lhs (stmt);
2937 if (TREE_CODE (def) != SSA_NAME)
2938 return;
2940 type = TREE_TYPE (def);
2941 if (!nowrap_type_p (type))
2942 return;
2944 ptr = gimple_assign_rhs1 (stmt);
2945 if (!expr_invariant_in_loop_p (loop, ptr))
2946 return;
2948 var = gimple_assign_rhs2 (stmt);
2949 if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var)))
2950 return;
2952 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
2953 if (chrec_contains_undetermined (scev))
2954 return;
2956 base = initial_condition_in_loop_num (scev, loop->num);
2957 step = evolution_part_in_loop_num (scev, loop->num);
2959 if (!base || !step
2960 || TREE_CODE (step) != INTEGER_CST
2961 || tree_contains_chrecs (base, NULL)
2962 || chrec_contains_symbols_defined_in_loop (base, loop->num))
2963 return;
2965 low = lower_bound_in_type (type, type);
2966 high = upper_bound_in_type (type, type);
2968 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot
2969 produce a NULL pointer. The contrary would mean NULL points to an object,
2970 while NULL is supposed to compare unequal with the address of all objects.
2971 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a
2972 NULL pointer since that would mean wrapping, which we assume here not to
2973 happen. So, we can exclude NULL from the valid range of pointer
2974 arithmetic. */
2975 if (flag_delete_null_pointer_checks && int_cst_value (low) == 0)
2976 low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type)));
2978 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
2981 /* Determine information about number of iterations of a LOOP from the fact
2982 that signed arithmetics in STMT does not overflow. */
2984 static void
2985 infer_loop_bounds_from_signedness (struct loop *loop, gimple stmt)
2987 tree def, base, step, scev, type, low, high;
2989 if (gimple_code (stmt) != GIMPLE_ASSIGN)
2990 return;
2992 def = gimple_assign_lhs (stmt);
2994 if (TREE_CODE (def) != SSA_NAME)
2995 return;
2997 type = TREE_TYPE (def);
2998 if (!INTEGRAL_TYPE_P (type)
2999 || !TYPE_OVERFLOW_UNDEFINED (type))
3000 return;
3002 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
3003 if (chrec_contains_undetermined (scev))
3004 return;
3006 base = initial_condition_in_loop_num (scev, loop->num);
3007 step = evolution_part_in_loop_num (scev, loop->num);
3009 if (!base || !step
3010 || TREE_CODE (step) != INTEGER_CST
3011 || tree_contains_chrecs (base, NULL)
3012 || chrec_contains_symbols_defined_in_loop (base, loop->num))
3013 return;
3015 low = lower_bound_in_type (type, type);
3016 high = upper_bound_in_type (type, type);
3018 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
3021 /* The following analyzers are extracting informations on the bounds
3022 of LOOP from the following undefined behaviors:
3024 - data references should not access elements over the statically
3025 allocated size,
3027 - signed variables should not overflow when flag_wrapv is not set.
3030 static void
3031 infer_loop_bounds_from_undefined (struct loop *loop)
3033 unsigned i;
3034 basic_block *bbs;
3035 gimple_stmt_iterator bsi;
3036 basic_block bb;
3037 bool reliable;
3039 bbs = get_loop_body (loop);
3041 for (i = 0; i < loop->num_nodes; i++)
3043 bb = bbs[i];
3045 /* If BB is not executed in each iteration of the loop, we cannot
3046 use the operations in it to infer reliable upper bound on the
3047 # of iterations of the loop. However, we can use it as a guess.
3048 Reliable guesses come only from array bounds. */
3049 reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb);
3051 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
3053 gimple stmt = gsi_stmt (bsi);
3055 infer_loop_bounds_from_array (loop, stmt);
3057 if (reliable)
3059 infer_loop_bounds_from_signedness (loop, stmt);
3060 infer_loop_bounds_from_pointer_arith (loop, stmt);
3066 free (bbs);
3069 /* Compare wide ints, callback for qsort. */
3071 static int
3072 wide_int_cmp (const void *p1, const void *p2)
3074 const widest_int *d1 = (const widest_int *) p1;
3075 const widest_int *d2 = (const widest_int *) p2;
3076 return wi::cmpu (*d1, *d2);
3079 /* Return index of BOUND in BOUNDS array sorted in increasing order.
