]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/tree-ssa-loop-niter.c
[multiple changes]
[thirdparty/gcc.git] / gcc / tree-ssa-loop-niter.c
1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 2, or (at your option) any
9 later version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING. If not, write to the Free
18 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
19 02110-1301, USA. */
20
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"
46
47 #define SWAP(X, Y) do { affine_iv *tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
48
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
53
54 /*
55
56 Analysis of number of iterations of an affine exit test.
57
58 */
59
60 /* Bounds on some value, BELOW <= X <= UP. */
61
62 typedef struct
63 {
64 mpz_t below, up;
65 } bounds;
66
67
68 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
69
70 static void
71 split_to_var_and_offset (tree expr, tree *var, mpz_t offset)
72 {
73 tree type = TREE_TYPE (expr);
74 tree op0, op1;
75 double_int off;
76 bool negate = false;
77
78 *var = expr;
79 mpz_set_ui (offset, 0);
80
81 switch (TREE_CODE (expr))
82 {
83 case MINUS_EXPR:
84 negate = true;
85 /* Fallthru. */
86
87 case PLUS_EXPR:
88 case POINTER_PLUS_EXPR:
89 op0 = TREE_OPERAND (expr, 0);
90 op1 = TREE_OPERAND (expr, 1);
91
92 if (TREE_CODE (op1) != INTEGER_CST)
93 break;
94
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;
101
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;
107
108 default:
109 break;
110 }
111 }
112
113 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
114 in TYPE to MIN and MAX. */
115
116 static void
117 determine_value_range (tree type, tree var, mpz_t off,
118 mpz_t min, mpz_t max)
119 {
120 /* If the expression is a constant, we know its value exactly. */
121 if (integer_zerop (var))
122 {
123 mpz_set (min, off);
124 mpz_set (max, off);
125 return;
126 }
127
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;
133
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);
140 }
141
142 /* Stores the bounds on the difference of the values of the expressions
143 (var + X) and (var + Y), computed in TYPE, to BNDS. */
144
145 static void
146 bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y,
147 bounds *bnds)
148 {
149 int rel = mpz_cmp (x, y);
150 bool may_wrap = !nowrap_type_p (type);
151 mpz_t m;
152
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). */
166
167 if (rel == 0)
168 {
169 mpz_set_ui (bnds->below, 0);
170 mpz_set_ui (bnds->up, 0);
171 return;
172 }
173
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);
179
180 if (may_wrap)
181 {
182 if (rel > 0)
183 mpz_sub (bnds->below, bnds->below, m);
184 else
185 mpz_add (bnds->up, bnds->up, m);
186 }
187
188 mpz_clear (m);
189 }
190
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. */
194
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)
200 {
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;
206
207 switch (cmp)
208 {
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 (!tree_ssa_useless_type_conversion_1 (ctype, type))
217 return;
218
219 break;
220
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;
226
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;
233
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);
241
242 if (TYPE_MIN_VALUE (type)
243 && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0))
244 {
245 cmp = GT_EXPR;
246 break;
247 }
248 if (TYPE_MAX_VALUE (type)
249 && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0))
250 {
251 cmp = LT_EXPR;
252 break;
253 }
254
255 return;
256 default:
257 return;
258 }
259
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);
264
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. */
268
269 if (operand_equal_p (varx, varc1, 0))
270 {
271 tmp = varc0; varc0 = varc1; varc1 = tmp;
272 mpz_swap (offc0, offc1);
273 cmp = swap_tree_comparison (cmp);
274 }
275
276 if (!operand_equal_p (varx, varc0, 0)
277 || !operand_equal_p (vary, varc1, 0))
278 goto end;
279
280 mpz_init_set (loffx, offx);
281 mpz_init_set (loffy, offy);
282
283 if (cmp == GT_EXPR || cmp == GE_EXPR)
284 {
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;
290 }
291
292 /* If there is no overflow, the condition implies that
293
294 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
295
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
300
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. */
307
308 if (no_wrap)
309 {
310 x_ok = true;
311 y_ok = true;
312 }
313 else
314 {
315 x_ok = (integer_zerop (varx)
316 || mpz_cmp (loffx, offc0) >= 0);
317 y_ok = (integer_zerop (vary)
318 || mpz_cmp (loffy, offc1) <= 0);
319 }
320
321 if (x_ok && y_ok)
322 {
323 mpz_init (bnd);
324 mpz_sub (bnd, loffx, loffy);
325 mpz_add (bnd, bnd, offc1);
326 mpz_sub (bnd, bnd, offc0);
327
328 if (cmp == LT_EXPR)
329 mpz_sub_ui (bnd, bnd, 1);
330
331 if (lbound)
332 {
333 mpz_neg (bnd, bnd);
334 if (mpz_cmp (bnds->below, bnd) < 0)
335 mpz_set (bnds->below, bnd);
336 }
337 else
338 {
339 if (mpz_cmp (bnd, bnds->up) < 0)
340 mpz_set (bnds->up, bnd);
341 }
342 mpz_clear (bnd);
343 }
344
345 mpz_clear (loffx);
346 mpz_clear (loffy);
347 end:
348 mpz_clear (offc0);
349 mpz_clear (offc1);
350 }
351
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.
355
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). */
360
361 static void
362 bound_difference (struct loop *loop, tree x, tree y, bounds *bnds)
363 {
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;
373
374 /* Get rid of unnecessary casts, but preserve the value of
375 the expressions. */
376 STRIP_SIGN_NOPS (x);
377 STRIP_SIGN_NOPS (y);
378
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);
385
386 if (!integer_zerop (varx)
387 && operand_equal_p (varx, vary, 0))
388 {
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);
393 }
394 else
395 {
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);
404
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);
411 }
412
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;
416
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))
422 {
423 if (!single_pred_p (bb))
424 continue;
425 e = single_pred_edge (bb);
426
427 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
428 continue;
429
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);
436
437 if (e->flags & EDGE_FALSE_VALUE)
438 cmp = invert_tree_comparison (cmp, false);
439
440 refine_bounds_using_guard (type, varx, offx, vary, offy,
441 c0, cmp, c1, bnds);
442 ++cnt;
443 }
444
445 end:
446 mpz_clear (offx);
447 mpz_clear (offy);
448 }
449
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. */
453
454 static void
455 bounds_add (bounds *bnds, double_int delta, tree type)
456 {
457 mpz_t mdelta, max;
458
459 mpz_init (mdelta);
460 mpz_set_double_int (mdelta, delta, false);
461
462 mpz_init (max);
463 mpz_set_double_int (max, double_int_mask (TYPE_PRECISION (type)), true);
464
465 mpz_add (bnds->up, bnds->up, mdelta);
466 mpz_add (bnds->below, bnds->below, mdelta);
467
468 if (mpz_cmp (bnds->up, max) > 0)
469 mpz_set (bnds->up, max);
470
471 mpz_neg (max, max);
472 if (mpz_cmp (bnds->below, max) < 0)
473 mpz_set (bnds->below, max);
474
475 mpz_clear (mdelta);
476 mpz_clear (max);
477 }
478
479 /* Update the bounds in BNDS that restrict the value of X to the bounds
480 that restrict the value of -X. */
481
482 static void
483 bounds_negate (bounds *bnds)
484 {
485 mpz_t tmp;
486
487 mpz_init_set (tmp, bnds->up);
488 mpz_neg (bnds->up, bnds->below);
489 mpz_neg (bnds->below, tmp);
490 mpz_clear (tmp);
491 }
492
493 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
494
495 static tree
496 inverse (tree x, tree mask)
497 {
498 tree type = TREE_TYPE (x);
499 tree rslt;
500 unsigned ctr = tree_floor_log2 (mask);
501
502 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT)
503 {
504 unsigned HOST_WIDE_INT ix;
505 unsigned HOST_WIDE_INT imask;
506 unsigned HOST_WIDE_INT irslt = 1;
507
508 gcc_assert (cst_and_fits_in_hwi (x));
509 gcc_assert (cst_and_fits_in_hwi (mask));
510
511 ix = int_cst_value (x);
512 imask = int_cst_value (mask);
513
514 for (; ctr; ctr--)
515 {
516 irslt *= ix;
517 ix *= ix;
518 }
519 irslt &= imask;
520
521 rslt = build_int_cst_type (type, irslt);
522 }
523 else
524 {
525 rslt = build_int_cst (type, 1);
526 for (; ctr; ctr--)
527 {
528 rslt = int_const_binop (MULT_EXPR, rslt, x, 0);
529 x = int_const_binop (MULT_EXPR, x, x, 0);
530 }
531 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask, 0);
532 }
533
534 return rslt;
535 }
536
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. */
542
543 static void
544 number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s,
545 bounds *bnds)
546 {
547 double_int max;
548 mpz_t d;
549
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))
554 {
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;
559 }
560
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);
569
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);
574 }
575
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. */
583
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)
588 {
589 tree niter_type = unsigned_type_for (type);
590 tree s, c, d, bits, assumption, tmp, bound;
591 mpz_t max;
592
593 niter->control = *iv;
594 niter->bound = final;
595 niter->cmp = NE_EXPR;
596
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))
603 {
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);
610 }
611 else
612 {
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));
617 }
618
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);
623
624 /* First the trivial cases -- when the step is 1. */
625 if (integer_onep (s))
626 {
627 niter->niter = c;
628 return true;
629 }
630
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)));
638
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);
642
643 if (!never_infinite)
644 {
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);
653 }
654
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;
659 }
660
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. */
669
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)
675 {
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;
685
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);
691
692 mpz_init (mmod);
693 mpz_set_double_int (mmod, tree_to_double_int (mod), true);
694 mpz_neg (mmod, mmod);
695
696 if (integer_nonzerop (iv0->step))
697 {
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))
702 {
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;
709 }
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));
717 }
718 else
719 {
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))
724 {
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;
731 }
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);
739 }
740
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);
751
752 ret = true;
753 end:
754 mpz_clear (mmod);
755 return ret;
756 }
757
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. */
762
763 static bool
764 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1,
765 struct tree_niter_desc *niter, tree step)
766 {
767 tree bound, d, assumption, diff;
768 tree niter_type = TREE_TYPE (step);
769
770 if (integer_nonzerop (iv0->step))
771 {
772 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
773 if (iv0->no_overflow)
774 return true;
775
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. */
779
780 if (TREE_CODE (iv0->base) == INTEGER_CST)
781 {
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);
786 }
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);
794 }
795 else
796 {
797 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
798 if (iv1->no_overflow)
799 return true;
800
801 if (TREE_CODE (iv1->base) == INTEGER_CST)
802 {
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);
807 }
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);
815 }
816
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);
822
823 iv0->no_overflow = true;
824 iv1->no_overflow = true;
825 return true;
826 }
827
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. */
831
832 static void
833 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1,
834 struct tree_niter_desc *niter, bounds *bnds)
835 {
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;
841
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
845
846 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
847
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).
