Concretize gimple_cond_set_{lhs|rhs}
[official-gcc.git] / gcc / tree-ssa-loop-niter.c
blob2179462000f8806b8b831a429d4418a490992859
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_phi_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 phi = gsi.phi ();
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 last;
1926 gimple_cond stmt;
1927 tree type;
1928 tree op0, op1;
1929 enum tree_code code;
1930 affine_iv iv0, iv1;
1931 bool safe;
1933 safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src);
1935 if (every_iteration && !safe)
1936 return false;
1938 niter->assumptions = boolean_false_node;
1939 last = last_stmt (exit->src);
1940 if (!last)
1941 return false;
1942 stmt = dyn_cast <gimple_cond> (last);
1943 if (!stmt)
1944 return false;
1946 /* We want the condition for staying inside loop. */
1947 code = gimple_cond_code (stmt);
1948 if (exit->flags & EDGE_TRUE_VALUE)
1949 code = invert_tree_comparison (code, false);
1951 switch (code)
1953 case GT_EXPR:
1954 case GE_EXPR:
1955 case LT_EXPR:
1956 case LE_EXPR:
1957 case NE_EXPR:
1958 break;
1960 default:
1961 return false;
1964 op0 = gimple_cond_lhs (stmt);
1965 op1 = gimple_cond_rhs (stmt);
1966 type = TREE_TYPE (op0);
1968 if (TREE_CODE (type) != INTEGER_TYPE
1969 && !POINTER_TYPE_P (type))
1970 return false;
1972 if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, false))
1973 return false;
1974 if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, false))
1975 return false;
1977 /* We don't want to see undefined signed overflow warnings while
1978 computing the number of iterations. */
1979 fold_defer_overflow_warnings ();
1981 iv0.base = expand_simple_operations (iv0.base);
1982 iv1.base = expand_simple_operations (iv1.base);
1983 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter,
1984 loop_only_exit_p (loop, exit), safe))
1986 fold_undefer_and_ignore_overflow_warnings ();
1987 return false;
1990 if (optimize >= 3)
1992 niter->assumptions = simplify_using_outer_evolutions (loop,
1993 niter->assumptions);
1994 niter->may_be_zero = simplify_using_outer_evolutions (loop,
1995 niter->may_be_zero);
1996 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
1999 niter->assumptions
2000 = simplify_using_initial_conditions (loop,
2001 niter->assumptions);
2002 niter->may_be_zero
2003 = simplify_using_initial_conditions (loop,
2004 niter->may_be_zero);
2006 fold_undefer_and_ignore_overflow_warnings ();
2008 /* If NITER has simplified into a constant, update MAX. */
2009 if (TREE_CODE (niter->niter) == INTEGER_CST)
2010 niter->max = wi::to_widest (niter->niter);
2012 if (integer_onep (niter->assumptions))
2013 return true;
2015 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
2016 But if we can prove that there is overflow or some other source of weird
2017 behavior, ignore the loop even with -funsafe-loop-optimizations. */
2018 if (integer_zerop (niter->assumptions) || !single_exit (loop))
2019 return false;
2021 if (flag_unsafe_loop_optimizations)
2022 niter->assumptions = boolean_true_node;
2024 if (warn)
2026 const char *wording;
2027 location_t loc = gimple_location (stmt);
2029 /* We can provide a more specific warning if one of the operator is
2030 constant and the other advances by +1 or -1. */
2031 if (!integer_zerop (iv1.step)
2032 ? (integer_zerop (iv0.step)
2033 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step)))
2034 : (integer_onep (iv0.step) || integer_all_onesp (iv0.step)))
2035 wording =
2036 flag_unsafe_loop_optimizations
2037 ? N_("assuming that the loop is not infinite")
2038 : N_("cannot optimize possibly infinite loops");
2039 else
2040 wording =
2041 flag_unsafe_loop_optimizations
2042 ? N_("assuming that the loop counter does not overflow")
2043 : N_("cannot optimize loop, the loop counter may overflow");
2045 warning_at ((LOCATION_LINE (loc) > 0) ? loc : input_location,
2046 OPT_Wunsafe_loop_optimizations, "%s", gettext (wording));
2049 return flag_unsafe_loop_optimizations;
2052 /* Try to determine the number of iterations of LOOP. If we succeed,
2053 expression giving number of iterations is returned and *EXIT is
2054 set to the edge from that the information is obtained. Otherwise
2055 chrec_dont_know is returned. */
2057 tree
2058 find_loop_niter (struct loop *loop, edge *exit)
2060 unsigned i;
2061 vec<edge> exits = get_loop_exit_edges (loop);
2062 edge ex;
2063 tree niter = NULL_TREE, aniter;
2064 struct tree_niter_desc desc;
2066 *exit = NULL;
2067 FOR_EACH_VEC_ELT (exits, i, ex)
2069 if (!number_of_iterations_exit (loop, ex, &desc, false))
2070 continue;
2072 if (integer_nonzerop (desc.may_be_zero))
2074 /* We exit in the first iteration through this exit.
2075 We won't find anything better. */
2076 niter = build_int_cst (unsigned_type_node, 0);
2077 *exit = ex;
2078 break;
2081 if (!integer_zerop (desc.may_be_zero))
2082 continue;
2084 aniter = desc.niter;
2086 if (!niter)
2088 /* Nothing recorded yet. */
2089 niter = aniter;
2090 *exit = ex;
2091 continue;
2094 /* Prefer constants, the lower the better. */
2095 if (TREE_CODE (aniter) != INTEGER_CST)
2096 continue;
2098 if (TREE_CODE (niter) != INTEGER_CST)
2100 niter = aniter;
2101 *exit = ex;
2102 continue;
2105 if (tree_int_cst_lt (aniter, niter))
2107 niter = aniter;
2108 *exit = ex;
2109 continue;
2112 exits.release ();
2114 return niter ? niter : chrec_dont_know;
2117 /* Return true if loop is known to have bounded number of iterations. */
2119 bool
2120 finite_loop_p (struct loop *loop)
2122 widest_int nit;
2123 int flags;
2125 if (flag_unsafe_loop_optimizations)
2126 return true;
2127 flags = flags_from_decl_or_type (current_function_decl);
2128 if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE))
2130 if (dump_file && (dump_flags & TDF_DETAILS))
2131 fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n",
2132 loop->num);
2133 return true;
2136 if (loop->any_upper_bound
2137 || max_loop_iterations (loop, &nit))
2139 if (dump_file && (dump_flags & TDF_DETAILS))
2140 fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n",
2141 loop->num);
2142 return true;
2144 return false;
2149 Analysis of a number of iterations of a loop by a brute-force evaluation.
2153 /* Bound on the number of iterations we try to evaluate. */
2155 #define MAX_ITERATIONS_TO_TRACK \
2156 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
2158 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
2159 result by a chain of operations such that all but exactly one of their
2160 operands are constants. */
2162 static gimple_phi
2163 chain_of_csts_start (struct loop *loop, tree x)
2165 gimple stmt = SSA_NAME_DEF_STMT (x);
2166 tree use;
2167 basic_block bb = gimple_bb (stmt);
2168 enum tree_code code;
2170 if (!bb
2171 || !flow_bb_inside_loop_p (loop, bb))
2172 return NULL;
2174 if (gimple_code (stmt) == GIMPLE_PHI)
2176 if (bb == loop->header)
2177 return as_a <gimple_phi> (stmt);
2179 return NULL;
2182 if (gimple_code (stmt) != GIMPLE_ASSIGN
2183 || gimple_assign_rhs_class (stmt) == GIMPLE_TERNARY_RHS)
2184 return NULL;
2186 code = gimple_assign_rhs_code (stmt);
2187 if (gimple_references_memory_p (stmt)
2188 || TREE_CODE_CLASS (code) == tcc_reference
2189 || (code == ADDR_EXPR
2190 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
2191 return NULL;
2193 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
2194 if (use == NULL_TREE)
2195 return NULL;
2197 return chain_of_csts_start (loop, use);
2200 /* Determines whether the expression X is derived from a result of a phi node
2201 in header of LOOP such that
2203 * the derivation of X consists only from operations with constants
2204 * the initial value of the phi node is constant
2205 * the value of the phi node in the next iteration can be derived from the
2206 value in the current iteration by a chain of operations with constants.
