2008-07-07 Richard Guenther <rguenther@suse.de>
[official-gcc.git] / gcc / tree-ssa-loop-niter.c
blob80b45c298b7bfa27b498fc9a3cefebb6ed71e39c
1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004, 2005, 2006, 2007, 2008 Free Software Foundation,
3 Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by the
9 Free Software Foundation; either version 3, or (at your option) any
10 later version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "basic-block.h"
30 #include "output.h"
31 #include "diagnostic.h"
32 #include "intl.h"
33 #include "tree-flow.h"
34 #include "tree-dump.h"
35 #include "cfgloop.h"
36 #include "tree-pass.h"
37 #include "ggc.h"
38 #include "tree-chrec.h"
39 #include "tree-scalar-evolution.h"
40 #include "tree-data-ref.h"
41 #include "params.h"
42 #include "flags.h"
43 #include "toplev.h"
44 #include "tree-inline.h"
45 #include "gmp.h"
47 #define SWAP(X, Y) do { affine_iv *tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
49 /* The maximum number of dominator BBs we search for conditions
50 of loop header copies we use for simplifying a conditional
51 expression. */
52 #define MAX_DOMINATORS_TO_WALK 8
56 Analysis of number of iterations of an affine exit test.
60 /* Bounds on some value, BELOW <= X <= UP. */
62 typedef struct
64 mpz_t below, up;
65 } bounds;
68 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
70 static void
71 split_to_var_and_offset (tree expr, tree *var, mpz_t offset)
73 tree type = TREE_TYPE (expr);
74 tree op0, op1;
75 double_int off;
76 bool negate = false;
78 *var = expr;
79 mpz_set_ui (offset, 0);
81 switch (TREE_CODE (expr))
83 case MINUS_EXPR:
84 negate = true;
85 /* Fallthru. */
87 case PLUS_EXPR:
88 case POINTER_PLUS_EXPR:
89 op0 = TREE_OPERAND (expr, 0);
90 op1 = TREE_OPERAND (expr, 1);
92 if (TREE_CODE (op1) != INTEGER_CST)
93 break;
95 *var = op0;
96 /* Always sign extend the offset. */
97 off = double_int_sext (tree_to_double_int (op1),
98 TYPE_PRECISION (type));
99 mpz_set_double_int (offset, off, false);
100 break;
102 case INTEGER_CST:
103 *var = build_int_cst_type (type, 0);
104 off = tree_to_double_int (expr);
105 mpz_set_double_int (offset, off, TYPE_UNSIGNED (type));
106 break;
108 default:
109 break;
113 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
114 in TYPE to MIN and MAX. */
116 static void
117 determine_value_range (tree type, tree var, mpz_t off,
118 mpz_t min, mpz_t max)
120 /* If the expression is a constant, we know its value exactly. */
121 if (integer_zerop (var))
123 mpz_set (min, off);
124 mpz_set (max, off);
125 return;
128 /* If the computation may wrap, we know nothing about the value, except for
129 the range of the type. */
130 get_type_static_bounds (type, min, max);
131 if (!nowrap_type_p (type))
132 return;
134 /* Since the addition of OFF does not wrap, if OFF is positive, then we may
135 add it to MIN, otherwise to MAX. */
136 if (mpz_sgn (off) < 0)
137 mpz_add (max, max, off);
138 else
139 mpz_add (min, min, off);
142 /* Stores the bounds on the difference of the values of the expressions
143 (var + X) and (var + Y), computed in TYPE, to BNDS. */
145 static void
146 bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y,
147 bounds *bnds)
149 int rel = mpz_cmp (x, y);
150 bool may_wrap = !nowrap_type_p (type);
151 mpz_t m;
153 /* If X == Y, then the expressions are always equal.
154 If X > Y, there are the following possibilities:
155 a) neither of var + X and var + Y overflow or underflow, or both of
156 them do. Then their difference is X - Y.
157 b) var + X overflows, and var + Y does not. Then the values of the
158 expressions are var + X - M and var + Y, where M is the range of
159 the type, and their difference is X - Y - M.
160 c) var + Y underflows and var + X does not. Their difference again
161 is M - X + Y.
162 Therefore, if the arithmetics in type does not overflow, then the
163 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y)
164 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or
165 (X - Y, X - Y + M). */
167 if (rel == 0)
169 mpz_set_ui (bnds->below, 0);
170 mpz_set_ui (bnds->up, 0);
171 return;
174 mpz_init (m);
175 mpz_set_double_int (m, double_int_mask (TYPE_PRECISION (type)), true);
176 mpz_add_ui (m, m, 1);
177 mpz_sub (bnds->up, x, y);
178 mpz_set (bnds->below, bnds->up);
180 if (may_wrap)
182 if (rel > 0)
183 mpz_sub (bnds->below, bnds->below, m);
184 else
185 mpz_add (bnds->up, bnds->up, m);
188 mpz_clear (m);
191 /* From condition C0 CMP C1 derives information regarding the
192 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE,
193 and stores it to BNDS. */
195 static void
196 refine_bounds_using_guard (tree type, tree varx, mpz_t offx,
197 tree vary, mpz_t offy,
198 tree c0, enum tree_code cmp, tree c1,
199 bounds *bnds)
201 tree varc0, varc1, tmp, ctype;
202 mpz_t offc0, offc1, loffx, loffy, bnd;
203 bool lbound = false;
204 bool no_wrap = nowrap_type_p (type);
205 bool x_ok, y_ok;
207 switch (cmp)
209 case LT_EXPR:
210 case LE_EXPR:
211 case GT_EXPR:
212 case GE_EXPR:
213 STRIP_SIGN_NOPS (c0);
214 STRIP_SIGN_NOPS (c1);
215 ctype = TREE_TYPE (c0);
216 if (!useless_type_conversion_p (ctype, type))
217 return;
219 break;
221 case EQ_EXPR:
222 /* We could derive quite precise information from EQ_EXPR, however, such
223 a guard is unlikely to appear, so we do not bother with handling
224 it. */
225 return;
227 case NE_EXPR:
228 /* NE_EXPR comparisons do not contain much of useful information, except for
229 special case of comparing with the bounds of the type. */
230 if (TREE_CODE (c1) != INTEGER_CST
231 || !INTEGRAL_TYPE_P (type))
232 return;
234 /* Ensure that the condition speaks about an expression in the same type
235 as X and Y. */
236 ctype = TREE_TYPE (c0);
237 if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type))
238 return;
239 c0 = fold_convert (type, c0);
240 c1 = fold_convert (type, c1);
242 if (TYPE_MIN_VALUE (type)
243 && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0))
245 cmp = GT_EXPR;
246 break;
248 if (TYPE_MAX_VALUE (type)
249 && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0))
251 cmp = LT_EXPR;
252 break;
255 return;
256 default:
257 return;
260 mpz_init (offc0);
261 mpz_init (offc1);
262 split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0);
263 split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1);
265 /* We are only interested in comparisons of expressions based on VARX and
266 VARY. TODO -- we might also be able to derive some bounds from
267 expressions containing just one of the variables. */
269 if (operand_equal_p (varx, varc1, 0))
271 tmp = varc0; varc0 = varc1; varc1 = tmp;
272 mpz_swap (offc0, offc1);
273 cmp = swap_tree_comparison (cmp);
276 if (!operand_equal_p (varx, varc0, 0)
277 || !operand_equal_p (vary, varc1, 0))
278 goto end;
280 mpz_init_set (loffx, offx);
281 mpz_init_set (loffy, offy);
283 if (cmp == GT_EXPR || cmp == GE_EXPR)
285 tmp = varx; varx = vary; vary = tmp;
286 mpz_swap (offc0, offc1);
287 mpz_swap (loffx, loffy);
288 cmp = swap_tree_comparison (cmp);
289 lbound = true;
292 /* If there is no overflow, the condition implies that
294 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
296 The overflows and underflows may complicate things a bit; each
297 overflow decreases the appropriate offset by M, and underflow
298 increases it by M. The above inequality would not necessarily be
299 true if
301 -- VARX + OFFX underflows and VARX + OFFC0 does not, or
302 VARX + OFFC0 overflows, but VARX + OFFX does not.
303 This may only happen if OFFX < OFFC0.
304 -- VARY + OFFY overflows and VARY + OFFC1 does not, or
305 VARY + OFFC1 underflows and VARY + OFFY does not.
306 This may only happen if OFFY > OFFC1. */
308 if (no_wrap)
310 x_ok = true;
311 y_ok = true;
313 else
315 x_ok = (integer_zerop (varx)
316 || mpz_cmp (loffx, offc0) >= 0);
317 y_ok = (integer_zerop (vary)
318 || mpz_cmp (loffy, offc1) <= 0);
321 if (x_ok && y_ok)
323 mpz_init (bnd);
324 mpz_sub (bnd, loffx, loffy);
325 mpz_add (bnd, bnd, offc1);
326 mpz_sub (bnd, bnd, offc0);
328 if (cmp == LT_EXPR)
329 mpz_sub_ui (bnd, bnd, 1);
331 if (lbound)
333 mpz_neg (bnd, bnd);
334 if (mpz_cmp (bnds->below, bnd) < 0)
335 mpz_set (bnds->below, bnd);
337 else
339 if (mpz_cmp (bnd, bnds->up) < 0)
340 mpz_set (bnds->up, bnd);
342 mpz_clear (bnd);
345 mpz_clear (loffx);
346 mpz_clear (loffy);
347 end:
348 mpz_clear (offc0);
349 mpz_clear (offc1);
352 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS.
353 The subtraction is considered to be performed in arbitrary precision,
354 without overflows.