3080 Lookup by binary search. */
3082 static int
3083 bound_index (vec<widest_int> bounds, const widest_int &bound)
3085 unsigned int end = bounds.length ();
3086 unsigned int begin = 0;
3088 /* Find a matching index by means of a binary search. */
3089 while (begin != end)
3091 unsigned int middle = (begin + end) / 2;
3092 widest_int index = bounds[middle];
3094 if (index == bound)
3095 return middle;
3096 else if (wi::ltu_p (index, bound))
3097 begin = middle + 1;
3098 else
3099 end = middle;
3101 gcc_unreachable ();
3104 /* We recorded loop bounds only for statements dominating loop latch (and thus
3105 executed each loop iteration). If there are any bounds on statements not
3106 dominating the loop latch we can improve the estimate by walking the loop
3107 body and seeing if every path from loop header to loop latch contains
3108 some bounded statement. */
3110 static void
3111 discover_iteration_bound_by_body_walk (struct loop *loop)
3113 struct nb_iter_bound *elt;
3114 vec<widest_int> bounds = vNULL;
3115 vec<vec<basic_block> > queues = vNULL;
3116 vec<basic_block> queue = vNULL;
3117 ptrdiff_t queue_index;
3118 ptrdiff_t latch_index = 0;
3120 /* Discover what bounds may interest us. */
3121 for (elt = loop->bounds; elt; elt = elt->next)
3123 widest_int bound = elt->bound;
3125 /* Exit terminates loop at given iteration, while non-exits produce undefined
3126 effect on the next iteration. */
3127 if (!elt->is_exit)
3129 bound += 1;
3130 /* If an overflow occurred, ignore the result. */
3131 if (bound == 0)
3132 continue;
3135 if (!loop->any_upper_bound
3136 || wi::ltu_p (bound, loop->nb_iterations_upper_bound))
3137 bounds.safe_push (bound);
3140 /* Exit early if there is nothing to do. */
3141 if (!bounds.exists ())
3142 return;
3144 if (dump_file && (dump_flags & TDF_DETAILS))
3145 fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n");
3147 /* Sort the bounds in decreasing order. */
3148 bounds.qsort (wide_int_cmp);
3150 /* For every basic block record the lowest bound that is guaranteed to
3151 terminate the loop. */
3153 hash_map<basic_block, ptrdiff_t> bb_bounds;
3154 for (elt = loop->bounds; elt; elt = elt->next)
3156 widest_int bound = elt->bound;
3157 if (!elt->is_exit)
3159 bound += 1;
3160 /* If an overflow occurred, ignore the result. */
3161 if (bound == 0)
3162 continue;
3165 if (!loop->any_upper_bound
3166 || wi::ltu_p (bound, loop->nb_iterations_upper_bound))
3168 ptrdiff_t index = bound_index (bounds, bound);
3169 ptrdiff_t *entry = bb_bounds.get (gimple_bb (elt->stmt));
3170 if (!entry)
3171 bb_bounds.put (gimple_bb (elt->stmt), index);
3172 else if ((ptrdiff_t)*entry > index)
3173 *entry = index;
3177 hash_map<basic_block, ptrdiff_t> block_priority;
3179 /* Perform shortest path discovery loop->header ... loop->latch.
3181 The "distance" is given by the smallest loop bound of basic block
3182 present in the path and we look for path with largest smallest bound
3183 on it.
3185 To avoid the need for fibonacci heap on double ints we simply compress
3186 double ints into indexes to BOUNDS array and then represent the queue
3187 as arrays of queues for every index.
3188 Index of BOUNDS.length() means that the execution of given BB has
3189 no bounds determined.
