tree-ssa-loop-niter.c: Include tree-ssanames.h.
[gcc.git] / gcc / tree-ssa-loop-niter.c
1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004-2013 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "tm_p.h"
26 #include "basic-block.h"
27 #include "gimple-pretty-print.h"
28 #include "intl.h"
29 #include "gimple.h"
30 #include "gimple-ssa.h"
31 #include "tree-cfg.h"
32 #include "tree-phinodes.h"
33 #include "ssa-iterators.h"
34 #include "tree-ssa-loop-ivopts.h"
35 #include "tree-ssa-loop-niter.h"
36 #include "tree-ssa-loop.h"
37 #include "dumpfile.h"
38 #include "cfgloop.h"
39 #include "ggc.h"
40 #include "tree-chrec.h"
41 #include "tree-scalar-evolution.h"
42 #include "tree-data-ref.h"
43 #include "params.h"
44 #include "flags.h"
45 #include "diagnostic-core.h"
46 #include "tree-inline.h"
47 #include "tree-pass.h"
48 #include "tree-ssanames.h"
49
50
51 #define SWAP(X, Y) do { affine_iv *tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
52
53 /* The maximum number of dominator BBs we search for conditions
54 of loop header copies we use for simplifying a conditional
55 expression. */
56 #define MAX_DOMINATORS_TO_WALK 8
57
58 /*
59
60 Analysis of number of iterations of an affine exit test.
61
62 */
63
64 /* Bounds on some value, BELOW <= X <= UP. */
65
66 typedef struct
67 {
68 mpz_t below, up;
69 } bounds;
70
71
72 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
73
74 static void
75 split_to_var_and_offset (tree expr, tree *var, mpz_t offset)
76 {
77 tree type = TREE_TYPE (expr);
78 tree op0, op1;
79 double_int off;
80 bool negate = false;
81
82 *var = expr;
83 mpz_set_ui (offset, 0);
84
85 switch (TREE_CODE (expr))
86 {
87 case MINUS_EXPR:
88 negate = true;
89 /* Fallthru. */
90
91 case PLUS_EXPR:
92 case POINTER_PLUS_EXPR:
93 op0 = TREE_OPERAND (expr, 0);
94 op1 = TREE_OPERAND (expr, 1);
95
96 if (TREE_CODE (op1) != INTEGER_CST)
97 break;
98
99 *var = op0;
100 /* Always sign extend the offset. */
101 off = tree_to_double_int (op1);
102 off = off.sext (TYPE_PRECISION (type));
103 mpz_set_double_int (offset, off, false);
104 if (negate)
105 mpz_neg (offset, offset);
106 break;
107
108 case INTEGER_CST:
109 *var = build_int_cst_type (type, 0);
110 off = tree_to_double_int (expr);
111 mpz_set_double_int (offset, off, TYPE_UNSIGNED (type));
112 break;
113
114 default:
115 break;
116 }
117 }
118
119 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
120 in TYPE to MIN and MAX. */
121
122 static void
123 determine_value_range (struct loop *loop, tree type, tree var, mpz_t off,
124 mpz_t min, mpz_t max)
125 {
126 double_int minv, maxv;
127 enum value_range_type rtype = VR_VARYING;
128
129 /* If the expression is a constant, we know its value exactly. */
130 if (integer_zerop (var))
131 {
132 mpz_set (min, off);
133 mpz_set (max, off);
134 return;
135 }
136
137 get_type_static_bounds (type, min, max);
138
139 /* See if we have some range info from VRP. */
140 if (TREE_CODE (var) == SSA_NAME && INTEGRAL_TYPE_P (type))
141 {
142 edge e = loop_preheader_edge (loop);
143 gimple_stmt_iterator gsi;
144
145 /* Either for VAR itself... */
146 rtype = get_range_info (var, &minv, &maxv);
147 /* Or for PHI results in loop->header where VAR is used as
148 PHI argument from the loop preheader edge. */
149 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi))
150 {
151 gimple phi = gsi_stmt (gsi);
152 double_int minc, maxc;
153 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var
154 && (get_range_info (gimple_phi_result (phi), &minc, &maxc)
155 == VR_RANGE))
156 {
157 if (rtype != VR_RANGE)
158 {
159 rtype = VR_RANGE;
160 minv = minc;
161 maxv = maxc;
162 }
163 else
164 {
165 minv = minv.max (minc, TYPE_UNSIGNED (type));
166 maxv = maxv.min (maxc, TYPE_UNSIGNED (type));
167 gcc_assert (minv.cmp (maxv, TYPE_UNSIGNED (type)) <= 0);
168 }
169 }
170 }
171 if (rtype == VR_RANGE)
172 {
173 mpz_t minm, maxm;
174 gcc_assert (minv.cmp (maxv, TYPE_UNSIGNED (type)) <= 0);
175 mpz_init (minm);
176 mpz_init (maxm);
177 mpz_set_double_int (minm, minv, TYPE_UNSIGNED (type));
178 mpz_set_double_int (maxm, maxv, TYPE_UNSIGNED (type));
179 mpz_add (minm, minm, off);
180 mpz_add (maxm, maxm, off);
181 /* If the computation may not wrap or off is zero, then this
182 is always fine. If off is negative and minv + off isn't
183 smaller than type's minimum, or off is positive and
184 maxv + off isn't bigger than type's maximum, use the more
185 precise range too. */
186 if (nowrap_type_p (type)
187 || mpz_sgn (off) == 0
188 || (mpz_sgn (off) < 0 && mpz_cmp (minm, min) >= 0)
189 || (mpz_sgn (off) > 0 && mpz_cmp (maxm, max) <= 0))
190 {
191 mpz_set (min, minm);
192 mpz_set (max, maxm);
193 mpz_clear (minm);
194 mpz_clear (maxm);
195 return;
196 }
197 mpz_clear (minm);
198 mpz_clear (maxm);
199 }
200 }
201
202 /* If the computation may wrap, we know nothing about the value, except for
203 the range of the type. */
204 if (!nowrap_type_p (type))
205 return;
206
207 /* Since the addition of OFF does not wrap, if OFF is positive, then we may
208 add it to MIN, otherwise to MAX. */
209 if (mpz_sgn (off) < 0)
210 mpz_add (max, max, off);
211 else
212 mpz_add (min, min, off);
213 }
214
215 /* Stores the bounds on the difference of the values of the expressions
216 (var + X) and (var + Y), computed in TYPE, to BNDS. */
217
218 static void
219 bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y,
220 bounds *bnds)
221 {
222 int rel = mpz_cmp (x, y);
223 bool may_wrap = !nowrap_type_p (type);
224 mpz_t m;
225
226 /* If X == Y, then the expressions are always equal.
227 If X > Y, there are the following possibilities:
228 a) neither of var + X and var + Y overflow or underflow, or both of
229 them do. Then their difference is X - Y.
230 b) var + X overflows, and var + Y does not. Then the values of the
231 expressions are var + X - M and var + Y, where M is the range of
232 the type, and their difference is X - Y - M.
233 c) var + Y underflows and var + X does not. Their difference again
234 is M - X + Y.
235 Therefore, if the arithmetics in type does not overflow, then the
236 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y)
237 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or
238 (X - Y, X - Y + M). */
239
240 if (rel == 0)
241 {
242 mpz_set_ui (bnds->below, 0);
243 mpz_set_ui (bnds->up, 0);
244 return;
245 }
246
247 mpz_init (m);
248 mpz_set_double_int (m, double_int::mask (TYPE_PRECISION (type)), true);
249 mpz_add_ui (m, m, 1);
250 mpz_sub (bnds->up, x, y);
251 mpz_set (bnds->below, bnds->up);
252
253 if (may_wrap)
254 {
255 if (rel > 0)
256 mpz_sub (bnds->below, bnds->below, m);
257 else
258 mpz_add (bnds->up, bnds->up, m);
259 }
260
261 mpz_clear (m);
262 }
263
264 /* From condition C0 CMP C1 derives information regarding the
265 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE,
266 and stores it to BNDS. */
267
268 static void
269 refine_bounds_using_guard (tree type, tree varx, mpz_t offx,
270 tree vary, mpz_t offy,
271 tree c0, enum tree_code cmp, tree c1,
272 bounds *bnds)
273 {
274 tree varc0, varc1, tmp, ctype;
275 mpz_t offc0, offc1, loffx, loffy, bnd;
276 bool lbound = false;
277 bool no_wrap = nowrap_type_p (type);
278 bool x_ok, y_ok;
279
280 switch (cmp)
281 {
282 case LT_EXPR:
283 case LE_EXPR:
284 case GT_EXPR:
285 case GE_EXPR:
286 STRIP_SIGN_NOPS (c0);
287 STRIP_SIGN_NOPS (c1);
288 ctype = TREE_TYPE (c0);
289 if (!useless_type_conversion_p (ctype, type))
290 return;
291
292 break;
293
294 case EQ_EXPR:
295 /* We could derive quite precise information from EQ_EXPR, however, such
296 a guard is unlikely to appear, so we do not bother with handling
297 it. */
298 return;
299
300 case NE_EXPR:
301 /* NE_EXPR comparisons do not contain much of useful information, except for
302 special case of comparing with the bounds of the type. */
303 if (TREE_CODE (c1) != INTEGER_CST
304 || !INTEGRAL_TYPE_P (type))
305 return;
306
307 /* Ensure that the condition speaks about an expression in the same type
308 as X and Y. */
309 ctype = TREE_TYPE (c0);
310 if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type))
311 return;
312 c0 = fold_convert (type, c0);
313 c1 = fold_convert (type, c1);
314
315 if (TYPE_MIN_VALUE (type)
316 && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0))
317 {
318 cmp = GT_EXPR;
319 break;
320 }
321 if (TYPE_MAX_VALUE (type)
322 && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0))
323 {
324 cmp = LT_EXPR;
325 break;
326 }
327
328 return;
329 default:
330 return;
331 }
332
333 mpz_init (offc0);
334 mpz_init (offc1);
335 split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0);
336 split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1);
337
338 /* We are only interested in comparisons of expressions based on VARX and
339 VARY. TODO -- we might also be able to derive some bounds from
340 expressions containing just one of the variables. */
341
342 if (operand_equal_p (varx, varc1, 0))
343 {
344 tmp = varc0; varc0 = varc1; varc1 = tmp;
345 mpz_swap (offc0, offc1);
346 cmp = swap_tree_comparison (cmp);
347 }
348
349 if (!operand_equal_p (varx, varc0, 0)
350 || !operand_equal_p (vary, varc1, 0))
351 goto end;
352
353 mpz_init_set (loffx, offx);
354 mpz_init_set (loffy, offy);
355
356 if (cmp == GT_EXPR || cmp == GE_EXPR)
357 {
358 tmp = varx; varx = vary; vary = tmp;
359 mpz_swap (offc0, offc1);
360 mpz_swap (loffx, loffy);
361 cmp = swap_tree_comparison (cmp);
362 lbound = true;
363 }
364
365 /* If there is no overflow, the condition implies that
366
367 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
368
369 The overflows and underflows may complicate things a bit; each
370 overflow decreases the appropriate offset by M, and underflow
371 increases it by M. The above inequality would not necessarily be
372 true if
373
374 -- VARX + OFFX underflows and VARX + OFFC0 does not, or
375 VARX + OFFC0 overflows, but VARX + OFFX does not.
376 This may only happen if OFFX < OFFC0.
377 -- VARY + OFFY overflows and VARY + OFFC1 does not, or
378 VARY + OFFC1 underflows and VARY + OFFY does not.
379 This may only happen if OFFY > OFFC1. */
380
381 if (no_wrap)
382 {
383 x_ok = true;
384 y_ok = true;
385 }
386 else
387 {
388 x_ok = (integer_zerop (varx)
389 || mpz_cmp (loffx, offc0) >= 0);
390 y_ok = (integer_zerop (vary)
391 || mpz_cmp (loffy, offc1) <= 0);
392 }
393
394 if (x_ok && y_ok)
395 {
396 mpz_init (bnd);
397 mpz_sub (bnd, loffx, loffy);
398 mpz_add (bnd, bnd, offc1);
399 mpz_sub (bnd, bnd, offc0);
400
401 if (cmp == LT_EXPR)
402 mpz_sub_ui (bnd, bnd, 1);
403
404 if (lbound)
405 {
406 mpz_neg (bnd, bnd);
407 if (mpz_cmp (bnds->below, bnd) < 0)
408 mpz_set (bnds->below, bnd);
409 }
410 else
411 {
412 if (mpz_cmp (bnd, bnds->up) < 0)
413 mpz_set (bnds->up, bnd);
414 }
415 mpz_clear (bnd);
416 }
417
418 mpz_clear (loffx);
419 mpz_clear (loffy);
420 end:
421 mpz_clear (offc0);
422 mpz_clear (offc1);
423 }
424
425 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS.
426 The subtraction is considered to be performed in arbitrary precision,
427 without overflows.
428
429 We do not attempt to be too clever regarding the value ranges of X and
430 Y; most of the time, they are just integers or ssa names offsetted by
431 integer. However, we try to use the information contained in the
432 comparisons before the loop (usually created by loop header copying). */
433
434 static void
435 bound_difference (struct loop *loop, tree x, tree y, bounds *bnds)
436 {
437 tree type = TREE_TYPE (x);
438 tree varx, vary;
439 mpz_t offx, offy;
440 mpz_t minx, maxx, miny, maxy;
441 int cnt = 0;
442 edge e;
443 basic_block bb;
444 tree c0, c1;
445 gimple cond;
446 enum tree_code cmp;
447
448 /* Get rid of unnecessary casts, but preserve the value of
449 the expressions. */
450 STRIP_SIGN_NOPS (x);
451 STRIP_SIGN_NOPS (y);
452
453 mpz_init (bnds->below);
454 mpz_init (bnds->up);
455 mpz_init (offx);
456 mpz_init (offy);
457 split_to_var_and_offset (x, &varx, offx);
458 split_to_var_and_offset (y, &vary, offy);
459
460 if (!integer_zerop (varx)
461 && operand_equal_p (varx, vary, 0))
462 {
463 /* Special case VARX == VARY -- we just need to compare the
464 offsets. The matters are a bit more complicated in the
465 case addition of offsets may wrap. */
466 bound_difference_of_offsetted_base (type, offx, offy, bnds);
467 }
468 else
469 {
470 /* Otherwise, use the value ranges to determine the initial
471 estimates on below and up. */
472 mpz_init (minx);
473 mpz_init (maxx);
474 mpz_init (miny);
475 mpz_init (maxy);
476 determine_value_range (loop, type, varx, offx, minx, maxx);
477 determine_value_range (loop, type, vary, offy, miny, maxy);
478
479 mpz_sub (bnds->below, minx, maxy);
480 mpz_sub (bnds->up, maxx, miny);
481 mpz_clear (minx);
482 mpz_clear (maxx);
483 mpz_clear (miny);
484 mpz_clear (maxy);
485 }
486
487 /* If both X and Y are constants, we cannot get any more precise. */
488 if (integer_zerop (varx) && integer_zerop (vary))
489 goto end;
490
491 /* Now walk the dominators of the loop header and use the entry
492 guards to refine the estimates. */
493 for (bb = loop->header;
494 bb != ENTRY_BLOCK_PTR && cnt < MAX_DOMINATORS_TO_WALK;
495 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
496 {
497 if (!single_pred_p (bb))
498 continue;
499 e = single_pred_edge (bb);
500
501 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
502 continue;
503
504 cond = last_stmt (e->src);
505 c0 = gimple_cond_lhs (cond);
506 cmp = gimple_cond_code (cond);
507 c1 = gimple_cond_rhs (cond);
508
509 if (e->flags & EDGE_FALSE_VALUE)
510 cmp = invert_tree_comparison (cmp, false);
511
512 refine_bounds_using_guard (type, varx, offx, vary, offy,
513 c0, cmp, c1, bnds);
514 ++cnt;
515 }
516
517 end:
518 mpz_clear (offx);
519 mpz_clear (offy);
520 }
521
522 /* Update the bounds in BNDS that restrict the value of X to the bounds
523 that restrict the value of X + DELTA. X can be obtained as a
524 difference of two values in TYPE. */
525
526 static void
527 bounds_add (bounds *bnds, double_int delta, tree type)
528 {
529 mpz_t mdelta, max;
530
531 mpz_init (mdelta);
532 mpz_set_double_int (mdelta, delta, false);
533
534 mpz_init (max);
535 mpz_set_double_int (max, double_int::mask (TYPE_PRECISION (type)), true);
536
537 mpz_add (bnds->up, bnds->up, mdelta);
538 mpz_add (bnds->below, bnds->below, mdelta);
539
540 if (mpz_cmp (bnds->up, max) > 0)
541 mpz_set (bnds->up, max);
542
543 mpz_neg (max, max);
544 if (mpz_cmp (bnds->below, max) < 0)
545 mpz_set (bnds->below, max);
546
547 mpz_clear (mdelta);
548 mpz_clear (max);
549 }
550
551 /* Update the bounds in BNDS that restrict the value of X to the bounds
552 that restrict the value of -X. */
553
554 static void
555 bounds_negate (bounds *bnds)
556 {
557 mpz_t tmp;
558
559 mpz_init_set (tmp, bnds->up);
560 mpz_neg (bnds->up, bnds->below);
561 mpz_neg (bnds->below, tmp);
562 mpz_clear (tmp);
563 }
564
565 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
566
567 static tree
568 inverse (tree x, tree mask)
569 {
570 tree type = TREE_TYPE (x);
571 tree rslt;
572 unsigned ctr = tree_floor_log2 (mask);
573
574 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT)
575 {
576 unsigned HOST_WIDE_INT ix;
577 unsigned HOST_WIDE_INT imask;
578 unsigned HOST_WIDE_INT irslt = 1;
579
580 gcc_assert (cst_and_fits_in_hwi (x));
581 gcc_assert (cst_and_fits_in_hwi (mask));
582
583 ix = int_cst_value (x);
584 imask = int_cst_value (mask);
585
586 for (; ctr; ctr--)
587 {
588 irslt *= ix;
589 ix *= ix;
590 }
591 irslt &= imask;
592
593 rslt = build_int_cst_type (type, irslt);
594 }
595 else
596 {
597 rslt = build_int_cst (type, 1);
598 for (; ctr; ctr--)
599 {
600 rslt = int_const_binop (MULT_EXPR, rslt, x);
601 x = int_const_binop (MULT_EXPR, x, x);
602 }
603 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask);
604 }
605
606 return rslt;
607 }
608
609 /* Derives the upper bound BND on the number of executions of loop with exit
610 condition S * i <> C. If NO_OVERFLOW is true, then the control variable of
611 the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed
612 that the loop ends through this exit, i.e., the induction variable ever
613 reaches the value of C.