2208 If such phi node exists, it is returned, otherwise NULL is returned. */
2210 static gimple_phi
2211 get_base_for (struct loop *loop, tree x)
2213 gimple_phi phi;
2214 tree init, next;
2216 if (is_gimple_min_invariant (x))
2217 return NULL;
2219 phi = chain_of_csts_start (loop, x);
2220 if (!phi)
2221 return NULL;
2223 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2224 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2226 if (TREE_CODE (next) != SSA_NAME)
2227 return NULL;
2229 if (!is_gimple_min_invariant (init))
2230 return NULL;
2232 if (chain_of_csts_start (loop, next) != phi)
2233 return NULL;
2235 return phi;
2238 /* Given an expression X, then
2240 * if X is NULL_TREE, we return the constant BASE.
2241 * otherwise X is a SSA name, whose value in the considered loop is derived
2242 by a chain of operations with constant from a result of a phi node in
2243 the header of the loop. Then we return value of X when the value of the
2244 result of this phi node is given by the constant BASE. */
2246 static tree
2247 get_val_for (tree x, tree base)
2249 gimple stmt;
2251 gcc_checking_assert (is_gimple_min_invariant (base));
2253 if (!x)
2254 return base;
2256 stmt = SSA_NAME_DEF_STMT (x);
2257 if (gimple_code (stmt) == GIMPLE_PHI)
2258 return base;
2260 gcc_checking_assert (is_gimple_assign (stmt));
2262 /* STMT must be either an assignment of a single SSA name or an
2263 expression involving an SSA name and a constant. Try to fold that
2264 expression using the value for the SSA name. */
2265 if (gimple_assign_ssa_name_copy_p (stmt))
2266 return get_val_for (gimple_assign_rhs1 (stmt), base);
2267 else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS
2268 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
2270 return fold_build1 (gimple_assign_rhs_code (stmt),
2271 gimple_expr_type (stmt),
2272 get_val_for (gimple_assign_rhs1 (stmt), base));
2274 else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS)
2276 tree rhs1 = gimple_assign_rhs1 (stmt);
2277 tree rhs2 = gimple_assign_rhs2 (stmt);
2278 if (TREE_CODE (rhs1) == SSA_NAME)
2279 rhs1 = get_val_for (rhs1, base);
2280 else if (TREE_CODE (rhs2) == SSA_NAME)
2281 rhs2 = get_val_for (rhs2, base);
2282 else
2283 gcc_unreachable ();
2284 return fold_build2 (gimple_assign_rhs_code (stmt),
2285 gimple_expr_type (stmt), rhs1, rhs2);
2287 else
2288 gcc_unreachable ();
2292 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2293 by brute force -- i.e. by determining the value of the operands of the
2294 condition at EXIT in first few iterations of the loop (assuming that
2295 these values are constant) and determining the first one in that the
2296 condition is not satisfied. Returns the constant giving the number
2297 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2299 tree
2300 loop_niter_by_eval (struct loop *loop, edge exit)
2302 tree acnd;
2303 tree op[2], val[2], next[2], aval[2];
2304 gimple_phi phi;
2305 gimple cond;
2306 unsigned i, j;
2307 enum tree_code cmp;
2309 cond = last_stmt (exit->src);
2310 if (!cond || gimple_code (cond) != GIMPLE_COND)
2311 return chrec_dont_know;
2313 cmp = gimple_cond_code (cond);
2314 if (exit->flags & EDGE_TRUE_VALUE)
2315 cmp = invert_tree_comparison (cmp, false);
2317 switch (cmp)
2319 case EQ_EXPR:
2320 case NE_EXPR:
2321 case GT_EXPR:
2322 case GE_EXPR:
2323 case LT_EXPR:
2324 case LE_EXPR:
2325 op[0] = gimple_cond_lhs (cond);
2326 op[1] = gimple_cond_rhs (cond);
2327 break;
2329 default:
2330 return chrec_dont_know;
2333 for (j = 0; j < 2; j++)
2335 if (is_gimple_min_invariant (op[j]))
2337 val[j] = op[j];
2338 next[j] = NULL_TREE;
2339 op[j] = NULL_TREE;
2341 else
2343 phi = get_base_for (loop, op[j]);
2344 if (!phi)
2345 return chrec_dont_know;
2346 val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2347 next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2351 /* Don't issue signed overflow warnings. */
2352 fold_defer_overflow_warnings ();
2354 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
2356 for (j = 0; j < 2; j++)
2357 aval[j] = get_val_for (op[j], val[j]);
2359 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
2360 if (acnd && integer_zerop (acnd))
2362 fold_undefer_and_ignore_overflow_warnings ();
2363 if (dump_file && (dump_flags & TDF_DETAILS))
2364 fprintf (dump_file,
2365 "Proved that loop %d iterates %d times using brute force.\n",
2366 loop->num, i);
2367 return build_int_cst (unsigned_type_node, i);
2370 for (j = 0; j < 2; j++)
2372 val[j] = get_val_for (next[j], val[j]);
2373 if (!is_gimple_min_invariant (val[j]))
2375 fold_undefer_and_ignore_overflow_warnings ();
2376 return chrec_dont_know;
2381 fold_undefer_and_ignore_overflow_warnings ();
2383 return chrec_dont_know;
2386 /* Finds the exit of the LOOP by that the loop exits after a constant
2387 number of iterations and stores the exit edge to *EXIT. The constant
2388 giving the number of iterations of LOOP is returned. The number of
2389 iterations is determined using loop_niter_by_eval (i.e. by brute force
2390 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2391 determines the number of iterations, chrec_dont_know is returned. */
2393 tree
2394 find_loop_niter_by_eval (struct loop *loop, edge *exit)
2396 unsigned i;
2397 vec<edge> exits = get_loop_exit_edges (loop);
2398 edge ex;
2399 tree niter = NULL_TREE, aniter;
2401 *exit = NULL;
2403 /* Loops with multiple exits are expensive to handle and less important. */
2404 if (!flag_expensive_optimizations
2405 && exits.length () > 1)
2407 exits.release ();
2408 return chrec_dont_know;
2411 FOR_EACH_VEC_ELT (exits, i, ex)
2413 if (!just_once_each_iteration_p (loop, ex->src))
2414 continue;
2416 aniter = loop_niter_by_eval (loop, ex);
2417 if (chrec_contains_undetermined (aniter))
2418 continue;
2420 if (niter
2421 && !tree_int_cst_lt (aniter, niter))
2422 continue;
2424 niter = aniter;
2425 *exit = ex;
2427 exits.release ();
2429 return niter ? niter : chrec_dont_know;