3191 VISITED is a pointer map translating basic block into smallest index
3192 it was inserted into the priority queue with. */
3193 latch_index = -1;
3195 /* Start walk in loop header with index set to infinite bound. */
3196 queue_index = bounds.length ();
3197 queues.safe_grow_cleared (queue_index + 1);
3198 queue.safe_push (loop->header);
3199 queues[queue_index] = queue;
3200 block_priority.put (loop->header, queue_index);
3202 for (; queue_index >= 0; queue_index--)
3204 if (latch_index < queue_index)
3206 while (queues[queue_index].length ())
3208 basic_block bb;
3209 ptrdiff_t bound_index = queue_index;
3210 edge e;
3211 edge_iterator ei;
3213 queue = queues[queue_index];
3214 bb = queue.pop ();
3216 /* OK, we later inserted the BB with lower priority, skip it. */
3217 if (*block_priority.get (bb) > queue_index)
3218 continue;
3220 /* See if we can improve the bound. */
3221 ptrdiff_t *entry = bb_bounds.get (bb);
3222 if (entry && *entry < bound_index)
3223 bound_index = *entry;
3225 /* Insert succesors into the queue, watch for latch edge
3226 and record greatest index we saw. */
3227 FOR_EACH_EDGE (e, ei, bb->succs)
3229 bool insert = false;
3231 if (loop_exit_edge_p (loop, e))
3232 continue;
3234 if (e == loop_latch_edge (loop)
3235 && latch_index < bound_index)
3236 latch_index = bound_index;
3237 else if (!(entry = block_priority.get (e->dest)))
3239 insert = true;
3240 block_priority.put (e->dest, bound_index);
3242 else if (*entry < bound_index)
3244 insert = true;
3245 *entry = bound_index;
3248 if (insert)
3249 queues[bound_index].safe_push (e->dest);
3253 queues[queue_index].release ();
3256 gcc_assert (latch_index >= 0);
3257 if ((unsigned)latch_index < bounds.length ())
3259 if (dump_file && (dump_flags & TDF_DETAILS))
3261 fprintf (dump_file, "Found better loop bound ");
3262 print_decu (bounds[latch_index], dump_file);
3263 fprintf (dump_file, "\n");
3265 record_niter_bound (loop, bounds[latch_index], false, true);
3268 queues.release ();
3269 bounds.release ();
3272 /* See if every path cross the loop goes through a statement that is known
3273 to not execute at the last iteration. In that case we can decrese iteration
3274 count by 1. */
3276 static void
3277 maybe_lower_iteration_bound (struct loop *loop)
3279 hash_set<gimple> *not_executed_last_iteration = NULL;
3280 struct nb_iter_bound *elt;
3281 bool found_exit = false;
3282 vec<basic_block> queue = vNULL;
3283 bitmap visited;
3285 /* Collect all statements with interesting (i.e. lower than
3286 nb_iterations_upper_bound) bound on them.
3288 TODO: Due to the way record_estimate choose estimates to store, the bounds
3289 will be always nb_iterations_upper_bound-1. We can change this to record
3290 also statements not dominating the loop latch and update the walk bellow
3291 to the shortest path algorthm. */
3292 for (elt = loop->bounds; elt; elt = elt->next)
3294 if (!elt->is_exit
3295 && wi::ltu_p (elt->bound, loop->nb_iterations_upper_bound))
3297 if (!not_executed_last_iteration)
3298 not_executed_last_iteration = new hash_set<gimple>;
3299 not_executed_last_iteration->add (elt->stmt);
3302 if (!not_executed_last_iteration)
3303 return;
3305 /* Start DFS walk in the loop header and see if we can reach the
3306 loop latch or any of the exits (including statements with side
3307 effects that may terminate the loop otherwise) without visiting
3308 any of the statements known to have undefined effect on the last
3309 iteration. */
3310 queue.safe_push (loop->header);
3311 visited = BITMAP_ALLOC (NULL);
3312 bitmap_set_bit (visited, loop->header->index);
3313 found_exit = false;
3317 basic_block bb = queue.pop ();
3318 gimple_stmt_iterator gsi;
3319 bool stmt_found = false;
3321 /* Loop for possible exits and statements bounding the execution. */
3322 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
3324 gimple stmt = gsi_stmt (gsi);
3325 if (not_executed_last_iteration->contains (stmt))
3327 stmt_found = true;
3328 break;
3330 if (gimple_has_side_effects (stmt))
3332 found_exit = true;
3333 break;
3336 if (found_exit)
3337 break;