614
615 The value C is equal to final - base, where final and base are the final and
616 initial value of the actual induction variable in the analysed loop. BNDS
617 bounds the value of this difference when computed in signed type with
618 unbounded range, while the computation of C is performed in an unsigned
619 type with the range matching the range of the type of the induction variable.
620 In particular, BNDS.up contains an upper bound on C in the following cases:
621 -- if the iv must reach its final value without overflow, i.e., if
622 NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or
623 -- if final >= base, which we know to hold when BNDS.below >= 0. */
624
625 static void
626 number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s,
627 bounds *bnds, bool exit_must_be_taken)
628 {
629 double_int max;
630 mpz_t d;
631 tree type = TREE_TYPE (c);
632 bool bnds_u_valid = ((no_overflow && exit_must_be_taken)
633 || mpz_sgn (bnds->below) >= 0);
634
635 if (integer_onep (s)
636 || (TREE_CODE (c) == INTEGER_CST
637 && TREE_CODE (s) == INTEGER_CST
638 && tree_to_double_int (c).mod (tree_to_double_int (s),
639 TYPE_UNSIGNED (type),
640 EXACT_DIV_EXPR).is_zero ())
641 || (TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (c))
642 && multiple_of_p (type, c, s)))
643 {
644 /* If C is an exact multiple of S, then its value will be reached before
645 the induction variable overflows (unless the loop is exited in some
646 other way before). Note that the actual induction variable in the
647 loop (which ranges from base to final instead of from 0 to C) may
648 overflow, in which case BNDS.up will not be giving a correct upper
649 bound on C; thus, BNDS_U_VALID had to be computed in advance. */
650 no_overflow = true;
651 exit_must_be_taken = true;
652 }
653
654 /* If the induction variable can overflow, the number of iterations is at
655 most the period of the control variable (or infinite, but in that case
656 the whole # of iterations analysis will fail). */
657 if (!no_overflow)
658 {
659 max = double_int::mask (TYPE_PRECISION (type)
660 - tree_low_cst (num_ending_zeros (s), 1));
661 mpz_set_double_int (bnd, max, true);
662 return;
663 }
664
665 /* Now we know that the induction variable does not overflow, so the loop
666 iterates at most (range of type / S) times. */
667 mpz_set_double_int (bnd, double_int::mask (TYPE_PRECISION (type)), true);
668
669 /* If the induction variable is guaranteed to reach the value of C before
670 overflow, ... */
671 if (exit_must_be_taken)
672 {
673 /* ... then we can strengthen this to C / S, and possibly we can use
674 the upper bound on C given by BNDS. */
675 if (TREE_CODE (c) == INTEGER_CST)
676 mpz_set_double_int (bnd, tree_to_double_int (c), true);
677 else if (bnds_u_valid)
678 mpz_set (bnd, bnds->up);
679 }
680
681 mpz_init (d);
682 mpz_set_double_int (d, tree_to_double_int (s), true);
683 mpz_fdiv_q (bnd, bnd, d);
684 mpz_clear (d);
685 }
686
687 /* Determines number of iterations of loop whose ending condition
688 is IV <> FINAL. TYPE is the type of the iv. The number of
689 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
690 we know that the exit must be taken eventually, i.e., that the IV
691 ever reaches the value FINAL (we derived this earlier, and possibly set
692 NITER->assumptions to make sure this is the case). BNDS contains the
693 bounds on the difference FINAL - IV->base. */
694
695 static bool
696 number_of_iterations_ne (tree type, affine_iv *iv, tree final,
697 struct tree_niter_desc *niter, bool exit_must_be_taken,
698 bounds *bnds)
699 {
700 tree niter_type = unsigned_type_for (type);
701 tree s, c, d, bits, assumption, tmp, bound;
702 mpz_t max;
703
704 niter->control = *iv;
705 niter->bound = final;
706 niter->cmp = NE_EXPR;
707
708 /* Rearrange the terms so that we get inequality S * i <> C, with S
709 positive. Also cast everything to the unsigned type. If IV does
710 not overflow, BNDS bounds the value of C. Also, this is the
711 case if the computation |FINAL - IV->base| does not overflow, i.e.,
712 if BNDS->below in the result is nonnegative. */
713 if (tree_int_cst_sign_bit (iv->step))
714 {
715 s = fold_convert (niter_type,
716 fold_build1 (NEGATE_EXPR, type, iv->step));
717 c = fold_build2 (MINUS_EXPR, niter_type,
718 fold_convert (niter_type, iv->base),
719 fold_convert (niter_type, final));
720 bounds_negate (bnds);
721 }
722 else
723 {
724 s = fold_convert (niter_type, iv->step);
725 c = fold_build2 (MINUS_EXPR, niter_type,
726 fold_convert (niter_type, final),
727 fold_convert (niter_type, iv->base));
728 }
729
730 mpz_init (max);
731 number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds,
732 exit_must_be_taken);
733 niter->max = mpz_get_double_int (niter_type, max, false);
734 mpz_clear (max);
735
736 /* First the trivial cases -- when the step is 1. */
737 if (integer_onep (s))
738 {
739 niter->niter = c;
740 return true;
741 }
742
743 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
744 is infinite. Otherwise, the number of iterations is
745 (inverse(s/d) * (c/d)) mod (size of mode/d). */
746 bits = num_ending_zeros (s);
747 bound = build_low_bits_mask (niter_type,
748 (TYPE_PRECISION (niter_type)
749 - tree_low_cst (bits, 1)));
750
751 d = fold_binary_to_constant (LSHIFT_EXPR, niter_type,
752 build_int_cst (niter_type, 1), bits);
753 s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits);
754
755 if (!exit_must_be_taken)
756 {
757 /* If we cannot assume that the exit is taken eventually, record the
758 assumptions for divisibility of c. */
759 assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d);
760 assumption = fold_build2 (EQ_EXPR, boolean_type_node,
761 assumption, build_int_cst (niter_type, 0));
762 if (!integer_nonzerop (assumption))
763 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
764 niter->assumptions, assumption);
765 }
766
767 c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d);
768 tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound));
769 niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound);
770 return true;
771 }
772
773 /* Checks whether we can determine the final value of the control variable
774 of the loop with ending condition IV0 < IV1 (computed in TYPE).
775 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
776 of the step. The assumptions necessary to ensure that the computation
777 of the final value does not overflow are recorded in NITER. If we
778 find the final value, we adjust DELTA and return TRUE. Otherwise
779 we return false. BNDS bounds the value of IV1->base - IV0->base,
780 and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is
781 true if we know that the exit must be taken eventually. */
782
783 static bool
784 number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1,
785 struct tree_niter_desc *niter,
786 tree *delta, tree step,
787 bool exit_must_be_taken, bounds *bnds)
788 {
789 tree niter_type = TREE_TYPE (step);
790 tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step);
791 tree tmod;
792 mpz_t mmod;
793 tree assumption = boolean_true_node, bound, noloop;
794 bool ret = false, fv_comp_no_overflow;
795 tree type1 = type;
796 if (POINTER_TYPE_P (type))
797 type1 = sizetype;
798
799 if (TREE_CODE (mod) != INTEGER_CST)
800 return false;
801 if (integer_nonzerop (mod))
802 mod = fold_build2 (MINUS_EXPR, niter_type, step, mod);
803 tmod = fold_convert (type1, mod);
804
805 mpz_init (mmod);
806 mpz_set_double_int (mmod, tree_to_double_int (mod), true);
807 mpz_neg (mmod, mmod);
808
809 /* If the induction variable does not overflow and the exit is taken,
810 then the computation of the final value does not overflow. This is
811 also obviously the case if the new final value is equal to the
812 current one. Finally, we postulate this for pointer type variables,
813 as the code cannot rely on the object to that the pointer points being
814 placed at the end of the address space (and more pragmatically,
815 TYPE_{MIN,MAX}_VALUE is not defined for pointers). */
816 if (integer_zerop (mod) || POINTER_TYPE_P (type))
817 fv_comp_no_overflow = true;
818 else if (!exit_must_be_taken)
819 fv_comp_no_overflow = false;
820 else
821 fv_comp_no_overflow =
822 (iv0->no_overflow && integer_nonzerop (iv0->step))
823 || (iv1->no_overflow && integer_nonzerop (iv1->step));
824
825 if (integer_nonzerop (iv0->step))
826 {
827 /* The final value of the iv is iv1->base + MOD, assuming that this
828 computation does not overflow, and that
829 iv0->base <= iv1->base + MOD. */
830 if (!fv_comp_no_overflow)
831 {
832 bound = fold_build2 (MINUS_EXPR, type1,
833 TYPE_MAX_VALUE (type1), tmod);
834 assumption = fold_build2 (LE_EXPR, boolean_type_node,
835 iv1->base, bound);
836 if (integer_zerop (assumption))
837 goto end;
838 }
839 if (mpz_cmp (mmod, bnds->below) < 0)
840 noloop = boolean_false_node;
841 else if (POINTER_TYPE_P (type))
842 noloop = fold_build2 (GT_EXPR, boolean_type_node,
843 iv0->base,
844 fold_build_pointer_plus (iv1->base, tmod));
845 else
846 noloop = fold_build2 (GT_EXPR, boolean_type_node,
847 iv0->base,
848 fold_build2 (PLUS_EXPR, type1,
849 iv1->base, tmod));
850 }
851 else
852 {
853 /* The final value of the iv is iv0->base - MOD, assuming that this
854 computation does not overflow, and that
855 iv0->base - MOD <= iv1->base. */
856 if (!fv_comp_no_overflow)
857 {
858 bound = fold_build2 (PLUS_EXPR, type1,
859 TYPE_MIN_VALUE (type1), tmod);
860 assumption = fold_build2 (GE_EXPR, boolean_type_node,
861 iv0->base, bound);
862 if (integer_zerop (assumption))
863 goto end;
864 }
865 if (mpz_cmp (mmod, bnds->below) < 0)
866 noloop = boolean_false_node;
867 else if (POINTER_TYPE_P (type))
868 noloop = fold_build2 (GT_EXPR, boolean_type_node,
869 fold_build_pointer_plus (iv0->base,
870 fold_build1 (NEGATE_EXPR,
871 type1, tmod)),
872 iv1->base);
873 else
874 noloop = fold_build2 (GT_EXPR, boolean_type_node,
875 fold_build2 (MINUS_EXPR, type1,
876 iv0->base, tmod),
877 iv1->base);
878 }
879
880 if (!integer_nonzerop (assumption))
881 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
882 niter->assumptions,
883 assumption);
884 if (!integer_zerop (noloop))
885 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
886 niter->may_be_zero,
887 noloop);
888 bounds_add (bnds, tree_to_double_int (mod), type);
889 *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod);
890
891 ret = true;
892 end:
893 mpz_clear (mmod);
894 return ret;
895 }
896
897 /* Add assertions to NITER that ensure that the control variable of the loop
898 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
899 are TYPE. Returns false if we can prove that there is an overflow, true
900 otherwise. STEP is the absolute value of the step. */
901
902 static bool
903 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1,
904 struct tree_niter_desc *niter, tree step)
905 {
906 tree bound, d, assumption, diff;
907 tree niter_type = TREE_TYPE (step);
908
909 if (integer_nonzerop (iv0->step))
910 {
911 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
912 if (iv0->no_overflow)
913 return true;
914
915 /* If iv0->base is a constant, we can determine the last value before
916 overflow precisely; otherwise we conservatively assume
917 MAX - STEP + 1. */
918
919 if (TREE_CODE (iv0->base) == INTEGER_CST)
920 {
921 d = fold_build2 (MINUS_EXPR, niter_type,
922 fold_convert (niter_type, TYPE_MAX_VALUE (type)),
923 fold_convert (niter_type, iv0->base));
924 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
925 }
926 else
927 diff = fold_build2 (MINUS_EXPR, niter_type, step,
928 build_int_cst (niter_type, 1));
929 bound = fold_build2 (MINUS_EXPR, type,
930 TYPE_MAX_VALUE (type), fold_convert (type, diff));
931 assumption = fold_build2 (LE_EXPR, boolean_type_node,
932 iv1->base, bound);
933 }
934 else
935 {
936 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
937 if (iv1->no_overflow)
938 return true;
939
940 if (TREE_CODE (iv1->base) == INTEGER_CST)
941 {
942 d = fold_build2 (MINUS_EXPR, niter_type,
943 fold_convert (niter_type, iv1->base),
944 fold_convert (niter_type, TYPE_MIN_VALUE (type)));
945 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
946 }
947 else
948 diff = fold_build2 (MINUS_EXPR, niter_type, step,
949 build_int_cst (niter_type, 1));
950 bound = fold_build2 (PLUS_EXPR, type,
951 TYPE_MIN_VALUE (type), fold_convert (type, diff));
952 assumption = fold_build2 (GE_EXPR, boolean_type_node,
953 iv0->base, bound);
954 }
955
956 if (integer_zerop (assumption))
957 return false;
958 if (!integer_nonzerop (assumption))
959 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
960 niter->assumptions, assumption);
961
962 iv0->no_overflow = true;
963 iv1->no_overflow = true;
964 return true;
965 }
966
967 /* Add an assumption to NITER that a loop whose ending condition
968 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS
969 bounds the value of IV1->base - IV0->base. */
970
971 static void
972 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1,
973 struct tree_niter_desc *niter, bounds *bnds)
974 {
975 tree assumption = boolean_true_node, bound, diff;
976 tree mbz, mbzl, mbzr, type1;
977 bool rolls_p, no_overflow_p;
978 double_int dstep;
979 mpz_t mstep, max;
980
981 /* We are going to compute the number of iterations as
982 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned
983 variant of TYPE. This formula only works if
984
985 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
986
987 (where MAX is the maximum value of the unsigned variant of TYPE, and
988 the computations in this formula are performed in full precision,
989 i.e., without overflows).