851
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.
857
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. */
861
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
867 {
868 dstep = double_int_sext (tree_to_double_int (iv1->step),
869 TYPE_PRECISION (type));
870 dstep = double_int_neg (dstep);
871 }
872
873 mpz_init (mstep);
874 mpz_set_double_int (mstep, dstep, true);
875 mpz_neg (mstep, mstep);
876 mpz_add_ui (mstep, mstep, 1);
877
878 rolls_p = mpz_cmp (mstep, bnds->below) <= 0;
879
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);
892
893 if (rolls_p && no_overflow_p)
894 return;
895
896 type1 = type;
897 if (POINTER_TYPE_P (type))
898 type1 = sizetype;
899
900 /* Now the hard part; we must formulate the assumption(s) as expressions, and
901 we must be careful not to introduce overflow. */
902
903 if (integer_nonzerop (iv0->step))
904 {
905 diff = fold_build2 (MINUS_EXPR, type1,
906 iv0->step, build_int_cst (type1, 1));
907
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))
912 {
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);
917 }
918
919 /* And then we can compute iv0->base - diff, and compare it with
920 iv1->base. */
921 mbzl = fold_build2 (MINUS_EXPR, type1, iv0->base, diff);
922 mbzr = iv1->base;
923 }
924 else
925 {
926 diff = fold_build2 (PLUS_EXPR, type1,
927 iv1->step, build_int_cst (type1, 1));
928
929 if (!POINTER_TYPE_P (type))
930 {
931 bound = fold_build2 (PLUS_EXPR, type1,
932 TYPE_MAX_VALUE (type), diff);
933 assumption = fold_build2 (LE_EXPR, boolean_type_node,
934 iv1->base, bound);
935 }
936
937 mbzl = iv0->base;
938 mbzr = fold_build2 (MINUS_EXPR, type1, iv1->base, diff);
939 }
940
941 if (!integer_nonzerop (assumption))
942 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
943 niter->assumptions, assumption);
944 if (!rolls_p)
945 {
946 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr);
947 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
948 niter->may_be_zero, mbz);
949 }
950 }
951
952 /* Determines number of iterations of loop whose ending condition
953 is IV0 < IV1. TYPE is the type of the iv. The number of
954 iterations is stored to NITER. BNDS bounds the difference
955 IV1->base - IV0->base. */
956
957 static bool
958 number_of_iterations_lt (tree type, affine_iv *iv0, affine_iv *iv1,
959 struct tree_niter_desc *niter,
960 bool never_infinite ATTRIBUTE_UNUSED,
961 bounds *bnds)
962 {
963 tree niter_type = unsigned_type_for (type);
964 tree delta, step, s;
965 mpz_t mstep, tmp;
966
967 if (integer_nonzerop (iv0->step))
968 {
969 niter->control = *iv0;
970 niter->cmp = LT_EXPR;
971 niter->bound = iv1->base;
972 }
973 else
974 {
975 niter->control = *iv1;
976 niter->cmp = GT_EXPR;
977 niter->bound = iv0->base;
978 }
979
980 delta = fold_build2 (MINUS_EXPR, niter_type,
981 fold_convert (niter_type, iv1->base),
982 fold_convert (niter_type, iv0->base));
983
984 /* First handle the special case that the step is +-1. */
985 if ((integer_onep (iv0->step) && integer_zerop (iv1->step))
986 || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step)))
987 {
988 /* for (i = iv0->base; i < iv1->base; i++)
989
990 or
991
992 for (i = iv1->base; i > iv0->base; i--).
993
994 In both cases # of iterations is iv1->base - iv0->base, assuming that
995 iv1->base >= iv0->base.
996
997 First try to derive a lower bound on the value of
998 iv1->base - iv0->base, computed in full precision. If the difference
999 is nonnegative, we are done, otherwise we must record the
1000 condition. */
1001
1002 if (mpz_sgn (bnds->below) < 0)
1003 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node,
1004 iv1->base, iv0->base);
1005 niter->niter = delta;
1006 niter->max = mpz_get_double_int (niter_type, bnds->up, false);
1007 return true;
1008 }
1009
1010 if (integer_nonzerop (iv0->step))
1011 step = fold_convert (niter_type, iv0->step);
1012 else
1013 step = fold_convert (niter_type,
1014 fold_build1 (NEGATE_EXPR, type, iv1->step));
1015
1016 /* If we can determine the final value of the control iv exactly, we can
1017 transform the condition to != comparison. In particular, this will be
1018 the case if DELTA is constant. */
1019 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step,
1020 bnds))
1021 {
1022 affine_iv zps;
1023
1024 zps.base = build_int_cst (niter_type, 0);
1025 zps.step = step;
1026 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1027 zps does not overflow. */
1028 zps.no_overflow = true;
1029
1030 return number_of_iterations_ne (type, &zps, delta, niter, true, bnds);
1031 }
1032
1033 /* Make sure that the control iv does not overflow. */
1034 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step))
1035 return false;
1036
1037 /* We determine the number of iterations as (delta + step - 1) / step. For
1038 this to work, we must know that iv1->base >= iv0->base - step + 1,
1039 otherwise the loop does not roll. */
1040 assert_loop_rolls_lt (type, iv0, iv1, niter, bnds);
1041
1042 s = fold_build2 (MINUS_EXPR, niter_type,
1043 step, build_int_cst (niter_type, 1));
1044 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s);
1045 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step);
1046
1047 mpz_init (mstep);
1048 mpz_init (tmp);
1049 mpz_set_double_int (mstep, tree_to_double_int (step), true);
1050 mpz_add (tmp, bnds->up, mstep);
1051 mpz_sub_ui (tmp, tmp, 1);
1052 mpz_fdiv_q (tmp, tmp, mstep);
1053 niter->max = mpz_get_double_int (niter_type, tmp, false);
1054 mpz_clear (mstep);
1055 mpz_clear (tmp);
1056
1057 return true;
1058 }
1059
1060 /* Determines number of iterations of loop whose ending condition
1061 is IV0 <= IV1. TYPE is the type of the iv. The number of
1062 iterations is stored to NITER. NEVER_INFINITE is true if
1063 we know that this condition must eventually become false (we derived this
1064 earlier, and possibly set NITER->assumptions to make sure this
1065 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1066
1067 static bool
1068 number_of_iterations_le (tree type, affine_iv *iv0, affine_iv *iv1,
1069 struct tree_niter_desc *niter, bool never_infinite,
1070 bounds *bnds)
1071 {
1072 tree assumption;
1073 tree type1 = type;
1074 if (POINTER_TYPE_P (type))
1075 type1 = sizetype;
1076
1077 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1078 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1079 value of the type. This we must know anyway, since if it is
1080 equal to this value, the loop rolls forever. */
1081
1082 if (!never_infinite)
1083 {
1084 if (integer_nonzerop (iv0->step))
1085 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1086 iv1->base, TYPE_MAX_VALUE (type1));
1087 else
1088 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1089 iv0->base, TYPE_MIN_VALUE (type1));
1090
1091 if (integer_zerop (assumption))
1092 return false;
1093 if (!integer_nonzerop (assumption))
1094 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1095 niter->assumptions, assumption);
1096 }
1097
1098 if (integer_nonzerop (iv0->step))
1099 iv1->base = fold_build2 (PLUS_EXPR, type1,
1100 iv1->base, build_int_cst (type1, 1));
1101 else
1102 iv0->base = fold_build2 (MINUS_EXPR, type1,
1103 iv0->base, build_int_cst (type1, 1));
1104
1105 bounds_add (bnds, double_int_one, type1);
1106
1107 return number_of_iterations_lt (type, iv0, iv1, niter, never_infinite, bnds);
1108 }
1109
1110 /* Dumps description of affine induction variable IV to FILE. */
1111
1112 static void
1113 dump_affine_iv (FILE *file, affine_iv *iv)
1114 {
1115 if (!integer_zerop (iv->step))
1116 fprintf (file, "[");
1117
1118 print_generic_expr (dump_file, iv->base, TDF_SLIM);
1119
1120 if (!integer_zerop (iv->step))
1121 {
1122 fprintf (file, ", + , ");
1123 print_generic_expr (dump_file, iv->step, TDF_SLIM);
1124 fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : "");