3339 /* If no bounding statement is found, continue the walk. */
3340 if (!stmt_found)
3342 edge e;
3343 edge_iterator ei;
3345 FOR_EACH_EDGE (e, ei, bb->succs)
3347 if (loop_exit_edge_p (loop, e)
3348 || e == loop_latch_edge (loop))
3350 found_exit = true;
3351 break;
3353 if (bitmap_set_bit (visited, e->dest->index))
3354 queue.safe_push (e->dest);
3358 while (queue.length () && !found_exit);
3360 /* If every path through the loop reach bounding statement before exit,
3361 then we know the last iteration of the loop will have undefined effect
3362 and we can decrease number of iterations. */
3364 if (!found_exit)
3366 if (dump_file && (dump_flags & TDF_DETAILS))
3367 fprintf (dump_file, "Reducing loop iteration estimate by 1; "
3368 "undefined statement must be executed at the last iteration.\n");
3369 record_niter_bound (loop, loop->nb_iterations_upper_bound - 1,
3370 false, true);
3372 BITMAP_FREE (visited);
3373 queue.release ();
3374 delete not_executed_last_iteration;
3377 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P
3378 is true also use estimates derived from undefined behavior. */
3380 static void
3381 estimate_numbers_of_iterations_loop (struct loop *loop)
3383 vec<edge> exits;
3384 tree niter, type;
3385 unsigned i;
3386 struct tree_niter_desc niter_desc;
3387 edge ex;
3388 widest_int bound;
3389 edge likely_exit;
3391 /* Give up if we already have tried to compute an estimation. */
3392 if (loop->estimate_state != EST_NOT_COMPUTED)
3393 return;
3395 loop->estimate_state = EST_AVAILABLE;
3396 /* Force estimate compuation but leave any existing upper bound in place. */
3397 loop->any_estimate = false;
3399 /* Ensure that loop->nb_iterations is computed if possible. If it turns out
3400 to be constant, we avoid undefined behavior implied bounds and instead
3401 diagnose those loops with -Waggressive-loop-optimizations. */
3402 number_of_latch_executions (loop);
3404 exits = get_loop_exit_edges (loop);
3405 likely_exit = single_likely_exit (loop);
3406 FOR_EACH_VEC_ELT (exits, i, ex)
3408 if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false))
3409 continue;
3411 niter = niter_desc.niter;
3412 type = TREE_TYPE (niter);
3413 if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST)
3414 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
3415 build_int_cst (type, 0),
3416 niter);
3417 record_estimate (loop, niter, niter_desc.max,
3418 last_stmt (ex->src),
3419 true, ex == likely_exit, true);
3421 exits.release ();
3423 if (flag_aggressive_loop_optimizations)
3424 infer_loop_bounds_from_undefined (loop);
3426 discover_iteration_bound_by_body_walk (loop);
3428 maybe_lower_iteration_bound (loop);
3430 /* If we have a measured profile, use it to estimate the number of
3431 iterations. */
3432 if (loop->header->count != 0)
3434 gcov_type nit = expected_loop_iterations_unbounded (loop) + 1;
3435 bound = gcov_type_to_wide_int (nit);
3436 record_niter_bound (loop, bound, true, false);
3439 /* If we know the exact number of iterations of this loop, try to
3440 not break code with undefined behavior by not recording smaller
3441 maximum number of iterations. */
3442 if (loop->nb_iterations
3443 && TREE_CODE (loop->nb_iterations) == INTEGER_CST)
3445 loop->any_upper_bound = true;
3446 loop->nb_iterations_upper_bound = wi::to_widest (loop->nb_iterations);
3450 /* Sets NIT to the estimated number of executions of the latch of the
3451 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
3452 large as the number of iterations. If we have no reliable estimate,
3453 the function returns false, otherwise returns true. */
3455 bool
3456 estimated_loop_iterations (struct loop *loop, widest_int *nit)
3458 /* When SCEV information is available, try to update loop iterations
3459 estimate. Otherwise just return whatever we recorded earlier. */
3460 if (scev_initialized_p ())
3461 estimate_numbers_of_iterations_loop (loop);
3463 return (get_estimated_loop_iterations (loop, nit));
3466 /* Similar to estimated_loop_iterations, but returns the estimate only
3467 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3468 on the number of iterations of LOOP could not be derived, returns -1. */
