Fix mistake in my previous commit (n_of_executions_at_least should be
[official-gcc.git] / gcc / tree-ssa-loop-niter.c
blobbc359ad309f101b0079f2350cc420f8c24483efe
1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004, 2005 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 2, or (at your option) any
9 later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING. If not, write to the Free
18 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
19 02110-1301, USA. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "basic-block.h"
30 #include "output.h"
31 #include "diagnostic.h"
32 #include "intl.h"
33 #include "tree-flow.h"
34 #include "tree-dump.h"
35 #include "cfgloop.h"
36 #include "tree-pass.h"
37 #include "ggc.h"
38 #include "tree-chrec.h"
39 #include "tree-scalar-evolution.h"
40 #include "tree-data-ref.h"
41 #include "params.h"
42 #include "flags.h"
43 #include "toplev.h"
44 #include "tree-inline.h"
46 #define SWAP(X, Y) do { void *tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
51 Analysis of number of iterations of an affine exit test.
55 /* Returns true if ARG is either NULL_TREE or constant zero. Unlike
56 integer_zerop, it does not care about overflow flags. */
58 bool
59 zero_p (tree arg)
61 if (!arg)
62 return true;
64 if (TREE_CODE (arg) != INTEGER_CST)
65 return false;
67 return (TREE_INT_CST_LOW (arg) == 0 && TREE_INT_CST_HIGH (arg) == 0);
70 /* Returns true if ARG a nonzero constant. Unlike integer_nonzerop, it does
71 not care about overflow flags. */
73 static bool
74 nonzero_p (tree arg)
76 if (!arg)
77 return false;
79 if (TREE_CODE (arg) != INTEGER_CST)
80 return false;
82 return (TREE_INT_CST_LOW (arg) != 0 || TREE_INT_CST_HIGH (arg) != 0);
85 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
87 static tree
88 inverse (tree x, tree mask)
90 tree type = TREE_TYPE (x);
91 tree rslt;
92 unsigned ctr = tree_floor_log2 (mask);
94 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT)
96 unsigned HOST_WIDE_INT ix;
97 unsigned HOST_WIDE_INT imask;
98 unsigned HOST_WIDE_INT irslt = 1;
100 gcc_assert (cst_and_fits_in_hwi (x));
101 gcc_assert (cst_and_fits_in_hwi (mask));
103 ix = int_cst_value (x);
104 imask = int_cst_value (mask);
106 for (; ctr; ctr--)
108 irslt *= ix;
109 ix *= ix;
111 irslt &= imask;
113 rslt = build_int_cst_type (type, irslt);
115 else
117 rslt = build_int_cst (type, 1);
118 for (; ctr; ctr--)
120 rslt = int_const_binop (MULT_EXPR, rslt, x, 0);
121 x = int_const_binop (MULT_EXPR, x, x, 0);
123 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask, 0);
126 return rslt;
129 /* Determines number of iterations of loop whose ending condition
130 is IV <> FINAL. TYPE is the type of the iv. The number of
131 iterations is stored to NITER. NEVER_INFINITE is true if
132 we know that the exit must be taken eventually, i.e., that the IV
133 ever reaches the value FINAL (we derived this earlier, and possibly set
134 NITER->assumptions to make sure this is the case). */
136 static bool
137 number_of_iterations_ne (tree type, affine_iv *iv, tree final,
138 struct tree_niter_desc *niter, bool never_infinite)
140 tree niter_type = unsigned_type_for (type);
141 tree s, c, d, bits, assumption, tmp, bound;
143 niter->control = *iv;
144 niter->bound = final;
145 niter->cmp = NE_EXPR;
147 /* Rearrange the terms so that we get inequality s * i <> c, with s
148 positive. Also cast everything to the unsigned type. */
149 if (tree_int_cst_sign_bit (iv->step))
151 s = fold_convert (niter_type,
152 fold_build1 (NEGATE_EXPR, type, iv->step));
153 c = fold_build2 (MINUS_EXPR, niter_type,
154 fold_convert (niter_type, iv->base),
155 fold_convert (niter_type, final));
157 else
159 s = fold_convert (niter_type, iv->step);
160 c = fold_build2 (MINUS_EXPR, niter_type,
161 fold_convert (niter_type, final),
162 fold_convert (niter_type, iv->base));
165 /* First the trivial cases -- when the step is 1. */
166 if (integer_onep (s))
168 niter->niter = c;
169 return true;
172 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
173 is infinite. Otherwise, the number of iterations is
174 (inverse(s/d) * (c/d)) mod (size of mode/d). */
175 bits = num_ending_zeros (s);
176 bound = build_low_bits_mask (niter_type,
177 (TYPE_PRECISION (niter_type)
178 - tree_low_cst (bits, 1)));
180 d = fold_binary_to_constant (LSHIFT_EXPR, niter_type,
181 build_int_cst (niter_type, 1), bits);
182 s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits);
184 if (!never_infinite)
186 /* If we cannot assume that the loop is not infinite, record the
187 assumptions for divisibility of c. */
188 assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d);
189 assumption = fold_build2 (EQ_EXPR, boolean_type_node,
190 assumption, build_int_cst (niter_type, 0));
191 if (!nonzero_p (assumption))
192 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
193 niter->assumptions, assumption);
196 c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d);
197 tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound));
198 niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound);
199 return true;
202 /* Checks whether we can determine the final value of the control variable
203 of the loop with ending condition IV0 < IV1 (computed in TYPE).