1125 }
1126 }
1127
1128 /* Determine the number of iterations according to condition (for staying
1129 inside loop) which compares two induction variables using comparison
1130 operator CODE. The induction variable on left side of the comparison
1131 is IV0, the right-hand side is IV1. Both induction variables must have
1132 type TYPE, which must be an integer or pointer type. The steps of the
1133 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1134
1135 LOOP is the loop whose number of iterations we are determining.
1136
1137 ONLY_EXIT is true if we are sure this is the only way the loop could be
1138 exited (including possibly non-returning function calls, exceptions, etc.)
1139 -- in this case we can use the information whether the control induction
1140 variables can overflow or not in a more efficient way.
1141
1142 The results (number of iterations and assumptions as described in
1143 comments at struct tree_niter_desc in tree-flow.h) are stored to NITER.
1144 Returns false if it fails to determine number of iterations, true if it
1145 was determined (possibly with some assumptions). */
1146
1147 static bool
1148 number_of_iterations_cond (struct loop *loop,
1149 tree type, affine_iv *iv0, enum tree_code code,
1150 affine_iv *iv1, struct tree_niter_desc *niter,
1151 bool only_exit)
1152 {
1153 bool never_infinite, ret;
1154 bounds bnds;
1155
1156 /* The meaning of these assumptions is this:
1157 if !assumptions
1158 then the rest of information does not have to be valid
1159 if may_be_zero then the loop does not roll, even if
1160 niter != 0. */
1161 niter->assumptions = boolean_true_node;
1162 niter->may_be_zero = boolean_false_node;
1163 niter->niter = NULL_TREE;
1164 niter->max = double_int_zero;
1165
1166 niter->bound = NULL_TREE;
1167 niter->cmp = ERROR_MARK;
1168
1169 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1170 the control variable is on lhs. */
1171 if (code == GE_EXPR || code == GT_EXPR
1172 || (code == NE_EXPR && integer_zerop (iv0->step)))
1173 {
1174 SWAP (iv0, iv1);
1175 code = swap_tree_comparison (code);
1176 }
1177
1178 if (!only_exit)
1179 {
1180 /* If this is not the only possible exit from the loop, the information
1181 that the induction variables cannot overflow as derived from
1182 signedness analysis cannot be relied upon. We use them e.g. in the
1183 following way: given loop for (i = 0; i <= n; i++), if i is
1184 signed, it cannot overflow, thus this loop is equivalent to
1185 for (i = 0; i < n + 1; i++); however, if n == MAX, but the loop
1186 is exited in some other way before i overflows, this transformation
1187 is incorrect (the new loop exits immediately). */
1188 iv0->no_overflow = false;
1189 iv1->no_overflow = false;
1190 }
1191
1192 if (POINTER_TYPE_P (type))
1193 {
1194 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1195 to the same object. If they do, the control variable cannot wrap
1196 (as wrap around the bounds of memory will never return a pointer
1197 that would be guaranteed to point to the same object, even if we
1198 avoid undefined behavior by casting to size_t and back). The
1199 restrictions on pointer arithmetics and comparisons of pointers
1200 ensure that using the no-overflow assumptions is correct in this
1201 case even if ONLY_EXIT is false. */
1202 iv0->no_overflow = true;
1203 iv1->no_overflow = true;
1204 }
1205
1206 /* If the control induction variable does not overflow, the loop obviously
1207 cannot be infinite. */
1208 if (!integer_zerop (iv0->step) && iv0->no_overflow)
1209 never_infinite = true;
1210 else if (!integer_zerop (iv1->step) && iv1->no_overflow)
1211 never_infinite = true;
1212 else
1213 never_infinite = false;
1214
1215 /* We can handle the case when neither of the sides of the comparison is
1216 invariant, provided that the test is NE_EXPR. This rarely occurs in
1217 practice, but it is simple enough to manage. */
1218 if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step))
1219 {
1220 if (code != NE_EXPR)
1221 return false;
1222
1223 iv0->step = fold_binary_to_constant (MINUS_EXPR, type,
1224 iv0->step, iv1->step);
1225 iv0->no_overflow = false;
1226 iv1->step = build_int_cst (type, 0);
1227 iv1->no_overflow = true;
1228 }
1229
1230 /* If the result of the comparison is a constant, the loop is weird. More
1231 precise handling would be possible, but the situation is not common enough
1232 to waste time on it. */
1233 if (integer_zerop (iv0->step) && integer_zerop (iv1->step))
1234 return false;
1235
1236 /* Ignore loops of while (i-- < 10) type. */
1237 if (code != NE_EXPR)
1238 {
1239 if (iv0->step && tree_int_cst_sign_bit (iv0->step))
1240 return false;
1241
1242 if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step))
1243 return false;
1244 }
1245
1246 /* If the loop exits immediately, there is nothing to do. */
1247 if (integer_zerop (fold_build2 (code, boolean_type_node, iv0->base, iv1->base)))
1248 {
1249 niter->niter = build_int_cst (unsigned_type_for (type), 0);
1250 niter->max = double_int_zero;
1251 return true;
1252 }
1253
1254 /* OK, now we know we have a senseful loop. Handle several cases, depending
1255 on what comparison operator is used. */
1256 bound_difference (loop, iv1->base, iv0->base, &bnds);
1257
1258 if (dump_file && (dump_flags & TDF_DETAILS))
1259 {
1260 fprintf (dump_file,
1261 "Analyzing # of iterations of loop %d\n", loop->num);
1262
1263 fprintf (dump_file, " exit condition ");
1264 dump_affine_iv (dump_file, iv0);
1265 fprintf (dump_file, " %s ",
1266 code == NE_EXPR ? "!="
1267 : code == LT_EXPR ? "<"
1268 : "<=");
1269 dump_affine_iv (dump_file, iv1);
1270 fprintf (dump_file, "\n");
1271
1272 fprintf (dump_file, " bounds on difference of bases: ");
1273 mpz_out_str (dump_file, 10, bnds.below);
1274 fprintf (dump_file, " ... ");
1275 mpz_out_str (dump_file, 10, bnds.up);
1276 fprintf (dump_file, "\n");
1277 }
1278
1279 switch (code)
1280 {
1281 case NE_EXPR:
1282 gcc_assert (integer_zerop (iv1->step));
1283 ret = number_of_iterations_ne (type, iv0, iv1->base, niter,
1284 never_infinite, &bnds);
1285 break;
1286
1287 case LT_EXPR:
1288 ret = number_of_iterations_lt (type, iv0, iv1, niter, never_infinite,
1289 &bnds);
1290 break;
1291
1292 case LE_EXPR:
1293 ret = number_of_iterations_le (type, iv0, iv1, niter, never_infinite,
1294 &bnds);
1295 break;
1296
1297 default:
1298 gcc_unreachable ();
1299 }
1300
1301 mpz_clear (bnds.up);
1302 mpz_clear (bnds.below);
1303
1304 if (dump_file && (dump_flags & TDF_DETAILS))
1305 {
1306 if (ret)
1307 {
1308 fprintf (dump_file, " result:\n");
1309 if (!integer_nonzerop (niter->assumptions))
1310 {
1311 fprintf (dump_file, " under assumptions ");
1312 print_generic_expr (dump_file, niter->assumptions, TDF_SLIM);
1313 fprintf (dump_file, "\n");
1314 }
1315
1316 if (!integer_zerop (niter->may_be_zero))
1317 {
1318 fprintf (dump_file, " zero if ");
1319 print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM);
1320 fprintf (dump_file, "\n");
1321 }
1322
1323 fprintf (dump_file, " # of iterations ");