356 We do not attempt to be too clever regarding the value ranges of X and
357 Y; most of the time, they are just integers or ssa names offsetted by
358 integer. However, we try to use the information contained in the
359 comparisons before the loop (usually created by loop header copying). */
361 static void
362 bound_difference (struct loop *loop, tree x, tree y, bounds *bnds)
364 tree type = TREE_TYPE (x);
365 tree varx, vary;
366 mpz_t offx, offy;
367 mpz_t minx, maxx, miny, maxy;
368 int cnt = 0;
369 edge e;
370 basic_block bb;
371 tree cond, c0, c1;
372 enum tree_code cmp;
374 /* Get rid of unnecessary casts, but preserve the value of
375 the expressions. */
376 STRIP_SIGN_NOPS (x);
377 STRIP_SIGN_NOPS (y);
379 mpz_init (bnds->below);
380 mpz_init (bnds->up);
381 mpz_init (offx);
382 mpz_init (offy);
383 split_to_var_and_offset (x, &varx, offx);
384 split_to_var_and_offset (y, &vary, offy);
386 if (!integer_zerop (varx)
387 && operand_equal_p (varx, vary, 0))
389 /* Special case VARX == VARY -- we just need to compare the
390 offsets. The matters are a bit more complicated in the
391 case addition of offsets may wrap. */
392 bound_difference_of_offsetted_base (type, offx, offy, bnds);
394 else
396 /* Otherwise, use the value ranges to determine the initial
397 estimates on below and up. */
398 mpz_init (minx);
399 mpz_init (maxx);
400 mpz_init (miny);
401 mpz_init (maxy);
402 determine_value_range (type, varx, offx, minx, maxx);
403 determine_value_range (type, vary, offy, miny, maxy);
405 mpz_sub (bnds->below, minx, maxy);
406 mpz_sub (bnds->up, maxx, miny);
407 mpz_clear (minx);
408 mpz_clear (maxx);
409 mpz_clear (miny);
410 mpz_clear (maxy);
413 /* If both X and Y are constants, we cannot get any more precise. */
414 if (integer_zerop (varx) && integer_zerop (vary))
415 goto end;
417 /* Now walk the dominators of the loop header and use the entry
418 guards to refine the estimates. */
419 for (bb = loop->header;
420 bb != ENTRY_BLOCK_PTR && cnt < MAX_DOMINATORS_TO_WALK;
421 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
423 if (!single_pred_p (bb))
424 continue;
425 e = single_pred_edge (bb);
427 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
428 continue;
430 cond = COND_EXPR_COND (last_stmt (e->src));
431 if (!COMPARISON_CLASS_P (cond))
432 continue;
433 c0 = TREE_OPERAND (cond, 0);
434 cmp = TREE_CODE (cond);
435 c1 = TREE_OPERAND (cond, 1);
437 if (e->flags & EDGE_FALSE_VALUE)
438 cmp = invert_tree_comparison (cmp, false);
440 refine_bounds_using_guard (type, varx, offx, vary, offy,
441 c0, cmp, c1, bnds);
442 ++cnt;
445 end:
446 mpz_clear (offx);
447 mpz_clear (offy);
450 /* Update the bounds in BNDS that restrict the value of X to the bounds
451 that restrict the value of X + DELTA. X can be obtained as a
452 difference of two values in TYPE. */
454 static void
455 bounds_add (bounds *bnds, double_int delta, tree type)
457 mpz_t mdelta, max;
459 mpz_init (mdelta);
460 mpz_set_double_int (mdelta, delta, false);
462 mpz_init (max);
463 mpz_set_double_int (max, double_int_mask (TYPE_PRECISION (type)), true);
465 mpz_add (bnds->up, bnds->up, mdelta);
466 mpz_add (bnds->below, bnds->below, mdelta);
468 if (mpz_cmp (bnds->up, max) > 0)
469 mpz_set (bnds->up, max);
471 mpz_neg (max, max);
472 if (mpz_cmp (bnds->below, max) < 0)
473 mpz_set (bnds->below, max);
475 mpz_clear (mdelta);
476 mpz_clear (max);
479 /* Update the bounds in BNDS that restrict the value of X to the bounds
480 that restrict the value of -X. */
482 static void
483 bounds_negate (bounds *bnds)
485 mpz_t tmp;
487 mpz_init_set (tmp, bnds->up);
488 mpz_neg (bnds->up, bnds->below);
489 mpz_neg (bnds->below, tmp);
490 mpz_clear (tmp);
493 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
495 static tree
496 inverse (tree x, tree mask)
498 tree type = TREE_TYPE (x);
499 tree rslt;
500 unsigned ctr = tree_floor_log2 (mask);
502 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT)
504 unsigned HOST_WIDE_INT ix;
505 unsigned HOST_WIDE_INT imask;
506 unsigned HOST_WIDE_INT irslt = 1;
508 gcc_assert (cst_and_fits_in_hwi (x));
509 gcc_assert (cst_and_fits_in_hwi (mask));
511 ix = int_cst_value (x);
512 imask = int_cst_value (mask);
514 for (; ctr; ctr--)
516 irslt *= ix;
517 ix *= ix;
519 irslt &= imask;
521 rslt = build_int_cst_type (type, irslt);
523 else
525 rslt = build_int_cst (type, 1);
526 for (; ctr; ctr--)
528 rslt = int_const_binop (MULT_EXPR, rslt, x, 0);
529 x = int_const_binop (MULT_EXPR, x, x, 0);
531 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask, 0);
534 return rslt;
537 /* Derives the upper bound BND on the number of executions of loop with exit
538 condition S * i <> C, assuming that the loop is not infinite. If
539 NO_OVERFLOW is true, then the control variable of the loop does not
540 overflow. If NO_OVERFLOW is true or BNDS.below >= 0, then BNDS.up
541 contains the upper bound on the value of C. */
543 static void
544 number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s,
545 bounds *bnds)
547 double_int max;
548 mpz_t d;
550 /* If the control variable does not overflow, the number of iterations is
551 at most c / s. Otherwise it is at most the period of the control
552 variable. */
553 if (!no_overflow && !multiple_of_p (TREE_TYPE (c), c, s))
555 max = double_int_mask (TYPE_PRECISION (TREE_TYPE (c))
556 - tree_low_cst (num_ending_zeros (s), 1));
557 mpz_set_double_int (bnd, max, true);
558 return;
561 /* Determine the upper bound on C. */
562 if (no_overflow || mpz_sgn (bnds->below) >= 0)
563 mpz_set (bnd, bnds->up);
564 else if (TREE_CODE (c) == INTEGER_CST)
565 mpz_set_double_int (bnd, tree_to_double_int (c), true);
566 else
567 mpz_set_double_int (bnd, double_int_mask (TYPE_PRECISION (TREE_TYPE (c))),
568 true);
570 mpz_init (d);
571 mpz_set_double_int (d, tree_to_double_int (s), true);
572 mpz_fdiv_q (bnd, bnd, d);
573 mpz_clear (d);
576 /* Determines number of iterations of loop whose ending condition
577 is IV <> FINAL. TYPE is the type of the iv. The number of
578 iterations is stored to NITER. NEVER_INFINITE is true if
579 we know that the exit must be taken eventually, i.e., that the IV
580 ever reaches the value FINAL (we derived this earlier, and possibly set
581 NITER->assumptions to make sure this is the case). BNDS contains the
582 bounds on the difference FINAL - IV->base. */
584 static bool
585 number_of_iterations_ne (tree type, affine_iv *iv, tree final,
586 struct tree_niter_desc *niter, bool never_infinite,
587 bounds *bnds)
589 tree niter_type = unsigned_type_for (type);
590 tree s, c, d, bits, assumption, tmp, bound;
591 mpz_t max;
593 niter->control = *iv;
594 niter->bound = final;
595 niter->cmp = NE_EXPR;
597 /* Rearrange the terms so that we get inequality S * i <> C, with S
598 positive. Also cast everything to the unsigned type. If IV does
599 not overflow, BNDS bounds the value of C. Also, this is the
600 case if the computation |FINAL - IV->base| does not overflow, i.e.,
601 if BNDS->below in the result is nonnegative. */
602 if (tree_int_cst_sign_bit (iv->step))
604 s = fold_convert (niter_type,
605 fold_build1 (NEGATE_EXPR, type, iv->step));
606 c = fold_build2 (MINUS_EXPR, niter_type,
607 fold_convert (niter_type, iv->base),
608 fold_convert (niter_type, final));
609 bounds_negate (bnds);
611 else
613 s = fold_convert (niter_type, iv->step);
614 c = fold_build2 (MINUS_EXPR, niter_type,
615 fold_convert (niter_type, final),
616 fold_convert (niter_type, iv->base));
619 mpz_init (max);
620 number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds);
621 niter->max = mpz_get_double_int (niter_type, max, false);
622 mpz_clear (max);
624 /* First the trivial cases -- when the step is 1. */
625 if (integer_onep (s))
627 niter->niter = c;
628 return true;
631 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
632 is infinite. Otherwise, the number of iterations is
633 (inverse(s/d) * (c/d)) mod (size of mode/d). */
634 bits = num_ending_zeros (s);
635 bound = build_low_bits_mask (niter_type,
636 (TYPE_PRECISION (niter_type)
637 - tree_low_cst (bits, 1)));
639 d = fold_binary_to_constant (LSHIFT_EXPR, niter_type,
640 build_int_cst (niter_type, 1), bits);
641 s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits);
643 if (!never_infinite)
645 /* If we cannot assume that the loop is not infinite, record the
646 assumptions for divisibility of c. */
647 assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d);
648 assumption = fold_build2 (EQ_EXPR, boolean_type_node,
649 assumption, build_int_cst (niter_type, 0));
650 if (!integer_nonzerop (assumption))
651 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
652 niter->assumptions, assumption);
655 c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d);
656 tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound));
657 niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound);
658 return true;
661 /* Checks whether we can determine the final value of the control variable
662 of the loop with ending condition IV0 < IV1 (computed in TYPE).
663 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
664 of the step. The assumptions necessary to ensure that the computation
665 of the final value does not overflow are recorded in NITER. If we
666 find the final value, we adjust DELTA and return TRUE. Otherwise
667 we return false. BNDS bounds the value of IV1->base - IV0->base,
668 and will be updated by the same amount as DELTA. */
670 static bool
671 number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1,
672 struct tree_niter_desc *niter,
673 tree *delta, tree step,
674 bounds *bnds)
676 tree niter_type = TREE_TYPE (step);
677 tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step);
678 tree tmod;
679 mpz_t mmod;
680 tree assumption = boolean_true_node, bound, noloop;
681 bool ret = false;
682 tree type1 = type;
683 if (POINTER_TYPE_P (type))
684 type1 = sizetype;
686 if (TREE_CODE (mod) != INTEGER_CST)
687 return false;
688 if (integer_nonzerop (mod))
689 mod = fold_build2 (MINUS_EXPR, niter_type, step, mod);
690 tmod = fold_convert (type1, mod);
692 mpz_init (mmod);
693 mpz_set_double_int (mmod, tree_to_double_int (mod), true);
694 mpz_neg (mmod, mmod);
696 if (integer_nonzerop (iv0->step))
698 /* The final value of the iv is iv1->base + MOD, assuming that this
699 computation does not overflow, and that
700 iv0->base <= iv1->base + MOD. */
701 if (!iv1->no_overflow && !integer_zerop (mod))
703 bound = fold_build2 (MINUS_EXPR, type,
704 TYPE_MAX_VALUE (type1), tmod);
705 assumption = fold_build2 (LE_EXPR, boolean_type_node,
706 iv1->base, bound);
707 if (integer_zerop (assumption))
708 goto end;
710 if (mpz_cmp (mmod, bnds->below) < 0)
711 noloop = boolean_false_node;
712 else
713 noloop = fold_build2 (GT_EXPR, boolean_type_node,
714 iv0->base,
715 fold_build2 (PLUS_EXPR, type1,
716 iv1->base, tmod));
718 else
720 /* The final value of the iv is iv0->base - MOD, assuming that this
721 computation does not overflow, and that
722 iv0->base - MOD <= iv1->base. */
723 if (!iv0->no_overflow && !integer_zerop (mod))
725 bound = fold_build2 (PLUS_EXPR, type1,
726 TYPE_MIN_VALUE (type1), tmod);
727 assumption = fold_build2 (GE_EXPR, boolean_type_node,
728 iv0->base, bound);
729 if (integer_zerop (assumption))
730 goto end;
732 if (mpz_cmp (mmod, bnds->below) < 0)
733 noloop = boolean_false_node;
734 else
735 noloop = fold_build2 (GT_EXPR, boolean_type_node,
736 fold_build2 (MINUS_EXPR, type1,
737 iv0->base, tmod),
738 iv1->base);
741 if (!integer_nonzerop (assumption))
742 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
743 niter->assumptions,
744 assumption);
745 if (!integer_zerop (noloop))
746 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
747 niter->may_be_zero,
748 noloop);
749 bounds_add (bnds, tree_to_double_int (mod), type);
750 *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod);
752 ret = true;
753 end:
754 mpz_clear (mmod);
755 return ret;
758 /* Add assertions to NITER that ensure that the control variable of the loop
759 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
760 are TYPE. Returns false if we can prove that there is an overflow, true
761 otherwise. STEP is the absolute value of the step. */
763 static bool
764 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1,
765 struct tree_niter_desc *niter, tree step)
767 tree bound, d, assumption, diff;
768 tree niter_type = TREE_TYPE (step);
770 if (integer_nonzerop (iv0->step))
772 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
773 if (iv0->no_overflow)
774 return true;
776 /* If iv0->base is a constant, we can determine the last value before
777 overflow precisely; otherwise we conservatively assume
778 MAX - STEP + 1. */
780 if (TREE_CODE (iv0->base) == INTEGER_CST)
782 d = fold_build2 (MINUS_EXPR, niter_type,
783 fold_convert (niter_type, TYPE_MAX_VALUE (type)),
784 fold_convert (niter_type, iv0->base));
785 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
787 else
788 diff = fold_build2 (MINUS_EXPR, niter_type, step,
789 build_int_cst (niter_type, 1));
790 bound = fold_build2 (MINUS_EXPR, type,
791 TYPE_MAX_VALUE (type), fold_convert (type, diff));
792 assumption = fold_build2 (LE_EXPR, boolean_type_node,
793 iv1->base, bound);
795 else
797 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
798 if (iv1->no_overflow)
799 return true;
801 if (TREE_CODE (iv1->base) == INTEGER_CST)
803 d = fold_build2 (MINUS_EXPR, niter_type,
804 fold_convert (niter_type, iv1->base),
805 fold_convert (niter_type, TYPE_MIN_VALUE (type)));
806 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
808 else
809 diff = fold_build2 (MINUS_EXPR, niter_type, step,
810 build_int_cst (niter_type, 1));
811 bound = fold_build2 (PLUS_EXPR, type,
812 TYPE_MIN_VALUE (type), fold_convert (type, diff));
813 assumption = fold_build2 (GE_EXPR, boolean_type_node,
814 iv0->base, bound);
817 if (integer_zerop (assumption))
818 return false;
819 if (!integer_nonzerop (assumption))
820 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
821 niter->assumptions, assumption);
823 iv0->no_overflow = true;
824 iv1->no_overflow = true;
825 return true;
828 /* Add an assumption to NITER that a loop whose ending condition
829 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS
830 bounds the value of IV1->base - IV0->base. */
832 static void
833 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1,
834 struct tree_niter_desc *niter, bounds *bnds)
836 tree assumption = boolean_true_node, bound, diff;
837 tree mbz, mbzl, mbzr, type1;
838 bool rolls_p, no_overflow_p;
839 double_int dstep;
840 mpz_t mstep, max;
842 /* We are going to compute the number of iterations as
843 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned
844 variant of TYPE. This formula only works if
846 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
848 (where MAX is the maximum value of the unsigned variant of TYPE, and
849 the computations in this formula are performed in full precision
850 (without overflows).