3470 HOST_WIDE_INT
3471 estimated_loop_iterations_int (struct loop *loop)
3473 widest_int nit;
3474 HOST_WIDE_INT hwi_nit;
3476 if (!estimated_loop_iterations (loop, &nit))
3477 return -1;
3479 if (!wi::fits_shwi_p (nit))
3480 return -1;
3481 hwi_nit = nit.to_shwi ();
3483 return hwi_nit < 0 ? -1 : hwi_nit;
3487 /* Sets NIT to an upper bound for the maximum number of executions of the
3488 latch of the LOOP. If we have no reliable estimate, the function returns
3489 false, otherwise returns true. */
3491 bool
3492 max_loop_iterations (struct loop *loop, widest_int *nit)
3494 /* When SCEV information is available, try to update loop iterations
3495 estimate. Otherwise just return whatever we recorded earlier. */
3496 if (scev_initialized_p ())
3497 estimate_numbers_of_iterations_loop (loop);
3499 return get_max_loop_iterations (loop, nit);
3502 /* Similar to max_loop_iterations, but returns the estimate only
3503 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3504 on the number of iterations of LOOP could not be derived, returns -1. */
3506 HOST_WIDE_INT
3507 max_loop_iterations_int (struct loop *loop)
3509 widest_int nit;
3510 HOST_WIDE_INT hwi_nit;
3512 if (!max_loop_iterations (loop, &nit))
3513 return -1;
3515 if (!wi::fits_shwi_p (nit))
3516 return -1;
3517 hwi_nit = nit.to_shwi ();
3519 return hwi_nit < 0 ? -1 : hwi_nit;
3522 /* Returns an estimate for the number of executions of statements
3523 in the LOOP. For statements before the loop exit, this exceeds
3524 the number of execution of the latch by one. */
3526 HOST_WIDE_INT
3527 estimated_stmt_executions_int (struct loop *loop)
3529 HOST_WIDE_INT nit = estimated_loop_iterations_int (loop);
3530 HOST_WIDE_INT snit;
3532 if (nit == -1)
3533 return -1;
3535 snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1);
3537 /* If the computation overflows, return -1. */
3538 return snit < 0 ? -1 : snit;
3541 /* Sets NIT to the estimated maximum number of executions of the latch of the
3542 LOOP, plus one. If we have no reliable estimate, the function returns
3543 false, otherwise returns true. */
3545 bool
3546 max_stmt_executions (struct loop *loop, widest_int *nit)
3548 widest_int nit_minus_one;
3550 if (!max_loop_iterations (loop, nit))
3551 return false;
3553 nit_minus_one = *nit;
3555 *nit += 1;
3557 return wi::gtu_p (*nit, nit_minus_one);
3560 /* Sets NIT to the estimated number of executions of the latch of the
3561 LOOP, plus one. If we have no reliable estimate, the function returns
3562 false, otherwise returns true. */
3564 bool
3565 estimated_stmt_executions (struct loop *loop, widest_int *nit)
3567 widest_int nit_minus_one;
3569 if (!estimated_loop_iterations (loop, nit))
3570 return false;
3572 nit_minus_one = *nit;
3574 *nit += 1;
3576 return wi::gtu_p (*nit, nit_minus_one);
3579 /* Records estimates on numbers of iterations of loops. */
3581 void
3582 estimate_numbers_of_iterations (void)
3584 struct loop *loop;
3586 /* We don't want to issue signed overflow warnings while getting
3587 loop iteration estimates. */
3588 fold_defer_overflow_warnings ();
3590 FOR_EACH_LOOP (loop, 0)
3592 estimate_numbers_of_iterations_loop (loop);
3595 fold_undefer_and_ignore_overflow_warnings ();
3598 /* Returns true if statement S1 dominates statement S2. */
3600 bool
3601 stmt_dominates_stmt_p (gimple s1, gimple s2)
3603 basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
3605 if (!bb1
3606 || s1 == s2)
3607 return true;
3609 if (bb1 == bb2)
3611 gimple_stmt_iterator bsi;
3613 if (gimple_code (s2) == GIMPLE_PHI)
3614 return false;
3616 if (gimple_code (s1) == GIMPLE_PHI)
3617 return true;
3619 for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi))
3620 if (gsi_stmt (bsi) == s1)
3621 return true;
3623 return false;
3626 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
3629 /* Returns true when we can prove that the number of executions of
3630 STMT in the loop is at most NITER, according to the bound on
3631 the number of executions of the statement NITER_BOUND->stmt recorded in
3632 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT.