990
991 Usually, for loops with exit condition iv0->base + step * i < iv1->base,
992 we have a condition of the form iv0->base - step < iv1->base before the loop,
993 and for loops iv0->base < iv1->base - step * i the condition
994 iv0->base < iv1->base + step, due to loop header copying, which enable us
995 to prove the lower bound.
996
997 The upper bound is more complicated. Unless the expressions for initial
998 and final value themselves contain enough information, we usually cannot
999 derive it from the context. */
1000
1001 /* First check whether the answer does not follow from the bounds we gathered
1002 before. */
1003 if (integer_nonzerop (iv0->step))
1004 dstep = tree_to_double_int (iv0->step);
1005 else
1006 {
1007 dstep = tree_to_double_int (iv1->step).sext (TYPE_PRECISION (type));
1008 dstep = -dstep;
1009 }
1010
1011 mpz_init (mstep);
1012 mpz_set_double_int (mstep, dstep, true);
1013 mpz_neg (mstep, mstep);
1014 mpz_add_ui (mstep, mstep, 1);
1015
1016 rolls_p = mpz_cmp (mstep, bnds->below) <= 0;
1017
1018 mpz_init (max);
1019 mpz_set_double_int (max, double_int::mask (TYPE_PRECISION (type)), true);
1020 mpz_add (max, max, mstep);
1021 no_overflow_p = (mpz_cmp (bnds->up, max) <= 0
1022 /* For pointers, only values lying inside a single object
1023 can be compared or manipulated by pointer arithmetics.
1024 Gcc in general does not allow or handle objects larger
1025 than half of the address space, hence the upper bound
1026 is satisfied for pointers. */
1027 || POINTER_TYPE_P (type));
1028 mpz_clear (mstep);
1029 mpz_clear (max);
1030
1031 if (rolls_p && no_overflow_p)
1032 return;
1033
1034 type1 = type;
1035 if (POINTER_TYPE_P (type))
1036 type1 = sizetype;
1037
1038 /* Now the hard part; we must formulate the assumption(s) as expressions, and
1039 we must be careful not to introduce overflow. */
1040
1041 if (integer_nonzerop (iv0->step))
1042 {
1043 diff = fold_build2 (MINUS_EXPR, type1,
1044 iv0->step, build_int_cst (type1, 1));
1045
1046 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
1047 0 address never belongs to any object, we can assume this for
1048 pointers. */
1049 if (!POINTER_TYPE_P (type))
1050 {
1051 bound = fold_build2 (PLUS_EXPR, type1,
1052 TYPE_MIN_VALUE (type), diff);
1053 assumption = fold_build2 (GE_EXPR, boolean_type_node,
1054 iv0->base, bound);
1055 }
1056
1057 /* And then we can compute iv0->base - diff, and compare it with
1058 iv1->base. */
1059 mbzl = fold_build2 (MINUS_EXPR, type1,
1060 fold_convert (type1, iv0->base), diff);
1061 mbzr = fold_convert (type1, iv1->base);
1062 }
1063 else
1064 {
1065 diff = fold_build2 (PLUS_EXPR, type1,
1066 iv1->step, build_int_cst (type1, 1));
1067
1068 if (!POINTER_TYPE_P (type))
1069 {
1070 bound = fold_build2 (PLUS_EXPR, type1,
1071 TYPE_MAX_VALUE (type), diff);
1072 assumption = fold_build2 (LE_EXPR, boolean_type_node,
1073 iv1->base, bound);
1074 }
1075
1076 mbzl = fold_convert (type1, iv0->base);
1077 mbzr = fold_build2 (MINUS_EXPR, type1,
1078 fold_convert (type1, iv1->base), diff);
1079 }
1080
1081 if (!integer_nonzerop (assumption))
1082 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1083 niter->assumptions, assumption);
1084 if (!rolls_p)
1085 {
1086 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr);
1087 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
1088 niter->may_be_zero, mbz);
1089 }
1090 }
1091
1092 /* Determines number of iterations of loop whose ending condition
1093 is IV0 < IV1. TYPE is the type of the iv. The number of
1094 iterations is stored to NITER. BNDS bounds the difference
1095 IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know
1096 that the exit must be taken eventually. */
1097
1098 static bool
1099 number_of_iterations_lt (tree type, affine_iv *iv0, affine_iv *iv1,
1100 struct tree_niter_desc *niter,
1101 bool exit_must_be_taken, bounds *bnds)
1102 {
1103 tree niter_type = unsigned_type_for (type);
1104 tree delta, step, s;
1105 mpz_t mstep, tmp;
1106
1107 if (integer_nonzerop (iv0->step))
1108 {
1109 niter->control = *iv0;
1110 niter->cmp = LT_EXPR;
1111 niter->bound = iv1->base;
1112 }
1113 else
1114 {
1115 niter->control = *iv1;
1116 niter->cmp = GT_EXPR;
1117 niter->bound = iv0->base;
1118 }
1119
1120 delta = fold_build2 (MINUS_EXPR, niter_type,
1121 fold_convert (niter_type, iv1->base),
1122 fold_convert (niter_type, iv0->base));
1123
1124 /* First handle the special case that the step is +-1. */
1125 if ((integer_onep (iv0->step) && integer_zerop (iv1->step))
1126 || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step)))
1127 {
1128 /* for (i = iv0->base; i < iv1->base; i++)
1129
1130 or
1131
1132 for (i = iv1->base; i > iv0->base; i--).
1133
1134 In both cases # of iterations is iv1->base - iv0->base, assuming that
1135 iv1->base >= iv0->base.
1136
1137 First try to derive a lower bound on the value of
1138 iv1->base - iv0->base, computed in full precision. If the difference
1139 is nonnegative, we are done, otherwise we must record the
1140 condition. */
1141
1142 if (mpz_sgn (bnds->below) < 0)
1143 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node,
1144 iv1->base, iv0->base);
1145 niter->niter = delta;
1146 niter->max = mpz_get_double_int (niter_type, bnds->up, false);
1147 return true;
1148 }
1149
1150 if (integer_nonzerop (iv0->step))
1151 step = fold_convert (niter_type, iv0->step);
1152 else
1153 step = fold_convert (niter_type,
1154 fold_build1 (NEGATE_EXPR, type, iv1->step));
1155
1156 /* If we can determine the final value of the control iv exactly, we can
1157 transform the condition to != comparison. In particular, this will be
1158 the case if DELTA is constant. */
1159 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step,
1160 exit_must_be_taken, bnds))
1161 {
1162 affine_iv zps;
1163
1164 zps.base = build_int_cst (niter_type, 0);
1165 zps.step = step;
1166 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1167 zps does not overflow. */
1168 zps.no_overflow = true;
1169
1170 return number_of_iterations_ne (type, &zps, delta, niter, true, bnds);
1171 }
1172
1173 /* Make sure that the control iv does not overflow. */
1174 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step))
1175 return false;
1176
1177 /* We determine the number of iterations as (delta + step - 1) / step. For
1178 this to work, we must know that iv1->base >= iv0->base - step + 1,
1179 otherwise the loop does not roll. */
1180 assert_loop_rolls_lt (type, iv0, iv1, niter, bnds);
1181
1182 s = fold_build2 (MINUS_EXPR, niter_type,
1183 step, build_int_cst (niter_type, 1));
1184 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s);
1185 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step);
1186
1187 mpz_init (mstep);
1188 mpz_init (tmp);
1189 mpz_set_double_int (mstep, tree_to_double_int (step), true);
1190 mpz_add (tmp, bnds->up, mstep);
1191 mpz_sub_ui (tmp, tmp, 1);
1192 mpz_fdiv_q (tmp, tmp, mstep);
1193 niter->max = mpz_get_double_int (niter_type, tmp, false);
1194 mpz_clear (mstep);
1195 mpz_clear (tmp);
1196
1197 return true;
1198 }
1199
1200 /* Determines number of iterations of loop whose ending condition
1201 is IV0 <= IV1. TYPE is the type of the iv. The number of
1202 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
1203 we know that this condition must eventually become false (we derived this
1204 earlier, and possibly set NITER->assumptions to make sure this
1205 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1206
1207 static bool
1208 number_of_iterations_le (tree type, affine_iv *iv0, affine_iv *iv1,
1209 struct tree_niter_desc *niter, bool exit_must_be_taken,
1210 bounds *bnds)
1211 {
1212 tree assumption;
1213 tree type1 = type;
1214 if (POINTER_TYPE_P (type))
1215 type1 = sizetype;
1216
1217 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1218 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1219 value of the type. This we must know anyway, since if it is
1220 equal to this value, the loop rolls forever. We do not check
1221 this condition for pointer type ivs, as the code cannot rely on
1222 the object to that the pointer points being placed at the end of
1223 the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is
1224 not defined for pointers). */
1225
1226 if (!exit_must_be_taken && !POINTER_TYPE_P (type))
1227 {
1228 if (integer_nonzerop (iv0->step))
1229 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1230 iv1->base, TYPE_MAX_VALUE (type));
1231 else
1232 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1233 iv0->base, TYPE_MIN_VALUE (type));
1234
1235 if (integer_zerop (assumption))
1236 return false;
1237 if (!integer_nonzerop (assumption))
1238 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1239 niter->assumptions, assumption);
1240 }
1241
1242 if (integer_nonzerop (iv0->step))
1243 {
1244 if (POINTER_TYPE_P (type))
1245 iv1->base = fold_build_pointer_plus_hwi (iv1->base, 1);
1246 else
1247 iv1->base = fold_build2 (PLUS_EXPR, type1, iv1->base,
1248 build_int_cst (type1, 1));
1249 }
1250 else if (POINTER_TYPE_P (type))
1251 iv0->base = fold_build_pointer_plus_hwi (iv0->base, -1);
1252 else
1253 iv0->base = fold_build2 (MINUS_EXPR, type1,
1254 iv0->base, build_int_cst (type1, 1));
1255
1256 bounds_add (bnds, double_int_one, type1);
1257
1258 return number_of_iterations_lt (type, iv0, iv1, niter, exit_must_be_taken,
1259 bnds);
1260 }
1261
1262 /* Dumps description of affine induction variable IV to FILE. */
1263
1264 static void
1265 dump_affine_iv (FILE *file, affine_iv *iv)
1266 {
1267 if (!integer_zerop (iv->step))
1268 fprintf (file, "[");
1269
1270 print_generic_expr (dump_file, iv->base, TDF_SLIM);
1271
1272 if (!integer_zerop (iv->step))
1273 {
1274 fprintf (file, ", + , ");
1275 print_generic_expr (dump_file, iv->step, TDF_SLIM);
1276 fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : "");
1277 }
1278 }
1279
1280 /* Determine the number of iterations according to condition (for staying
1281 inside loop) which compares two induction variables using comparison
1282 operator CODE. The induction variable on left side of the comparison
1283 is IV0, the right-hand side is IV1. Both induction variables must have
1284 type TYPE, which must be an integer or pointer type. The steps of the
1285 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1286
1287 LOOP is the loop whose number of iterations we are determining.
1288
1289 ONLY_EXIT is true if we are sure this is the only way the loop could be
1290 exited (including possibly non-returning function calls, exceptions, etc.)
1291 -- in this case we can use the information whether the control induction
1292 variables can overflow or not in a more efficient way.
1293
1294 if EVERY_ITERATION is true, we know the test is executed on every iteration.
1295
1296 The results (number of iterations and assumptions as described in
1297 comments at struct tree_niter_desc in tree-flow.h) are stored to NITER.
1298 Returns false if it fails to determine number of iterations, true if it
1299 was determined (possibly with some assumptions). */
1300
1301 static bool
1302 number_of_iterations_cond (struct loop *loop,
1303 tree type, affine_iv *iv0, enum tree_code code,
1304 affine_iv *iv1, struct tree_niter_desc *niter,
1305 bool only_exit, bool every_iteration)
1306 {
1307 bool exit_must_be_taken = false, ret;
1308 bounds bnds;
1309
1310 /* If the test is not executed every iteration, wrapping may make the test
1311 to pass again.