852 Usually, for loops with exit condition iv0->base + step * i < iv1->base,
853 we have a condition of form iv0->base - step < iv1->base before the loop,
854 and for loops iv0->base < iv1->base - step * i the condition
855 iv0->base < iv1->base + step, due to loop header copying, which enable us
856 to prove the lower bound.
858 The upper bound is more complicated. Unless the expressions for initial
859 and final value themselves contain enough information, we usually cannot
860 derive it from the context. */
862 /* First check whether the answer does not follow from the bounds we gathered
863 before. */
864 if (integer_nonzerop (iv0->step))
865 dstep = tree_to_double_int (iv0->step);
866 else
868 dstep = double_int_sext (tree_to_double_int (iv1->step),
869 TYPE_PRECISION (type));
870 dstep = double_int_neg (dstep);
873 mpz_init (mstep);
874 mpz_set_double_int (mstep, dstep, true);
875 mpz_neg (mstep, mstep);
876 mpz_add_ui (mstep, mstep, 1);
878 rolls_p = mpz_cmp (mstep, bnds->below) <= 0;
880 mpz_init (max);
881 mpz_set_double_int (max, double_int_mask (TYPE_PRECISION (type)), true);
882 mpz_add (max, max, mstep);
883 no_overflow_p = (mpz_cmp (bnds->up, max) <= 0
884 /* For pointers, only values lying inside a single object
885 can be compared or manipulated by pointer arithmetics.
886 Gcc in general does not allow or handle objects larger
887 than half of the address space, hence the upper bound
888 is satisfied for pointers. */
889 || POINTER_TYPE_P (type));
890 mpz_clear (mstep);
891 mpz_clear (max);
893 if (rolls_p && no_overflow_p)
894 return;
896 type1 = type;
897 if (POINTER_TYPE_P (type))
898 type1 = sizetype;
900 /* Now the hard part; we must formulate the assumption(s) as expressions, and
901 we must be careful not to introduce overflow. */
903 if (integer_nonzerop (iv0->step))
905 diff = fold_build2 (MINUS_EXPR, type1,
906 iv0->step, build_int_cst (type1, 1));
908 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
909 0 address never belongs to any object, we can assume this for
910 pointers. */
911 if (!POINTER_TYPE_P (type))
913 bound = fold_build2 (PLUS_EXPR, type1,
914 TYPE_MIN_VALUE (type), diff);
915 assumption = fold_build2 (GE_EXPR, boolean_type_node,
916 iv0->base, bound);
919 /* And then we can compute iv0->base - diff, and compare it with
920 iv1->base. */
921 mbzl = fold_build2 (MINUS_EXPR, type1,
922 fold_convert (type1, iv0->base), diff);
923 mbzr = fold_convert (type1, iv1->base);
925 else
927 diff = fold_build2 (PLUS_EXPR, type1,
928 iv1->step, build_int_cst (type1, 1));
930 if (!POINTER_TYPE_P (type))
932 bound = fold_build2 (PLUS_EXPR, type1,
933 TYPE_MAX_VALUE (type), diff);
934 assumption = fold_build2 (LE_EXPR, boolean_type_node,
935 iv1->base, bound);
938 mbzl = fold_convert (type1, iv0->base);
939 mbzr = fold_build2 (MINUS_EXPR, type1,
940 fold_convert (type1, iv1->base), diff);
943 if (!integer_nonzerop (assumption))
944 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
945 niter->assumptions, assumption);
946 if (!rolls_p)
948 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr);
949 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
950 niter->may_be_zero, mbz);
954 /* Determines number of iterations of loop whose ending condition
955 is IV0 < IV1. TYPE is the type of the iv. The number of
956 iterations is stored to NITER. BNDS bounds the difference
957 IV1->base - IV0->base. */
959 static bool
960 number_of_iterations_lt (tree type, affine_iv *iv0, affine_iv *iv1,
961 struct tree_niter_desc *niter,
962 bool never_infinite ATTRIBUTE_UNUSED,
963 bounds *bnds)
965 tree niter_type = unsigned_type_for (type);
966 tree delta, step, s;
967 mpz_t mstep, tmp;
969 if (integer_nonzerop (iv0->step))
971 niter->control = *iv0;
972 niter->cmp = LT_EXPR;
973 niter->bound = iv1->base;
975 else
977 niter->control = *iv1;
978 niter->cmp = GT_EXPR;
979 niter->bound = iv0->base;
982 delta = fold_build2 (MINUS_EXPR, niter_type,
983 fold_convert (niter_type, iv1->base),
984 fold_convert (niter_type, iv0->base));
986 /* First handle the special case that the step is +-1. */
987 if ((integer_onep (iv0->step) && integer_zerop (iv1->step))
988 || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step)))
990 /* for (i = iv0->base; i < iv1->base; i++)
994 for (i = iv1->base; i > iv0->base; i--).
996 In both cases # of iterations is iv1->base - iv0->base, assuming that
997 iv1->base >= iv0->base.
999 First try to derive a lower bound on the value of
1000 iv1->base - iv0->base, computed in full precision. If the difference
1001 is nonnegative, we are done, otherwise we must record the
1002 condition. */
1004 if (mpz_sgn (bnds->below) < 0)
1005 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node,
1006 iv1->base, iv0->base);
1007 niter->niter = delta;
1008 niter->max = mpz_get_double_int (niter_type, bnds->up, false);
1009 return true;
1012 if (integer_nonzerop (iv0->step))
1013 step = fold_convert (niter_type, iv0->step);
1014 else
1015 step = fold_convert (niter_type,
1016 fold_build1 (NEGATE_EXPR, type, iv1->step));
1018 /* If we can determine the final value of the control iv exactly, we can
1019 transform the condition to != comparison. In particular, this will be
1020 the case if DELTA is constant. */
1021 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step,
1022 bnds))
1024 affine_iv zps;
1026 zps.base = build_int_cst (niter_type, 0);
1027 zps.step = step;
1028 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1029 zps does not overflow. */
1030 zps.no_overflow = true;
1032 return number_of_iterations_ne (type, &zps, delta, niter, true, bnds);
1035 /* Make sure that the control iv does not overflow. */
1036 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step))
1037 return false;
1039 /* We determine the number of iterations as (delta + step - 1) / step. For
1040 this to work, we must know that iv1->base >= iv0->base - step + 1,
1041 otherwise the loop does not roll. */
1042 assert_loop_rolls_lt (type, iv0, iv1, niter, bnds);
1044 s = fold_build2 (MINUS_EXPR, niter_type,
1045 step, build_int_cst (niter_type, 1));
1046 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s);
1047 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step);
1049 mpz_init (mstep);
1050 mpz_init (tmp);
1051 mpz_set_double_int (mstep, tree_to_double_int (step), true);
1052 mpz_add (tmp, bnds->up, mstep);
1053 mpz_sub_ui (tmp, tmp, 1);
1054 mpz_fdiv_q (tmp, tmp, mstep);
1055 niter->max = mpz_get_double_int (niter_type, tmp, false);
1056 mpz_clear (mstep);
1057 mpz_clear (tmp);
1059 return true;
1062 /* Determines number of iterations of loop whose ending condition
1063 is IV0 <= IV1. TYPE is the type of the iv. The number of
1064 iterations is stored to NITER. NEVER_INFINITE is true if
1065 we know that this condition must eventually become false (we derived this
1066 earlier, and possibly set NITER->assumptions to make sure this
1067 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1069 static bool
1070 number_of_iterations_le (tree type, affine_iv *iv0, affine_iv *iv1,
1071 struct tree_niter_desc *niter, bool never_infinite,
1072 bounds *bnds)
1074 tree assumption;
1075 tree type1 = type;
1076 if (POINTER_TYPE_P (type))
1077 type1 = sizetype;
1079 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1080 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1081 value of the type. This we must know anyway, since if it is
1082 equal to this value, the loop rolls forever. */
1084 if (!never_infinite)
1086 if (integer_nonzerop (iv0->step))
1087 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1088 iv1->base, TYPE_MAX_VALUE (type1));
1089 else
1090 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1091 iv0->base, TYPE_MIN_VALUE (type1));
1093 if (integer_zerop (assumption))
1094 return false;
1095 if (!integer_nonzerop (assumption))
1096 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1097 niter->assumptions, assumption);
1100 if (integer_nonzerop (iv0->step))
1101 iv1->base = fold_build2 (PLUS_EXPR, type1,
1102 iv1->base, build_int_cst (type1, 1));
1103 else
1104 iv0->base = fold_build2 (MINUS_EXPR, type1,
1105 iv0->base, build_int_cst (type1, 1));
1107 bounds_add (bnds, double_int_one, type1);
1109 return number_of_iterations_lt (type, iv0, iv1, niter, never_infinite, bnds);
1112 /* Dumps description of affine induction variable IV to FILE. */
1114 static void
1115 dump_affine_iv (FILE *file, affine_iv *iv)
1117 if (!integer_zerop (iv->step))
1118 fprintf (file, "[");
1120 print_generic_expr (dump_file, iv->base, TDF_SLIM);
1122 if (!integer_zerop (iv->step))
1124 fprintf (file, ", + , ");
1125 print_generic_expr (dump_file, iv->step, TDF_SLIM);
1126 fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : "");
1130 /* Determine the number of iterations according to condition (for staying
1131 inside loop) which compares two induction variables using comparison
1132 operator CODE. The induction variable on left side of the comparison
1133 is IV0, the right-hand side is IV1. Both induction variables must have
1134 type TYPE, which must be an integer or pointer type. The steps of the
1135 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1137 LOOP is the loop whose number of iterations we are determining.
1139 ONLY_EXIT is true if we are sure this is the only way the loop could be
1140 exited (including possibly non-returning function calls, exceptions, etc.)
1141 -- in this case we can use the information whether the control induction
1142 variables can overflow or not in a more efficient way.