2434 Analysis of upper bounds on number of iterations of a loop.
2438 static widest_int derive_constant_upper_bound_ops (tree, tree,
2439 enum tree_code, tree);
2441 /* Returns a constant upper bound on the value of the right-hand side of
2442 an assignment statement STMT. */
2444 static widest_int
2445 derive_constant_upper_bound_assign (gimple stmt)
2447 enum tree_code code = gimple_assign_rhs_code (stmt);
2448 tree op0 = gimple_assign_rhs1 (stmt);
2449 tree op1 = gimple_assign_rhs2 (stmt);
2451 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)),
2452 op0, code, op1);
2455 /* Returns a constant upper bound on the value of expression VAL. VAL
2456 is considered to be unsigned. If its type is signed, its value must
2457 be nonnegative. */
2459 static widest_int
2460 derive_constant_upper_bound (tree val)
2462 enum tree_code code;
2463 tree op0, op1;
2465 extract_ops_from_tree (val, &code, &op0, &op1);
2466 return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1);
2469 /* Returns a constant upper bound on the value of expression OP0 CODE OP1,
2470 whose type is TYPE. The expression is considered to be unsigned. If
2471 its type is signed, its value must be nonnegative. */
2473 static widest_int
2474 derive_constant_upper_bound_ops (tree type, tree op0,
2475 enum tree_code code, tree op1)
2477 tree subtype, maxt;
2478 widest_int bnd, max, mmax, cst;
2479 gimple stmt;
2481 if (INTEGRAL_TYPE_P (type))
2482 maxt = TYPE_MAX_VALUE (type);
2483 else
2484 maxt = upper_bound_in_type (type, type);
2486 max = wi::to_widest (maxt);
2488 switch (code)
2490 case INTEGER_CST:
2491 return wi::to_widest (op0);
2493 CASE_CONVERT:
2494 subtype = TREE_TYPE (op0);
2495 if (!TYPE_UNSIGNED (subtype)
2496 /* If TYPE is also signed, the fact that VAL is nonnegative implies
2497 that OP0 is nonnegative. */
2498 && TYPE_UNSIGNED (type)
2499 && !tree_expr_nonnegative_p (op0))
2501 /* If we cannot prove that the casted expression is nonnegative,
2502 we cannot establish more useful upper bound than the precision
2503 of the type gives us. */
2504 return max;
2507 /* We now know that op0 is an nonnegative value. Try deriving an upper
2508 bound for it. */
2509 bnd = derive_constant_upper_bound (op0);
2511 /* If the bound does not fit in TYPE, max. value of TYPE could be
2512 attained. */
2513 if (wi::ltu_p (max, bnd))
2514 return max;
2516 return bnd;
2518 case PLUS_EXPR:
2519 case POINTER_PLUS_EXPR:
2520 case MINUS_EXPR:
2521 if (TREE_CODE (op1) != INTEGER_CST
2522 || !tree_expr_nonnegative_p (op0))
2523 return max;
2525 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
2526 choose the most logical way how to treat this constant regardless
2527 of the signedness of the type. */
2528 cst = wi::sext (wi::to_widest (op1), TYPE_PRECISION (type));
2529 if (code != MINUS_EXPR)
2530 cst = -cst;
2532 bnd = derive_constant_upper_bound (op0);
2534 if (wi::neg_p (cst))
2536 cst = -cst;
2537 /* Avoid CST == 0x80000... */
2538 if (wi::neg_p (cst))
2539 return max;;
2541 /* OP0 + CST. We need to check that
2542 BND <= MAX (type) - CST. */
2544 mmax -= cst;
2545 if (wi::ltu_p (bnd, max))
2546 return max;
2548 return bnd + cst;
2550 else
2552 /* OP0 - CST, where CST >= 0.
2554 If TYPE is signed, we have already verified that OP0 >= 0, and we
2555 know that the result is nonnegative. This implies that
2556 VAL <= BND - CST.
2558 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2559 otherwise the operation underflows.
2562 /* This should only happen if the type is unsigned; however, for
2563 buggy programs that use overflowing signed arithmetics even with
2564 -fno-wrapv, this condition may also be true for signed values. */
2565 if (wi::ltu_p (bnd, cst))
2566 return max;
2568 if (TYPE_UNSIGNED (type))
2570 tree tem = fold_binary (GE_EXPR, boolean_type_node, op0,
2571 wide_int_to_tree (type, cst));
2572 if (!tem || integer_nonzerop (tem))
2573 return max;
2576 bnd -= cst;
2579 return bnd;
2581 case FLOOR_DIV_EXPR:
2582 case EXACT_DIV_EXPR:
2583 if (TREE_CODE (op1) != INTEGER_CST
2584 || tree_int_cst_sign_bit (op1))
2585 return max;
2587 bnd = derive_constant_upper_bound (op0);
2588 return wi::udiv_floor (bnd, wi::to_widest (op1));
2590 case BIT_AND_EXPR:
2591 if (TREE_CODE (op1) != INTEGER_CST
2592 || tree_int_cst_sign_bit (op1))
2593 return max;
2594 return wi::to_widest (op1);
2596 case SSA_NAME:
2597 stmt = SSA_NAME_DEF_STMT (op0);
2598 if (gimple_code (stmt) != GIMPLE_ASSIGN
2599 || gimple_assign_lhs (stmt) != op0)
2600 return max;
2601 return derive_constant_upper_bound_assign (stmt);
2603 default:
2604 return max;
2608 /* Emit a -Waggressive-loop-optimizations warning if needed. */
2610 static void
2611 do_warn_aggressive_loop_optimizations (struct loop *loop,
2612 widest_int i_bound, gimple stmt)
2614 /* Don't warn if the loop doesn't have known constant bound. */
2615 if (!loop->nb_iterations
2616 || TREE_CODE (loop->nb_iterations) != INTEGER_CST
2617 || !warn_aggressive_loop_optimizations
2618 /* To avoid warning multiple times for the same loop,
2619 only start warning when we preserve loops. */
2620 || (cfun->curr_properties & PROP_loops) == 0
2621 /* Only warn once per loop. */
2622 || loop->warned_aggressive_loop_optimizations
2623 /* Only warn if undefined behavior gives us lower estimate than the
2624 known constant bound. */
2625 || wi::cmpu (i_bound, wi::to_widest (loop->nb_iterations)) >= 0
2626 /* And undefined behavior happens unconditionally. */
2627 || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt)))
2628 return;
2630 edge e = single_exit (loop);
2631 if (e == NULL)
2632 return;
2634 gimple estmt = last_stmt (e->src);
2635 if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations,
2636 "iteration %E invokes undefined behavior",
2637 wide_int_to_tree (TREE_TYPE (loop->nb_iterations),
2638 i_bound)))
2639 inform (gimple_location (estmt), "containing loop");
2640 loop->warned_aggressive_loop_optimizations = true;
2643 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
2644 is true if the loop is exited immediately after STMT, and this exit
2645 is taken at last when the STMT is executed BOUND + 1 times.