1312 TODO: the overflow case can be still used as unreliable estimate of upper
1313 bound. But we have no API to pass it down to number of iterations code
1314 and, at present, it will not use it anyway. */
1315 if (!every_iteration
1316 && (!iv0->no_overflow || !iv1->no_overflow
1317 || code == NE_EXPR || code == EQ_EXPR))
1318 return false;
1319
1320 /* The meaning of these assumptions is this:
1321 if !assumptions
1322 then the rest of information does not have to be valid
1323 if may_be_zero then the loop does not roll, even if
1324 niter != 0. */
1325 niter->assumptions = boolean_true_node;
1326 niter->may_be_zero = boolean_false_node;
1327 niter->niter = NULL_TREE;
1328 niter->max = double_int_zero;
1329
1330 niter->bound = NULL_TREE;
1331 niter->cmp = ERROR_MARK;
1332
1333 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1334 the control variable is on lhs. */
1335 if (code == GE_EXPR || code == GT_EXPR
1336 || (code == NE_EXPR && integer_zerop (iv0->step)))
1337 {
1338 SWAP (iv0, iv1);
1339 code = swap_tree_comparison (code);
1340 }
1341
1342 if (POINTER_TYPE_P (type))
1343 {
1344 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1345 to the same object. If they do, the control variable cannot wrap
1346 (as wrap around the bounds of memory will never return a pointer
1347 that would be guaranteed to point to the same object, even if we
1348 avoid undefined behavior by casting to size_t and back). */
1349 iv0->no_overflow = true;
1350 iv1->no_overflow = true;
1351 }
1352
1353 /* If the control induction variable does not overflow and the only exit
1354 from the loop is the one that we analyze, we know it must be taken
1355 eventually. */
1356 if (only_exit)
1357 {
1358 if (!integer_zerop (iv0->step) && iv0->no_overflow)
1359 exit_must_be_taken = true;
1360 else if (!integer_zerop (iv1->step) && iv1->no_overflow)
1361 exit_must_be_taken = true;
1362 }
1363
1364 /* We can handle the case when neither of the sides of the comparison is
1365 invariant, provided that the test is NE_EXPR. This rarely occurs in
1366 practice, but it is simple enough to manage. */
1367 if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step))
1368 {
1369 tree step_type = POINTER_TYPE_P (type) ? sizetype : type;
1370 if (code != NE_EXPR)
1371 return false;
1372
1373 iv0->step = fold_binary_to_constant (MINUS_EXPR, step_type,
1374 iv0->step, iv1->step);
1375 iv0->no_overflow = false;
1376 iv1->step = build_int_cst (step_type, 0);
1377 iv1->no_overflow = true;
1378 }
1379
1380 /* If the result of the comparison is a constant, the loop is weird. More
1381 precise handling would be possible, but the situation is not common enough
1382 to waste time on it. */
1383 if (integer_zerop (iv0->step) && integer_zerop (iv1->step))
1384 return false;
1385
1386 /* Ignore loops of while (i-- < 10) type. */
1387 if (code != NE_EXPR)
1388 {
1389 if (iv0->step && tree_int_cst_sign_bit (iv0->step))
1390 return false;
1391
1392 if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step))
1393 return false;
1394 }
1395
1396 /* If the loop exits immediately, there is nothing to do. */
1397 tree tem = fold_binary (code, boolean_type_node, iv0->base, iv1->base);
1398 if (tem && integer_zerop (tem))
1399 {
1400 niter->niter = build_int_cst (unsigned_type_for (type), 0);
1401 niter->max = double_int_zero;
1402 return true;
1403 }
1404
1405 /* OK, now we know we have a senseful loop. Handle several cases, depending
1406 on what comparison operator is used. */
1407 bound_difference (loop, iv1->base, iv0->base, &bnds);
1408
1409 if (dump_file && (dump_flags & TDF_DETAILS))
1410 {
1411 fprintf (dump_file,
1412 "Analyzing # of iterations of loop %d\n", loop->num);
1413
1414 fprintf (dump_file, " exit condition ");
1415 dump_affine_iv (dump_file, iv0);
1416 fprintf (dump_file, " %s ",
1417 code == NE_EXPR ? "!="
1418 : code == LT_EXPR ? "<"
1419 : "<=");
1420 dump_affine_iv (dump_file, iv1);
1421 fprintf (dump_file, "\n");
1422
1423 fprintf (dump_file, " bounds on difference of bases: ");
1424 mpz_out_str (dump_file, 10, bnds.below);
1425 fprintf (dump_file, " ... ");
1426 mpz_out_str (dump_file, 10, bnds.up);
1427 fprintf (dump_file, "\n");
1428 }
1429
1430 switch (code)
1431 {
1432 case NE_EXPR:
1433 gcc_assert (integer_zerop (iv1->step));
1434 ret = number_of_iterations_ne (type, iv0, iv1->base, niter,
1435 exit_must_be_taken, &bnds);
1436 break;
1437
1438 case LT_EXPR:
1439 ret = number_of_iterations_lt (type, iv0, iv1, niter, exit_must_be_taken,
1440 &bnds);
1441 break;
1442
1443 case LE_EXPR:
1444 ret = number_of_iterations_le (type, iv0, iv1, niter, exit_must_be_taken,
1445 &bnds);
1446 break;
1447
1448 default:
1449 gcc_unreachable ();
1450 }
1451
1452 mpz_clear (bnds.up);
1453 mpz_clear (bnds.below);
1454
1455 if (dump_file && (dump_flags & TDF_DETAILS))
1456 {
1457 if (ret)
1458 {
1459 fprintf (dump_file, " result:\n");
1460 if (!integer_nonzerop (niter->assumptions))
1461 {
1462 fprintf (dump_file, " under assumptions ");
1463 print_generic_expr (dump_file, niter->assumptions, TDF_SLIM);
1464 fprintf (dump_file, "\n");
1465 }
1466
1467 if (!integer_zerop (niter->may_be_zero))
1468 {
1469 fprintf (dump_file, " zero if ");
1470 print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM);
1471 fprintf (dump_file, "\n");
1472 }
1473
1474 fprintf (dump_file, " # of iterations ");
1475 print_generic_expr (dump_file, niter->niter, TDF_SLIM);
1476 fprintf (dump_file, ", bounded by ");
1477 dump_double_int (dump_file, niter->max, true);
1478 fprintf (dump_file, "\n");
1479 }
1480 else
1481 fprintf (dump_file, " failed\n\n");
1482 }
1483 return ret;
1484 }
1485
1486 /* Substitute NEW for OLD in EXPR and fold the result. */
1487
1488 static tree
1489 simplify_replace_tree (tree expr, tree old, tree new_tree)
1490 {
1491 unsigned i, n;
1492 tree ret = NULL_TREE, e, se;
1493
1494 if (!expr)
1495 return NULL_TREE;
1496
1497 /* Do not bother to replace constants. */
1498 if (CONSTANT_CLASS_P (old))
1499 return expr;
1500
1501 if (expr == old
1502 || operand_equal_p (expr, old, 0))
1503 return unshare_expr (new_tree);
1504
1505 if (!EXPR_P (expr))
1506 return expr;
1507
1508 n = TREE_OPERAND_LENGTH (expr);
1509 for (i = 0; i < n; i++)
1510 {
1511 e = TREE_OPERAND (expr, i);
1512 se = simplify_replace_tree (e, old, new_tree);
1513 if (e == se)
1514 continue;
1515
1516 if (!ret)
1517 ret = copy_node (expr);
1518
1519 TREE_OPERAND (ret, i) = se;
1520 }
1521
1522 return (ret ? fold (ret) : expr);
1523 }
1524
1525 /* Expand definitions of ssa names in EXPR as long as they are simple
1526 enough, and return the new expression. */
1527
1528 tree
1529 expand_simple_operations (tree expr)
1530 {
1531 unsigned i, n;
1532 tree ret = NULL_TREE, e, ee, e1;
1533 enum tree_code code;
1534 gimple stmt;
1535
1536 if (expr == NULL_TREE)
1537 return expr;
1538
1539 if (is_gimple_min_invariant (expr))
1540 return expr;
1541
1542 code = TREE_CODE (expr);
1543 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code)))
1544 {
1545 n = TREE_OPERAND_LENGTH (expr);
1546 for (i = 0; i < n; i++)
1547 {
1548 e = TREE_OPERAND (expr, i);
1549 ee = expand_simple_operations (e);
1550 if (e == ee)
1551 continue;
1552
1553 if (!ret)
1554 ret = copy_node (expr);
1555
1556 TREE_OPERAND (ret, i) = ee;
1557 }
1558
1559 if (!ret)
1560 return expr;
1561
1562 fold_defer_overflow_warnings ();
1563 ret = fold (ret);
1564 fold_undefer_and_ignore_overflow_warnings ();
1565 return ret;
1566 }
1567
1568 if (TREE_CODE (expr) != SSA_NAME)
1569 return expr;
1570
1571 stmt = SSA_NAME_DEF_STMT (expr);
1572 if (gimple_code (stmt) == GIMPLE_PHI)
1573 {
1574 basic_block src, dest;
1575
1576 if (gimple_phi_num_args (stmt) != 1)
1577 return expr;
1578 e = PHI_ARG_DEF (stmt, 0);
1579
1580 /* Avoid propagating through loop exit phi nodes, which
1581 could break loop-closed SSA form restrictions. */
1582 dest = gimple_bb (stmt);
1583 src = single_pred (dest);
1584 if (TREE_CODE (e) == SSA_NAME
1585 && src->loop_father != dest->loop_father)
1586 return expr;
1587
1588 return expand_simple_operations (e);
1589 }
1590 if (gimple_code (stmt) != GIMPLE_ASSIGN)
1591 return expr;
1592
1593 /* Avoid expanding to expressions that contain SSA names that need
1594 to take part in abnormal coalescing. */
1595 ssa_op_iter iter;
1596 FOR_EACH_SSA_TREE_OPERAND (e, stmt, iter, SSA_OP_USE)
1597 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e))
1598 return expr;
1599
1600 e = gimple_assign_rhs1 (stmt);
1601 code = gimple_assign_rhs_code (stmt);
1602 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1603 {
1604 if (is_gimple_min_invariant (e))
1605 return e;
1606
1607 if (code == SSA_NAME)
1608 return expand_simple_operations (e);
1609
1610 return expr;
1611 }
1612
1613 switch (code)
1614 {
1615 CASE_CONVERT:
1616 /* Casts are simple. */
1617 ee = expand_simple_operations (e);
1618 return fold_build1 (code, TREE_TYPE (expr), ee);
1619
1620 case PLUS_EXPR:
1621 case MINUS_EXPR:
1622 case POINTER_PLUS_EXPR:
1623 /* And increments and decrements by a constant are simple. */
1624 e1 = gimple_assign_rhs2 (stmt);
1625 if (!is_gimple_min_invariant (e1))
1626 return expr;
1627
1628 ee = expand_simple_operations (e);
1629 return fold_build2 (code, TREE_TYPE (expr), ee, e1);
1630
1631 default:
1632 return expr;
1633 }
1634 }
1635
1636 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1637 expression (or EXPR unchanged, if no simplification was possible). */
1638
1639 static tree
1640 tree_simplify_using_condition_1 (tree cond, tree expr)
1641 {
1642 bool changed;
1643 tree e, te, e0, e1, e2, notcond;
1644 enum tree_code code = TREE_CODE (expr);
1645
1646 if (code == INTEGER_CST)
1647 return expr;
1648
1649 if (code == TRUTH_OR_EXPR
1650 || code == TRUTH_AND_EXPR
1651 || code == COND_EXPR)
1652 {
1653 changed = false;
1654
1655 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0));
1656 if (TREE_OPERAND (expr, 0) != e0)
1657 changed = true;
1658
1659 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1));
1660 if (TREE_OPERAND (expr, 1) != e1)
1661 changed = true;
1662
1663 if (code == COND_EXPR)
1664 {
1665 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2));
1666 if (TREE_OPERAND (expr, 2) != e2)
1667 changed = true;
1668 }
1669 else
1670 e2 = NULL_TREE;
1671
1672 if (changed)
1673 {
1674 if (code == COND_EXPR)
1675 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1676 else
1677 expr = fold_build2 (code, boolean_type_node, e0, e1);
1678 }
1679
1680 return expr;
1681 }
1682
1683 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
1684 propagation, and vice versa. Fold does not handle this, since it is
1685 considered too expensive. */
1686 if (TREE_CODE (cond) == EQ_EXPR)
1687 {
1688 e0 = TREE_OPERAND (cond, 0);
1689 e1 = TREE_OPERAND (cond, 1);
1690
1691 /* We know that e0 == e1. Check whether we cannot simplify expr
1692 using this fact. */
1693 e = simplify_replace_tree (expr, e0, e1);
1694 if (integer_zerop (e) || integer_nonzerop (e))
1695 return e;
1696
1697 e = simplify_replace_tree (expr, e1, e0);
1698 if (integer_zerop (e) || integer_nonzerop (e))
1699 return e;
1700 }
1701 if (TREE_CODE (expr) == EQ_EXPR)
1702 {
1703 e0 = TREE_OPERAND (expr, 0);
1704 e1 = TREE_OPERAND (expr, 1);
1705
1706 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
1707 e = simplify_replace_tree (cond, e0, e1);
1708 if (integer_zerop (e))
1709 return e;
1710 e = simplify_replace_tree (cond, e1, e0);
1711 if (integer_zerop (e))
1712 return e;
1713 }
1714 if (TREE_CODE (expr) == NE_EXPR)
1715 {
1716 e0 = TREE_OPERAND (expr, 0);
1717 e1 = TREE_OPERAND (expr, 1);
1718
1719 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
1720 e = simplify_replace_tree (cond, e0, e1);
1721 if (integer_zerop (e))
1722 return boolean_true_node;
1723 e = simplify_replace_tree (cond, e1, e0);
1724 if (integer_zerop (e))
1725 return boolean_true_node;
1726 }
1727
1728 te = expand_simple_operations (expr);
1729
1730 /* Check whether COND ==> EXPR. */
1731 notcond = invert_truthvalue (cond);
1732 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, te);
1733 if (e && integer_nonzerop (e))
1734 return e;
1735
1736 /* Check whether COND ==> not EXPR. */
1737 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, te);
1738 if (e && integer_zerop (e))
1739 return e;
1740
1741 return expr;
1742 }
1743
1744 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1745 expression (or EXPR unchanged, if no simplification was possible).
1746 Wrapper around tree_simplify_using_condition_1 that ensures that chains
1747 of simple operations in definitions of ssa names in COND are expanded,
1748 so that things like casts or incrementing the value of the bound before
1749 the loop do not cause us to fail. */
1750
1751 static tree
1752 tree_simplify_using_condition (tree cond, tree expr)
1753 {
1754 cond = expand_simple_operations (cond);
1755
1756 return tree_simplify_using_condition_1 (cond, expr);
1757 }
1758
1759 /* Tries to simplify EXPR using the conditions on entry to LOOP.
1760 Returns the simplified expression (or EXPR unchanged, if no
1761 simplification was possible).*/
1762
1763 static tree
1764 simplify_using_initial_conditions (struct loop *loop, tree expr)
1765 {
1766 edge e;
1767 basic_block bb;
1768 gimple stmt;
1769 tree cond;
1770 int cnt = 0;
1771
1772 if (TREE_CODE (expr) == INTEGER_CST)
1773 return expr;
1774
1775 /* Limit walking the dominators to avoid quadraticness in
1776 the number of BBs times the number of loops in degenerate
1777 cases. */
1778 for (bb = loop->header;
1779 bb != ENTRY_BLOCK_PTR && cnt < MAX_DOMINATORS_TO_WALK;
1780 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
1781 {
1782 if (!single_pred_p (bb))
1783 continue;
1784 e = single_pred_edge (bb);
1785
1786 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
1787 continue;
1788
1789 stmt = last_stmt (e->src);
1790 cond = fold_build2 (gimple_cond_code (stmt),
1791 boolean_type_node,
1792 gimple_cond_lhs (stmt),
1793 gimple_cond_rhs (stmt));
1794 if (e->flags & EDGE_FALSE_VALUE)
1795 cond = invert_truthvalue (cond);
1796 expr = tree_simplify_using_condition (cond, expr);
1797 ++cnt;
1798 }
1799
1800 return expr;
1801 }
1802
1803 /* Tries to simplify EXPR using the evolutions of the loop invariants
1804 in the superloops of LOOP. Returns the simplified expression
1805 (or EXPR unchanged, if no simplification was possible). */
1806
1807 static tree
1808 simplify_using_outer_evolutions (struct loop *loop, tree expr)
1809 {
1810 enum tree_code code = TREE_CODE (expr);
1811 bool changed;
1812 tree e, e0, e1, e2;
1813
1814 if (is_gimple_min_invariant (expr))
1815 return expr;
1816
1817 if (code == TRUTH_OR_EXPR
1818 || code == TRUTH_AND_EXPR
1819 || code == COND_EXPR)
1820 {
1821 changed = false;
1822
1823 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0));
1824 if (TREE_OPERAND (expr, 0) != e0)
1825 changed = true;
1826
1827 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1));
1828 if (TREE_OPERAND (expr, 1) != e1)
1829 changed = true;
1830
1831 if (code == COND_EXPR)
1832 {
1833 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2));
1834 if (TREE_OPERAND (expr, 2) != e2)
1835 changed = true;
1836 }
1837 else
1838 e2 = NULL_TREE;
1839
1840 if (changed)
1841 {
1842 if (code == COND_EXPR)
1843 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1844 else
1845 expr = fold_build2 (code, boolean_type_node, e0, e1);
1846 }
1847
1848 return expr;
1849 }
1850
1851 e = instantiate_parameters (loop, expr);
1852 if (is_gimple_min_invariant (e))
1853 return e;
1854
1855 return expr;
1856 }
1857
1858 /* Returns true if EXIT is the only possible exit from LOOP. */
1859
1860 bool
1861 loop_only_exit_p (const struct loop *loop, const_edge exit)
1862 {
1863 basic_block *body;
1864 gimple_stmt_iterator bsi;
1865 unsigned i;
1866 gimple call;
1867
1868 if (exit != single_exit (loop))
1869 return false;
1870
1871 body = get_loop_body (loop);
1872 for (i = 0; i < loop->num_nodes; i++)
1873 {
1874 for (bsi = gsi_start_bb (body[i]); !gsi_end_p (bsi); gsi_next (&bsi))
1875 {
1876 call = gsi_stmt (bsi);
1877 if (gimple_code (call) != GIMPLE_CALL)
1878 continue;
1879
1880 if (gimple_has_side_effects (call))
1881 {
1882 free (body);
1883 return false;
1884 }
1885 }
1886 }
1887
1888 free (body);
1889 return true;
1890 }
1891
1892 /* Stores description of number of iterations of LOOP derived from
1893 EXIT (an exit edge of the LOOP) in NITER. Returns true if some
1894 useful information could be derived (and fields of NITER has
1895 meaning described in comments at struct tree_niter_desc
1896 declaration), false otherwise. If WARN is true and
1897 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use
1898 potentially unsafe assumptions.
1899 When EVERY_ITERATION is true, only tests that are known to be executed
1900 every iteration are considered (i.e. only test that alone bounds the loop).