3634 ??? This code can become quite a CPU hog - we can have many bounds,
3635 and large basic block forcing stmt_dominates_stmt_p to be queried
3636 many times on a large basic blocks, so the whole thing is O(n^2)
3637 for scev_probably_wraps_p invocation (that can be done n times).
3639 It would make more sense (and give better answers) to remember BB
3640 bounds computed by discover_iteration_bound_by_body_walk. */
3642 static bool
3643 n_of_executions_at_most (gimple stmt,
3644 struct nb_iter_bound *niter_bound,
3645 tree niter)
3647 widest_int bound = niter_bound->bound;
3648 tree nit_type = TREE_TYPE (niter), e;
3649 enum tree_code cmp;
3651 gcc_assert (TYPE_UNSIGNED (nit_type));
3653 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
3654 the number of iterations is small. */
3655 if (!wi::fits_to_tree_p (bound, nit_type))
3656 return false;
3658 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
3659 times. This means that:
3661 -- if NITER_BOUND->is_exit is true, then everything after
3662 it at most NITER_BOUND->bound times.
3664 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
3665 is executed, then NITER_BOUND->stmt is executed as well in the same
3666 iteration then STMT is executed at most NITER_BOUND->bound + 1 times.
3668 If we can determine that NITER_BOUND->stmt is always executed
3669 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times.
3670 We conclude that if both statements belong to the same
3671 basic block and STMT is before NITER_BOUND->stmt and there are no
3672 statements with side effects in between. */
3674 if (niter_bound->is_exit)
3676 if (stmt == niter_bound->stmt
3677 || !stmt_dominates_stmt_p (niter_bound->stmt, stmt))
3678 return false;
3679 cmp = GE_EXPR;
3681 else
3683 if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt))
3685 gimple_stmt_iterator bsi;
3686 if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt)
3687 || gimple_code (stmt) == GIMPLE_PHI
3688 || gimple_code (niter_bound->stmt) == GIMPLE_PHI)
3689 return false;
3691 /* By stmt_dominates_stmt_p we already know that STMT appears
3692 before NITER_BOUND->STMT. Still need to test that the loop
3693 can not be terinated by a side effect in between. */
3694 for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt;
3695 gsi_next (&bsi))
3696 if (gimple_has_side_effects (gsi_stmt (bsi)))
3697 return false;
3698 bound += 1;
3699 if (bound == 0
3700 || !wi::fits_to_tree_p (bound, nit_type))
3701 return false;
3703 cmp = GT_EXPR;
3706 e = fold_binary (cmp, boolean_type_node,
3707 niter, wide_int_to_tree (nit_type, bound));
3708 return e && integer_nonzerop (e);
3711 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
3713 bool
3714 nowrap_type_p (tree type)
3716 if (INTEGRAL_TYPE_P (type)
3717 && TYPE_OVERFLOW_UNDEFINED (type))
3718 return true;
3720 if (POINTER_TYPE_P (type))
3721 return true;
3723 return false;
3726 /* Return false only when the induction variable BASE + STEP * I is
3727 known to not overflow: i.e. when the number of iterations is small
3728 enough with respect to the step and initial condition in order to
3729 keep the evolution confined in TYPEs bounds. Return true when the
3730 iv is known to overflow or when the property is not computable.