2646 REALISTIC is true if BOUND is expected to be close to the real number
2647 of iterations. UPPER is true if we are sure the loop iterates at most
2648 BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */
2650 static void
2651 record_estimate (struct loop *loop, tree bound, const widest_int &i_bound,
2652 gimple at_stmt, bool is_exit, bool realistic, bool upper)
2654 widest_int delta;
2656 if (dump_file && (dump_flags & TDF_DETAILS))
2658 fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : "");
2659 print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM);
2660 fprintf (dump_file, " is %sexecuted at most ",
2661 upper ? "" : "probably ");
2662 print_generic_expr (dump_file, bound, TDF_SLIM);
2663 fprintf (dump_file, " (bounded by ");
2664 print_decu (i_bound, dump_file);
2665 fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num);
2668 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
2669 real number of iterations. */
2670 if (TREE_CODE (bound) != INTEGER_CST)
2671 realistic = false;
2672 else
2673 gcc_checking_assert (i_bound == wi::to_widest (bound));
2674 if (!upper && !realistic)
2675 return;
2677 /* If we have a guaranteed upper bound, record it in the appropriate
2678 list, unless this is an !is_exit bound (i.e. undefined behavior in
2679 at_stmt) in a loop with known constant number of iterations. */
2680 if (upper
2681 && (is_exit
2682 || loop->nb_iterations == NULL_TREE
2683 || TREE_CODE (loop->nb_iterations) != INTEGER_CST))
2685 struct nb_iter_bound *elt = ggc_alloc<nb_iter_bound> ();
2687 elt->bound = i_bound;
2688 elt->stmt = at_stmt;
2689 elt->is_exit = is_exit;
2690 elt->next = loop->bounds;
2691 loop->bounds = elt;
2694 /* If statement is executed on every path to the loop latch, we can directly
2695 infer the upper bound on the # of iterations of the loop. */
2696 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt)))
2697 return;
2699 /* Update the number of iteration estimates according to the bound.
2700 If at_stmt is an exit then the loop latch is executed at most BOUND times,
2701 otherwise it can be executed BOUND + 1 times. We will lower the estimate
2702 later if such statement must be executed on last iteration */
2703 if (is_exit)
2704 delta = 0;
2705 else
2706 delta = 1;
2707 widest_int new_i_bound = i_bound + delta;
2709 /* If an overflow occurred, ignore the result. */
2710 if (wi::ltu_p (new_i_bound, delta))
2711 return;
2713 if (upper && !is_exit)
2714 do_warn_aggressive_loop_optimizations (loop, new_i_bound, at_stmt);
2715 record_niter_bound (loop, new_i_bound, realistic, upper);
2718 /* Record the estimate on number of iterations of LOOP based on the fact that
2719 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
2720 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
2721 estimated number of iterations is expected to be close to the real one.
2722 UPPER is true if we are sure the induction variable does not wrap. */
2724 static void
2725 record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple stmt,
2726 tree low, tree high, bool realistic, bool upper)
2728 tree niter_bound, extreme, delta;
2729 tree type = TREE_TYPE (base), unsigned_type;
2731 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step))
2732 return;
2734 if (dump_file && (dump_flags & TDF_DETAILS))
2736 fprintf (dump_file, "Induction variable (");
2737 print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM);
2738 fprintf (dump_file, ") ");
2739 print_generic_expr (dump_file, base, TDF_SLIM);
2740 fprintf (dump_file, " + ");
2741 print_generic_expr (dump_file, step, TDF_SLIM);
2742 fprintf (dump_file, " * iteration does not wrap in statement ");
2743 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2744 fprintf (dump_file, " in loop %d.\n", loop->num);
2747 unsigned_type = unsigned_type_for (type);
2748 base = fold_convert (unsigned_type, base);
2749 step = fold_convert (unsigned_type, step);
2751 if (tree_int_cst_sign_bit (step))
2753 extreme = fold_convert (unsigned_type, low);
2754 if (TREE_CODE (base) != INTEGER_CST)
2755 base = fold_convert (unsigned_type, high);
2756 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
2757 step = fold_build1 (NEGATE_EXPR, unsigned_type, step);
2759 else
2761 extreme = fold_convert (unsigned_type, high);
2762 if (TREE_CODE (base) != INTEGER_CST)
2763 base = fold_convert (unsigned_type, low);
2764 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
2767 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
2768 would get out of the range. */
2769 niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step);
2770 widest_int max = derive_constant_upper_bound (niter_bound);
2771 record_estimate (loop, niter_bound, max, stmt, false, realistic, upper);
2774 /* Determine information about number of iterations a LOOP from the index
2775 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
2776 guaranteed to be executed in every iteration of LOOP. Callback for
2777 for_each_index. */
2779 struct ilb_data
2781 struct loop *loop;
2782 gimple stmt;
2785 static bool
2786 idx_infer_loop_bounds (tree base, tree *idx, void *dta)
2788 struct ilb_data *data = (struct ilb_data *) dta;
2789 tree ev, init, step;
2790 tree low, high, type, next;
2791 bool sign, upper = true, at_end = false;
2792 struct loop *loop = data->loop;
2793 bool reliable = true;
2795 if (TREE_CODE (base) != ARRAY_REF)
2796 return true;
2798 /* For arrays at the end of the structure, we are not guaranteed that they
2799 do not really extend over their declared size. However, for arrays of
2800 size greater than one, this is unlikely to be intended. */
2801 if (array_at_struct_end_p (base))
2803 at_end = true;
2804 upper = false;
2807 struct loop *dloop = loop_containing_stmt (data->stmt);
2808 if (!dloop)
2809 return true;
2811 ev = analyze_scalar_evolution (dloop, *idx);
2812 ev = instantiate_parameters (loop, ev);
2813 init = initial_condition (ev);
2814 step = evolution_part_in_loop_num (ev, loop->num);
2816 if (!init
2817 || !step
2818 || TREE_CODE (step) != INTEGER_CST
2819 || integer_zerop (step)
2820 || tree_contains_chrecs (init, NULL)
2821 || chrec_contains_symbols_defined_in_loop (init, loop->num))
2822 return true;
2824 low = array_ref_low_bound (base);
2825 high = array_ref_up_bound (base);
2827 /* The case of nonconstant bounds could be handled, but it would be
2828 complicated. */
2829 if (TREE_CODE (low) != INTEGER_CST
2830 || !high
2831 || TREE_CODE (high) != INTEGER_CST)
2832 return true;
2833 sign = tree_int_cst_sign_bit (step);
2834 type = TREE_TYPE (step);
2836 /* The array of length 1 at the end of a structure most likely extends
2837 beyond its bounds. */
2838 if (at_end
2839 && operand_equal_p (low, high, 0))
2840 return true;
2842 /* In case the relevant bound of the array does not fit in type, or
2843 it does, but bound + step (in type) still belongs into the range of the
2844 array, the index may wrap and still stay within the range of the array
2845 (consider e.g. if the array is indexed by the full range of
2846 unsigned char).