1901 */
1902
1903 bool
1904 number_of_iterations_exit (struct loop *loop, edge exit,
1905 struct tree_niter_desc *niter,
1906 bool warn, bool every_iteration)
1907 {
1908 gimple stmt;
1909 tree type;
1910 tree op0, op1;
1911 enum tree_code code;
1912 affine_iv iv0, iv1;
1913 bool safe;
1914
1915 safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src);
1916
1917 if (every_iteration && !safe)
1918 return false;
1919
1920 niter->assumptions = boolean_false_node;
1921 stmt = last_stmt (exit->src);
1922 if (!stmt || gimple_code (stmt) != GIMPLE_COND)
1923 return false;
1924
1925 /* We want the condition for staying inside loop. */
1926 code = gimple_cond_code (stmt);
1927 if (exit->flags & EDGE_TRUE_VALUE)
1928 code = invert_tree_comparison (code, false);
1929
1930 switch (code)
1931 {
1932 case GT_EXPR:
1933 case GE_EXPR:
1934 case LT_EXPR:
1935 case LE_EXPR:
1936 case NE_EXPR:
1937 break;
1938
1939 default:
1940 return false;
1941 }
1942
1943 op0 = gimple_cond_lhs (stmt);
1944 op1 = gimple_cond_rhs (stmt);
1945 type = TREE_TYPE (op0);
1946
1947 if (TREE_CODE (type) != INTEGER_TYPE
1948 && !POINTER_TYPE_P (type))
1949 return false;
1950
1951 if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, false))
1952 return false;
1953 if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, false))
1954 return false;
1955
1956 /* We don't want to see undefined signed overflow warnings while
1957 computing the number of iterations. */
1958 fold_defer_overflow_warnings ();
1959
1960 iv0.base = expand_simple_operations (iv0.base);
1961 iv1.base = expand_simple_operations (iv1.base);
1962 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter,
1963 loop_only_exit_p (loop, exit), safe))
1964 {
1965 fold_undefer_and_ignore_overflow_warnings ();
1966 return false;
1967 }
1968
1969 if (optimize >= 3)
1970 {
1971 niter->assumptions = simplify_using_outer_evolutions (loop,
1972 niter->assumptions);
1973 niter->may_be_zero = simplify_using_outer_evolutions (loop,
1974 niter->may_be_zero);
1975 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
1976 }
1977
1978 niter->assumptions
1979 = simplify_using_initial_conditions (loop,
1980 niter->assumptions);
1981 niter->may_be_zero
1982 = simplify_using_initial_conditions (loop,
1983 niter->may_be_zero);
1984
1985 fold_undefer_and_ignore_overflow_warnings ();
1986
1987 /* If NITER has simplified into a constant, update MAX. */
1988 if (TREE_CODE (niter->niter) == INTEGER_CST)
1989 niter->max = tree_to_double_int (niter->niter);
1990
1991 if (integer_onep (niter->assumptions))
1992 return true;
1993
1994 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
1995 But if we can prove that there is overflow or some other source of weird
1996 behavior, ignore the loop even with -funsafe-loop-optimizations. */
1997 if (integer_zerop (niter->assumptions) || !single_exit (loop))
1998 return false;
1999
2000 if (flag_unsafe_loop_optimizations)
2001 niter->assumptions = boolean_true_node;
2002
2003 if (warn)
2004 {
2005 const char *wording;
2006 location_t loc = gimple_location (stmt);
2007
2008 /* We can provide a more specific warning if one of the operator is
2009 constant and the other advances by +1 or -1. */
2010 if (!integer_zerop (iv1.step)
2011 ? (integer_zerop (iv0.step)
2012 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step)))
2013 : (integer_onep (iv0.step) || integer_all_onesp (iv0.step)))
2014 wording =
2015 flag_unsafe_loop_optimizations
2016 ? N_("assuming that the loop is not infinite")
2017 : N_("cannot optimize possibly infinite loops");
2018 else
2019 wording =
2020 flag_unsafe_loop_optimizations
2021 ? N_("assuming that the loop counter does not overflow")
2022 : N_("cannot optimize loop, the loop counter may overflow");
2023
2024 warning_at ((LOCATION_LINE (loc) > 0) ? loc : input_location,
2025 OPT_Wunsafe_loop_optimizations, "%s", gettext (wording));
2026 }
2027
2028 return flag_unsafe_loop_optimizations;
2029 }
2030
2031 /* Try to determine the number of iterations of LOOP. If we succeed,
2032 expression giving number of iterations is returned and *EXIT is
2033 set to the edge from that the information is obtained. Otherwise
2034 chrec_dont_know is returned. */
2035
2036 tree
2037 find_loop_niter (struct loop *loop, edge *exit)
2038 {
2039 unsigned i;
2040 vec<edge> exits = get_loop_exit_edges (loop);
2041 edge ex;
2042 tree niter = NULL_TREE, aniter;
2043 struct tree_niter_desc desc;
2044
2045 *exit = NULL;
2046 FOR_EACH_VEC_ELT (exits, i, ex)
2047 {
2048 if (!number_of_iterations_exit (loop, ex, &desc, false))
2049 continue;
2050
2051 if (integer_nonzerop (desc.may_be_zero))
2052 {
2053 /* We exit in the first iteration through this exit.
2054 We won't find anything better. */
2055 niter = build_int_cst (unsigned_type_node, 0);
2056 *exit = ex;
2057 break;
2058 }
2059
2060 if (!integer_zerop (desc.may_be_zero))
2061 continue;
2062
2063 aniter = desc.niter;
2064
2065 if (!niter)
2066 {
2067 /* Nothing recorded yet. */
2068 niter = aniter;
2069 *exit = ex;
2070 continue;
2071 }
2072
2073 /* Prefer constants, the lower the better. */
2074 if (TREE_CODE (aniter) != INTEGER_CST)
2075 continue;
2076
2077 if (TREE_CODE (niter) != INTEGER_CST)
2078 {
2079 niter = aniter;
2080 *exit = ex;
2081 continue;
2082 }
2083
2084 if (tree_int_cst_lt (aniter, niter))
2085 {
2086 niter = aniter;
2087 *exit = ex;
2088 continue;
2089 }
2090 }
2091 exits.release ();
2092
2093 return niter ? niter : chrec_dont_know;
2094 }
2095
2096 /* Return true if loop is known to have bounded number of iterations. */
2097
2098 bool
2099 finite_loop_p (struct loop *loop)
2100 {
2101 double_int nit;
2102 int flags;
2103
2104 if (flag_unsafe_loop_optimizations)
2105 return true;
2106 flags = flags_from_decl_or_type (current_function_decl);
2107 if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE))
2108 {
2109 if (dump_file && (dump_flags & TDF_DETAILS))
2110 fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n",
2111 loop->num);
2112 return true;
2113 }
2114
2115 if (loop->any_upper_bound
2116 || max_loop_iterations (loop, &nit))
2117 {
2118 if (dump_file && (dump_flags & TDF_DETAILS))
2119 fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n",
2120 loop->num);
2121 return true;
2122 }
2123 return false;
2124 }
2125
2126 /*
2127
2128 Analysis of a number of iterations of a loop by a brute-force evaluation.
2129
2130 */
2131
2132 /* Bound on the number of iterations we try to evaluate. */
2133
2134 #define MAX_ITERATIONS_TO_TRACK \
2135 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
2136
2137 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
2138 result by a chain of operations such that all but exactly one of their
2139 operands are constants. */
2140
2141 static gimple
2142 chain_of_csts_start (struct loop *loop, tree x)
2143 {
2144 gimple stmt = SSA_NAME_DEF_STMT (x);
2145 tree use;
2146 basic_block bb = gimple_bb (stmt);
2147 enum tree_code code;
2148
2149 if (!bb
2150 || !flow_bb_inside_loop_p (loop, bb))
2151 return NULL;
2152
2153 if (gimple_code (stmt) == GIMPLE_PHI)
2154 {
2155 if (bb == loop->header)
2156 return stmt;
2157
2158 return NULL;
2159 }
2160
2161 if (gimple_code (stmt) != GIMPLE_ASSIGN)
2162 return NULL;
2163
2164 code = gimple_assign_rhs_code (stmt);
2165 if (gimple_references_memory_p (stmt)
2166 || TREE_CODE_CLASS (code) == tcc_reference
2167 || (code == ADDR_EXPR
2168 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
2169 return NULL;
2170
2171 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
2172 if (use == NULL_TREE)
2173 return NULL;
2174
2175 return chain_of_csts_start (loop, use);
2176 }
2177
2178 /* Determines whether the expression X is derived from a result of a phi node
2179 in header of LOOP such that
2180
2181 * the derivation of X consists only from operations with constants
2182 * the initial value of the phi node is constant
2183 * the value of the phi node in the next iteration can be derived from the
2184 value in the current iteration by a chain of operations with constants.
2185
2186 If such phi node exists, it is returned, otherwise NULL is returned. */
2187
2188 static gimple
2189 get_base_for (struct loop *loop, tree x)
2190 {
2191 gimple phi;
2192 tree init, next;
2193
2194 if (is_gimple_min_invariant (x))
2195 return NULL;
2196
2197 phi = chain_of_csts_start (loop, x);
2198 if (!phi)
2199 return NULL;
2200
2201 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2202 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2203
2204 if (TREE_CODE (next) != SSA_NAME)
2205 return NULL;
2206
2207 if (!is_gimple_min_invariant (init))
2208 return NULL;
2209
2210 if (chain_of_csts_start (loop, next) != phi)
2211 return NULL;
2212
2213 return phi;
2214 }
2215
2216 /* Given an expression X, then
2217
2218 * if X is NULL_TREE, we return the constant BASE.
2219 * otherwise X is a SSA name, whose value in the considered loop is derived
2220 by a chain of operations with constant from a result of a phi node in
2221 the header of the loop. Then we return value of X when the value of the
2222 result of this phi node is given by the constant BASE. */
2223
2224 static tree
2225 get_val_for (tree x, tree base)
2226 {
2227 gimple stmt;
2228
2229 gcc_assert (is_gimple_min_invariant (base));
2230
2231 if (!x)
2232 return base;
2233
2234 stmt = SSA_NAME_DEF_STMT (x);
2235 if (gimple_code (stmt) == GIMPLE_PHI)
2236 return base;
2237
2238 gcc_assert (is_gimple_assign (stmt));
2239
2240 /* STMT must be either an assignment of a single SSA name or an
2241 expression involving an SSA name and a constant. Try to fold that
2242 expression using the value for the SSA name. */
2243 if (gimple_assign_ssa_name_copy_p (stmt))
2244 return get_val_for (gimple_assign_rhs1 (stmt), base);
2245 else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS
2246 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
2247 {
2248 return fold_build1 (gimple_assign_rhs_code (stmt),
2249 gimple_expr_type (stmt),
2250 get_val_for (gimple_assign_rhs1 (stmt), base));
2251 }
2252 else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS)
2253 {
2254 tree rhs1 = gimple_assign_rhs1 (stmt);
2255 tree rhs2 = gimple_assign_rhs2 (stmt);
2256 if (TREE_CODE (rhs1) == SSA_NAME)
2257 rhs1 = get_val_for (rhs1, base);
2258 else if (TREE_CODE (rhs2) == SSA_NAME)
2259 rhs2 = get_val_for (rhs2, base);
2260 else
2261 gcc_unreachable ();
2262 return fold_build2 (gimple_assign_rhs_code (stmt),
2263 gimple_expr_type (stmt), rhs1, rhs2);
2264 }
2265 else
2266 gcc_unreachable ();
2267 }
2268
2269
2270 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2271 by brute force -- i.e. by determining the value of the operands of the
2272 condition at EXIT in first few iterations of the loop (assuming that
2273 these values are constant) and determining the first one in that the
2274 condition is not satisfied. Returns the constant giving the number
2275 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2276
2277 tree
2278 loop_niter_by_eval (struct loop *loop, edge exit)
2279 {
2280 tree acnd;
2281 tree op[2], val[2], next[2], aval[2];
2282 gimple phi, cond;
2283 unsigned i, j;
2284 enum tree_code cmp;
2285
2286 cond = last_stmt (exit->src);
2287 if (!cond || gimple_code (cond) != GIMPLE_COND)
2288 return chrec_dont_know;
2289
2290 cmp = gimple_cond_code (cond);
2291 if (exit->flags & EDGE_TRUE_VALUE)
2292 cmp = invert_tree_comparison (cmp, false);
2293
2294 switch (cmp)
2295 {
2296 case EQ_EXPR:
2297 case NE_EXPR:
2298 case GT_EXPR:
2299 case GE_EXPR:
2300 case LT_EXPR:
2301 case LE_EXPR:
2302 op[0] = gimple_cond_lhs (cond);
2303 op[1] = gimple_cond_rhs (cond);
2304 break;
2305
2306 default:
2307 return chrec_dont_know;
2308 }
2309
2310 for (j = 0; j < 2; j++)
2311 {
2312 if (is_gimple_min_invariant (op[j]))
2313 {
2314 val[j] = op[j];
2315 next[j] = NULL_TREE;
2316 op[j] = NULL_TREE;
2317 }
2318 else
2319 {
2320 phi = get_base_for (loop, op[j]);
2321 if (!phi)
2322 return chrec_dont_know;
2323 val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2324 next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2325 }
2326 }
2327
2328 /* Don't issue signed overflow warnings. */
2329 fold_defer_overflow_warnings ();
2330
2331 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
2332 {
2333 for (j = 0; j < 2; j++)
2334 aval[j] = get_val_for (op[j], val[j]);
2335
2336 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
2337 if (acnd && integer_zerop (acnd))
2338 {
2339 fold_undefer_and_ignore_overflow_warnings ();
2340 if (dump_file && (dump_flags & TDF_DETAILS))
2341 fprintf (dump_file,
2342 "Proved that loop %d iterates %d times using brute force.\n",
2343 loop->num, i);
2344 return build_int_cst (unsigned_type_node, i);
2345 }
2346
2347 for (j = 0; j < 2; j++)
2348 {
2349 val[j] = get_val_for (next[j], val[j]);
2350 if (!is_gimple_min_invariant (val[j]))
2351 {
2352 fold_undefer_and_ignore_overflow_warnings ();
2353 return chrec_dont_know;
2354 }
2355 }
2356 }
2357
2358 fold_undefer_and_ignore_overflow_warnings ();
2359
2360 return chrec_dont_know;
2361 }
2362
2363 /* Finds the exit of the LOOP by that the loop exits after a constant
2364 number of iterations and stores the exit edge to *EXIT. The constant
2365 giving the number of iterations of LOOP is returned. The number of
2366 iterations is determined using loop_niter_by_eval (i.e. by brute force
2367 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2368 determines the number of iterations, chrec_dont_know is returned. */
2369
2370 tree
2371 find_loop_niter_by_eval (struct loop *loop, edge *exit)
2372 {
2373 unsigned i;
2374 vec<edge> exits = get_loop_exit_edges (loop);
2375 edge ex;
2376 tree niter = NULL_TREE, aniter;
2377
2378 *exit = NULL;
2379
2380 /* Loops with multiple exits are expensive to handle and less important. */
2381 if (!flag_expensive_optimizations
2382 && exits.length () > 1)
2383 {
2384 exits.release ();
2385 return chrec_dont_know;
2386 }
2387
2388 FOR_EACH_VEC_ELT (exits, i, ex)
2389 {
2390 if (!just_once_each_iteration_p (loop, ex->src))
2391 continue;
2392
2393 aniter = loop_niter_by_eval (loop, ex);
2394 if (chrec_contains_undetermined (aniter))
2395 continue;
2396
2397 if (niter
2398 && !tree_int_cst_lt (aniter, niter))
2399 continue;
2400
2401 niter = aniter;
2402 *exit = ex;
2403 }
2404 exits.release ();
2405
2406 return niter ? niter : chrec_dont_know;