3732 USE_OVERFLOW_SEMANTICS is true if this function should assume that
3733 the rules for overflow of the given language apply (e.g., that signed
3734 arithmetics in C does not overflow). */
3736 bool
3737 scev_probably_wraps_p (tree base, tree step,
3738 gimple at_stmt, struct loop *loop,
3739 bool use_overflow_semantics)
3741 tree delta, step_abs;
3742 tree unsigned_type, valid_niter;
3743 tree type = TREE_TYPE (step);
3744 tree e;
3745 widest_int niter;
3746 struct nb_iter_bound *bound;
3748 /* FIXME: We really need something like
3749 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
3751 We used to test for the following situation that frequently appears
3752 during address arithmetics:
3754 D.1621_13 = (long unsigned intD.4) D.1620_12;
3755 D.1622_14 = D.1621_13 * 8;
3756 D.1623_15 = (doubleD.29 *) D.1622_14;
3758 And derived that the sequence corresponding to D_14
3759 can be proved to not wrap because it is used for computing a
3760 memory access; however, this is not really the case -- for example,
3761 if D_12 = (unsigned char) [254,+,1], then D_14 has values
3762 2032, 2040, 0, 8, ..., but the code is still legal. */
3764 if (chrec_contains_undetermined (base)
3765 || chrec_contains_undetermined (step))
3766 return true;
3768 if (integer_zerop (step))
3769 return false;
3771 /* If we can use the fact that signed and pointer arithmetics does not
3772 wrap, we are done. */
3773 if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base)))
3774 return false;
3776 /* To be able to use estimates on number of iterations of the loop,
3777 we must have an upper bound on the absolute value of the step. */
3778 if (TREE_CODE (step) != INTEGER_CST)
3779 return true;
3781 /* Don't issue signed overflow warnings. */
3782 fold_defer_overflow_warnings ();
3784 /* Otherwise, compute the number of iterations before we reach the
3785 bound of the type, and verify that the loop is exited before this
3786 occurs. */
3787 unsigned_type = unsigned_type_for (type);
3788 base = fold_convert (unsigned_type, base);
3790 if (tree_int_cst_sign_bit (step))
3792 tree extreme = fold_convert (unsigned_type,
3793 lower_bound_in_type (type, type));
3794 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
3795 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type,
3796 fold_convert (unsigned_type, step));
3798 else
3800 tree extreme = fold_convert (unsigned_type,
3801 upper_bound_in_type (type, type));
3802 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
3803 step_abs = fold_convert (unsigned_type, step);
3806 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
3808 estimate_numbers_of_iterations_loop (loop);
3810 if (max_loop_iterations (loop, &niter)
3811 && wi::fits_to_tree_p (niter, TREE_TYPE (valid_niter))
3812 && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter,
3813 wide_int_to_tree (TREE_TYPE (valid_niter),
3814 niter))) != NULL
3815 && integer_nonzerop (e))
3817 fold_undefer_and_ignore_overflow_warnings ();
3818 return false;
3820 if (at_stmt)
3821 for (bound = loop->bounds; bound; bound = bound->next)
3823 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
3825 fold_undefer_and_ignore_overflow_warnings ();
3826 return false;
3830 fold_undefer_and_ignore_overflow_warnings ();
3832 /* At this point we still don't have a proof that the iv does not
3833 overflow: give up. */
3834 return true;
3837 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
3839 void
3840 free_numbers_of_iterations_estimates_loop (struct loop *loop)
3842 struct nb_iter_bound *bound, *next;
3844 loop->nb_iterations = NULL;
3845 loop->estimate_state = EST_NOT_COMPUTED;
3846 for (bound = loop->bounds; bound; bound = next)
3848 next = bound->next;
3849 ggc_free (bound);
3852 loop->bounds = NULL;
3855 /* Frees the information on upper bounds on numbers of iterations of loops. */
3857 void
3858 free_numbers_of_iterations_estimates (void)
3860 struct loop *loop;
3862 FOR_EACH_LOOP (loop, 0)
3864 free_numbers_of_iterations_estimates_loop (loop);
3868 /* Substitute value VAL for ssa name NAME inside expressions held
3869 at LOOP. */
3871 void
3872 substitute_in_loop_info (struct loop *loop, tree name, tree val)
3874 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);