1144 The results (number of iterations and assumptions as described in
1145 comments at struct tree_niter_desc in tree-flow.h) are stored to NITER.
1146 Returns false if it fails to determine number of iterations, true if it
1147 was determined (possibly with some assumptions). */
1149 static bool
1150 number_of_iterations_cond (struct loop *loop,
1151 tree type, affine_iv *iv0, enum tree_code code,
1152 affine_iv *iv1, struct tree_niter_desc *niter,
1153 bool only_exit)
1155 bool never_infinite, ret;
1156 bounds bnds;
1158 /* The meaning of these assumptions is this:
1159 if !assumptions
1160 then the rest of information does not have to be valid
1161 if may_be_zero then the loop does not roll, even if
1162 niter != 0. */
1163 niter->assumptions = boolean_true_node;
1164 niter->may_be_zero = boolean_false_node;
1165 niter->niter = NULL_TREE;
1166 niter->max = double_int_zero;
1168 niter->bound = NULL_TREE;
1169 niter->cmp = ERROR_MARK;
1171 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1172 the control variable is on lhs. */
1173 if (code == GE_EXPR || code == GT_EXPR
1174 || (code == NE_EXPR && integer_zerop (iv0->step)))
1176 SWAP (iv0, iv1);
1177 code = swap_tree_comparison (code);
1180 if (!only_exit)
1182 /* If this is not the only possible exit from the loop, the information
1183 that the induction variables cannot overflow as derived from
1184 signedness analysis cannot be relied upon. We use them e.g. in the
1185 following way: given loop for (i = 0; i <= n; i++), if i is
1186 signed, it cannot overflow, thus this loop is equivalent to
1187 for (i = 0; i < n + 1; i++); however, if n == MAX, but the loop
1188 is exited in some other way before i overflows, this transformation
1189 is incorrect (the new loop exits immediately). */
1190 iv0->no_overflow = false;
1191 iv1->no_overflow = false;
1194 if (POINTER_TYPE_P (type))
1196 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1197 to the same object. If they do, the control variable cannot wrap
1198 (as wrap around the bounds of memory will never return a pointer
1199 that would be guaranteed to point to the same object, even if we
1200 avoid undefined behavior by casting to size_t and back). The
1201 restrictions on pointer arithmetics and comparisons of pointers
1202 ensure that using the no-overflow assumptions is correct in this
1203 case even if ONLY_EXIT is false. */
1204 iv0->no_overflow = true;
1205 iv1->no_overflow = true;
1208 /* If the control induction variable does not overflow, the loop obviously
1209 cannot be infinite. */
1210 if (!integer_zerop (iv0->step) && iv0->no_overflow)
1211 never_infinite = true;
1212 else if (!integer_zerop (iv1->step) && iv1->no_overflow)
1213 never_infinite = true;
1214 else
1215 never_infinite = false;
1217 /* We can handle the case when neither of the sides of the comparison is
1218 invariant, provided that the test is NE_EXPR. This rarely occurs in
1219 practice, but it is simple enough to manage. */
1220 if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step))
1222 if (code != NE_EXPR)
1223 return false;
1225 iv0->step = fold_binary_to_constant (MINUS_EXPR, type,
1226 iv0->step, iv1->step);
1227 iv0->no_overflow = false;
1228 iv1->step = build_int_cst (type, 0);
1229 iv1->no_overflow = true;
1232 /* If the result of the comparison is a constant, the loop is weird. More
1233 precise handling would be possible, but the situation is not common enough
1234 to waste time on it. */
1235 if (integer_zerop (iv0->step) && integer_zerop (iv1->step))
1236 return false;
1238 /* Ignore loops of while (i-- < 10) type. */
1239 if (code != NE_EXPR)
1241 if (iv0->step && tree_int_cst_sign_bit (iv0->step))
1242 return false;
1244 if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step))
1245 return false;
1248 /* If the loop exits immediately, there is nothing to do. */
1249 if (integer_zerop (fold_build2 (code, boolean_type_node, iv0->base, iv1->base)))
1251 niter->niter = build_int_cst (unsigned_type_for (type), 0);
1252 niter->max = double_int_zero;
1253 return true;
1256 /* OK, now we know we have a senseful loop. Handle several cases, depending
1257 on what comparison operator is used. */
1258 bound_difference (loop, iv1->base, iv0->base, &bnds);
1260 if (dump_file && (dump_flags & TDF_DETAILS))
1262 fprintf (dump_file,
1263 "Analyzing # of iterations of loop %d\n", loop->num);
1265 fprintf (dump_file, " exit condition ");
1266 dump_affine_iv (dump_file, iv0);
1267 fprintf (dump_file, " %s ",
1268 code == NE_EXPR ? "!="
1269 : code == LT_EXPR ? "<"
1270 : "<=");
1271 dump_affine_iv (dump_file, iv1);
1272 fprintf (dump_file, "\n");
1274 fprintf (dump_file, " bounds on difference of bases: ");
1275 mpz_out_str (dump_file, 10, bnds.below);
1276 fprintf (dump_file, " ... ");
1277 mpz_out_str (dump_file, 10, bnds.up);
1278 fprintf (dump_file, "\n");
1281 switch (code)
1283 case NE_EXPR:
1284 gcc_assert (integer_zerop (iv1->step));
1285 ret = number_of_iterations_ne (type, iv0, iv1->base, niter,
1286 never_infinite, &bnds);
1287 break;
1289 case LT_EXPR:
1290 ret = number_of_iterations_lt (type, iv0, iv1, niter, never_infinite,
1291 &bnds);
1292 break;
1294 case LE_EXPR:
1295 ret = number_of_iterations_le (type, iv0, iv1, niter, never_infinite,
1296 &bnds);
1297 break;
1299 default:
1300 gcc_unreachable ();
1303 mpz_clear (bnds.up);
1304 mpz_clear (bnds.below);
1306 if (dump_file && (dump_flags & TDF_DETAILS))
1308 if (ret)
1310 fprintf (dump_file, " result:\n");
1311 if (!integer_nonzerop (niter->assumptions))
1313 fprintf (dump_file, " under assumptions ");
1314 print_generic_expr (dump_file, niter->assumptions, TDF_SLIM);
1315 fprintf (dump_file, "\n");
1318 if (!integer_zerop (niter->may_be_zero))
1320 fprintf (dump_file, " zero if ");
1321 print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM);
1322 fprintf (dump_file, "\n");
1325 fprintf (dump_file, " # of iterations ");
1326 print_generic_expr (dump_file, niter->niter, TDF_SLIM);
1327 fprintf (dump_file, ", bounded by ");
1328 dump_double_int (dump_file, niter->max, true);
1329 fprintf (dump_file, "\n");
1331 else
1332 fprintf (dump_file, " failed\n\n");
1334 return ret;
1337 /* Substitute NEW for OLD in EXPR and fold the result. */
1339 static tree
1340 simplify_replace_tree (tree expr, tree old, tree new_tree)
1342 unsigned i, n;
1343 tree ret = NULL_TREE, e, se;
1345 if (!expr)
1346 return NULL_TREE;
1348 if (expr == old
1349 || operand_equal_p (expr, old, 0))
1350 return unshare_expr (new_tree);
1352 if (!EXPR_P (expr) && !GIMPLE_STMT_P (expr))
1353 return expr;
1355 n = TREE_OPERAND_LENGTH (expr);
1356 for (i = 0; i < n; i++)
1358 e = TREE_OPERAND (expr, i);
1359 se = simplify_replace_tree (e, old, new_tree);
1360 if (e == se)
1361 continue;
1363 if (!ret)
1364 ret = copy_node (expr);
1366 TREE_OPERAND (ret, i) = se;
1369 return (ret ? fold (ret) : expr);
1372 /* Expand definitions of ssa names in EXPR as long as they are simple
1373 enough, and return the new expression. */
1375 tree
1376 expand_simple_operations (tree expr)
1378 unsigned i, n;
1379 tree ret = NULL_TREE, e, ee, stmt;
1380 enum tree_code code;
1382 if (expr == NULL_TREE)
1383 return expr;
1385 if (is_gimple_min_invariant (expr))
1386 return expr;
1388 code = TREE_CODE (expr);
1389 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code)))
1391 n = TREE_OPERAND_LENGTH (expr);
1392 for (i = 0; i < n; i++)
1394 e = TREE_OPERAND (expr, i);
1395 ee = expand_simple_operations (e);
1396 if (e == ee)
1397 continue;
1399 if (!ret)
1400 ret = copy_node (expr);
1402 TREE_OPERAND (ret, i) = ee;
1405 if (!ret)
1406 return expr;
1408 fold_defer_overflow_warnings ();
1409 ret = fold (ret);
1410 fold_undefer_and_ignore_overflow_warnings ();
1411 return ret;
1414 if (TREE_CODE (expr) != SSA_NAME)
1415 return expr;
1417 stmt = SSA_NAME_DEF_STMT (expr);
1418 if (TREE_CODE (stmt) == PHI_NODE)
1420 basic_block src, dest;
1422 if (PHI_NUM_ARGS (stmt) != 1)
1423 return expr;
1424 e = PHI_ARG_DEF (stmt, 0);
1426 /* Avoid propagating through loop exit phi nodes, which
1427 could break loop-closed SSA form restrictions. */
1428 dest = bb_for_stmt (stmt);
1429 src = single_pred (dest);
1430 if (TREE_CODE (e) == SSA_NAME
1431 && src->loop_father != dest->loop_father)
1432 return expr;
1434 return expand_simple_operations (e);
1436 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
1437 return expr;
1439 e = GIMPLE_STMT_OPERAND (stmt, 1);
1440 if (/* Casts are simple. */
1441 !CONVERT_EXPR_P (e)
1442 /* Copies are simple. */
1443 && TREE_CODE (e) != SSA_NAME
1444 /* Assignments of invariants are simple. */
1445 && !is_gimple_min_invariant (e)
1446 /* And increments and decrements by a constant are simple. */
1447 && !((TREE_CODE (e) == PLUS_EXPR
1448 || TREE_CODE (e) == MINUS_EXPR
1449 || TREE_CODE (e) == POINTER_PLUS_EXPR)
1450 && is_gimple_min_invariant (TREE_OPERAND (e, 1))))
1451 return expr;
1453 return expand_simple_operations (e);
1456 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1457 expression (or EXPR unchanged, if no simplification was possible). */
1459 static tree
1460 tree_simplify_using_condition_1 (tree cond, tree expr)
1462 bool changed;
1463 tree e, te, e0, e1, e2, notcond;
1464 enum tree_code code = TREE_CODE (expr);
1466 if (code == INTEGER_CST)
1467 return expr;
1469 if (code == TRUTH_OR_EXPR
1470 || code == TRUTH_AND_EXPR
1471 || code == COND_EXPR)
1473 changed = false;
1475 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0));
1476 if (TREE_OPERAND (expr, 0) != e0)
1477 changed = true;
1479 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1));
1480 if (TREE_OPERAND (expr, 1) != e1)
1481 changed = true;
1483 if (code == COND_EXPR)
1485 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2));
1486 if (TREE_OPERAND (expr, 2) != e2)
1487 changed = true;
1489 else
1490 e2 = NULL_TREE;
1492 if (changed)
1494 if (code == COND_EXPR)
1495 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1496 else
1497 expr = fold_build2 (code, boolean_type_node, e0, e1);
1500 return expr;
1503 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
1504 propagation, and vice versa. Fold does not handle this, since it is
1505 considered too expensive. */
1506 if (TREE_CODE (cond) == EQ_EXPR)
1508 e0 = TREE_OPERAND (cond, 0);
1509 e1 = TREE_OPERAND (cond, 1);
1511 /* We know that e0 == e1. Check whether we cannot simplify expr
1512 using this fact. */
1513 e = simplify_replace_tree (expr, e0, e1);
1514 if (integer_zerop (e) || integer_nonzerop (e))
1515 return e;
1517 e = simplify_replace_tree (expr, e1, e0);
1518 if (integer_zerop (e) || integer_nonzerop (e))
1519 return e;
1521 if (TREE_CODE (expr) == EQ_EXPR)
1523 e0 = TREE_OPERAND (expr, 0);
1524 e1 = TREE_OPERAND (expr, 1);
1526 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
1527 e = simplify_replace_tree (cond, e0, e1);
1528 if (integer_zerop (e))
1529 return e;
1530 e = simplify_replace_tree (cond, e1, e0);
1531 if (integer_zerop (e))
1532 return e;
1534 if (TREE_CODE (expr) == NE_EXPR)
1536 e0 = TREE_OPERAND (expr, 0);
1537 e1 = TREE_OPERAND (expr, 1);
1539 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
1540 e = simplify_replace_tree (cond, e0, e1);
1541 if (integer_zerop (e))
1542 return boolean_true_node;
1543 e = simplify_replace_tree (cond, e1, e0);
1544 if (integer_zerop (e))
1545 return boolean_true_node;
1548 te = expand_simple_operations (expr);
1550 /* Check whether COND ==> EXPR. */
1551 notcond = invert_truthvalue (cond);
1552 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, te);
1553 if (e && integer_nonzerop (e))
1554 return e;
1556 /* Check whether COND ==> not EXPR. */
1557 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, te);
1558 if (e && integer_zerop (e))
1559 return e;
1561 return expr;
1564 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1565 expression (or EXPR unchanged, if no simplification was possible).