2407 }
2408
2409 /*
2410
2411 Analysis of upper bounds on number of iterations of a loop.
2412
2413 */
2414
2415 static double_int derive_constant_upper_bound_ops (tree, tree,
2416 enum tree_code, tree);
2417
2418 /* Returns a constant upper bound on the value of the right-hand side of
2419 an assignment statement STMT. */
2420
2421 static double_int
2422 derive_constant_upper_bound_assign (gimple stmt)
2423 {
2424 enum tree_code code = gimple_assign_rhs_code (stmt);
2425 tree op0 = gimple_assign_rhs1 (stmt);
2426 tree op1 = gimple_assign_rhs2 (stmt);
2427
2428 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)),
2429 op0, code, op1);
2430 }
2431
2432 /* Returns a constant upper bound on the value of expression VAL. VAL
2433 is considered to be unsigned. If its type is signed, its value must
2434 be nonnegative. */
2435
2436 static double_int
2437 derive_constant_upper_bound (tree val)
2438 {
2439 enum tree_code code;
2440 tree op0, op1;
2441
2442 extract_ops_from_tree (val, &code, &op0, &op1);
2443 return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1);
2444 }
2445
2446 /* Returns a constant upper bound on the value of expression OP0 CODE OP1,
2447 whose type is TYPE. The expression is considered to be unsigned. If
2448 its type is signed, its value must be nonnegative. */
2449
2450 static double_int
2451 derive_constant_upper_bound_ops (tree type, tree op0,
2452 enum tree_code code, tree op1)
2453 {
2454 tree subtype, maxt;
2455 double_int bnd, max, mmax, cst;
2456 gimple stmt;
2457
2458 if (INTEGRAL_TYPE_P (type))
2459 maxt = TYPE_MAX_VALUE (type);
2460 else
2461 maxt = upper_bound_in_type (type, type);
2462
2463 max = tree_to_double_int (maxt);
2464
2465 switch (code)
2466 {
2467 case INTEGER_CST:
2468 return tree_to_double_int (op0);
2469
2470 CASE_CONVERT:
2471 subtype = TREE_TYPE (op0);
2472 if (!TYPE_UNSIGNED (subtype)
2473 /* If TYPE is also signed, the fact that VAL is nonnegative implies
2474 that OP0 is nonnegative. */
2475 && TYPE_UNSIGNED (type)
2476 && !tree_expr_nonnegative_p (op0))
2477 {
2478 /* If we cannot prove that the casted expression is nonnegative,
2479 we cannot establish more useful upper bound than the precision
2480 of the type gives us. */
2481 return max;
2482 }
2483
2484 /* We now know that op0 is an nonnegative value. Try deriving an upper
2485 bound for it. */
2486 bnd = derive_constant_upper_bound (op0);
2487
2488 /* If the bound does not fit in TYPE, max. value of TYPE could be
2489 attained. */
2490 if (max.ult (bnd))
2491 return max;
2492
2493 return bnd;
2494
2495 case PLUS_EXPR:
2496 case POINTER_PLUS_EXPR:
2497 case MINUS_EXPR:
2498 if (TREE_CODE (op1) != INTEGER_CST
2499 || !tree_expr_nonnegative_p (op0))
2500 return max;
2501
2502 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
2503 choose the most logical way how to treat this constant regardless
2504 of the signedness of the type. */
2505 cst = tree_to_double_int (op1);
2506 cst = cst.sext (TYPE_PRECISION (type));
2507 if (code != MINUS_EXPR)
2508 cst = -cst;
2509
2510 bnd = derive_constant_upper_bound (op0);
2511
2512 if (cst.is_negative ())
2513 {
2514 cst = -cst;
2515 /* Avoid CST == 0x80000... */
2516 if (cst.is_negative ())
2517 return max;;
2518
2519 /* OP0 + CST. We need to check that
2520 BND <= MAX (type) - CST. */
2521
2522 mmax -= cst;
2523 if (bnd.ugt (mmax))
2524 return max;
2525
2526 return bnd + cst;
2527 }
2528 else
2529 {
2530 /* OP0 - CST, where CST >= 0.
2531
2532 If TYPE is signed, we have already verified that OP0 >= 0, and we
2533 know that the result is nonnegative. This implies that
2534 VAL <= BND - CST.
2535
2536 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2537 otherwise the operation underflows.
2538 */
2539
2540 /* This should only happen if the type is unsigned; however, for
2541 buggy programs that use overflowing signed arithmetics even with
2542 -fno-wrapv, this condition may also be true for signed values. */
2543 if (bnd.ult (cst))
2544 return max;
2545
2546 if (TYPE_UNSIGNED (type))
2547 {
2548 tree tem = fold_binary (GE_EXPR, boolean_type_node, op0,
2549 double_int_to_tree (type, cst));
2550 if (!tem || integer_nonzerop (tem))
2551 return max;
2552 }
2553
2554 bnd -= cst;
2555 }
2556
2557 return bnd;
2558
2559 case FLOOR_DIV_EXPR:
2560 case EXACT_DIV_EXPR:
2561 if (TREE_CODE (op1) != INTEGER_CST
2562 || tree_int_cst_sign_bit (op1))
2563 return max;
2564
2565 bnd = derive_constant_upper_bound (op0);
2566 return bnd.udiv (tree_to_double_int (op1), FLOOR_DIV_EXPR);
2567
2568 case BIT_AND_EXPR:
2569 if (TREE_CODE (op1) != INTEGER_CST
2570 || tree_int_cst_sign_bit (op1))
2571 return max;
2572 return tree_to_double_int (op1);
2573
2574 case SSA_NAME:
2575 stmt = SSA_NAME_DEF_STMT (op0);
2576 if (gimple_code (stmt) != GIMPLE_ASSIGN
2577 || gimple_assign_lhs (stmt) != op0)
2578 return max;
2579 return derive_constant_upper_bound_assign (stmt);
2580
2581 default:
2582 return max;
2583 }
2584 }
2585
2586 /* Emit a -Waggressive-loop-optimizations warning if needed. */
2587
2588 static void
2589 do_warn_aggressive_loop_optimizations (struct loop *loop,
2590 double_int i_bound, gimple stmt)
2591 {
2592 /* Don't warn if the loop doesn't have known constant bound. */
2593 if (!loop->nb_iterations
2594 || TREE_CODE (loop->nb_iterations) != INTEGER_CST
2595 || !warn_aggressive_loop_optimizations
2596 /* To avoid warning multiple times for the same loop,
2597 only start warning when we preserve loops. */
2598 || (cfun->curr_properties & PROP_loops) == 0
2599 /* Only warn once per loop. */
2600 || loop->warned_aggressive_loop_optimizations
2601 /* Only warn if undefined behavior gives us lower estimate than the
2602 known constant bound. */
2603 || i_bound.ucmp (tree_to_double_int (loop->nb_iterations)) >= 0
2604 /* And undefined behavior happens unconditionally. */
2605 || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt)))
2606 return;
2607
2608 edge e = single_exit (loop);
2609 if (e == NULL)
2610 return;
2611
2612 gimple estmt = last_stmt (e->src);
2613 if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations,
2614 "iteration %E invokes undefined behavior",
2615 double_int_to_tree (TREE_TYPE (loop->nb_iterations),
2616 i_bound)))
2617 inform (gimple_location (estmt), "containing loop");
2618 loop->warned_aggressive_loop_optimizations = true;
2619 }
2620
2621 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
2622 is true if the loop is exited immediately after STMT, and this exit
2623 is taken at last when the STMT is executed BOUND + 1 times.
2624 REALISTIC is true if BOUND is expected to be close to the real number
2625 of iterations. UPPER is true if we are sure the loop iterates at most
2626 BOUND times. I_BOUND is an unsigned double_int upper estimate on BOUND. */
2627
2628 static void
2629 record_estimate (struct loop *loop, tree bound, double_int i_bound,
2630 gimple at_stmt, bool is_exit, bool realistic, bool upper)
2631 {
2632 double_int delta;
2633
2634 if (dump_file && (dump_flags & TDF_DETAILS))
2635 {
2636 fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : "");
2637 print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM);
2638 fprintf (dump_file, " is %sexecuted at most ",
2639 upper ? "" : "probably ");
2640 print_generic_expr (dump_file, bound, TDF_SLIM);
2641 fprintf (dump_file, " (bounded by ");
2642 dump_double_int (dump_file, i_bound, true);
2643 fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num);
2644 }
2645
2646 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
2647 real number of iterations. */
2648 if (TREE_CODE (bound) != INTEGER_CST)
2649 realistic = false;
2650 else
2651 gcc_checking_assert (i_bound == tree_to_double_int (bound));
2652 if (!upper && !realistic)
2653 return;
2654
2655 /* If we have a guaranteed upper bound, record it in the appropriate
2656 list, unless this is an !is_exit bound (i.e. undefined behavior in
2657 at_stmt) in a loop with known constant number of iterations. */
2658 if (upper
2659 && (is_exit
2660 || loop->nb_iterations == NULL_TREE
2661 || TREE_CODE (loop->nb_iterations) != INTEGER_CST))
2662 {
2663 struct nb_iter_bound *elt = ggc_alloc_nb_iter_bound ();
2664
2665 elt->bound = i_bound;
2666 elt->stmt = at_stmt;
2667 elt->is_exit = is_exit;
2668 elt->next = loop->bounds;
2669 loop->bounds = elt;
2670 }
2671
2672 /* If statement is executed on every path to the loop latch, we can directly
2673 infer the upper bound on the # of iterations of the loop. */
2674 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt)))
2675 return;
2676
2677 /* Update the number of iteration estimates according to the bound.
2678 If at_stmt is an exit then the loop latch is executed at most BOUND times,
2679 otherwise it can be executed BOUND + 1 times. We will lower the estimate
2680 later if such statement must be executed on last iteration */
2681 if (is_exit)
2682 delta = double_int_zero;
2683 else
2684 delta = double_int_one;
2685 i_bound += delta;
2686
2687 /* If an overflow occurred, ignore the result. */
2688 if (i_bound.ult (delta))
2689 return;
2690
2691 if (upper && !is_exit)
2692 do_warn_aggressive_loop_optimizations (loop, i_bound, at_stmt);
2693 record_niter_bound (loop, i_bound, realistic, upper);
2694 }
2695
2696 /* Record the estimate on number of iterations of LOOP based on the fact that
2697 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
2698 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
2699 estimated number of iterations is expected to be close to the real one.
2700 UPPER is true if we are sure the induction variable does not wrap. */
2701
2702 static void
2703 record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple stmt,
2704 tree low, tree high, bool realistic, bool upper)
2705 {
2706 tree niter_bound, extreme, delta;
2707 tree type = TREE_TYPE (base), unsigned_type;
2708 double_int max;
2709
2710 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step))
2711 return;
2712
2713 if (dump_file && (dump_flags & TDF_DETAILS))
2714 {
2715 fprintf (dump_file, "Induction variable (");
2716 print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM);
2717 fprintf (dump_file, ") ");
2718 print_generic_expr (dump_file, base, TDF_SLIM);
2719 fprintf (dump_file, " + ");
2720 print_generic_expr (dump_file, step, TDF_SLIM);
2721 fprintf (dump_file, " * iteration does not wrap in statement ");
2722 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2723 fprintf (dump_file, " in loop %d.\n", loop->num);
2724 }
2725
2726 unsigned_type = unsigned_type_for (type);
2727 base = fold_convert (unsigned_type, base);
2728 step = fold_convert (unsigned_type, step);
2729
2730 if (tree_int_cst_sign_bit (step))
2731 {
2732 extreme = fold_convert (unsigned_type, low);
2733 if (TREE_CODE (base) != INTEGER_CST)
2734 base = fold_convert (unsigned_type, high);
2735 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
2736 step = fold_build1 (NEGATE_EXPR, unsigned_type, step);
2737 }
2738 else
2739 {
2740 extreme = fold_convert (unsigned_type, high);
2741 if (TREE_CODE (base) != INTEGER_CST)
2742 base = fold_convert (unsigned_type, low);
2743 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
2744 }
2745
2746 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
2747 would get out of the range. */
2748 niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step);
2749 max = derive_constant_upper_bound (niter_bound);
2750 record_estimate (loop, niter_bound, max, stmt, false, realistic, upper);
2751 }
2752
2753 /* Determine information about number of iterations a LOOP from the index
2754 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
2755 guaranteed to be executed in every iteration of LOOP. Callback for
2756 for_each_index. */
2757
2758 struct ilb_data
2759 {
2760 struct loop *loop;
2761 gimple stmt;
2762 };
2763
2764 static bool
2765 idx_infer_loop_bounds (tree base, tree *idx, void *dta)
2766 {
2767 struct ilb_data *data = (struct ilb_data *) dta;
2768 tree ev, init, step;
2769 tree low, high, type, next;
2770 bool sign, upper = true, at_end = false;
2771 struct loop *loop = data->loop;
2772 bool reliable = true;
2773
2774 if (TREE_CODE (base) != ARRAY_REF)
2775 return true;
2776
2777 /* For arrays at the end of the structure, we are not guaranteed that they
2778 do not really extend over their declared size. However, for arrays of
2779 size greater than one, this is unlikely to be intended. */
2780 if (array_at_struct_end_p (base))
2781 {
2782 at_end = true;
2783 upper = false;
2784 }
2785
2786 struct loop *dloop = loop_containing_stmt (data->stmt);
2787 if (!dloop)
2788 return true;
2789
2790 ev = analyze_scalar_evolution (dloop, *idx);
2791 ev = instantiate_parameters (loop, ev);
2792 init = initial_condition (ev);
2793 step = evolution_part_in_loop_num (ev, loop->num);
2794
2795 if (!init
2796 || !step
2797 || TREE_CODE (step) != INTEGER_CST
2798 || integer_zerop (step)
2799 || tree_contains_chrecs (init, NULL)
2800 || chrec_contains_symbols_defined_in_loop (init, loop->num))
2801 return true;
2802
2803 low = array_ref_low_bound (base);
2804 high = array_ref_up_bound (base);
2805
2806 /* The case of nonconstant bounds could be handled, but it would be
2807 complicated. */
2808 if (TREE_CODE (low) != INTEGER_CST
2809 || !high
2810 || TREE_CODE (high) != INTEGER_CST)
2811 return true;
2812 sign = tree_int_cst_sign_bit (step);
2813 type = TREE_TYPE (step);
2814
2815 /* The array of length 1 at the end of a structure most likely extends
2816 beyond its bounds. */
2817 if (at_end
2818 && operand_equal_p (low, high, 0))
2819 return true;
2820
2821 /* In case the relevant bound of the array does not fit in type, or
2822 it does, but bound + step (in type) still belongs into the range of the
2823 array, the index may wrap and still stay within the range of the array
2824 (consider e.g. if the array is indexed by the full range of
2825 unsigned char).
2826
2827 To make things simpler, we require both bounds to fit into type, although
2828 there are cases where this would not be strictly necessary. */
2829 if (!int_fits_type_p (high, type)
2830 || !int_fits_type_p (low, type))
2831 return true;
2832 low = fold_convert (type, low);
2833 high = fold_convert (type, high);
2834
2835 if (sign)
2836 next = fold_binary (PLUS_EXPR, type, low, step);
2837 else
2838 next = fold_binary (PLUS_EXPR, type, high, step);
2839
2840 if (tree_int_cst_compare (low, next) <= 0
2841 && tree_int_cst_compare (next, high) <= 0)
2842 return true;
2843
2844 /* If access is not executed on every iteration, we must ensure that overlow may
2845 not make the access valid later. */
2846 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt))
2847 && scev_probably_wraps_p (initial_condition_in_loop_num (ev, loop->num),
2848 step, data->stmt, loop, true))
2849 reliable = false;
2850
2851 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, reliable, upper);
2852 return true;
2853 }
2854
2855 /* Determine information about number of iterations a LOOP from the bounds
2856 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
2857 STMT is guaranteed to be executed in every iteration of LOOP.*/
2858
2859 static void
2860 infer_loop_bounds_from_ref (struct loop *loop, gimple stmt, tree ref)
2861 {
2862 struct ilb_data data;
2863
2864 data.loop = loop;
2865 data.stmt = stmt;
2866 for_each_index (&ref, idx_infer_loop_bounds, &data);
2867 }
2868
2869 /* Determine information about number of iterations of a LOOP from the way
2870 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
2871 executed in every iteration of LOOP. */
2872
2873 static void
2874 infer_loop_bounds_from_array (struct loop *loop, gimple stmt)
2875 {
2876 if (is_gimple_assign (stmt))
2877 {
2878 tree op0 = gimple_assign_lhs (stmt);
2879 tree op1 = gimple_assign_rhs1 (stmt);
2880
2881 /* For each memory access, analyze its access function
2882 and record a bound on the loop iteration domain. */
2883 if (REFERENCE_CLASS_P (op0))
2884 infer_loop_bounds_from_ref (loop, stmt, op0);
2885
2886 if (REFERENCE_CLASS_P (op1))
2887 infer_loop_bounds_from_ref (loop, stmt, op1);
2888 }
2889 else if (is_gimple_call (stmt))
2890 {
2891 tree arg, lhs;
2892 unsigned i, n = gimple_call_num_args (stmt);
2893
2894 lhs = gimple_call_lhs (stmt);
2895 if (lhs && REFERENCE_CLASS_P (lhs))
2896 infer_loop_bounds_from_ref (loop, stmt, lhs);
2897
2898 for (i = 0; i < n; i++)
2899 {
2900 arg = gimple_call_arg (stmt, i);
2901 if (REFERENCE_CLASS_P (arg))
2902 infer_loop_bounds_from_ref (loop, stmt, arg);
2903 }
2904 }
2905 }
2906
2907 /* Determine information about number of iterations of a LOOP from the fact
2908 that pointer arithmetics in STMT does not overflow. */
2909
2910 static void
2911 infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple stmt)
2912 {
2913 tree def, base, step, scev, type, low, high;
2914 tree var, ptr;
2915
2916 if (!is_gimple_assign (stmt)
2917 || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR)
2918 return;
2919
2920 def = gimple_assign_lhs (stmt);
2921 if (TREE_CODE (def) != SSA_NAME)
2922 return;
2923
2924 type = TREE_TYPE (def);
2925 if (!nowrap_type_p (type))
2926 return;
2927
2928 ptr = gimple_assign_rhs1 (stmt);
2929 if (!expr_invariant_in_loop_p (loop, ptr))
2930 return;
2931
2932 var = gimple_assign_rhs2 (stmt);
2933 if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var)))
2934 return;
2935
2936 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
2937 if (chrec_contains_undetermined (scev))
2938 return;
2939
2940 base = initial_condition_in_loop_num (scev, loop->num);
2941 step = evolution_part_in_loop_num (scev, loop->num);
2942
2943 if (!base || !step
2944 || TREE_CODE (step) != INTEGER_CST
2945 || tree_contains_chrecs (base, NULL)
2946 || chrec_contains_symbols_defined_in_loop (base, loop->num))
2947 return;
2948
2949 low = lower_bound_in_type (type, type);
2950 high = upper_bound_in_type (type, type);
2951
2952 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot
2953 produce a NULL pointer. The contrary would mean NULL points to an object,
2954 while NULL is supposed to compare unequal with the address of all objects.