1324 print_generic_expr (dump_file, niter->niter, TDF_SLIM);
1325 fprintf (dump_file, ", bounded by ");
1326 dump_double_int (dump_file, niter->max, true);
1327 fprintf (dump_file, "\n");
1328 }
1329 else
1330 fprintf (dump_file, " failed\n\n");
1331 }
1332 return ret;
1333 }
1334
1335 /* Substitute NEW for OLD in EXPR and fold the result. */
1336
1337 static tree
1338 simplify_replace_tree (tree expr, tree old, tree new_tree)
1339 {
1340 unsigned i, n;
1341 tree ret = NULL_TREE, e, se;
1342
1343 if (!expr)
1344 return NULL_TREE;
1345
1346 if (expr == old
1347 || operand_equal_p (expr, old, 0))
1348 return unshare_expr (new_tree);
1349
1350 if (!EXPR_P (expr) && !GIMPLE_STMT_P (expr))
1351 return expr;
1352
1353 n = TREE_OPERAND_LENGTH (expr);
1354 for (i = 0; i < n; i++)
1355 {
1356 e = TREE_OPERAND (expr, i);
1357 se = simplify_replace_tree (e, old, new_tree);
1358 if (e == se)
1359 continue;
1360
1361 if (!ret)
1362 ret = copy_node (expr);
1363
1364 TREE_OPERAND (ret, i) = se;
1365 }
1366
1367 return (ret ? fold (ret) : expr);
1368 }
1369
1370 /* Expand definitions of ssa names in EXPR as long as they are simple
1371 enough, and return the new expression. */
1372
1373 tree
1374 expand_simple_operations (tree expr)
1375 {
1376 unsigned i, n;
1377 tree ret = NULL_TREE, e, ee, stmt;
1378 enum tree_code code;
1379
1380 if (expr == NULL_TREE)
1381 return expr;
1382
1383 if (is_gimple_min_invariant (expr))
1384 return expr;
1385
1386 code = TREE_CODE (expr);
1387 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code)))
1388 {
1389 n = TREE_OPERAND_LENGTH (expr);
1390 for (i = 0; i < n; i++)
1391 {
1392 e = TREE_OPERAND (expr, i);
1393 ee = expand_simple_operations (e);
1394 if (e == ee)
1395 continue;
1396
1397 if (!ret)
1398 ret = copy_node (expr);
1399
1400 TREE_OPERAND (ret, i) = ee;
1401 }
1402
1403 if (!ret)
1404 return expr;
1405
1406 fold_defer_overflow_warnings ();
1407 ret = fold (ret);
1408 fold_undefer_and_ignore_overflow_warnings ();
1409 return ret;
1410 }
1411
1412 if (TREE_CODE (expr) != SSA_NAME)
1413 return expr;
1414
1415 stmt = SSA_NAME_DEF_STMT (expr);
1416 if (TREE_CODE (stmt) == PHI_NODE)
1417 {
1418 basic_block src, dest;
1419
1420 if (PHI_NUM_ARGS (stmt) != 1)
1421 return expr;
1422 e = PHI_ARG_DEF (stmt, 0);
1423
1424 /* Avoid propagating through loop exit phi nodes, which
1425 could break loop-closed SSA form restrictions. */
1426 dest = bb_for_stmt (stmt);
1427 src = single_pred (dest);
1428 if (TREE_CODE (e) == SSA_NAME
1429 && src->loop_father != dest->loop_father)
1430 return expr;
1431
1432 return expand_simple_operations (e);
1433 }
1434 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
1435 return expr;
1436
1437 e = GIMPLE_STMT_OPERAND (stmt, 1);
1438 if (/* Casts are simple. */
1439 TREE_CODE (e) != NOP_EXPR
1440 && TREE_CODE (e) != CONVERT_EXPR
1441 /* Copies are simple. */
1442 && TREE_CODE (e) != SSA_NAME
1443 /* Assignments of invariants are simple. */
1444 && !is_gimple_min_invariant (e)
1445 /* And increments and decrements by a constant are simple. */
1446 && !((TREE_CODE (e) == PLUS_EXPR
1447 || TREE_CODE (e) == MINUS_EXPR
1448 || TREE_CODE (e) == POINTER_PLUS_EXPR)
1449 && is_gimple_min_invariant (TREE_OPERAND (e, 1))))
1450 return expr;
1451
1452 return expand_simple_operations (e);
1453 }
1454
1455 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1456 expression (or EXPR unchanged, if no simplification was possible). */
1457
1458 static tree
1459 tree_simplify_using_condition_1 (tree cond, tree expr)
1460 {
1461 bool changed;
1462 tree e, te, e0, e1, e2, notcond;
1463 enum tree_code code = TREE_CODE (expr);
1464
1465 if (code == INTEGER_CST)
1466 return expr;
1467
1468 if (code == TRUTH_OR_EXPR
1469 || code == TRUTH_AND_EXPR
1470 || code == COND_EXPR)
1471 {
1472 changed = false;
1473
1474 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0));
1475 if (TREE_OPERAND (expr, 0) != e0)
1476 changed = true;
1477
1478 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1));
1479 if (TREE_OPERAND (expr, 1) != e1)
1480 changed = true;
1481
1482 if (code == COND_EXPR)
1483 {
1484 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2));
1485 if (TREE_OPERAND (expr, 2) != e2)
1486 changed = true;
1487 }
1488 else
1489 e2 = NULL_TREE;
1490
1491 if (changed)
1492 {
1493 if (code == COND_EXPR)
1494 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1495 else
1496 expr = fold_build2 (code, boolean_type_node, e0, e1);
1497 }
1498
1499 return expr;
1500 }
1501
1502 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
1503 propagation, and vice versa. Fold does not handle this, since it is
1504 considered too expensive. */
1505 if (TREE_CODE (cond) == EQ_EXPR)
1506 {
1507 e0 = TREE_OPERAND (cond, 0);
1508 e1 = TREE_OPERAND (cond, 1);
1509
1510 /* We know that e0 == e1. Check whether we cannot simplify expr
1511 using this fact. */
1512 e = simplify_replace_tree (expr, e0, e1);
1513 if (integer_zerop (e) || integer_nonzerop (e))
1514 return e;
1515
1516 e = simplify_replace_tree (expr, e1, e0);
1517 if (integer_zerop (e) || integer_nonzerop (e))
1518 return e;
1519 }
1520 if (TREE_CODE (expr) == EQ_EXPR)
1521 {
1522 e0 = TREE_OPERAND (expr, 0);
1523 e1 = TREE_OPERAND (expr, 1);
1524
1525 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
1526 e = simplify_replace_tree (cond, e0, e1);
1527 if (integer_zerop (e))
1528 return e;
1529 e = simplify_replace_tree (cond, e1, e0);
1530 if (integer_zerop (e))
1531 return e;
1532 }
1533 if (TREE_CODE (expr) == NE_EXPR)
1534 {
1535 e0 = TREE_OPERAND (expr, 0);
1536 e1 = TREE_OPERAND (expr, 1);
1537
1538 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
1539 e = simplify_replace_tree (cond, e0, e1);
1540 if (integer_zerop (e))
1541 return boolean_true_node;
1542 e = simplify_replace_tree (cond, e1, e0);
1543 if (integer_zerop (e))
1544 return boolean_true_node;
1545 }
1546
1547 te = expand_simple_operations (expr);
1548
1549 /* Check whether COND ==> EXPR. */
1550 notcond = invert_truthvalue (cond);
1551 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, te);
1552 if (e && integer_nonzerop (e))
1553 return e;
1554
1555 /* Check whether COND ==> not EXPR. */
1556 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, te);
1557 if (e && integer_zerop (e))
1558 return e;
1559
1560 return expr;
1561 }
1562
1563 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1564 expression (or EXPR unchanged, if no simplification was possible).
1565 Wrapper around tree_simplify_using_condition_1 that ensures that chains
1566 of simple operations in definitions of ssa names in COND are expanded,
1567 so that things like casts or incrementing the value of the bound before
1568 the loop do not cause us to fail. */
1569
1570 static tree
1571 tree_simplify_using_condition (tree cond, tree expr)
1572 {
1573 cond = expand_simple_operations (cond);
1574
1575 return tree_simplify_using_condition_1 (cond, expr);
1576 }
1577
1578 /* Tries to simplify EXPR using the conditions on entry to LOOP.