2848 To make things simpler, we require both bounds to fit into type, although
2849 there are cases where this would not be strictly necessary. */
2850 if (!int_fits_type_p (high, type)
2851 || !int_fits_type_p (low, type))
2852 return true;
2853 low = fold_convert (type, low);
2854 high = fold_convert (type, high);
2856 if (sign)
2857 next = fold_binary (PLUS_EXPR, type, low, step);
2858 else
2859 next = fold_binary (PLUS_EXPR, type, high, step);
2861 if (tree_int_cst_compare (low, next) <= 0
2862 && tree_int_cst_compare (next, high) <= 0)
2863 return true;
2865 /* If access is not executed on every iteration, we must ensure that overlow may
2866 not make the access valid later. */
2867 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt))
2868 && scev_probably_wraps_p (initial_condition_in_loop_num (ev, loop->num),
2869 step, data->stmt, loop, true))
2870 reliable = false;
2872 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, reliable, upper);
2873 return true;
2876 /* Determine information about number of iterations a LOOP from the bounds
2877 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
2878 STMT is guaranteed to be executed in every iteration of LOOP.*/
2880 static void
2881 infer_loop_bounds_from_ref (struct loop *loop, gimple stmt, tree ref)
2883 struct ilb_data data;
2885 data.loop = loop;
2886 data.stmt = stmt;
2887 for_each_index (&ref, idx_infer_loop_bounds, &data);
2890 /* Determine information about number of iterations of a LOOP from the way
2891 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
2892 executed in every iteration of LOOP. */
2894 static void
2895 infer_loop_bounds_from_array (struct loop *loop, gimple stmt)
2897 if (is_gimple_assign (stmt))
2899 tree op0 = gimple_assign_lhs (stmt);
2900 tree op1 = gimple_assign_rhs1 (stmt);
2902 /* For each memory access, analyze its access function
2903 and record a bound on the loop iteration domain. */
2904 if (REFERENCE_CLASS_P (op0))
2905 infer_loop_bounds_from_ref (loop, stmt, op0);
2907 if (REFERENCE_CLASS_P (op1))
2908 infer_loop_bounds_from_ref (loop, stmt, op1);
2910 else if (is_gimple_call (stmt))
2912 tree arg, lhs;
2913 unsigned i, n = gimple_call_num_args (stmt);
2915 lhs = gimple_call_lhs (stmt);
2916 if (lhs && REFERENCE_CLASS_P (lhs))
2917 infer_loop_bounds_from_ref (loop, stmt, lhs);
2919 for (i = 0; i < n; i++)
2921 arg = gimple_call_arg (stmt, i);
2922 if (REFERENCE_CLASS_P (arg))
2923 infer_loop_bounds_from_ref (loop, stmt, arg);
2928 /* Determine information about number of iterations of a LOOP from the fact
2929 that pointer arithmetics in STMT does not overflow. */
2931 static void
2932 infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple stmt)
2934 tree def, base, step, scev, type, low, high;
2935 tree var, ptr;
2937 if (!is_gimple_assign (stmt)
2938 || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR)
2939 return;
2941 def = gimple_assign_lhs (stmt);
2942 if (TREE_CODE (def) != SSA_NAME)
2943 return;
2945 type = TREE_TYPE (def);
2946 if (!nowrap_type_p (type))
2947 return;
2949 ptr = gimple_assign_rhs1 (stmt);
2950 if (!expr_invariant_in_loop_p (loop, ptr))
2951 return;
2953 var = gimple_assign_rhs2 (stmt);
2954 if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var)))
2955 return;
2957 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
2958 if (chrec_contains_undetermined (scev))
2959 return;
2961 base = initial_condition_in_loop_num (scev, loop->num);
2962 step = evolution_part_in_loop_num (scev, loop->num);
2964 if (!base || !step
2965 || TREE_CODE (step) != INTEGER_CST
2966 || tree_contains_chrecs (base, NULL)
2967 || chrec_contains_symbols_defined_in_loop (base, loop->num))
2968 return;
2970 low = lower_bound_in_type (type, type);
2971 high = upper_bound_in_type (type, type);
2973 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot
2974 produce a NULL pointer. The contrary would mean NULL points to an object,
2975 while NULL is supposed to compare unequal with the address of all objects.
2976 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a
2977 NULL pointer since that would mean wrapping, which we assume here not to
2978 happen. So, we can exclude NULL from the valid range of pointer
2979 arithmetic. */
2980 if (flag_delete_null_pointer_checks && int_cst_value (low) == 0)
2981 low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type)));
2983 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
2986 /* Determine information about number of iterations of a LOOP from the fact
2987 that signed arithmetics in STMT does not overflow. */
2989 static void
2990 infer_loop_bounds_from_signedness (struct loop *loop, gimple stmt)
2992 tree def, base, step, scev, type, low, high;
2994 if (gimple_code (stmt) != GIMPLE_ASSIGN)
2995 return;
2997 def = gimple_assign_lhs (stmt);
2999 if (TREE_CODE (def) != SSA_NAME)
3000 return;
3002 type = TREE_TYPE (def);
3003 if (!INTEGRAL_TYPE_P (type)
3004 || !TYPE_OVERFLOW_UNDEFINED (type))
3005 return;
3007 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
3008 if (chrec_contains_undetermined (scev))
3009 return;
3011 base = initial_condition_in_loop_num (scev, loop->num);
3012 step = evolution_part_in_loop_num (scev, loop->num);
3014 if (!base || !step
3015 || TREE_CODE (step) != INTEGER_CST
3016 || tree_contains_chrecs (base, NULL)
3017 || chrec_contains_symbols_defined_in_loop (base, loop->num))
3018 return;
3020 low = lower_bound_in_type (type, type);
3021 high = upper_bound_in_type (type, type);
3023 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
3026 /* The following analyzers are extracting informations on the bounds
3027 of LOOP from the following undefined behaviors:
3029 - data references should not access elements over the statically
3030 allocated size,
3032 - signed variables should not overflow when flag_wrapv is not set.
3035 static void
3036 infer_loop_bounds_from_undefined (struct loop *loop)
3038 unsigned i;
3039 basic_block *bbs;
3040 gimple_stmt_iterator bsi;
3041 basic_block bb;
3042 bool reliable;
3044 bbs = get_loop_body (loop);
3046 for (i = 0; i < loop->num_nodes; i++)
3048 bb = bbs[i];
3050 /* If BB is not executed in each iteration of the loop, we cannot
3051 use the operations in it to infer reliable upper bound on the
3052 # of iterations of the loop. However, we can use it as a guess.
3053 Reliable guesses come only from array bounds. */
3054 reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb);
3056 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
3058 gimple stmt = gsi_stmt (bsi);
3060 infer_loop_bounds_from_array (loop, stmt);
3062 if (reliable)
3064 infer_loop_bounds_from_signedness (loop, stmt);
3065 infer_loop_bounds_from_pointer_arith (loop, stmt);
3071 free (bbs);
3074 /* Compare wide ints, callback for qsort. */
3076 static int
3077 wide_int_cmp (const void *p1, const void *p2)
3079 const widest_int *d1 = (const widest_int *) p1;
3080 const widest_int *d2 = (const widest_int *) p2;
3081 return wi::cmpu (*d1, *d2);
3084 /* Return index of BOUND in BOUNDS array sorted in increasing order.