2955 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a
2956 NULL pointer since that would mean wrapping, which we assume here not to
2957 happen. So, we can exclude NULL from the valid range of pointer
2958 arithmetic. */
2959 if (flag_delete_null_pointer_checks && int_cst_value (low) == 0)
2960 low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type)));
2961
2962 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
2963 }
2964
2965 /* Determine information about number of iterations of a LOOP from the fact
2966 that signed arithmetics in STMT does not overflow. */
2967
2968 static void
2969 infer_loop_bounds_from_signedness (struct loop *loop, gimple stmt)
2970 {
2971 tree def, base, step, scev, type, low, high;
2972
2973 if (gimple_code (stmt) != GIMPLE_ASSIGN)
2974 return;
2975
2976 def = gimple_assign_lhs (stmt);
2977
2978 if (TREE_CODE (def) != SSA_NAME)
2979 return;
2980
2981 type = TREE_TYPE (def);
2982 if (!INTEGRAL_TYPE_P (type)
2983 || !TYPE_OVERFLOW_UNDEFINED (type))
2984 return;
2985
2986 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
2987 if (chrec_contains_undetermined (scev))
2988 return;
2989
2990 base = initial_condition_in_loop_num (scev, loop->num);
2991 step = evolution_part_in_loop_num (scev, loop->num);
2992
2993 if (!base || !step
2994 || TREE_CODE (step) != INTEGER_CST
2995 || tree_contains_chrecs (base, NULL)
2996 || chrec_contains_symbols_defined_in_loop (base, loop->num))
2997 return;
2998
2999 low = lower_bound_in_type (type, type);
3000 high = upper_bound_in_type (type, type);
3001
3002 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
3003 }
3004
3005 /* The following analyzers are extracting informations on the bounds
3006 of LOOP from the following undefined behaviors:
3007
3008 - data references should not access elements over the statically
3009 allocated size,
3010
3011 - signed variables should not overflow when flag_wrapv is not set.
3012 */
3013
3014 static void
3015 infer_loop_bounds_from_undefined (struct loop *loop)
3016 {
3017 unsigned i;
3018 basic_block *bbs;
3019 gimple_stmt_iterator bsi;
3020 basic_block bb;
3021 bool reliable;
3022
3023 bbs = get_loop_body (loop);
3024
3025 for (i = 0; i < loop->num_nodes; i++)
3026 {
3027 bb = bbs[i];
3028
3029 /* If BB is not executed in each iteration of the loop, we cannot
3030 use the operations in it to infer reliable upper bound on the
3031 # of iterations of the loop. However, we can use it as a guess.
3032 Reliable guesses come only from array bounds. */
3033 reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb);
3034
3035 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
3036 {
3037 gimple stmt = gsi_stmt (bsi);
3038
3039 infer_loop_bounds_from_array (loop, stmt);
3040
3041 if (reliable)
3042 {
3043 infer_loop_bounds_from_signedness (loop, stmt);
3044 infer_loop_bounds_from_pointer_arith (loop, stmt);
3045 }
3046 }
3047
3048 }
3049
3050 free (bbs);
3051 }
3052
3053
3054
3055 /* Compare double ints, callback for qsort. */
3056
3057 static int
3058 double_int_cmp (const void *p1, const void *p2)
3059 {
3060 const double_int *d1 = (const double_int *)p1;
3061 const double_int *d2 = (const double_int *)p2;
3062 if (*d1 == *d2)
3063 return 0;
3064 if (d1->ult (*d2))
3065 return -1;
3066 return 1;
3067 }
3068
3069 /* Return index of BOUND in BOUNDS array sorted in increasing order.
3070 Lookup by binary search. */
3071
3072 static int
3073 bound_index (vec<double_int> bounds, double_int bound)
3074 {
3075 unsigned int end = bounds.length ();
3076 unsigned int begin = 0;
3077
3078 /* Find a matching index by means of a binary search. */
3079 while (begin != end)
3080 {
3081 unsigned int middle = (begin + end) / 2;
3082 double_int index = bounds[middle];
3083
3084 if (index == bound)
3085 return middle;
3086 else if (index.ult (bound))
3087 begin = middle + 1;
3088 else
3089 end = middle;
3090 }
3091 gcc_unreachable ();
3092 }
3093
3094 /* We recorded loop bounds only for statements dominating loop latch (and thus
3095 executed each loop iteration). If there are any bounds on statements not
3096 dominating the loop latch we can improve the estimate by walking the loop
3097 body and seeing if every path from loop header to loop latch contains
3098 some bounded statement. */
3099
3100 static void
3101 discover_iteration_bound_by_body_walk (struct loop *loop)
3102 {
3103 pointer_map_t *bb_bounds;
3104 struct nb_iter_bound *elt;
3105 vec<double_int> bounds = vNULL;
3106 vec<vec<basic_block> > queues = vNULL;
3107 vec<basic_block> queue = vNULL;
3108 ptrdiff_t queue_index;
3109 ptrdiff_t latch_index = 0;
3110 pointer_map_t *block_priority;
3111
3112 /* Discover what bounds may interest us. */
3113 for (elt = loop->bounds; elt; elt = elt->next)
3114 {
3115 double_int bound = elt->bound;
3116
3117 /* Exit terminates loop at given iteration, while non-exits produce undefined
3118 effect on the next iteration. */
3119 if (!elt->is_exit)
3120 {
3121 bound += double_int_one;
3122 /* If an overflow occurred, ignore the result. */
3123 if (bound.is_zero ())
3124 continue;
3125 }
3126
3127 if (!loop->any_upper_bound
3128 || bound.ult (loop->nb_iterations_upper_bound))
3129 bounds.safe_push (bound);
3130 }
3131
3132 /* Exit early if there is nothing to do. */
3133 if (!bounds.exists ())
3134 return;
3135
3136 if (dump_file && (dump_flags & TDF_DETAILS))
3137 fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n");
3138
3139 /* Sort the bounds in decreasing order. */
3140 qsort (bounds.address (), bounds.length (),
3141 sizeof (double_int), double_int_cmp);
3142
3143 /* For every basic block record the lowest bound that is guaranteed to
3144 terminate the loop. */
3145
3146 bb_bounds = pointer_map_create ();
3147 for (elt = loop->bounds; elt; elt = elt->next)
3148 {
3149 double_int bound = elt->bound;
3150 if (!elt->is_exit)
3151 {
3152 bound += double_int_one;
3153 /* If an overflow occurred, ignore the result. */
3154 if (bound.is_zero ())
3155 continue;
3156 }
3157
3158 if (!loop->any_upper_bound
3159 || bound.ult (loop->nb_iterations_upper_bound))
3160 {
3161 ptrdiff_t index = bound_index (bounds, bound);
3162 void **entry = pointer_map_contains (bb_bounds,
3163 gimple_bb (elt->stmt));
3164 if (!entry)
3165 *pointer_map_insert (bb_bounds,
3166 gimple_bb (elt->stmt)) = (void *)index;
3167 else if ((ptrdiff_t)*entry > index)
3168 *entry = (void *)index;
3169 }
3170 }
3171
3172 block_priority = pointer_map_create ();
3173
3174 /* Perform shortest path discovery loop->header ... loop->latch.
3175
3176 The "distance" is given by the smallest loop bound of basic block
3177 present in the path and we look for path with largest smallest bound
3178 on it.
3179
3180 To avoid the need for fibonacci heap on double ints we simply compress
3181 double ints into indexes to BOUNDS array and then represent the queue
3182 as arrays of queues for every index.
3183 Index of BOUNDS.length() means that the execution of given BB has
3184 no bounds determined.
3185
3186 VISITED is a pointer map translating basic block into smallest index
3187 it was inserted into the priority queue with. */
3188 latch_index = -1;
3189
3190 /* Start walk in loop header with index set to infinite bound. */
3191 queue_index = bounds.length ();
3192 queues.safe_grow_cleared (queue_index + 1);
3193 queue.safe_push (loop->header);
3194 queues[queue_index] = queue;
3195 *pointer_map_insert (block_priority, loop->header) = (void *)queue_index;
3196
3197 for (; queue_index >= 0; queue_index--)
3198 {
3199 if (latch_index < queue_index)
3200 {
3201 while (queues[queue_index].length ())
3202 {
3203 basic_block bb;
3204 ptrdiff_t bound_index = queue_index;
3205 void **entry;
3206 edge e;
3207 edge_iterator ei;
3208
3209 queue = queues[queue_index];
3210 bb = queue.pop ();
3211
3212 /* OK, we later inserted the BB with lower priority, skip it. */
3213 if ((ptrdiff_t)*pointer_map_contains (block_priority, bb) > queue_index)
3214 continue;
3215
3216 /* See if we can improve the bound. */
3217 entry = pointer_map_contains (bb_bounds, bb);
3218 if (entry && (ptrdiff_t)*entry < bound_index)
3219 bound_index = (ptrdiff_t)*entry;
3220
3221 /* Insert succesors into the queue, watch for latch edge
3222 and record greatest index we saw. */
3223 FOR_EACH_EDGE (e, ei, bb->succs)
3224 {
3225 bool insert = false;
3226 void **entry;
3227
3228 if (loop_exit_edge_p (loop, e))
3229 continue;
3230
3231 if (e == loop_latch_edge (loop)
3232 && latch_index < bound_index)
3233 latch_index = bound_index;
3234 else if (!(entry = pointer_map_contains (block_priority, e->dest)))
3235 {
3236 insert = true;
3237 *pointer_map_insert (block_priority, e->dest) = (void *)bound_index;
3238 }
3239 else if ((ptrdiff_t)*entry < bound_index)
3240 {
3241 insert = true;
3242 *entry = (void *)bound_index;
3243 }
3244
3245 if (insert)
3246 queues[bound_index].safe_push (e->dest);
3247 }
3248 }
3249 }
3250 queues[queue_index].release ();
3251 }
3252
3253 gcc_assert (latch_index >= 0);
3254 if ((unsigned)latch_index < bounds.length ())
3255 {
3256 if (dump_file && (dump_flags & TDF_DETAILS))
3257 {
3258 fprintf (dump_file, "Found better loop bound ");
3259 dump_double_int (dump_file, bounds[latch_index], true);
3260 fprintf (dump_file, "\n");
3261 }
3262 record_niter_bound (loop, bounds[latch_index], false, true);
3263 }
3264
3265 queues.release ();
3266 bounds.release ();
3267 pointer_map_destroy (bb_bounds);
3268 pointer_map_destroy (block_priority);
3269 }
3270
3271 /* See if every path cross the loop goes through a statement that is known
3272 to not execute at the last iteration. In that case we can decrese iteration
3273 count by 1. */
3274
3275 static void
3276 maybe_lower_iteration_bound (struct loop *loop)
3277 {
3278 pointer_set_t *not_executed_last_iteration = NULL;
3279 struct nb_iter_bound *elt;
3280 bool found_exit = false;
3281 vec<basic_block> queue = vNULL;
3282 bitmap visited;
3283
3284 /* Collect all statements with interesting (i.e. lower than
3285 nb_iterations_upper_bound) bound on them.