1566 Wrapper around tree_simplify_using_condition_1 that ensures that chains
1567 of simple operations in definitions of ssa names in COND are expanded,
1568 so that things like casts or incrementing the value of the bound before
1569 the loop do not cause us to fail. */
1571 static tree
1572 tree_simplify_using_condition (tree cond, tree expr)
1574 cond = expand_simple_operations (cond);
1576 return tree_simplify_using_condition_1 (cond, expr);
1579 /* Tries to simplify EXPR using the conditions on entry to LOOP.
1580 Returns the simplified expression (or EXPR unchanged, if no
1581 simplification was possible).*/
1583 static tree
1584 simplify_using_initial_conditions (struct loop *loop, tree expr)
1586 edge e;
1587 basic_block bb;
1588 tree cond;
1589 int cnt = 0;
1591 if (TREE_CODE (expr) == INTEGER_CST)
1592 return expr;
1594 /* Limit walking the dominators to avoid quadraticness in
1595 the number of BBs times the number of loops in degenerate
1596 cases. */
1597 for (bb = loop->header;
1598 bb != ENTRY_BLOCK_PTR && cnt < MAX_DOMINATORS_TO_WALK;
1599 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
1601 if (!single_pred_p (bb))
1602 continue;
1603 e = single_pred_edge (bb);
1605 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
1606 continue;
1608 cond = COND_EXPR_COND (last_stmt (e->src));
1609 if (e->flags & EDGE_FALSE_VALUE)
1610 cond = invert_truthvalue (cond);
1611 expr = tree_simplify_using_condition (cond, expr);
1612 ++cnt;
1615 return expr;
1618 /* Tries to simplify EXPR using the evolutions of the loop invariants
1619 in the superloops of LOOP. Returns the simplified expression
1620 (or EXPR unchanged, if no simplification was possible). */
1622 static tree
1623 simplify_using_outer_evolutions (struct loop *loop, tree expr)
1625 enum tree_code code = TREE_CODE (expr);
1626 bool changed;
1627 tree e, e0, e1, e2;
1629 if (is_gimple_min_invariant (expr))
1630 return expr;
1632 if (code == TRUTH_OR_EXPR
1633 || code == TRUTH_AND_EXPR
1634 || code == COND_EXPR)
1636 changed = false;
1638 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0));
1639 if (TREE_OPERAND (expr, 0) != e0)
1640 changed = true;
1642 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1));
1643 if (TREE_OPERAND (expr, 1) != e1)
1644 changed = true;
1646 if (code == COND_EXPR)
1648 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2));
1649 if (TREE_OPERAND (expr, 2) != e2)
1650 changed = true;
1652 else
1653 e2 = NULL_TREE;
1655 if (changed)
1657 if (code == COND_EXPR)
1658 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1659 else
1660 expr = fold_build2 (code, boolean_type_node, e0, e1);
1663 return expr;
1666 e = instantiate_parameters (loop, expr);
1667 if (is_gimple_min_invariant (e))
1668 return e;
1670 return expr;
1673 /* Returns true if EXIT is the only possible exit from LOOP. */
1675 bool
1676 loop_only_exit_p (const struct loop *loop, const_edge exit)
1678 basic_block *body;
1679 block_stmt_iterator bsi;
1680 unsigned i;
1681 tree call;
1683 if (exit != single_exit (loop))
1684 return false;
1686 body = get_loop_body (loop);
1687 for (i = 0; i < loop->num_nodes; i++)
1689 for (bsi = bsi_start (body[0]); !bsi_end_p (bsi); bsi_next (&bsi))
1691 call = get_call_expr_in (bsi_stmt (bsi));
1692 if (call && TREE_SIDE_EFFECTS (call))
1694 free (body);
1695 return false;
1700 free (body);
1701 return true;
1704 /* Stores description of number of iterations of LOOP derived from
1705 EXIT (an exit edge of the LOOP) in NITER. Returns true if some
1706 useful information could be derived (and fields of NITER has
1707 meaning described in comments at struct tree_niter_desc
1708 declaration), false otherwise. If WARN is true and
1709 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use
1710 potentially unsafe assumptions. */
1712 bool
1713 number_of_iterations_exit (struct loop *loop, edge exit,
1714 struct tree_niter_desc *niter,
1715 bool warn)
1717 tree stmt, cond, type;
1718 tree op0, op1;
1719 enum tree_code code;
1720 affine_iv iv0, iv1;
1722 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src))
1723 return false;
1725 niter->assumptions = boolean_false_node;
1726 stmt = last_stmt (exit->src);
1727 if (!stmt || TREE_CODE (stmt) != COND_EXPR)
1728 return false;
1730 /* We want the condition for staying inside loop. */
1731 cond = COND_EXPR_COND (stmt);
1732 if (exit->flags & EDGE_TRUE_VALUE)
1733 cond = invert_truthvalue (cond);
1735 code = TREE_CODE (cond);
1736 switch (code)
1738 case GT_EXPR:
1739 case GE_EXPR:
1740 case NE_EXPR:
1741 case LT_EXPR:
1742 case LE_EXPR:
1743 break;
1745 default:
1746 return false;
1749 op0 = TREE_OPERAND (cond, 0);
1750 op1 = TREE_OPERAND (cond, 1);
1751 type = TREE_TYPE (op0);
1753 if (TREE_CODE (type) != INTEGER_TYPE
1754 && !POINTER_TYPE_P (type))
1755 return false;
1757 if (!simple_iv (loop, stmt, op0, &iv0, false))
1758 return false;
1759 if (!simple_iv (loop, stmt, op1, &iv1, false))
1760 return false;
1762 /* We don't want to see undefined signed overflow warnings while
1763 computing the number of iterations. */
1764 fold_defer_overflow_warnings ();
1766 iv0.base = expand_simple_operations (iv0.base);
1767 iv1.base = expand_simple_operations (iv1.base);
1768 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter,
1769 loop_only_exit_p (loop, exit)))
1771 fold_undefer_and_ignore_overflow_warnings ();
1772 return false;
1775 if (optimize >= 3)
1777 niter->assumptions = simplify_using_outer_evolutions (loop,
1778 niter->assumptions);
1779 niter->may_be_zero = simplify_using_outer_evolutions (loop,
1780 niter->may_be_zero);
1781 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
1784 niter->assumptions
1785 = simplify_using_initial_conditions (loop,
1786 niter->assumptions);
1787 niter->may_be_zero
1788 = simplify_using_initial_conditions (loop,
1789 niter->may_be_zero);
1791 fold_undefer_and_ignore_overflow_warnings ();
1793 if (integer_onep (niter->assumptions))
1794 return true;
1796 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
1797 But if we can prove that there is overflow or some other source of weird
1798 behavior, ignore the loop even with -funsafe-loop-optimizations. */
1799 if (integer_zerop (niter->assumptions))
1800 return false;
1802 if (flag_unsafe_loop_optimizations)
1803 niter->assumptions = boolean_true_node;
1805 if (warn)
1807 const char *wording;
1808 location_t loc = EXPR_LOCATION (stmt);
1810 /* We can provide a more specific warning if one of the operator is
1811 constant and the other advances by +1 or -1. */
1812 if (!integer_zerop (iv1.step)
1813 ? (integer_zerop (iv0.step)
1814 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step)))
1815 : (integer_onep (iv0.step) || integer_all_onesp (iv0.step)))
1816 wording =
1817 flag_unsafe_loop_optimizations
1818 ? N_("assuming that the loop is not infinite")
1819 : N_("cannot optimize possibly infinite loops");
1820 else
1821 wording =
1822 flag_unsafe_loop_optimizations
1823 ? N_("assuming that the loop counter does not overflow")
1824 : N_("cannot optimize loop, the loop counter may overflow");
1826 if (LOCATION_LINE (loc) > 0)
1827 warning (OPT_Wunsafe_loop_optimizations, "%H%s", &loc, gettext (wording));
1828 else
1829 warning (OPT_Wunsafe_loop_optimizations, "%s", gettext (wording));
1832 return flag_unsafe_loop_optimizations;
1835 /* Try to determine the number of iterations of LOOP. If we succeed,
1836 expression giving number of iterations is returned and *EXIT is
1837 set to the edge from that the information is obtained. Otherwise
1838 chrec_dont_know is returned. */
1840 tree
1841 find_loop_niter (struct loop *loop, edge *exit)
1843 unsigned i;
1844 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1845 edge ex;
1846 tree niter = NULL_TREE, aniter;
1847 struct tree_niter_desc desc;
1849 *exit = NULL;
1850 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
1852 if (!just_once_each_iteration_p (loop, ex->src))
1853 continue;
1855 if (!number_of_iterations_exit (loop, ex, &desc, false))
1856 continue;
1858 if (integer_nonzerop (desc.may_be_zero))
1860 /* We exit in the first iteration through this exit.
1861 We won't find anything better. */
1862 niter = build_int_cst (unsigned_type_node, 0);
1863 *exit = ex;
1864 break;
1867 if (!integer_zerop (desc.may_be_zero))
1868 continue;
1870 aniter = desc.niter;
1872 if (!niter)
1874 /* Nothing recorded yet. */
1875 niter = aniter;
1876 *exit = ex;
1877 continue;
1880 /* Prefer constants, the lower the better. */
1881 if (TREE_CODE (aniter) != INTEGER_CST)
1882 continue;
1884 if (TREE_CODE (niter) != INTEGER_CST)
1886 niter = aniter;
1887 *exit = ex;
1888 continue;
1891 if (tree_int_cst_lt (aniter, niter))
1893 niter = aniter;
1894 *exit = ex;
1895 continue;
1898 VEC_free (edge, heap, exits);
1900 return niter ? niter : chrec_dont_know;
1905 Analysis of a number of iterations of a loop by a brute-force evaluation.
1909 /* Bound on the number of iterations we try to evaluate. */
1911 #define MAX_ITERATIONS_TO_TRACK \
1912 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
1914 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
1915 result by a chain of operations such that all but exactly one of their
1916 operands are constants. */
1918 static tree
1919 chain_of_csts_start (struct loop *loop, tree x)
1921 tree stmt = SSA_NAME_DEF_STMT (x);
1922 tree use;
1923 basic_block bb = bb_for_stmt (stmt);
1925 if (!bb
1926 || !flow_bb_inside_loop_p (loop, bb))
1927 return NULL_TREE;
1929 if (TREE_CODE (stmt) == PHI_NODE)
1931 if (bb == loop->header)
1932 return stmt;
1934 return NULL_TREE;
1937 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
1938 return NULL_TREE;
1940 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
1941 return NULL_TREE;
1942 if (SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF) == NULL_DEF_OPERAND_P)
1943 return NULL_TREE;
1945 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
1946 if (use == NULL_USE_OPERAND_P)
1947 return NULL_TREE;
1949 return chain_of_csts_start (loop, use);
1952 /* Determines whether the expression X is derived from a result of a phi node
1953 in header of LOOP such that
1955 * the derivation of X consists only from operations with constants
1956 * the initial value of the phi node is constant
1957 * the value of the phi node in the next iteration can be derived from the
1958 value in the current iteration by a chain of operations with constants.