1579 Returns the simplified expression (or EXPR unchanged, if no
1580 simplification was possible).*/
1581
1582 static tree
1583 simplify_using_initial_conditions (struct loop *loop, tree expr)
1584 {
1585 edge e;
1586 basic_block bb;
1587 tree cond;
1588 int cnt = 0;
1589
1590 if (TREE_CODE (expr) == INTEGER_CST)
1591 return expr;
1592
1593 /* Limit walking the dominators to avoid quadraticness in
1594 the number of BBs times the number of loops in degenerate
1595 cases. */
1596 for (bb = loop->header;
1597 bb != ENTRY_BLOCK_PTR && cnt < MAX_DOMINATORS_TO_WALK;
1598 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
1599 {
1600 if (!single_pred_p (bb))
1601 continue;
1602 e = single_pred_edge (bb);
1603
1604 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
1605 continue;
1606
1607 cond = COND_EXPR_COND (last_stmt (e->src));
1608 if (e->flags & EDGE_FALSE_VALUE)
1609 cond = invert_truthvalue (cond);
1610 expr = tree_simplify_using_condition (cond, expr);
1611 ++cnt;
1612 }
1613
1614 return expr;
1615 }
1616
1617 /* Tries to simplify EXPR using the evolutions of the loop invariants
1618 in the superloops of LOOP. Returns the simplified expression
1619 (or EXPR unchanged, if no simplification was possible). */
1620
1621 static tree
1622 simplify_using_outer_evolutions (struct loop *loop, tree expr)
1623 {
1624 enum tree_code code = TREE_CODE (expr);
1625 bool changed;
1626 tree e, e0, e1, e2;
1627
1628 if (is_gimple_min_invariant (expr))
1629 return expr;
1630
1631 if (code == TRUTH_OR_EXPR
1632 || code == TRUTH_AND_EXPR
1633 || code == COND_EXPR)
1634 {
1635 changed = false;
1636
1637 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0));
1638 if (TREE_OPERAND (expr, 0) != e0)
1639 changed = true;
1640
1641 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1));
1642 if (TREE_OPERAND (expr, 1) != e1)
1643 changed = true;
1644
1645 if (code == COND_EXPR)
1646 {
1647 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2));
1648 if (TREE_OPERAND (expr, 2) != e2)
1649 changed = true;
1650 }
1651 else
1652 e2 = NULL_TREE;
1653
1654 if (changed)
1655 {
1656 if (code == COND_EXPR)
1657 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1658 else
1659 expr = fold_build2 (code, boolean_type_node, e0, e1);
1660 }
1661
1662 return expr;
1663 }
1664
1665 e = instantiate_parameters (loop, expr);
1666 if (is_gimple_min_invariant (e))
1667 return e;
1668
1669 return expr;
1670 }
1671
1672 /* Returns true if EXIT is the only possible exit from LOOP. */
1673
1674 static bool
1675 loop_only_exit_p (struct loop *loop, edge exit)
1676 {
1677 basic_block *body;
1678 block_stmt_iterator bsi;
1679 unsigned i;
1680 tree call;
1681
1682 if (exit != single_exit (loop))
1683 return false;
1684
1685 body = get_loop_body (loop);
1686 for (i = 0; i < loop->num_nodes; i++)
1687 {
1688 for (bsi = bsi_start (body[0]); !bsi_end_p (bsi); bsi_next (&bsi))
1689 {
1690 call = get_call_expr_in (bsi_stmt (bsi));
1691 if (call && TREE_SIDE_EFFECTS (call))
1692 {
1693 free (body);
1694 return false;
1695 }
1696 }
1697 }
1698
1699 free (body);
1700 return true;
1701 }
1702
1703 /* Stores description of number of iterations of LOOP derived from
1704 EXIT (an exit edge of the LOOP) in NITER. Returns true if some
1705 useful information could be derived (and fields of NITER has
1706 meaning described in comments at struct tree_niter_desc
1707 declaration), false otherwise. If WARN is true and
1708 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use
1709 potentially unsafe assumptions. */
1710
1711 bool
1712 number_of_iterations_exit (struct loop *loop, edge exit,
1713 struct tree_niter_desc *niter,
1714 bool warn)
1715 {
1716 tree stmt, cond, type;
1717 tree op0, op1;
1718 enum tree_code code;
1719 affine_iv iv0, iv1;
1720
1721 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src))
1722 return false;
1723
1724 niter->assumptions = boolean_false_node;
1725 stmt = last_stmt (exit->src);
1726 if (!stmt || TREE_CODE (stmt) != COND_EXPR)
1727 return false;
1728
1729 /* We want the condition for staying inside loop. */
1730 cond = COND_EXPR_COND (stmt);
1731 if (exit->flags & EDGE_TRUE_VALUE)
1732 cond = invert_truthvalue (cond);
1733
1734 code = TREE_CODE (cond);
1735 switch (code)
1736 {
1737 case GT_EXPR:
1738 case GE_EXPR:
1739 case NE_EXPR:
1740 case LT_EXPR:
1741 case LE_EXPR:
1742 break;
1743
1744 default:
1745 return false;
1746 }
1747
1748 op0 = TREE_OPERAND (cond, 0);
1749 op1 = TREE_OPERAND (cond, 1);
1750 type = TREE_TYPE (op0);
1751
1752 if (TREE_CODE (type) != INTEGER_TYPE
1753 && !POINTER_TYPE_P (type))
1754 return false;
1755
1756 if (!simple_iv (loop, stmt, op0, &iv0, false))
1757 return false;
1758 if (!simple_iv (loop, stmt, op1, &iv1, false))
1759 return false;
1760
1761 /* We don't want to see undefined signed overflow warnings while
1762 computing the number of iterations. */
1763 fold_defer_overflow_warnings ();
1764
1765 iv0.base = expand_simple_operations (iv0.base);
1766 iv1.base = expand_simple_operations (iv1.base);
1767 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter,
1768 loop_only_exit_p (loop, exit)))
1769 {
1770 fold_undefer_and_ignore_overflow_warnings ();
1771 return false;
1772 }
1773
1774 if (optimize >= 3)
1775 {
1776 niter->assumptions = simplify_using_outer_evolutions (loop,
1777 niter->assumptions);
1778 niter->may_be_zero = simplify_using_outer_evolutions (loop,
1779 niter->may_be_zero);
1780 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
1781 }
1782
1783 niter->assumptions
1784 = simplify_using_initial_conditions (loop,
1785 niter->assumptions);
1786 niter->may_be_zero
1787 = simplify_using_initial_conditions (loop,
1788 niter->may_be_zero);
1789
1790 fold_undefer_and_ignore_overflow_warnings ();
1791
1792 if (integer_onep (niter->assumptions))
1793 return true;
1794
1795 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
1796 But if we can prove that there is overflow or some other source of weird
1797 behavior, ignore the loop even with -funsafe-loop-optimizations. */
1798 if (integer_zerop (niter->assumptions))
1799 return false;
1800
1801 if (flag_unsafe_loop_optimizations)
1802 niter->assumptions = boolean_true_node;
1803
1804 if (warn)
1805 {
1806 const char *wording;
1807 location_t loc = EXPR_LOCATION (stmt);
1808
1809 /* We can provide a more specific warning if one of the operator is
1810 constant and the other advances by +1 or -1. */
1811 if (!integer_zerop (iv1.step)
1812 ? (integer_zerop (iv0.step)
1813 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step)))
1814 : (integer_onep (iv0.step) || integer_all_onesp (iv0.step)))
1815 wording =
1816 flag_unsafe_loop_optimizations
1817 ? N_("assuming that the loop is not infinite")
1818 : N_("cannot optimize possibly infinite loops");
1819 else
1820 wording =
1821 flag_unsafe_loop_optimizations
1822 ? N_("assuming that the loop counter does not overflow")
1823 : N_("cannot optimize loop, the loop counter may overflow");
1824
1825 if (LOCATION_LINE (loc) > 0)
1826 warning (OPT_Wunsafe_loop_optimizations, "%H%s", &loc, gettext (wording));
1827 else
1828 warning (OPT_Wunsafe_loop_optimizations, "%s", gettext (wording));
1829 }
1830
1831 return flag_unsafe_loop_optimizations;
1832 }
1833
1834 /* Try to determine the number of iterations of LOOP. If we succeed,
1835 expression giving number of iterations is returned and *EXIT is
1836 set to the edge from that the information is obtained. Otherwise
1837 chrec_dont_know is returned. */
1838
1839 tree
1840 find_loop_niter (struct loop *loop, edge *exit)
1841 {
1842 unsigned i;
1843 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1844 edge ex;
1845 tree niter = NULL_TREE, aniter;
1846 struct tree_niter_desc desc;
1847
1848 *exit = NULL;
1849 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
1850 {
1851 if (!just_once_each_iteration_p (loop, ex->src))
1852 continue;
1853
1854 if (!number_of_iterations_exit (loop, ex, &desc, false))
1855 continue;
1856
1857 if (integer_nonzerop (desc.may_be_zero))
1858 {
1859 /* We exit in the first iteration through this exit.