3085 Lookup by binary search. */
3087 static int
3088 bound_index (vec<widest_int> bounds, const widest_int &bound)
3090 unsigned int end = bounds.length ();
3091 unsigned int begin = 0;
3093 /* Find a matching index by means of a binary search. */
3094 while (begin != end)
3096 unsigned int middle = (begin + end) / 2;
3097 widest_int index = bounds[middle];
3099 if (index == bound)
3100 return middle;
3101 else if (wi::ltu_p (index, bound))
3102 begin = middle + 1;
3103 else
3104 end = middle;
3106 gcc_unreachable ();
3109 /* We recorded loop bounds only for statements dominating loop latch (and thus
3110 executed each loop iteration). If there are any bounds on statements not
3111 dominating the loop latch we can improve the estimate by walking the loop
3112 body and seeing if every path from loop header to loop latch contains
3113 some bounded statement. */
3115 static void
3116 discover_iteration_bound_by_body_walk (struct loop *loop)
3118 struct nb_iter_bound *elt;
3119 vec<widest_int> bounds = vNULL;
3120 vec<vec<basic_block> > queues = vNULL;
3121 vec<basic_block> queue = vNULL;
3122 ptrdiff_t queue_index;
3123 ptrdiff_t latch_index = 0;
3125 /* Discover what bounds may interest us. */
3126 for (elt = loop->bounds; elt; elt = elt->next)
3128 widest_int bound = elt->bound;
3130 /* Exit terminates loop at given iteration, while non-exits produce undefined
3131 effect on the next iteration. */
3132 if (!elt->is_exit)
3134 bound += 1;
3135 /* If an overflow occurred, ignore the result. */
3136 if (bound == 0)
3137 continue;
3140 if (!loop->any_upper_bound
3141 || wi::ltu_p (bound, loop->nb_iterations_upper_bound))
3142 bounds.safe_push (bound);
3145 /* Exit early if there is nothing to do. */
3146 if (!bounds.exists ())
3147 return;
3149 if (dump_file && (dump_flags & TDF_DETAILS))
3150 fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n");
3152 /* Sort the bounds in decreasing order. */
3153 bounds.qsort (wide_int_cmp);
3155 /* For every basic block record the lowest bound that is guaranteed to
3156 terminate the loop. */
3158 hash_map<basic_block, ptrdiff_t> bb_bounds;
3159 for (elt = loop->bounds; elt; elt = elt->next)
3161 widest_int bound = elt->bound;
3162 if (!elt->is_exit)
3164 bound += 1;
3165 /* If an overflow occurred, ignore the result. */
3166 if (bound == 0)
3167 continue;
3170 if (!loop->any_upper_bound
3171 || wi::ltu_p (bound, loop->nb_iterations_upper_bound))
3173 ptrdiff_t index = bound_index (bounds, bound);
3174 ptrdiff_t *entry = bb_bounds.get (gimple_bb (elt->stmt));
3175 if (!entry)
3176 bb_bounds.put (gimple_bb (elt->stmt), index);
3177 else if ((ptrdiff_t)*entry > index)
3178 *entry = index;
3182 hash_map<basic_block, ptrdiff_t> block_priority;
3184 /* Perform shortest path discovery loop->header ... loop->latch.
3186 The "distance" is given by the smallest loop bound of basic block
3187 present in the path and we look for path with largest smallest bound
3188 on it.
3190 To avoid the need for fibonacci heap on double ints we simply compress
3191 double ints into indexes to BOUNDS array and then represent the queue
3192 as arrays of queues for every index.
3193 Index of BOUNDS.length() means that the execution of given BB has
3194 no bounds determined.
3196 VISITED is a pointer map translating basic block into smallest index
3197 it was inserted into the priority queue with. */
3198 latch_index = -1;
3200 /* Start walk in loop header with index set to infinite bound. */
3201 queue_index = bounds.length ();
3202 queues.safe_grow_cleared (queue_index + 1);
3203 queue.safe_push (loop->header);
3204 queues[queue_index] = queue;
3205 block_priority.put (loop->header, queue_index);
3207 for (; queue_index >= 0; queue_index--)
3209 if (latch_index < queue_index)
3211 while (queues[queue_index].length ())
3213 basic_block bb;
3214 ptrdiff_t bound_index = queue_index;
3215 edge e;
3216 edge_iterator ei;
3218 queue = queues[queue_index];
3219 bb = queue.pop ();
3221 /* OK, we later inserted the BB with lower priority, skip it. */
3222 if (*block_priority.get (bb) > queue_index)
3223 continue;
3225 /* See if we can improve the bound. */
3226 ptrdiff_t *entry = bb_bounds.get (bb);
3227 if (entry && *entry < bound_index)
3228 bound_index = *entry;
3230 /* Insert succesors into the queue, watch for latch edge
3231 and record greatest index we saw. */
3232 FOR_EACH_EDGE (e, ei, bb->succs)
3234 bool insert = false;
3236 if (loop_exit_edge_p (loop, e))
3237 continue;
3239 if (e == loop_latch_edge (loop)
3240 && latch_index < bound_index)
3241 latch_index = bound_index;
3242 else if (!(entry = block_priority.get (e->dest)))
3244 insert = true;
3245 block_priority.put (e->dest, bound_index);
3247 else if (*entry < bound_index)
3249 insert = true;
3250 *entry = bound_index;
3253 if (insert)
3254 queues[bound_index].safe_push (e->dest);
3258 queues[queue_index].release ();
3261 gcc_assert (latch_index >= 0);
3262 if ((unsigned)latch_index < bounds.length ())
3264 if (dump_file && (dump_flags & TDF_DETAILS))
3266 fprintf (dump_file, "Found better loop bound ");
3267 print_decu (bounds[latch_index], dump_file);
3268 fprintf (dump_file, "\n");
3270 record_niter_bound (loop, bounds[latch_index], false, true);
3273 queues.release ();
3274 bounds.release ();
3277 /* See if every path cross the loop goes through a statement that is known
3278 to not execute at the last iteration. In that case we can decrese iteration
3279 count by 1. */
3281 static void
3282 maybe_lower_iteration_bound (struct loop *loop)
3284 hash_set<gimple> *not_executed_last_iteration = NULL;
3285 struct nb_iter_bound *elt;
3286 bool found_exit = false;
3287 vec<basic_block> queue = vNULL;
3288 bitmap visited;
3290 /* Collect all statements with interesting (i.e. lower than
3291 nb_iterations_upper_bound) bound on them.
3293 TODO: Due to the way record_estimate choose estimates to store, the bounds
3294 will be always nb_iterations_upper_bound-1. We can change this to record
3295 also statements not dominating the loop latch and update the walk bellow
3296 to the shortest path algorthm. */
3297 for (elt = loop->bounds; elt; elt = elt->next)
3299 if (!elt->is_exit
3300 && wi::ltu_p (elt->bound, loop->nb_iterations_upper_bound))
3302 if (!not_executed_last_iteration)
3303 not_executed_last_iteration = new hash_set<gimple>;
3304 not_executed_last_iteration->add (elt->stmt);
3307 if (!not_executed_last_iteration)
3308 return;
3310 /* Start DFS walk in the loop header and see if we can reach the
3311 loop latch or any of the exits (including statements with side
3312 effects that may terminate the loop otherwise) without visiting
3313 any of the statements known to have undefined effect on the last
3314 iteration. */
3315 queue.safe_push (loop->header);
3316 visited = BITMAP_ALLOC (NULL);
3317 bitmap_set_bit (visited, loop->header->index);
3318 found_exit = false;
3322 basic_block bb = queue.pop ();
3323 gimple_stmt_iterator gsi;
3324 bool stmt_found = false;
3326 /* Loop for possible exits and statements bounding the execution. */
3327 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
3329 gimple stmt = gsi_stmt (gsi);
3330 if (not_executed_last_iteration->contains (stmt))
3332 stmt_found = true;
3333 break;
3335 if (gimple_has_side_effects (stmt))
3337 found_exit = true;
3338 break;
3341 if (found_exit)
3342 break;
3344 /* If no bounding statement is found, continue the walk. */
3345 if (!stmt_found)
3347 edge e;
3348 edge_iterator ei;
3350 FOR_EACH_EDGE (e, ei, bb->succs)
3352 if (loop_exit_edge_p (loop, e)
3353 || e == loop_latch_edge (loop))
3355 found_exit = true;
3356 break;
3358 if (bitmap_set_bit (visited, e->dest->index))
3359 queue.safe_push (e->dest);
3363 while (queue.length () && !found_exit);
3365 /* If every path through the loop reach bounding statement before exit,