1960 If such phi node exists, it is returned. If X is a constant, X is returned
1961 unchanged. Otherwise NULL_TREE is returned. */
1963 static tree
1964 get_base_for (struct loop *loop, tree x)
1966 tree phi, init, next;
1968 if (is_gimple_min_invariant (x))
1969 return x;
1971 phi = chain_of_csts_start (loop, x);
1972 if (!phi)
1973 return NULL_TREE;
1975 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
1976 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
1978 if (TREE_CODE (next) != SSA_NAME)
1979 return NULL_TREE;
1981 if (!is_gimple_min_invariant (init))
1982 return NULL_TREE;
1984 if (chain_of_csts_start (loop, next) != phi)
1985 return NULL_TREE;
1987 return phi;
1990 /* Given an expression X, then
1992 * if X is NULL_TREE, we return the constant BASE.
1993 * otherwise X is a SSA name, whose value in the considered loop is derived
1994 by a chain of operations with constant from a result of a phi node in
1995 the header of the loop. Then we return value of X when the value of the
1996 result of this phi node is given by the constant BASE. */
1998 static tree
1999 get_val_for (tree x, tree base)
2001 tree stmt, nx, val;
2002 use_operand_p op;
2003 ssa_op_iter iter;
2005 gcc_assert (is_gimple_min_invariant (base));
2007 if (!x)
2008 return base;
2010 stmt = SSA_NAME_DEF_STMT (x);
2011 if (TREE_CODE (stmt) == PHI_NODE)
2012 return base;
2014 FOR_EACH_SSA_USE_OPERAND (op, stmt, iter, SSA_OP_USE)
2016 nx = USE_FROM_PTR (op);
2017 val = get_val_for (nx, base);
2018 SET_USE (op, val);
2019 val = fold (GIMPLE_STMT_OPERAND (stmt, 1));
2020 SET_USE (op, nx);
2021 /* only iterate loop once. */
2022 return val;
2025 /* Should never reach here. */
2026 gcc_unreachable ();
2029 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2030 by brute force -- i.e. by determining the value of the operands of the
2031 condition at EXIT in first few iterations of the loop (assuming that
2032 these values are constant) and determining the first one in that the
2033 condition is not satisfied. Returns the constant giving the number
2034 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2036 tree
2037 loop_niter_by_eval (struct loop *loop, edge exit)
2039 tree cond, cnd, acnd;
2040 tree op[2], val[2], next[2], aval[2], phi[2];
2041 unsigned i, j;
2042 enum tree_code cmp;
2044 cond = last_stmt (exit->src);
2045 if (!cond || TREE_CODE (cond) != COND_EXPR)
2046 return chrec_dont_know;
2048 cnd = COND_EXPR_COND (cond);
2049 if (exit->flags & EDGE_TRUE_VALUE)
2050 cnd = invert_truthvalue (cnd);
2052 cmp = TREE_CODE (cnd);
2053 switch (cmp)
2055 case EQ_EXPR:
2056 case NE_EXPR:
2057 case GT_EXPR:
2058 case GE_EXPR:
2059 case LT_EXPR:
2060 case LE_EXPR:
2061 for (j = 0; j < 2; j++)
2062 op[j] = TREE_OPERAND (cnd, j);
2063 break;
2065 default:
2066 return chrec_dont_know;
2069 for (j = 0; j < 2; j++)
2071 phi[j] = get_base_for (loop, op[j]);
2072 if (!phi[j])
2073 return chrec_dont_know;
2076 for (j = 0; j < 2; j++)
2078 if (TREE_CODE (phi[j]) == PHI_NODE)
2080 val[j] = PHI_ARG_DEF_FROM_EDGE (phi[j], loop_preheader_edge (loop));
2081 next[j] = PHI_ARG_DEF_FROM_EDGE (phi[j], loop_latch_edge (loop));
2083 else
2085 val[j] = phi[j];
2086 next[j] = NULL_TREE;
2087 op[j] = NULL_TREE;
2091 /* Don't issue signed overflow warnings. */
2092 fold_defer_overflow_warnings ();
2094 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
2096 for (j = 0; j < 2; j++)
2097 aval[j] = get_val_for (op[j], val[j]);
2099 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
2100 if (acnd && integer_zerop (acnd))
2102 fold_undefer_and_ignore_overflow_warnings ();
2103 if (dump_file && (dump_flags & TDF_DETAILS))
2104 fprintf (dump_file,
2105 "Proved that loop %d iterates %d times using brute force.\n",
2106 loop->num, i);
2107 return build_int_cst (unsigned_type_node, i);
2110 for (j = 0; j < 2; j++)
2112 val[j] = get_val_for (next[j], val[j]);
2113 if (!is_gimple_min_invariant (val[j]))
2115 fold_undefer_and_ignore_overflow_warnings ();
2116 return chrec_dont_know;
2121 fold_undefer_and_ignore_overflow_warnings ();
2123 return chrec_dont_know;
2126 /* Finds the exit of the LOOP by that the loop exits after a constant
2127 number of iterations and stores the exit edge to *EXIT. The constant
2128 giving the number of iterations of LOOP is returned. The number of
2129 iterations is determined using loop_niter_by_eval (i.e. by brute force
2130 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2131 determines the number of iterations, chrec_dont_know is returned. */
2133 tree
2134 find_loop_niter_by_eval (struct loop *loop, edge *exit)
2136 unsigned i;
2137 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
2138 edge ex;
2139 tree niter = NULL_TREE, aniter;
2141 *exit = NULL;
2142 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
2144 if (!just_once_each_iteration_p (loop, ex->src))
2145 continue;
2147 aniter = loop_niter_by_eval (loop, ex);
2148 if (chrec_contains_undetermined (aniter))
2149 continue;
2151 if (niter
2152 && !tree_int_cst_lt (aniter, niter))
2153 continue;
2155 niter = aniter;
2156 *exit = ex;
2158 VEC_free (edge, heap, exits);
2160 return niter ? niter : chrec_dont_know;
2165 Analysis of upper bounds on number of iterations of a loop.
2169 /* Returns a constant upper bound on the value of expression VAL. VAL
2170 is considered to be unsigned. If its type is signed, its value must
2171 be nonnegative. */
2173 static double_int
2174 derive_constant_upper_bound (const_tree val)
2176 tree type = TREE_TYPE (val);
2177 tree op0, op1, subtype, maxt;
2178 double_int bnd, max, mmax, cst;
2179 tree stmt;
2181 if (INTEGRAL_TYPE_P (type))
2182 maxt = TYPE_MAX_VALUE (type);
2183 else
2184 maxt = upper_bound_in_type (type, type);
2186 max = tree_to_double_int (maxt);
2188 switch (TREE_CODE (val))
2190 case INTEGER_CST:
2191 return tree_to_double_int (val);
2193 CASE_CONVERT:
2194 op0 = TREE_OPERAND (val, 0);
2195 subtype = TREE_TYPE (op0);
2196 if (!TYPE_UNSIGNED (subtype)
2197 /* If TYPE is also signed, the fact that VAL is nonnegative implies
2198 that OP0 is nonnegative. */
2199 && TYPE_UNSIGNED (type)
2200 && !tree_expr_nonnegative_p (op0))
2202 /* If we cannot prove that the casted expression is nonnegative,
2203 we cannot establish more useful upper bound than the precision
2204 of the type gives us. */
2205 return max;
2208 /* We now know that op0 is an nonnegative value. Try deriving an upper
2209 bound for it. */
2210 bnd = derive_constant_upper_bound (op0);
2212 /* If the bound does not fit in TYPE, max. value of TYPE could be
2213 attained. */
2214 if (double_int_ucmp (max, bnd) < 0)
2215 return max;
2217 return bnd;
2219 case PLUS_EXPR:
2220 case POINTER_PLUS_EXPR:
2221 case MINUS_EXPR:
2222 op0 = TREE_OPERAND (val, 0);
2223 op1 = TREE_OPERAND (val, 1);
2225 if (TREE_CODE (op1) != INTEGER_CST
2226 || !tree_expr_nonnegative_p (op0))
2227 return max;
2229 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
2230 choose the most logical way how to treat this constant regardless
2231 of the signedness of the type. */
2232 cst = tree_to_double_int (op1);
2233 cst = double_int_sext (cst, TYPE_PRECISION (type));
2234 if (TREE_CODE (val) == PLUS_EXPR)
2235 cst = double_int_neg (cst);
2237 bnd = derive_constant_upper_bound (op0);
2239 if (double_int_negative_p (cst))
2241 cst = double_int_neg (cst);
2242 /* Avoid CST == 0x80000... */
2243 if (double_int_negative_p (cst))
2244 return max;;
2246 /* OP0 + CST. We need to check that
2247 BND <= MAX (type) - CST. */
2249 mmax = double_int_add (max, double_int_neg (cst));
2250 if (double_int_ucmp (bnd, mmax) > 0)
2251 return max;
2253 return double_int_add (bnd, cst);
2255 else
2257 /* OP0 - CST, where CST >= 0.
2259 If TYPE is signed, we have already verified that OP0 >= 0, and we
2260 know that the result is nonnegative. This implies that
2261 VAL <= BND - CST.
2263 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2264 otherwise the operation underflows.
2267 /* This should only happen if the type is unsigned; however, for
2268 buggy programs that use overflowing signed arithmetics even with
2269 -fno-wrapv, this condition may also be true for signed values. */
2270 if (double_int_ucmp (bnd, cst) < 0)
2271 return max;
2273 if (TYPE_UNSIGNED (type))
2275 tree tem = fold_binary (GE_EXPR, boolean_type_node, op0,
2276 double_int_to_tree (type, cst));
2277 if (!tem || integer_nonzerop (tem))
2278 return max;
2281 bnd = double_int_add (bnd, double_int_neg (cst));
2284 return bnd;
2286 case FLOOR_DIV_EXPR:
2287 case EXACT_DIV_EXPR:
2288 op0 = TREE_OPERAND (val, 0);
2289 op1 = TREE_OPERAND (val, 1);
2290 if (TREE_CODE (op1) != INTEGER_CST
2291 || tree_int_cst_sign_bit (op1))
2292 return max;
2294 bnd = derive_constant_upper_bound (op0);
2295 return double_int_udiv (bnd, tree_to_double_int (op1), FLOOR_DIV_EXPR);
2297 case BIT_AND_EXPR:
2298 op1 = TREE_OPERAND (val, 1);
2299 if (TREE_CODE (op1) != INTEGER_CST
2300 || tree_int_cst_sign_bit (op1))
2301 return max;
2302 return tree_to_double_int (op1);
2304 case SSA_NAME:
2305 stmt = SSA_NAME_DEF_STMT (val);
2306 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT
2307 || GIMPLE_STMT_OPERAND (stmt, 0) != val)
2308 return max;
2309 return derive_constant_upper_bound (GIMPLE_STMT_OPERAND (stmt, 1));
2311 default:
2312 return max;
2316 /* Records that every statement in LOOP is executed I_BOUND times.