1860 We won't find anything better. */
1861 niter = build_int_cst (unsigned_type_node, 0);
1862 *exit = ex;
1863 break;
1864 }
1865
1866 if (!integer_zerop (desc.may_be_zero))
1867 continue;
1868
1869 aniter = desc.niter;
1870
1871 if (!niter)
1872 {
1873 /* Nothing recorded yet. */
1874 niter = aniter;
1875 *exit = ex;
1876 continue;
1877 }
1878
1879 /* Prefer constants, the lower the better. */
1880 if (TREE_CODE (aniter) != INTEGER_CST)
1881 continue;
1882
1883 if (TREE_CODE (niter) != INTEGER_CST)
1884 {
1885 niter = aniter;
1886 *exit = ex;
1887 continue;
1888 }
1889
1890 if (tree_int_cst_lt (aniter, niter))
1891 {
1892 niter = aniter;
1893 *exit = ex;
1894 continue;
1895 }
1896 }
1897 VEC_free (edge, heap, exits);
1898
1899 return niter ? niter : chrec_dont_know;
1900 }
1901
1902 /*
1903
1904 Analysis of a number of iterations of a loop by a brute-force evaluation.
1905
1906 */
1907
1908 /* Bound on the number of iterations we try to evaluate. */
1909
1910 #define MAX_ITERATIONS_TO_TRACK \
1911 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
1912
1913 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
1914 result by a chain of operations such that all but exactly one of their
1915 operands are constants. */
1916
1917 static tree
1918 chain_of_csts_start (struct loop *loop, tree x)
1919 {
1920 tree stmt = SSA_NAME_DEF_STMT (x);
1921 tree use;
1922 basic_block bb = bb_for_stmt (stmt);
1923
1924 if (!bb
1925 || !flow_bb_inside_loop_p (loop, bb))
1926 return NULL_TREE;
1927
1928 if (TREE_CODE (stmt) == PHI_NODE)
1929 {
1930 if (bb == loop->header)
1931 return stmt;
1932
1933 return NULL_TREE;
1934 }
1935
1936 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
1937 return NULL_TREE;
1938
1939 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
1940 return NULL_TREE;
1941 if (SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF) == NULL_DEF_OPERAND_P)
1942 return NULL_TREE;
1943
1944 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
1945 if (use == NULL_USE_OPERAND_P)
1946 return NULL_TREE;
1947
1948 return chain_of_csts_start (loop, use);
1949 }
1950
1951 /* Determines whether the expression X is derived from a result of a phi node
1952 in header of LOOP such that
1953
1954 * the derivation of X consists only from operations with constants
1955 * the initial value of the phi node is constant
1956 * the value of the phi node in the next iteration can be derived from the
1957 value in the current iteration by a chain of operations with constants.
1958
1959 If such phi node exists, it is returned. If X is a constant, X is returned
1960 unchanged. Otherwise NULL_TREE is returned. */
1961
1962 static tree
1963 get_base_for (struct loop *loop, tree x)
1964 {
1965 tree phi, init, next;
1966
1967 if (is_gimple_min_invariant (x))
1968 return x;
1969
1970 phi = chain_of_csts_start (loop, x);
1971 if (!phi)
1972 return NULL_TREE;
1973
1974 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
1975 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
1976
1977 if (TREE_CODE (next) != SSA_NAME)
1978 return NULL_TREE;
1979
1980 if (!is_gimple_min_invariant (init))
1981 return NULL_TREE;
1982
1983 if (chain_of_csts_start (loop, next) != phi)
1984 return NULL_TREE;
1985
1986 return phi;
1987 }
1988
1989 /* Given an expression X, then
1990
1991 * if X is NULL_TREE, we return the constant BASE.
1992 * otherwise X is a SSA name, whose value in the considered loop is derived
1993 by a chain of operations with constant from a result of a phi node in
1994 the header of the loop. Then we return value of X when the value of the
1995 result of this phi node is given by the constant BASE. */
1996
1997 static tree
1998 get_val_for (tree x, tree base)
1999 {
2000 tree stmt, nx, val;
2001 use_operand_p op;
2002 ssa_op_iter iter;
2003
2004 gcc_assert (is_gimple_min_invariant (base));
2005
2006 if (!x)
2007 return base;
2008
2009 stmt = SSA_NAME_DEF_STMT (x);
2010 if (TREE_CODE (stmt) == PHI_NODE)
2011 return base;
2012
2013 FOR_EACH_SSA_USE_OPERAND (op, stmt, iter, SSA_OP_USE)
2014 {
2015 nx = USE_FROM_PTR (op);
2016 val = get_val_for (nx, base);
2017 SET_USE (op, val);
2018 val = fold (GIMPLE_STMT_OPERAND (stmt, 1));
2019 SET_USE (op, nx);
2020 /* only iterate loop once. */
2021 return val;
2022 }
2023
2024 /* Should never reach here. */
2025 gcc_unreachable ();
2026 }
2027
2028 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2029 by brute force -- i.e. by determining the value of the operands of the
2030 condition at EXIT in first few iterations of the loop (assuming that
2031 these values are constant) and determining the first one in that the
2032 condition is not satisfied. Returns the constant giving the number
2033 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2034
2035 tree
2036 loop_niter_by_eval (struct loop *loop, edge exit)
2037 {
2038 tree cond, cnd, acnd;
2039 tree op[2], val[2], next[2], aval[2], phi[2];
2040 unsigned i, j;
2041 enum tree_code cmp;
2042
2043 cond = last_stmt (exit->src);
2044 if (!cond || TREE_CODE (cond) != COND_EXPR)
2045 return chrec_dont_know;
2046
2047 cnd = COND_EXPR_COND (cond);
2048 if (exit->flags & EDGE_TRUE_VALUE)
2049 cnd = invert_truthvalue (cnd);
2050
2051 cmp = TREE_CODE (cnd);
2052 switch (cmp)
2053 {
2054 case EQ_EXPR:
2055 case NE_EXPR:
2056 case GT_EXPR:
2057 case GE_EXPR:
2058 case LT_EXPR:
2059 case LE_EXPR:
2060 for (j = 0; j < 2; j++)
2061 op[j] = TREE_OPERAND (cnd, j);
2062 break;
2063
2064 default:
2065 return chrec_dont_know;
2066 }
2067
2068 for (j = 0; j < 2; j++)
2069 {
2070 phi[j] = get_base_for (loop, op[j]);
2071 if (!phi[j])
2072 return chrec_dont_know;
2073 }
2074
2075 for (j = 0; j < 2; j++)
2076 {
2077 if (TREE_CODE (phi[j]) == PHI_NODE)
2078 {
2079 val[j] = PHI_ARG_DEF_FROM_EDGE (phi[j], loop_preheader_edge (loop));
2080 next[j] = PHI_ARG_DEF_FROM_EDGE (phi[j], loop_latch_edge (loop));
2081 }
2082 else
2083 {
2084 val[j] = phi[j];
2085 next[j] = NULL_TREE;
2086 op[j] = NULL_TREE;
2087 }
2088 }
2089
2090 /* Don't issue signed overflow warnings. */
2091 fold_defer_overflow_warnings ();
2092
2093 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
2094 {
2095 for (j = 0; j < 2; j++)
2096 aval[j] = get_val_for (op[j], val[j]);
2097
2098 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
2099 if (acnd && integer_zerop (acnd))
2100 {
2101 fold_undefer_and_ignore_overflow_warnings ();
2102 if (dump_file && (dump_flags & TDF_DETAILS))
2103 fprintf (dump_file,
2104 "Proved that loop %d iterates %d times using brute force.\n",
2105 loop->num, i);
2106 return build_int_cst (unsigned_type_node, i);
2107 }
2108
2109 for (j = 0; j < 2; j++)
2110 {
2111 val[j] = get_val_for (next[j], val[j]);
2112 if (!is_gimple_min_invariant (val[j]))
2113 {
2114 fold_undefer_and_ignore_overflow_warnings ();
2115 return chrec_dont_know;
2116 }
2117 }
2118 }
2119
2120 fold_undefer_and_ignore_overflow_warnings ();
2121
2122 return chrec_dont_know;
2123 }
2124
2125 /* Finds the exit of the LOOP by that the loop exits after a constant
2126 number of iterations and stores the exit edge to *EXIT. The constant
2127 giving the number of iterations of LOOP is returned. The number of
2128 iterations is determined using loop_niter_by_eval (i.e. by brute force
2129 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2130 determines the number of iterations, chrec_dont_know is returned. */
2131
2132 tree
2133 find_loop_niter_by_eval (struct loop *loop, edge *exit)
2134 {
2135 unsigned i;
2136 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
2137 edge ex;
2138 tree niter = NULL_TREE, aniter;
2139
2140 *exit = NULL;
2141 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
2142 {
2143 if (!just_once_each_iteration_p (loop, ex->src))
2144 continue;
2145
2146 aniter = loop_niter_by_eval (loop, ex);
2147 if (chrec_contains_undetermined (aniter))
2148 continue;
2149
2150 if (niter
2151 && !tree_int_cst_lt (aniter, niter))
2152 continue;
2153
2154 niter = aniter;
2155 *exit = ex;
2156 }
2157 VEC_free (edge, heap, exits);
2158
2159 return niter ? niter : chrec_dont_know;