3366 then we know the last iteration of the loop will have undefined effect
3367 and we can decrease number of iterations. */
3369 if (!found_exit)
3371 if (dump_file && (dump_flags & TDF_DETAILS))
3372 fprintf (dump_file, "Reducing loop iteration estimate by 1; "
3373 "undefined statement must be executed at the last iteration.\n");
3374 record_niter_bound (loop, loop->nb_iterations_upper_bound - 1,
3375 false, true);
3377 BITMAP_FREE (visited);
3378 queue.release ();
3379 delete not_executed_last_iteration;
3382 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P
3383 is true also use estimates derived from undefined behavior. */
3385 static void
3386 estimate_numbers_of_iterations_loop (struct loop *loop)
3388 vec<edge> exits;
3389 tree niter, type;
3390 unsigned i;
3391 struct tree_niter_desc niter_desc;
3392 edge ex;
3393 widest_int bound;
3394 edge likely_exit;
3396 /* Give up if we already have tried to compute an estimation. */
3397 if (loop->estimate_state != EST_NOT_COMPUTED)
3398 return;
3400 loop->estimate_state = EST_AVAILABLE;
3401 /* Force estimate compuation but leave any existing upper bound in place. */
3402 loop->any_estimate = false;
3404 /* Ensure that loop->nb_iterations is computed if possible. If it turns out
3405 to be constant, we avoid undefined behavior implied bounds and instead
3406 diagnose those loops with -Waggressive-loop-optimizations. */
3407 number_of_latch_executions (loop);
3409 exits = get_loop_exit_edges (loop);
3410 likely_exit = single_likely_exit (loop);
3411 FOR_EACH_VEC_ELT (exits, i, ex)
3413 if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false))
3414 continue;
3416 niter = niter_desc.niter;
3417 type = TREE_TYPE (niter);
3418 if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST)
3419 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
3420 build_int_cst (type, 0),
3421 niter);
3422 record_estimate (loop, niter, niter_desc.max,
3423 last_stmt (ex->src),
3424 true, ex == likely_exit, true);
3426 exits.release ();
3428 if (flag_aggressive_loop_optimizations)
3429 infer_loop_bounds_from_undefined (loop);
3431 discover_iteration_bound_by_body_walk (loop);
3433 maybe_lower_iteration_bound (loop);
3435 /* If we have a measured profile, use it to estimate the number of
3436 iterations. */
3437 if (loop->header->count != 0)
3439 gcov_type nit = expected_loop_iterations_unbounded (loop) + 1;
3440 bound = gcov_type_to_wide_int (nit);
3441 record_niter_bound (loop, bound, true, false);
3444 /* If we know the exact number of iterations of this loop, try to
3445 not break code with undefined behavior by not recording smaller
3446 maximum number of iterations. */
3447 if (loop->nb_iterations
3448 && TREE_CODE (loop->nb_iterations) == INTEGER_CST)
3450 loop->any_upper_bound = true;
3451 loop->nb_iterations_upper_bound = wi::to_widest (loop->nb_iterations);
3455 /* Sets NIT to the estimated number of executions of the latch of the
3456 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
3457 large as the number of iterations. If we have no reliable estimate,
3458 the function returns false, otherwise returns true. */
3460 bool
3461 estimated_loop_iterations (struct loop *loop, widest_int *nit)
3463 /* When SCEV information is available, try to update loop iterations
3464 estimate. Otherwise just return whatever we recorded earlier. */
3465 if (scev_initialized_p ())
3466 estimate_numbers_of_iterations_loop (loop);
3468 return (get_estimated_loop_iterations (loop, nit));
3471 /* Similar to estimated_loop_iterations, but returns the estimate only
3472 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3473 on the number of iterations of LOOP could not be derived, returns -1. */
3475 HOST_WIDE_INT
3476 estimated_loop_iterations_int (struct loop *loop)
3478 widest_int nit;
3479 HOST_WIDE_INT hwi_nit;
3481 if (!estimated_loop_iterations (loop, &nit))
3482 return -1;
3484 if (!wi::fits_shwi_p (nit))
3485 return -1;
3486 hwi_nit = nit.to_shwi ();
3488 return hwi_nit < 0 ? -1 : hwi_nit;
3492 /* Sets NIT to an upper bound for the maximum number of executions of the
3493 latch of the LOOP. If we have no reliable estimate, the function returns
3494 false, otherwise returns true. */
3496 bool
3497 max_loop_iterations (struct loop *loop, widest_int *nit)
3499 /* When SCEV information is available, try to update loop iterations
3500 estimate. Otherwise just return whatever we recorded earlier. */
3501 if (scev_initialized_p ())
3502 estimate_numbers_of_iterations_loop (loop);
3504 return get_max_loop_iterations (loop, nit);
3507 /* Similar to max_loop_iterations, but returns the estimate only
3508 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3509 on the number of iterations of LOOP could not be derived, returns -1. */
3511 HOST_WIDE_INT
3512 max_loop_iterations_int (struct loop *loop)
3514 widest_int nit;
3515 HOST_WIDE_INT hwi_nit;
3517 if (!max_loop_iterations (loop, &nit))
3518 return -1;
3520 if (!wi::fits_shwi_p (nit))
3521 return -1;
3522 hwi_nit = nit.to_shwi ();
3524 return hwi_nit < 0 ? -1 : hwi_nit;
3527 /* Returns an estimate for the number of executions of statements
3528 in the LOOP. For statements before the loop exit, this exceeds
3529 the number of execution of the latch by one. */
3531 HOST_WIDE_INT
3532 estimated_stmt_executions_int (struct loop *loop)
3534 HOST_WIDE_INT nit = estimated_loop_iterations_int (loop);
3535 HOST_WIDE_INT snit;
3537 if (nit == -1)
3538 return -1;
3540 snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1);
3542 /* If the computation overflows, return -1. */
3543 return snit < 0 ? -1 : snit;
3546 /* Sets NIT to the estimated maximum number of executions of the latch of the
3547 LOOP, plus one. If we have no reliable estimate, the function returns
3548 false, otherwise returns true. */
3550 bool
3551 max_stmt_executions (struct loop *loop, widest_int *nit)
3553 widest_int nit_minus_one;
3555 if (!max_loop_iterations (loop, nit))
3556 return false;
3558 nit_minus_one = *nit;
3560 *nit += 1;
3562 return wi::gtu_p (*nit, nit_minus_one);
3565 /* Sets NIT to the estimated number of executions of the latch of the
3566 LOOP, plus one. If we have no reliable estimate, the function returns
3567 false, otherwise returns true. */
3569 bool
3570 estimated_stmt_executions (struct loop *loop, widest_int *nit)
3572 widest_int nit_minus_one;
3574 if (!estimated_loop_iterations (loop, nit))
3575 return false;
3577 nit_minus_one = *nit;
3579 *nit += 1;
3581 return wi::gtu_p (*nit, nit_minus_one);
3584 /* Records estimates on numbers of iterations of loops. */
3586 void
3587 estimate_numbers_of_iterations (void)
3589 struct loop *loop;
3591 /* We don't want to issue signed overflow warnings while getting
3592 loop iteration estimates. */
3593 fold_defer_overflow_warnings ();
3595 FOR_EACH_LOOP (loop, 0)
3597 estimate_numbers_of_iterations_loop (loop);
3600 fold_undefer_and_ignore_overflow_warnings ();
3603 /* Returns true if statement S1 dominates statement S2. */
3605 bool
3606 stmt_dominates_stmt_p (gimple s1, gimple s2)
3608 basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
3610 if (!bb1
3611 || s1 == s2)
3612 return true;
3614 if (bb1 == bb2)
3616 gimple_stmt_iterator bsi;
3618 if (gimple_code (s2) == GIMPLE_PHI)
3619 return false;
3621 if (gimple_code (s1) == GIMPLE_PHI)
3622 return true;
3624 for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi))
3625 if (gsi_stmt (bsi) == s1)
3626 return true;
3628 return false;
3631 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
3634 /* Returns true when we can prove that the number of executions of
3635 STMT in the loop is at most NITER, according to the bound on
3636 the number of executions of the statement NITER_BOUND->stmt recorded in
3637 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT.