2317 REALISTIC is true if I_BOUND is expected to be close to the real number
2318 of iterations. UPPER is true if we are sure the loop iterates at most
2319 I_BOUND times. */
2321 static void
2322 record_niter_bound (struct loop *loop, double_int i_bound, bool realistic,
2323 bool upper)
2325 /* Update the bounds only when there is no previous estimation, or when the current
2326 estimation is smaller. */
2327 if (upper
2328 && (!loop->any_upper_bound
2329 || double_int_ucmp (i_bound, loop->nb_iterations_upper_bound) < 0))
2331 loop->any_upper_bound = true;
2332 loop->nb_iterations_upper_bound = i_bound;
2334 if (realistic
2335 && (!loop->any_estimate
2336 || double_int_ucmp (i_bound, loop->nb_iterations_estimate) < 0))
2338 loop->any_estimate = true;
2339 loop->nb_iterations_estimate = i_bound;
2343 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
2344 is true if the loop is exited immediately after STMT, and this exit
2345 is taken at last when the STMT is executed BOUND + 1 times.
2346 REALISTIC is true if BOUND is expected to be close to the real number
2347 of iterations. UPPER is true if we are sure the loop iterates at most
2348 BOUND times. I_BOUND is an unsigned double_int upper estimate on BOUND. */
2350 static void
2351 record_estimate (struct loop *loop, tree bound, double_int i_bound,
2352 tree at_stmt, bool is_exit, bool realistic, bool upper)
2354 double_int delta;
2355 edge exit;
2357 if (dump_file && (dump_flags & TDF_DETAILS))
2359 fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : "");
2360 print_generic_expr (dump_file, at_stmt, TDF_SLIM);
2361 fprintf (dump_file, " is %sexecuted at most ",
2362 upper ? "" : "probably ");
2363 print_generic_expr (dump_file, bound, TDF_SLIM);
2364 fprintf (dump_file, " (bounded by ");
2365 dump_double_int (dump_file, i_bound, true);
2366 fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num);
2369 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
2370 real number of iterations. */
2371 if (TREE_CODE (bound) != INTEGER_CST)
2372 realistic = false;
2373 if (!upper && !realistic)
2374 return;
2376 /* If we have a guaranteed upper bound, record it in the appropriate
2377 list. */
2378 if (upper)
2380 struct nb_iter_bound *elt = GGC_NEW (struct nb_iter_bound);
2382 elt->bound = i_bound;
2383 elt->stmt = at_stmt;
2384 elt->is_exit = is_exit;
2385 elt->next = loop->bounds;
2386 loop->bounds = elt;
2389 /* Update the number of iteration estimates according to the bound.
2390 If at_stmt is an exit, then every statement in the loop is
2391 executed at most BOUND + 1 times. If it is not an exit, then
2392 some of the statements before it could be executed BOUND + 2
2393 times, if an exit of LOOP is before stmt. */
2394 exit = single_exit (loop);
2395 if (is_exit
2396 || (exit != NULL
2397 && dominated_by_p (CDI_DOMINATORS,
2398 exit->src, bb_for_stmt (at_stmt))))
2399 delta = double_int_one;
2400 else
2401 delta = double_int_two;
2402 i_bound = double_int_add (i_bound, delta);
2404 /* If an overflow occurred, ignore the result. */
2405 if (double_int_ucmp (i_bound, delta) < 0)
2406 return;
2408 record_niter_bound (loop, i_bound, realistic, upper);
2411 /* Record the estimate on number of iterations of LOOP based on the fact that
2412 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
2413 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
2414 estimated number of iterations is expected to be close to the real one.
2415 UPPER is true if we are sure the induction variable does not wrap. */
2417 static void
2418 record_nonwrapping_iv (struct loop *loop, tree base, tree step, tree stmt,
2419 tree low, tree high, bool realistic, bool upper)
2421 tree niter_bound, extreme, delta;
2422 tree type = TREE_TYPE (base), unsigned_type;
2423 double_int max;
2425 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step))
2426 return;
2428 if (dump_file && (dump_flags & TDF_DETAILS))
2430 fprintf (dump_file, "Induction variable (");
2431 print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM);
2432 fprintf (dump_file, ") ");
2433 print_generic_expr (dump_file, base, TDF_SLIM);
2434 fprintf (dump_file, " + ");
2435 print_generic_expr (dump_file, step, TDF_SLIM);
2436 fprintf (dump_file, " * iteration does not wrap in statement ");
2437 print_generic_expr (dump_file, stmt, TDF_SLIM);
2438 fprintf (dump_file, " in loop %d.\n", loop->num);
2441 unsigned_type = unsigned_type_for (type);
2442 base = fold_convert (unsigned_type, base);
2443 step = fold_convert (unsigned_type, step);
2445 if (tree_int_cst_sign_bit (step))
2447 extreme = fold_convert (unsigned_type, low);
2448 if (TREE_CODE (base) != INTEGER_CST)
2449 base = fold_convert (unsigned_type, high);
2450 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
2451 step = fold_build1 (NEGATE_EXPR, unsigned_type, step);
2453 else
2455 extreme = fold_convert (unsigned_type, high);
2456 if (TREE_CODE (base) != INTEGER_CST)
2457 base = fold_convert (unsigned_type, low);
2458 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
2461 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
2462 would get out of the range. */
2463 niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step);
2464 max = derive_constant_upper_bound (niter_bound);
2465 record_estimate (loop, niter_bound, max, stmt, false, realistic, upper);
2468 /* Returns true if REF is a reference to an array at the end of a dynamically
2469 allocated structure. If this is the case, the array may be allocated larger
2470 than its upper bound implies. */
2472 static bool
2473 array_at_struct_end_p (tree ref)
2475 tree base = get_base_address (ref);
2476 tree parent, field;
2478 /* Unless the reference is through a pointer, the size of the array matches
2479 its declaration. */
2480 if (!base || !INDIRECT_REF_P (base))
2481 return false;
2483 for (;handled_component_p (ref); ref = parent)
2485 parent = TREE_OPERAND (ref, 0);
2487 if (TREE_CODE (ref) == COMPONENT_REF)
2489 /* All fields of a union are at its end. */
2490 if (TREE_CODE (TREE_TYPE (parent)) == UNION_TYPE)
2491 continue;
2493 /* Unless the field is at the end of the struct, we are done. */
2494 field = TREE_OPERAND (ref, 1);
2495 if (TREE_CHAIN (field))
2496 return false;
2499 /* The other options are ARRAY_REF, ARRAY_RANGE_REF, VIEW_CONVERT_EXPR.
2500 In all these cases, we might be accessing the last element, and
2501 although in practice this will probably never happen, it is legal for
2502 the indices of this last element to exceed the bounds of the array.
2503 Therefore, continue checking. */
2506 gcc_assert (INDIRECT_REF_P (ref));
2507 return true;
2510 /* Determine information about number of iterations a LOOP from the index
2511 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
2512 guaranteed to be executed in every iteration of LOOP. Callback for
2513 for_each_index. */
2515 struct ilb_data
2517 struct loop *loop;
2518 tree stmt;
2519 bool reliable;
2522 static bool
2523 idx_infer_loop_bounds (tree base, tree *idx, void *dta)
2525 struct ilb_data *data = (struct ilb_data *) dta;
2526 tree ev, init, step;
2527 tree low, high, type, next;
2528 bool sign, upper = data->reliable, at_end = false;
2529 struct loop *loop = data->loop;
2531 if (TREE_CODE (base) != ARRAY_REF)
2532 return true;
2534 /* For arrays at the end of the structure, we are not guaranteed that they
2535 do not really extend over their declared size. However, for arrays of
2536 size greater than one, this is unlikely to be intended. */
2537 if (array_at_struct_end_p (base))
2539 at_end = true;
2540 upper = false;
2543 ev = instantiate_parameters (loop, analyze_scalar_evolution (loop, *idx));
2544 init = initial_condition (ev);
2545 step = evolution_part_in_loop_num (ev, loop->num);
2547 if (!init
2548 || !step
2549 || TREE_CODE (step) != INTEGER_CST
2550 || integer_zerop (step)
2551 || tree_contains_chrecs (init, NULL)
2552 || chrec_contains_symbols_defined_in_loop (init, loop->num))
2553 return true;
2555 low = array_ref_low_bound (base);
2556 high = array_ref_up_bound (base);
2558 /* The case of nonconstant bounds could be handled, but it would be
2559 complicated. */
2560 if (TREE_CODE (low) != INTEGER_CST
2561 || !high
2562 || TREE_CODE (high) != INTEGER_CST)
2563 return true;
2564 sign = tree_int_cst_sign_bit (step);
2565 type = TREE_TYPE (step);
2567 /* The array of length 1 at the end of a structure most likely extends
2568 beyond its bounds. */
2569 if (at_end
2570 && operand_equal_p (low, high, 0))
2571 return true;
2573 /* In case the relevant bound of the array does not fit in type, or
2574 it does, but bound + step (in type) still belongs into the range of the
2575 array, the index may wrap and still stay within the range of the array
2576 (consider e.g. if the array is indexed by the full range of
2577 unsigned char).
2579 To make things simpler, we require both bounds to fit into type, although
2580 there are cases where this would not be strictly necessary. */
2581 if (!int_fits_type_p (high, type)
2582 || !int_fits_type_p (low, type))
2583 return true;
2584 low = fold_convert (type, low);
2585 high = fold_convert (type, high);
2587 if (sign)
2588 next = fold_binary (PLUS_EXPR, type, low, step);
2589 else
2590 next = fold_binary (PLUS_EXPR, type, high, step);
2592 if (tree_int_cst_compare (low, next) <= 0
2593 && tree_int_cst_compare (next, high) <= 0)
2594 return true;
2596 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, true, upper);
2597 return true;
2600 /* Determine information about number of iterations a LOOP from the bounds
2601 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
2602 STMT is guaranteed to be executed in every iteration of LOOP.*/
2604 static void
2605 infer_loop_bounds_from_ref (struct loop *loop, tree stmt, tree ref,
2606 bool reliable)
2608 struct ilb_data data;
2610 data.loop = loop;
2611 data.stmt = stmt;
2612 data.reliable = reliable;
2613 for_each_index (&ref, idx_infer_loop_bounds, &data);
2616 /* Determine information about number of iterations of a LOOP from the way
2617 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
2618 executed in every iteration of LOOP. */
2620 static void
2621 infer_loop_bounds_from_array (struct loop *loop, tree stmt, bool reliable)
2623 tree call;
2625 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
2627 tree op0 = GIMPLE_STMT_OPERAND (stmt, 0);
2628 tree op1 = GIMPLE_STMT_OPERAND (stmt, 1);
2630 /* For each memory access, analyze its access function
2631 and record a bound on the loop iteration domain. */
2632 if (REFERENCE_CLASS_P (op0))
2633 infer_loop_bounds_from_ref (loop, stmt, op0, reliable);
2635 if (REFERENCE_CLASS_P (op1))
2636 infer_loop_bounds_from_ref (loop, stmt, op1, reliable);
2640 call = get_call_expr_in (stmt);
2641 if (call)
2643 tree arg;
2644 call_expr_arg_iterator iter;
2646 FOR_EACH_CALL_EXPR_ARG (arg, iter, call)
2647 if (REFERENCE_CLASS_P (arg))
2648 infer_loop_bounds_from_ref (loop, stmt, arg, reliable);
2652 /* Determine information about number of iterations of a LOOP from the fact
2653 that signed arithmetics in STMT does not overflow. */
2655 static void
2656 infer_loop_bounds_from_signedness (struct loop *loop, tree stmt)
2658 tree def, base, step, scev, type, low, high;
2660 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
2661 return;
2663 def = GIMPLE_STMT_OPERAND (stmt, 0);
2665 if (TREE_CODE (def) != SSA_NAME)
2666 return;
2668 type = TREE_TYPE (def);
2669 if (!INTEGRAL_TYPE_P (type)
2670 || !TYPE_OVERFLOW_UNDEFINED (type))
2671 return;
2673 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
2674 if (chrec_contains_undetermined (scev))
2675 return;
2677 base = initial_condition_in_loop_num (scev, loop->num);
2678 step = evolution_part_in_loop_num (scev, loop->num);
2680 if (!base || !step
2681 || TREE_CODE (step) != INTEGER_CST
2682 || tree_contains_chrecs (base, NULL)
2683 || chrec_contains_symbols_defined_in_loop (base, loop->num))
2684 return;
2686 low = lower_bound_in_type (type, type);
2687 high = upper_bound_in_type (type, type);
2689 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
2692 /* The following analyzers are extracting informations on the bounds
2693 of LOOP from the following undefined behaviors:
2695 - data references should not access elements over the statically
2696 allocated size,
2698 - signed variables should not overflow when flag_wrapv is not set.