2160 }
2161
2162 /*
2163
2164 Analysis of upper bounds on number of iterations of a loop.
2165
2166 */
2167
2168 /* Returns a constant upper bound on the value of expression VAL. VAL
2169 is considered to be unsigned. If its type is signed, its value must
2170 be nonnegative. */
2171
2172 static double_int
2173 derive_constant_upper_bound (tree val)
2174 {
2175 tree type = TREE_TYPE (val);
2176 tree op0, op1, subtype, maxt;
2177 double_int bnd, max, mmax, cst;
2178 tree stmt;
2179
2180 if (INTEGRAL_TYPE_P (type))
2181 maxt = TYPE_MAX_VALUE (type);
2182 else
2183 maxt = upper_bound_in_type (type, type);
2184
2185 max = tree_to_double_int (maxt);
2186
2187 switch (TREE_CODE (val))
2188 {
2189 case INTEGER_CST:
2190 return tree_to_double_int (val);
2191
2192 case NOP_EXPR:
2193 case CONVERT_EXPR:
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))
2201 {
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;
2206 }
2207
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);
2211
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;
2216
2217 return bnd;
2218
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);
2224
2225 if (TREE_CODE (op1) != INTEGER_CST
2226 || !tree_expr_nonnegative_p (op0))
2227 return max;
2228
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);
2236
2237 bnd = derive_constant_upper_bound (op0);
2238
2239 if (double_int_negative_p (cst))
2240 {
2241 cst = double_int_neg (cst);
2242 /* Avoid CST == 0x80000... */
2243 if (double_int_negative_p (cst))
2244 return max;;
2245
2246 /* OP0 + CST. We need to check that
2247 BND <= MAX (type) - CST. */
2248
2249 mmax = double_int_add (max, double_int_neg (cst));
2250 if (double_int_ucmp (bnd, mmax) > 0)
2251 return max;
2252
2253 return double_int_add (bnd, cst);
2254 }
2255 else
2256 {
2257 /* OP0 - CST, where CST >= 0.
2258
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.
2262
2263 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2264 otherwise the operation underflows.
2265 */
2266
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;
2272
2273 if (TYPE_UNSIGNED (type))
2274 {
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;
2279 }
2280
2281 bnd = double_int_add (bnd, double_int_neg (cst));
2282 }
2283
2284 return bnd;
2285
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;
2293
2294 bnd = derive_constant_upper_bound (op0);
2295 return double_int_udiv (bnd, tree_to_double_int (op1), FLOOR_DIV_EXPR);
2296
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);
2303
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));
2310
2311 default:
2312 return max;
2313 }
2314 }
2315
2316 /* Records that every statement in LOOP is executed I_BOUND times.
2317 REALISTIC is true if I_BOUND is expected to be close the the real number
2318 of iterations. UPPER is true if we are sure the loop iterates at most
2319 I_BOUND times. */
2320
2321 static void
2322 record_niter_bound (struct loop *loop, double_int i_bound, bool realistic,
2323 bool upper)
2324 {
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))
2330 {
2331 loop->any_upper_bound = true;
2332 loop->nb_iterations_upper_bound = i_bound;
2333 }
2334 if (realistic
2335 && (!loop->any_estimate
2336 || double_int_ucmp (i_bound, loop->nb_iterations_estimate) < 0))
2337 {
2338 loop->any_estimate = true;
2339 loop->nb_iterations_estimate = i_bound;
2340 }
2341 }
2342
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 the 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. */
2349
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)
2353 {
2354 double_int delta;
2355 edge exit;
2356
2357 if (dump_file && (dump_flags & TDF_DETAILS))
2358 {
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);
2367 }
2368
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;
2375
2376 /* If we have a guaranteed upper bound, record it in the appropriate
2377 list. */
2378 if (upper)
2379 {
2380 struct nb_iter_bound *elt = GGC_NEW (struct nb_iter_bound);
2381
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;
2387 }
2388
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);
2403
2404 /* If an overflow occurred, ignore the result. */
2405 if (double_int_ucmp (i_bound, delta) < 0)
2406 return;
2407
2408 record_niter_bound (loop, i_bound, realistic, upper);
2409 }
2410
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. */
2416
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)
2420 {
2421 tree niter_bound, extreme, delta;
2422 tree type = TREE_TYPE (base), unsigned_type;
2423 double_int max;
2424
2425 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step))
2426 return;
2427
2428 if (dump_file && (dump_flags & TDF_DETAILS))
2429 {
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);
2439 }
2440
2441 unsigned_type = unsigned_type_for (type);
2442 base = fold_convert (unsigned_type, base);
2443 step = fold_convert (unsigned_type, step);
2444
2445 if (tree_int_cst_sign_bit (step))
2446 {
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);
2452 }
2453 else
2454 {
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);
2459 }
2460
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);
2466 }
2467
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. */
2471
2472 static bool
2473 array_at_struct_end_p (tree ref)
2474 {
2475 tree base = get_base_address (ref);
2476 tree parent, field;
2477
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;
2482
2483 for (;handled_component_p (ref); ref = parent)
2484 {
2485 parent = TREE_OPERAND (ref, 0);
2486
2487 if (TREE_CODE (ref) == COMPONENT_REF)
2488 {
2489 /* All fields of a union are at its end. */
2490 if (TREE_CODE (TREE_TYPE (parent)) == UNION_TYPE)
2491 continue;
2492
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;
2497 }
2498
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. */
2504 }
2505
2506 gcc_assert (INDIRECT_REF_P (ref));
2507 return true;
2508 }
2509
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. */
2514
2515 struct ilb_data
2516 {
2517 struct loop *loop;
2518 tree stmt;
2519 bool reliable;
2520 };
2521
2522 static bool
2523 idx_infer_loop_bounds (tree base, tree *idx, void *dta)
2524 {
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;
2530
2531 if (TREE_CODE (base) != ARRAY_REF)
2532 return true;
2533
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))
2538 {
2539 at_end = true;
2540 upper = false;
2541 }
2542
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);
2546
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;
2554
2555 low = array_ref_low_bound (base);
2556 high = array_ref_up_bound (base);
2557
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);
2566
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;
2572
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).
2578
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);
2586
2587 if (sign)
2588 next = fold_binary (PLUS_EXPR, type, low, step);
2589 else
2590 next = fold_binary (PLUS_EXPR, type, high, step);
2591
2592 if (tree_int_cst_compare (low, next) <= 0
2593 && tree_int_cst_compare (next, high) <= 0)
2594 return true;
2595
2596 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, true, upper);
2597 return true;
2598 }
2599
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.*/
2603
2604 static void
2605 infer_loop_bounds_from_ref (struct loop *loop, tree stmt, tree ref,
2606 bool reliable)
2607 {
2608 struct ilb_data data;
2609
2610 data.loop = loop;
2611 data.stmt = stmt;
2612 data.reliable = reliable;
2613 for_each_index (&ref, idx_infer_loop_bounds, &data);
2614 }
2615
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. */
2619
2620 static void
2621 infer_loop_bounds_from_array (struct loop *loop, tree stmt, bool reliable)
2622 {
2623 tree call;
2624
2625 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
2626 {
2627 tree op0 = GIMPLE_STMT_OPERAND (stmt, 0);
2628 tree op1 = GIMPLE_STMT_OPERAND (stmt, 1);
2629
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);
2634
2635 if (REFERENCE_CLASS_P (op1))
2636 infer_loop_bounds_from_ref (loop, stmt, op1, reliable);
2637 }
2638
2639
2640 call = get_call_expr_in (stmt);
2641 if (call)
2642 {
2643 tree arg;
2644 call_expr_arg_iterator iter;
2645
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);
2649 }
2650 }
2651
2652 /* Determine information about number of iterations of a LOOP from the fact
2653 that signed arithmetics in STMT does not overflow. */
2654
2655 static void
2656 infer_loop_bounds_from_signedness (struct loop *loop, tree stmt)
2657 {
2658 tree def, base, step, scev, type, low, high;
2659
2660 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
2661 return;
2662
2663 def = GIMPLE_STMT_OPERAND (stmt, 0);
2664
2665 if (TREE_CODE (def) != SSA_NAME)
2666 return;
2667
2668 type = TREE_TYPE (def);
2669 if (!INTEGRAL_TYPE_P (type)
2670 || !TYPE_OVERFLOW_UNDEFINED (type))
2671 return;
2672
2673 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
2674 if (chrec_contains_undetermined (scev))
2675 return;
2676
2677 base = initial_condition_in_loop_num (scev, loop->num);
2678 step = evolution_part_in_loop_num (scev, loop->num);
2679
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;
2685
2686 low = lower_bound_in_type (type, type);
2687 high = upper_bound_in_type (type, type);
2688
2689 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
2690 }
2691
2692 /* The following analyzers are extracting informations on the bounds
2693 of LOOP from the following undefined behaviors:
2694
2695 - data references should not access elements over the statically
2696 allocated size,
2697
2698 - signed variables should not overflow when flag_wrapv is not set.