3639 ??? This code can become quite a CPU hog - we can have many bounds,
3640 and large basic block forcing stmt_dominates_stmt_p to be queried
3641 many times on a large basic blocks, so the whole thing is O(n^2)
3642 for scev_probably_wraps_p invocation (that can be done n times).
3644 It would make more sense (and give better answers) to remember BB
3645 bounds computed by discover_iteration_bound_by_body_walk. */
3647 static bool
3648 n_of_executions_at_most (gimple stmt,
3649 struct nb_iter_bound *niter_bound,
3650 tree niter)
3652 widest_int bound = niter_bound->bound;
3653 tree nit_type = TREE_TYPE (niter), e;
3654 enum tree_code cmp;
3656 gcc_assert (TYPE_UNSIGNED (nit_type));
3658 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
3659 the number of iterations is small. */
3660 if (!wi::fits_to_tree_p (bound, nit_type))
3661 return false;
3663 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
3664 times. This means that:
3666 -- if NITER_BOUND->is_exit is true, then everything after
3667 it at most NITER_BOUND->bound times.
3669 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
3670 is executed, then NITER_BOUND->stmt is executed as well in the same
3671 iteration then STMT is executed at most NITER_BOUND->bound + 1 times.
3673 If we can determine that NITER_BOUND->stmt is always executed
3674 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times.
3675 We conclude that if both statements belong to the same
3676 basic block and STMT is before NITER_BOUND->stmt and there are no
3677 statements with side effects in between. */
3679 if (niter_bound->is_exit)
3681 if (stmt == niter_bound->stmt
3682 || !stmt_dominates_stmt_p (niter_bound->stmt, stmt))
3683 return false;
3684 cmp = GE_EXPR;
3686 else
3688 if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt))
3690 gimple_stmt_iterator bsi;
3691 if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt)
3692 || gimple_code (stmt) == GIMPLE_PHI
3693 || gimple_code (niter_bound->stmt) == GIMPLE_PHI)
3694 return false;
3696 /* By stmt_dominates_stmt_p we already know that STMT appears
3697 before NITER_BOUND->STMT. Still need to test that the loop
3698 can not be terinated by a side effect in between. */
3699 for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt;
3700 gsi_next (&bsi))
3701 if (gimple_has_side_effects (gsi_stmt (bsi)))
3702 return false;
3703 bound += 1;
3704 if (bound == 0
3705 || !wi::fits_to_tree_p (bound, nit_type))
3706 return false;
3708 cmp = GT_EXPR;
3711 e = fold_binary (cmp, boolean_type_node,
3712 niter, wide_int_to_tree (nit_type, bound));
3713 return e && integer_nonzerop (e);
3716 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
3718 bool
3719 nowrap_type_p (tree type)
3721 if (INTEGRAL_TYPE_P (type)
3722 && TYPE_OVERFLOW_UNDEFINED (type))
3723 return true;
3725 if (POINTER_TYPE_P (type))
3726 return true;
3728 return false;
3731 /* Return false only when the induction variable BASE + STEP * I is
3732 known to not overflow: i.e. when the number of iterations is small
3733 enough with respect to the step and initial condition in order to
3734 keep the evolution confined in TYPEs bounds. Return true when the
3735 iv is known to overflow or when the property is not computable.
3737 USE_OVERFLOW_SEMANTICS is true if this function should assume that
3738 the rules for overflow of the given language apply (e.g., that signed
3739 arithmetics in C does not overflow). */
3741 bool
3742 scev_probably_wraps_p (tree base, tree step,
3743 gimple at_stmt, struct loop *loop,
3744 bool use_overflow_semantics)
3746 tree delta, step_abs;
3747 tree unsigned_type, valid_niter;
3748 tree type = TREE_TYPE (step);
3749 tree e;
3750 widest_int niter;
3751 struct nb_iter_bound *bound;
3753 /* FIXME: We really need something like
3754 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
3756 We used to test for the following situation that frequently appears
3757 during address arithmetics:
3759 D.1621_13 = (long unsigned intD.4) D.1620_12;
3760 D.1622_14 = D.1621_13 * 8;
3761 D.1623_15 = (doubleD.29 *) D.1622_14;
3763 And derived that the sequence corresponding to D_14
3764 can be proved to not wrap because it is used for computing a
3765 memory access; however, this is not really the case -- for example,
3766 if D_12 = (unsigned char) [254,+,1], then D_14 has values
3767 2032, 2040, 0, 8, ..., but the code is still legal. */
3769 if (chrec_contains_undetermined (base)
3770 || chrec_contains_undetermined (step))
3771 return true;
3773 if (integer_zerop (step))
3774 return false;
3776 /* If we can use the fact that signed and pointer arithmetics does not
3777 wrap, we are done. */
3778 if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base)))
3779 return false;
3781 /* To be able to use estimates on number of iterations of the loop,
3782 we must have an upper bound on the absolute value of the step. */
3783 if (TREE_CODE (step) != INTEGER_CST)
3784 return true;
3786 /* Don't issue signed overflow warnings. */
3787 fold_defer_overflow_warnings ();
3789 /* Otherwise, compute the number of iterations before we reach the
3790 bound of the type, and verify that the loop is exited before this
3791 occurs. */
3792 unsigned_type = unsigned_type_for (type);
3793 base = fold_convert (unsigned_type, base);
3795 if (tree_int_cst_sign_bit (step))
3797 tree extreme = fold_convert (unsigned_type,
3798 lower_bound_in_type (type, type));
3799 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
3800 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type,
3801 fold_convert (unsigned_type, step));
3803 else
3805 tree extreme = fold_convert (unsigned_type,
3806 upper_bound_in_type (type, type));
3807 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
3808 step_abs = fold_convert (unsigned_type, step);
3811 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
3813 estimate_numbers_of_iterations_loop (loop);
3815 if (max_loop_iterations (loop, &niter)
3816 && wi::fits_to_tree_p (niter, TREE_TYPE (valid_niter))
3817 && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter,
3818 wide_int_to_tree (TREE_TYPE (valid_niter),
3819 niter))) != NULL
3820 && integer_nonzerop (e))
3822 fold_undefer_and_ignore_overflow_warnings ();
3823 return false;
3825 if (at_stmt)
3826 for (bound = loop->bounds; bound; bound = bound->next)
3828 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
3830 fold_undefer_and_ignore_overflow_warnings ();
3831 return false;
3835 fold_undefer_and_ignore_overflow_warnings ();
3837 /* At this point we still don't have a proof that the iv does not
3838 overflow: give up. */
3839 return true;
3842 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
3844 void
3845 free_numbers_of_iterations_estimates_loop (struct loop *loop)
3847 struct nb_iter_bound *bound, *next;
3849 loop->nb_iterations = NULL;
3850 loop->estimate_state = EST_NOT_COMPUTED;
3851 for (bound = loop->bounds; bound; bound = next)
3853 next = bound->next;
3854 ggc_free (bound);
3857 loop->bounds = NULL;
3860 /* Frees the information on upper bounds on numbers of iterations of loops. */
3862 void
3863 free_numbers_of_iterations_estimates (void)
3865 struct loop *loop;
3867 FOR_EACH_LOOP (loop, 0)
3869 free_numbers_of_iterations_estimates_loop (loop);
3873 /* Substitute value VAL for ssa name NAME inside expressions held
3874 at LOOP. */
3876 void
3877 substitute_in_loop_info (struct loop *loop, tree name, tree val)
3879 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);