2701 static void
2702 infer_loop_bounds_from_undefined (struct loop *loop)
2704 unsigned i;
2705 basic_block *bbs;
2706 block_stmt_iterator bsi;
2707 basic_block bb;
2708 bool reliable;
2710 bbs = get_loop_body (loop);
2712 for (i = 0; i < loop->num_nodes; i++)
2714 bb = bbs[i];
2716 /* If BB is not executed in each iteration of the loop, we cannot
2717 use the operations in it to infer reliable upper bound on the
2718 # of iterations of the loop. However, we can use it as a guess. */
2719 reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb);
2721 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2723 tree stmt = bsi_stmt (bsi);
2725 infer_loop_bounds_from_array (loop, stmt, reliable);
2727 if (reliable)
2728 infer_loop_bounds_from_signedness (loop, stmt);
2733 free (bbs);
2736 /* Converts VAL to double_int. */
2738 static double_int
2739 gcov_type_to_double_int (gcov_type val)
2741 double_int ret;
2743 ret.low = (unsigned HOST_WIDE_INT) val;
2744 /* If HOST_BITS_PER_WIDE_INT == HOST_BITS_PER_WIDEST_INT, avoid shifting by
2745 the size of type. */
2746 val >>= HOST_BITS_PER_WIDE_INT - 1;
2747 val >>= 1;
2748 ret.high = (unsigned HOST_WIDE_INT) val;
2750 return ret;
2753 /* Records estimates on numbers of iterations of LOOP. */
2755 void
2756 estimate_numbers_of_iterations_loop (struct loop *loop)
2758 VEC (edge, heap) *exits;
2759 tree niter, type;
2760 unsigned i;
2761 struct tree_niter_desc niter_desc;
2762 edge ex;
2763 double_int bound;
2765 /* Give up if we already have tried to compute an estimation. */
2766 if (loop->estimate_state != EST_NOT_COMPUTED)
2767 return;
2768 loop->estimate_state = EST_AVAILABLE;
2769 loop->any_upper_bound = false;
2770 loop->any_estimate = false;
2772 exits = get_loop_exit_edges (loop);
2773 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
2775 if (!number_of_iterations_exit (loop, ex, &niter_desc, false))
2776 continue;
2778 niter = niter_desc.niter;
2779 type = TREE_TYPE (niter);
2780 if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST)
2781 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
2782 build_int_cst (type, 0),
2783 niter);
2784 record_estimate (loop, niter, niter_desc.max,
2785 last_stmt (ex->src),
2786 true, true, true);
2788 VEC_free (edge, heap, exits);
2790 infer_loop_bounds_from_undefined (loop);
2792 /* If we have a measured profile, use it to estimate the number of
2793 iterations. */
2794 if (loop->header->count != 0)
2796 gcov_type nit = expected_loop_iterations_unbounded (loop) + 1;
2797 bound = gcov_type_to_double_int (nit);
2798 record_niter_bound (loop, bound, true, false);
2801 /* If an upper bound is smaller than the realistic estimate of the
2802 number of iterations, use the upper bound instead. */
2803 if (loop->any_upper_bound
2804 && loop->any_estimate
2805 && double_int_ucmp (loop->nb_iterations_upper_bound,
2806 loop->nb_iterations_estimate) < 0)
2807 loop->nb_iterations_estimate = loop->nb_iterations_upper_bound;
2810 /* Records estimates on numbers of iterations of loops. */
2812 void
2813 estimate_numbers_of_iterations (void)
2815 loop_iterator li;
2816 struct loop *loop;
2818 /* We don't want to issue signed overflow warnings while getting
2819 loop iteration estimates. */
2820 fold_defer_overflow_warnings ();
2822 FOR_EACH_LOOP (li, loop, 0)
2824 estimate_numbers_of_iterations_loop (loop);
2827 fold_undefer_and_ignore_overflow_warnings ();
2830 /* Returns true if statement S1 dominates statement S2. */
2832 bool
2833 stmt_dominates_stmt_p (tree s1, tree s2)
2835 basic_block bb1 = bb_for_stmt (s1), bb2 = bb_for_stmt (s2);
2837 if (!bb1
2838 || s1 == s2)
2839 return true;
2841 if (bb1 == bb2)
2843 block_stmt_iterator bsi;
2845 for (bsi = bsi_start (bb1); bsi_stmt (bsi) != s2; bsi_next (&bsi))
2846 if (bsi_stmt (bsi) == s1)
2847 return true;
2849 return false;
2852 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
2855 /* Returns true when we can prove that the number of executions of
2856 STMT in the loop is at most NITER, according to the bound on
2857 the number of executions of the statement NITER_BOUND->stmt recorded in
2858 NITER_BOUND. If STMT is NULL, we must prove this bound for all
2859 statements in the loop. */
2861 static bool
2862 n_of_executions_at_most (tree stmt,
2863 struct nb_iter_bound *niter_bound,
2864 tree niter)
2866 double_int bound = niter_bound->bound;
2867 tree nit_type = TREE_TYPE (niter), e;
2868 enum tree_code cmp;
2870 gcc_assert (TYPE_UNSIGNED (nit_type));
2872 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
2873 the number of iterations is small. */
2874 if (!double_int_fits_to_tree_p (nit_type, bound))
2875 return false;
2877 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
2878 times. This means that:
2880 -- if NITER_BOUND->is_exit is true, then everything before
2881 NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
2882 times, and everything after it at most NITER_BOUND->bound times.
2884 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
2885 is executed, then NITER_BOUND->stmt is executed as well in the same
2886 iteration (we conclude that if both statements belong to the same
2887 basic block, or if STMT is after NITER_BOUND->stmt), then STMT
2888 is executed at most NITER_BOUND->bound + 1 times. Otherwise STMT is
2889 executed at most NITER_BOUND->bound + 2 times. */
2891 if (niter_bound->is_exit)
2893 if (stmt
2894 && stmt != niter_bound->stmt
2895 && stmt_dominates_stmt_p (niter_bound->stmt, stmt))
2896 cmp = GE_EXPR;
2897 else
2898 cmp = GT_EXPR;
2900 else
2902 if (!stmt
2903 || (bb_for_stmt (stmt) != bb_for_stmt (niter_bound->stmt)
2904 && !stmt_dominates_stmt_p (niter_bound->stmt, stmt)))
2906 bound = double_int_add (bound, double_int_one);
2907 if (double_int_zero_p (bound)
2908 || !double_int_fits_to_tree_p (nit_type, bound))
2909 return false;
2911 cmp = GT_EXPR;
2914 e = fold_binary (cmp, boolean_type_node,
2915 niter, double_int_to_tree (nit_type, bound));
2916 return e && integer_nonzerop (e);
2919 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
2921 bool
2922 nowrap_type_p (tree type)
2924 if (INTEGRAL_TYPE_P (type)
2925 && TYPE_OVERFLOW_UNDEFINED (type))
2926 return true;
2928 if (POINTER_TYPE_P (type))
2929 return true;
2931 return false;
2934 /* Return false only when the induction variable BASE + STEP * I is
2935 known to not overflow: i.e. when the number of iterations is small
2936 enough with respect to the step and initial condition in order to
2937 keep the evolution confined in TYPEs bounds. Return true when the
2938 iv is known to overflow or when the property is not computable.
2940 USE_OVERFLOW_SEMANTICS is true if this function should assume that
2941 the rules for overflow of the given language apply (e.g., that signed
2942 arithmetics in C does not overflow). */
2944 bool
2945 scev_probably_wraps_p (tree base, tree step,
2946 tree at_stmt, struct loop *loop,
2947 bool use_overflow_semantics)
2949 struct nb_iter_bound *bound;
2950 tree delta, step_abs;
2951 tree unsigned_type, valid_niter;
2952 tree type = TREE_TYPE (step);
2954 /* FIXME: We really need something like
2955 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
2957 We used to test for the following situation that frequently appears
2958 during address arithmetics:
2960 D.1621_13 = (long unsigned intD.4) D.1620_12;
2961 D.1622_14 = D.1621_13 * 8;
2962 D.1623_15 = (doubleD.29 *) D.1622_14;
2964 And derived that the sequence corresponding to D_14
2965 can be proved to not wrap because it is used for computing a
2966 memory access; however, this is not really the case -- for example,
2967 if D_12 = (unsigned char) [254,+,1], then D_14 has values
2968 2032, 2040, 0, 8, ..., but the code is still legal. */
2970 if (chrec_contains_undetermined (base)
2971 || chrec_contains_undetermined (step))
2972 return true;
2974 if (integer_zerop (step))
2975 return false;
2977 /* If we can use the fact that signed and pointer arithmetics does not
2978 wrap, we are done. */
2979 if (use_overflow_semantics && nowrap_type_p (type))
2980 return false;
2982 /* To be able to use estimates on number of iterations of the loop,
2983 we must have an upper bound on the absolute value of the step. */
2984 if (TREE_CODE (step) != INTEGER_CST)
2985 return true;
2987 /* Don't issue signed overflow warnings. */
2988 fold_defer_overflow_warnings ();
2990 /* Otherwise, compute the number of iterations before we reach the
2991 bound of the type, and verify that the loop is exited before this
2992 occurs. */
2993 unsigned_type = unsigned_type_for (type);
2994 base = fold_convert (unsigned_type, base);
2996 if (tree_int_cst_sign_bit (step))
2998 tree extreme = fold_convert (unsigned_type,
2999 lower_bound_in_type (type, type));
3000 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
3001 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type,
3002 fold_convert (unsigned_type, step));
3004 else
3006 tree extreme = fold_convert (unsigned_type,
3007 upper_bound_in_type (type, type));
3008 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
3009 step_abs = fold_convert (unsigned_type, step);
3012 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
3014 estimate_numbers_of_iterations_loop (loop);
3015 for (bound = loop->bounds; bound; bound = bound->next)
3017 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
3019 fold_undefer_and_ignore_overflow_warnings ();
3020 return false;
3024 fold_undefer_and_ignore_overflow_warnings ();
3026 /* At this point we still don't have a proof that the iv does not
3027 overflow: give up. */
3028 return true;
3031 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
3033 void
3034 free_numbers_of_iterations_estimates_loop (struct loop *loop)
3036 struct nb_iter_bound *bound, *next;
3038 loop->nb_iterations = NULL;
3039 loop->estimate_state = EST_NOT_COMPUTED;
3040 for (bound = loop->bounds; bound; bound = next)
3042 next = bound->next;
3043 ggc_free (bound);
3046 loop->bounds = NULL;
3049 /* Frees the information on upper bounds on numbers of iterations of loops. */
3051 void
3052 free_numbers_of_iterations_estimates (void)
3054 loop_iterator li;
3055 struct loop *loop;
3057 FOR_EACH_LOOP (li, loop, 0)
3059 free_numbers_of_iterations_estimates_loop (loop);
3063 /* Substitute value VAL for ssa name NAME inside expressions held
3064 at LOOP. */
3066 void
3067 substitute_in_loop_info (struct loop *loop, tree name, tree val)
3069 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);