2699 */
2700
2701 static void
2702 infer_loop_bounds_from_undefined (struct loop *loop)
2703 {
2704 unsigned i;
2705 basic_block *bbs;
2706 block_stmt_iterator bsi;
2707 basic_block bb;
2708 bool reliable;
2709
2710 bbs = get_loop_body (loop);
2711
2712 for (i = 0; i < loop->num_nodes; i++)
2713 {
2714 bb = bbs[i];
2715
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);
2720
2721 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2722 {
2723 tree stmt = bsi_stmt (bsi);
2724
2725 infer_loop_bounds_from_array (loop, stmt, reliable);
2726
2727 if (reliable)
2728 infer_loop_bounds_from_signedness (loop, stmt);
2729 }
2730
2731 }
2732
2733 free (bbs);
2734 }
2735
2736 /* Converts VAL to double_int. */
2737
2738 static double_int
2739 gcov_type_to_double_int (gcov_type val)
2740 {
2741 double_int ret;
2742
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;
2749
2750 return ret;
2751 }
2752
2753 /* Records estimates on numbers of iterations of LOOP. */
2754
2755 void
2756 estimate_numbers_of_iterations_loop (struct loop *loop)
2757 {
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;
2764
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;
2771
2772 exits = get_loop_exit_edges (loop);
2773 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
2774 {
2775 if (!number_of_iterations_exit (loop, ex, &niter_desc, false))
2776 continue;
2777
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);
2787 }
2788 VEC_free (edge, heap, exits);
2789
2790 infer_loop_bounds_from_undefined (loop);
2791
2792 /* If we have a measured profile, use it to estimate the number of
2793 iterations. */
2794 if (loop->header->count != 0)
2795 {
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);
2799 }
2800
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;
2808 }
2809
2810 /* Records estimates on numbers of iterations of loops. */
2811
2812 void
2813 estimate_numbers_of_iterations (void)
2814 {
2815 loop_iterator li;
2816 struct loop *loop;
2817
2818 /* We don't want to issue signed overflow warnings while getting
2819 loop iteration estimates. */
2820 fold_defer_overflow_warnings ();
2821
2822 FOR_EACH_LOOP (li, loop, 0)
2823 {
2824 estimate_numbers_of_iterations_loop (loop);
2825 }
2826
2827 fold_undefer_and_ignore_overflow_warnings ();
2828 }
2829
2830 /* Returns true if statement S1 dominates statement S2. */
2831
2832 bool
2833 stmt_dominates_stmt_p (tree s1, tree s2)
2834 {
2835 basic_block bb1 = bb_for_stmt (s1), bb2 = bb_for_stmt (s2);
2836
2837 if (!bb1
2838 || s1 == s2)
2839 return true;
2840
2841 if (bb1 == bb2)
2842 {
2843 block_stmt_iterator bsi;
2844
2845 for (bsi = bsi_start (bb1); bsi_stmt (bsi) != s2; bsi_next (&bsi))
2846 if (bsi_stmt (bsi) == s1)
2847 return true;
2848
2849 return false;
2850 }
2851
2852 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
2853 }
2854
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. */
2860
2861 static bool
2862 n_of_executions_at_most (tree stmt,
2863 struct nb_iter_bound *niter_bound,
2864 tree niter)
2865 {
2866 double_int bound = niter_bound->bound;
2867 tree nit_type = TREE_TYPE (niter), e;
2868 enum tree_code cmp;
2869
2870 gcc_assert (TYPE_UNSIGNED (nit_type));
2871
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;
2876
2877 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
2878 times. This means that:
2879
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.
2883
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. */
2890
2891 if (niter_bound->is_exit)
2892 {
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;
2899 }
2900 else
2901 {
2902 if (!stmt
2903 || (bb_for_stmt (stmt) != bb_for_stmt (niter_bound->stmt)
2904 && !stmt_dominates_stmt_p (niter_bound->stmt, stmt)))
2905 {
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;
2910 }
2911 cmp = GT_EXPR;
2912 }
2913
2914 e = fold_binary (cmp, boolean_type_node,
2915 niter, double_int_to_tree (nit_type, bound));
2916 return e && integer_nonzerop (e);
2917 }
2918
2919 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
2920
2921 bool
2922 nowrap_type_p (tree type)
2923 {
2924 if (INTEGRAL_TYPE_P (type)
2925 && TYPE_OVERFLOW_UNDEFINED (type))
2926 return true;
2927
2928 if (POINTER_TYPE_P (type))
2929 return true;
2930
2931 return false;
2932 }
2933
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.
2939
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). */
2943
2944 bool
2945 scev_probably_wraps_p (tree base, tree step,
2946 tree at_stmt, struct loop *loop,
2947 bool use_overflow_semantics)
2948 {
2949 struct nb_iter_bound *bound;
2950 tree delta, step_abs;
2951 tree unsigned_type, valid_niter;
2952 tree type = TREE_TYPE (step);
2953
2954 /* FIXME: We really need something like
2955 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
2956
2957 We used to test for the following situation that frequently appears
2958 during address arithmetics:
2959
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;
2963
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. */
2969
2970 if (chrec_contains_undetermined (base)
2971 || chrec_contains_undetermined (step)
2972 || TREE_CODE (step) != INTEGER_CST)
2973 return true;
2974
2975 if (integer_zerop (step))
2976 return false;
2977
2978 /* If we can use the fact that signed and pointer arithmetics does not
2979 wrap, we are done. */
2980 if (use_overflow_semantics && nowrap_type_p (type))
2981 return false;
2982
2983 /* Don't issue signed overflow warnings. */
2984 fold_defer_overflow_warnings ();
2985
2986 /* Otherwise, compute the number of iterations before we reach the
2987 bound of the type, and verify that the loop is exited before this
2988 occurs. */
2989 unsigned_type = unsigned_type_for (type);
2990 base = fold_convert (unsigned_type, base);
2991
2992 if (tree_int_cst_sign_bit (step))
2993 {
2994 tree extreme = fold_convert (unsigned_type,
2995 lower_bound_in_type (type, type));
2996 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
2997 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type,
2998 fold_convert (unsigned_type, step));
2999 }
3000 else
3001 {
3002 tree extreme = fold_convert (unsigned_type,
3003 upper_bound_in_type (type, type));
3004 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
3005 step_abs = fold_convert (unsigned_type, step);
3006 }
3007
3008 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
3009
3010 estimate_numbers_of_iterations_loop (loop);
3011 for (bound = loop->bounds; bound; bound = bound->next)
3012 {
3013 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
3014 {
3015 fold_undefer_and_ignore_overflow_warnings ();
3016 return false;
3017 }
3018 }
3019
3020 fold_undefer_and_ignore_overflow_warnings ();
3021
3022 /* At this point we still don't have a proof that the iv does not
3023 overflow: give up. */
3024 return true;
3025 }
3026
3027 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
3028
3029 void
3030 free_numbers_of_iterations_estimates_loop (struct loop *loop)
3031 {
3032 struct nb_iter_bound *bound, *next;
3033
3034 loop->nb_iterations = NULL;
3035 loop->estimate_state = EST_NOT_COMPUTED;
3036 for (bound = loop->bounds; bound; bound = next)
3037 {
3038 next = bound->next;
3039 ggc_free (bound);
3040 }
3041
3042 loop->bounds = NULL;
3043 }
3044
3045 /* Frees the information on upper bounds on numbers of iterations of loops. */
3046
3047 void
3048 free_numbers_of_iterations_estimates (void)
3049 {
3050 loop_iterator li;
3051 struct loop *loop;
3052
3053 FOR_EACH_LOOP (li, loop, 0)
3054 {
3055 free_numbers_of_iterations_estimates_loop (loop);
3056 }
3057 }
3058
3059 /* Substitute value VAL for ssa name NAME inside expressions held
3060 at LOOP. */
3061
3062 void
3063 substitute_in_loop_info (struct loop *loop, tree name, tree val)
3064 {
3065 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);
3066 }