3286
3287 TODO: Due to the way record_estimate choose estimates to store, the bounds
3288 will be always nb_iterations_upper_bound-1. We can change this to record
3289 also statements not dominating the loop latch and update the walk bellow
3290 to the shortest path algorthm. */
3291 for (elt = loop->bounds; elt; elt = elt->next)
3292 {
3293 if (!elt->is_exit
3294 && elt->bound.ult (loop->nb_iterations_upper_bound))
3295 {
3296 if (!not_executed_last_iteration)
3297 not_executed_last_iteration = pointer_set_create ();
3298 pointer_set_insert (not_executed_last_iteration, elt->stmt);
3299 }
3300 }
3301 if (!not_executed_last_iteration)
3302 return;
3303
3304 /* Start DFS walk in the loop header and see if we can reach the
3305 loop latch or any of the exits (including statements with side
3306 effects that may terminate the loop otherwise) without visiting
3307 any of the statements known to have undefined effect on the last
3308 iteration. */
3309 queue.safe_push (loop->header);
3310 visited = BITMAP_ALLOC (NULL);
3311 bitmap_set_bit (visited, loop->header->index);
3312 found_exit = false;
3313
3314 do
3315 {
3316 basic_block bb = queue.pop ();
3317 gimple_stmt_iterator gsi;
3318 bool stmt_found = false;
3319
3320 /* Loop for possible exits and statements bounding the execution. */
3321 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
3322 {
3323 gimple stmt = gsi_stmt (gsi);
3324 if (pointer_set_contains (not_executed_last_iteration, stmt))
3325 {
3326 stmt_found = true;
3327 break;
3328 }
3329 if (gimple_has_side_effects (stmt))
3330 {
3331 found_exit = true;
3332 break;
3333 }
3334 }
3335 if (found_exit)
3336 break;
3337
3338 /* If no bounding statement is found, continue the walk. */
3339 if (!stmt_found)
3340 {
3341 edge e;
3342 edge_iterator ei;
3343
3344 FOR_EACH_EDGE (e, ei, bb->succs)
3345 {
3346 if (loop_exit_edge_p (loop, e)
3347 || e == loop_latch_edge (loop))
3348 {
3349 found_exit = true;
3350 break;
3351 }
3352 if (bitmap_set_bit (visited, e->dest->index))
3353 queue.safe_push (e->dest);
3354 }
3355 }
3356 }
3357 while (queue.length () && !found_exit);
3358
3359 /* If every path through the loop reach bounding statement before exit,
3360 then we know the last iteration of the loop will have undefined effect
3361 and we can decrease number of iterations. */
3362
3363 if (!found_exit)
3364 {
3365 if (dump_file && (dump_flags & TDF_DETAILS))
3366 fprintf (dump_file, "Reducing loop iteration estimate by 1; "
3367 "undefined statement must be executed at the last iteration.\n");
3368 record_niter_bound (loop, loop->nb_iterations_upper_bound - double_int_one,
3369 false, true);
3370 }
3371 BITMAP_FREE (visited);
3372 queue.release ();
3373 pointer_set_destroy (not_executed_last_iteration);
3374 }
3375
3376 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P
3377 is true also use estimates derived from undefined behavior. */
3378
3379 static void
3380 estimate_numbers_of_iterations_loop (struct loop *loop)
3381 {
3382 vec<edge> exits;
3383 tree niter, type;
3384 unsigned i;
3385 struct tree_niter_desc niter_desc;
3386 edge ex;
3387 double_int bound;
3388 edge likely_exit;
3389
3390 /* Give up if we already have tried to compute an estimation. */
3391 if (loop->estimate_state != EST_NOT_COMPUTED)
3392 return;
3393
3394 loop->estimate_state = EST_AVAILABLE;
3395 /* Force estimate compuation but leave any existing upper bound in place. */
3396 loop->any_estimate = false;
3397
3398 /* Ensure that loop->nb_iterations is computed if possible. If it turns out
3399 to be constant, we avoid undefined behavior implied bounds and instead
3400 diagnose those loops with -Waggressive-loop-optimizations. */
3401 number_of_latch_executions (loop);
3402
3403 exits = get_loop_exit_edges (loop);
3404 likely_exit = single_likely_exit (loop);
3405 FOR_EACH_VEC_ELT (exits, i, ex)
3406 {
3407 if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false))
3408 continue;
3409
3410 niter = niter_desc.niter;
3411 type = TREE_TYPE (niter);
3412 if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST)
3413 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
3414 build_int_cst (type, 0),
3415 niter);
3416 record_estimate (loop, niter, niter_desc.max,
3417 last_stmt (ex->src),
3418 true, ex == likely_exit, true);
3419 }
3420 exits.release ();
3421
3422 if (flag_aggressive_loop_optimizations)
3423 infer_loop_bounds_from_undefined (loop);
3424
3425 discover_iteration_bound_by_body_walk (loop);
3426
3427 maybe_lower_iteration_bound (loop);
3428
3429 /* If we have a measured profile, use it to estimate the number of
3430 iterations. */
3431 if (loop->header->count != 0)
3432 {
3433 gcov_type nit = expected_loop_iterations_unbounded (loop) + 1;
3434 bound = gcov_type_to_double_int (nit);
3435 record_niter_bound (loop, bound, true, false);
3436 }
3437
3438 /* If we know the exact number of iterations of this loop, try to
3439 not break code with undefined behavior by not recording smaller
3440 maximum number of iterations. */
3441 if (loop->nb_iterations
3442 && TREE_CODE (loop->nb_iterations) == INTEGER_CST)
3443 {
3444 loop->any_upper_bound = true;
3445 loop->nb_iterations_upper_bound
3446 = tree_to_double_int (loop->nb_iterations);
3447 }
3448 }
3449
3450 /* Sets NIT to the estimated number of executions of the latch of the
3451 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
3452 large as the number of iterations. If we have no reliable estimate,
3453 the function returns false, otherwise returns true. */
3454
3455 bool
3456 estimated_loop_iterations (struct loop *loop, double_int *nit)
3457 {
3458 /* When SCEV information is available, try to update loop iterations
3459 estimate. Otherwise just return whatever we recorded earlier. */
3460 if (scev_initialized_p ())
3461 estimate_numbers_of_iterations_loop (loop);
3462
3463 return (get_estimated_loop_iterations (loop, nit));
3464 }
3465
3466 /* Similar to estimated_loop_iterations, but returns the estimate only
3467 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3468 on the number of iterations of LOOP could not be derived, returns -1. */
3469
3470 HOST_WIDE_INT
3471 estimated_loop_iterations_int (struct loop *loop)
3472 {
3473 double_int nit;
3474 HOST_WIDE_INT hwi_nit;
3475
3476 if (!estimated_loop_iterations (loop, &nit))
3477 return -1;
3478
3479 if (!nit.fits_shwi ())
3480 return -1;
3481 hwi_nit = nit.to_shwi ();
3482
3483 return hwi_nit < 0 ? -1 : hwi_nit;
3484 }
3485
3486
3487 /* Sets NIT to an upper bound for the maximum number of executions of the
3488 latch of the LOOP. If we have no reliable estimate, the function returns
3489 false, otherwise returns true. */
3490
3491 bool
3492 max_loop_iterations (struct loop *loop, double_int *nit)
3493 {
3494 /* When SCEV information is available, try to update loop iterations
3495 estimate. Otherwise just return whatever we recorded earlier. */
3496 if (scev_initialized_p ())
3497 estimate_numbers_of_iterations_loop (loop);
3498
3499 return get_max_loop_iterations (loop, nit);
3500 }
3501
3502 /* Similar to max_loop_iterations, but returns the estimate only
3503 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3504 on the number of iterations of LOOP could not be derived, returns -1. */
3505
3506 HOST_WIDE_INT
3507 max_loop_iterations_int (struct loop *loop)
3508 {
3509 double_int nit;
3510 HOST_WIDE_INT hwi_nit;
3511
3512 if (!max_loop_iterations (loop, &nit))
3513 return -1;
3514
3515 if (!nit.fits_shwi ())
3516 return -1;
3517 hwi_nit = nit.to_shwi ();
3518
3519 return hwi_nit < 0 ? -1 : hwi_nit;
3520 }
3521
3522 /* Returns an estimate for the number of executions of statements
3523 in the LOOP. For statements before the loop exit, this exceeds
3524 the number of execution of the latch by one. */
3525
3526 HOST_WIDE_INT
3527 estimated_stmt_executions_int (struct loop *loop)
3528 {
3529 HOST_WIDE_INT nit = estimated_loop_iterations_int (loop);
3530 HOST_WIDE_INT snit;
3531
3532 if (nit == -1)
3533 return -1;
3534
3535 snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1);
3536
3537 /* If the computation overflows, return -1. */
3538 return snit < 0 ? -1 : snit;
3539 }
3540
3541 /* Sets NIT to the estimated maximum number of executions of the latch of the
3542 LOOP, plus one. If we have no reliable estimate, the function returns
3543 false, otherwise returns true. */
3544
3545 bool
3546 max_stmt_executions (struct loop *loop, double_int *nit)
3547 {
3548 double_int nit_minus_one;
3549
3550 if (!max_loop_iterations (loop, nit))
3551 return false;
3552
3553 nit_minus_one = *nit;
3554
3555 *nit += double_int_one;
3556
3557 return (*nit).ugt (nit_minus_one);
3558 }
3559
3560 /* Sets NIT to the estimated number of executions of the latch of the
3561 LOOP, plus one. If we have no reliable estimate, the function returns
3562 false, otherwise returns true. */
3563
3564 bool
3565 estimated_stmt_executions (struct loop *loop, double_int *nit)
3566 {
3567 double_int nit_minus_one;
3568
3569 if (!estimated_loop_iterations (loop, nit))
3570 return false;
3571
3572 nit_minus_one = *nit;
3573
3574 *nit += double_int_one;
3575
3576 return (*nit).ugt (nit_minus_one);
3577 }
3578
3579 /* Records estimates on numbers of iterations of loops. */
3580
3581 void
3582 estimate_numbers_of_iterations (void)
3583 {
3584 loop_iterator li;
3585 struct loop *loop;
3586
3587 /* We don't want to issue signed overflow warnings while getting
3588 loop iteration estimates. */
3589 fold_defer_overflow_warnings ();
3590
3591 FOR_EACH_LOOP (li, loop, 0)
3592 {
3593 estimate_numbers_of_iterations_loop (loop);
3594 }
3595
3596 fold_undefer_and_ignore_overflow_warnings ();
3597 }
3598
3599 /* Returns true if statement S1 dominates statement S2. */
3600
3601 bool
3602 stmt_dominates_stmt_p (gimple s1, gimple s2)
3603 {
3604 basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
3605
3606 if (!bb1
3607 || s1 == s2)
3608 return true;
3609
3610 if (bb1 == bb2)
3611 {
3612 gimple_stmt_iterator bsi;
3613
3614 if (gimple_code (s2) == GIMPLE_PHI)
3615 return false;
3616
3617 if (gimple_code (s1) == GIMPLE_PHI)
3618 return true;
3619
3620 for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi))
3621 if (gsi_stmt (bsi) == s1)
3622 return true;
3623
3624 return false;
3625 }
3626
3627 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
3628 }
3629
3630 /* Returns true when we can prove that the number of executions of
3631 STMT in the loop is at most NITER, according to the bound on
3632 the number of executions of the statement NITER_BOUND->stmt recorded in
3633 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT.
3634
3635 ??? This code can become quite a CPU hog - we can have many bounds,
3636 and large basic block forcing stmt_dominates_stmt_p to be queried
3637 many times on a large basic blocks, so the whole thing is O(n^2)
3638 for scev_probably_wraps_p invocation (that can be done n times).
3639
3640 It would make more sense (and give better answers) to remember BB
3641 bounds computed by discover_iteration_bound_by_body_walk. */
3642
3643 static bool
3644 n_of_executions_at_most (gimple stmt,
3645 struct nb_iter_bound *niter_bound,
3646 tree niter)
3647 {
3648 double_int bound = niter_bound->bound;
3649 tree nit_type = TREE_TYPE (niter), e;
3650 enum tree_code cmp;
3651
3652 gcc_assert (TYPE_UNSIGNED (nit_type));
3653
3654 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
3655 the number of iterations is small. */
3656 if (!double_int_fits_to_tree_p (nit_type, bound))
3657 return false;
3658
3659 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
3660 times. This means that:
3661
3662 -- if NITER_BOUND->is_exit is true, then everything after
3663 it at most NITER_BOUND->bound times.
3664
3665 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
3666 is executed, then NITER_BOUND->stmt is executed as well in the same
3667 iteration then STMT is executed at most NITER_BOUND->bound + 1 times.
3668
3669 If we can determine that NITER_BOUND->stmt is always executed
3670 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times.
3671 We conclude that if both statements belong to the same
3672 basic block and STMT is before NITER_BOUND->stmt and there are no
3673 statements with side effects in between. */
3674
3675 if (niter_bound->is_exit)
3676 {
3677 if (stmt == niter_bound->stmt
3678 || !stmt_dominates_stmt_p (niter_bound->stmt, stmt))
3679 return false;
3680 cmp = GE_EXPR;
3681 }
3682 else
3683 {
3684 if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt))
3685 {
3686 gimple_stmt_iterator bsi;
3687 if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt)
3688 || gimple_code (stmt) == GIMPLE_PHI
3689 || gimple_code (niter_bound->stmt) == GIMPLE_PHI)
3690 return false;
3691
3692 /* By stmt_dominates_stmt_p we already know that STMT appears
3693 before NITER_BOUND->STMT. Still need to test that the loop
3694 can not be terinated by a side effect in between. */
3695 for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt;
3696 gsi_next (&bsi))
3697 if (gimple_has_side_effects (gsi_stmt (bsi)))
3698 return false;
3699 bound += double_int_one;
3700 if (bound.is_zero ()
3701 || !double_int_fits_to_tree_p (nit_type, bound))
3702 return false;
3703 }
3704 cmp = GT_EXPR;
3705 }
3706
3707 e = fold_binary (cmp, boolean_type_node,
3708 niter, double_int_to_tree (nit_type, bound));
3709 return e && integer_nonzerop (e);
3710 }
3711
3712 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
3713
3714 bool
3715 nowrap_type_p (tree type)
3716 {
3717 if (INTEGRAL_TYPE_P (type)
3718 && TYPE_OVERFLOW_UNDEFINED (type))
3719 return true;
3720
3721 if (POINTER_TYPE_P (type))
3722 return true;
3723
3724 return false;
3725 }
3726
3727 /* Return false only when the induction variable BASE + STEP * I is
3728 known to not overflow: i.e. when the number of iterations is small
3729 enough with respect to the step and initial condition in order to
3730 keep the evolution confined in TYPEs bounds. Return true when the
3731 iv is known to overflow or when the property is not computable.
3732
3733 USE_OVERFLOW_SEMANTICS is true if this function should assume that
3734 the rules for overflow of the given language apply (e.g., that signed
3735 arithmetics in C does not overflow). */
3736
3737 bool
3738 scev_probably_wraps_p (tree base, tree step,
3739 gimple at_stmt, struct loop *loop,
3740 bool use_overflow_semantics)
3741 {
3742 tree delta, step_abs;
3743 tree unsigned_type, valid_niter;
3744 tree type = TREE_TYPE (step);
3745 tree e;
3746 double_int niter;
3747 struct nb_iter_bound *bound;
3748
3749 /* FIXME: We really need something like
3750 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
3751
3752 We used to test for the following situation that frequently appears
3753 during address arithmetics:
3754
3755 D.1621_13 = (long unsigned intD.4) D.1620_12;
3756 D.1622_14 = D.1621_13 * 8;
3757 D.1623_15 = (doubleD.29 *) D.1622_14;
3758
3759 And derived that the sequence corresponding to D_14
3760 can be proved to not wrap because it is used for computing a
3761 memory access; however, this is not really the case -- for example,
3762 if D_12 = (unsigned char) [254,+,1], then D_14 has values
3763 2032, 2040, 0, 8, ..., but the code is still legal. */
3764
3765 if (chrec_contains_undetermined (base)
3766 || chrec_contains_undetermined (step))
3767 return true;
3768
3769 if (integer_zerop (step))
3770 return false;
3771
3772 /* If we can use the fact that signed and pointer arithmetics does not
3773 wrap, we are done. */
3774 if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base)))
3775 return false;
3776
3777 /* To be able to use estimates on number of iterations of the loop,
3778 we must have an upper bound on the absolute value of the step. */
3779 if (TREE_CODE (step) != INTEGER_CST)
3780 return true;
3781
3782 /* Don't issue signed overflow warnings. */
3783 fold_defer_overflow_warnings ();
3784
3785 /* Otherwise, compute the number of iterations before we reach the
3786 bound of the type, and verify that the loop is exited before this
3787 occurs. */
3788 unsigned_type = unsigned_type_for (type);
3789 base = fold_convert (unsigned_type, base);
3790
3791 if (tree_int_cst_sign_bit (step))
3792 {
3793 tree extreme = fold_convert (unsigned_type,
3794 lower_bound_in_type (type, type));
3795 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
3796 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type,
3797 fold_convert (unsigned_type, step));
3798 }
3799 else
3800 {
3801 tree extreme = fold_convert (unsigned_type,
3802 upper_bound_in_type (type, type));
3803 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
3804 step_abs = fold_convert (unsigned_type, step);
3805 }
3806
3807 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
3808
3809 estimate_numbers_of_iterations_loop (loop);
3810
3811 if (max_loop_iterations (loop, &niter)
3812 && double_int_fits_to_tree_p (TREE_TYPE (valid_niter), niter)
3813 && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter,
3814 double_int_to_tree (TREE_TYPE (valid_niter),
3815 niter))) != NULL
3816 && integer_nonzerop (e))
3817 {
3818 fold_undefer_and_ignore_overflow_warnings ();
3819 return false;
3820 }
3821 if (at_stmt)
3822 for (bound = loop->bounds; bound; bound = bound->next)
3823 {
3824 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
3825 {
3826 fold_undefer_and_ignore_overflow_warnings ();
3827 return false;
3828 }
3829 }
3830
3831 fold_undefer_and_ignore_overflow_warnings ();
3832
3833 /* At this point we still don't have a proof that the iv does not
3834 overflow: give up. */
3835 return true;
3836 }
3837
3838 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
3839
3840 void
3841 free_numbers_of_iterations_estimates_loop (struct loop *loop)
3842 {
3843 struct nb_iter_bound *bound, *next;
3844
3845 loop->nb_iterations = NULL;
3846 loop->estimate_state = EST_NOT_COMPUTED;
3847 for (bound = loop->bounds; bound; bound = next)
3848 {
3849 next = bound->next;
3850 ggc_free (bound);
3851 }
3852
3853 loop->bounds = NULL;
3854 }
3855
3856 /* Frees the information on upper bounds on numbers of iterations of loops. */
3857
3858 void
3859 free_numbers_of_iterations_estimates (void)
3860 {
3861 loop_iterator li;
3862 struct loop *loop;
3863
3864 FOR_EACH_LOOP (li, loop, 0)
3865 {
3866 free_numbers_of_iterations_estimates_loop (loop);
3867 }
3868 }
3869
3870 /* Substitute value VAL for ssa name NAME inside expressions held
3871 at LOOP. */
3872
3873 void
3874 substitute_in_loop_info (struct loop *loop, tree name, tree val)
3875 {
3876 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);
3877 }