204 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
205 of the step. The assumptions necessary to ensure that the computation
206 of the final value does not overflow are recorded in NITER. If we
207 find the final value, we adjust DELTA and return TRUE. Otherwise
208 we return false. */
210 static bool
211 number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1,
212 struct tree_niter_desc *niter,
213 tree *delta, tree step)
215 tree niter_type = TREE_TYPE (step);
216 tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step);
217 tree tmod;
218 tree assumption = boolean_true_node, bound, noloop;
220 if (TREE_CODE (mod) != INTEGER_CST)
221 return false;
222 if (nonzero_p (mod))
223 mod = fold_build2 (MINUS_EXPR, niter_type, step, mod);
224 tmod = fold_convert (type, mod);
226 if (nonzero_p (iv0->step))
228 /* The final value of the iv is iv1->base + MOD, assuming that this
229 computation does not overflow, and that
230 iv0->base <= iv1->base + MOD. */
231 if (!iv1->no_overflow && !zero_p (mod))
233 bound = fold_build2 (MINUS_EXPR, type,
234 TYPE_MAX_VALUE (type), tmod);
235 assumption = fold_build2 (LE_EXPR, boolean_type_node,
236 iv1->base, bound);
237 if (zero_p (assumption))
238 return false;
240 noloop = fold_build2 (GT_EXPR, boolean_type_node,
241 iv0->base,
242 fold_build2 (PLUS_EXPR, type,
243 iv1->base, tmod));
245 else
247 /* The final value of the iv is iv0->base - MOD, assuming that this
248 computation does not overflow, and that
249 iv0->base - MOD <= iv1->base. */
250 if (!iv0->no_overflow && !zero_p (mod))
252 bound = fold_build2 (PLUS_EXPR, type,
253 TYPE_MIN_VALUE (type), tmod);
254 assumption = fold_build2 (GE_EXPR, boolean_type_node,
255 iv0->base, bound);
256 if (zero_p (assumption))
257 return false;
259 noloop = fold_build2 (GT_EXPR, boolean_type_node,
260 fold_build2 (MINUS_EXPR, type,
261 iv0->base, tmod),
262 iv1->base);
265 if (!nonzero_p (assumption))
266 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
267 niter->assumptions,
268 assumption);
269 if (!zero_p (noloop))
270 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
271 niter->may_be_zero,
272 noloop);
273 *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod);
274 return true;
277 /* Add assertions to NITER that ensure that the control variable of the loop
278 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
279 are TYPE. Returns false if we can prove that there is an overflow, true
280 otherwise. STEP is the absolute value of the step. */
282 static bool
283 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1,
284 struct tree_niter_desc *niter, tree step)
286 tree bound, d, assumption, diff;
287 tree niter_type = TREE_TYPE (step);
289 if (nonzero_p (iv0->step))
291 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
292 if (iv0->no_overflow)
293 return true;
295 /* If iv0->base is a constant, we can determine the last value before
296 overflow precisely; otherwise we conservatively assume
297 MAX - STEP + 1. */
299 if (TREE_CODE (iv0->base) == INTEGER_CST)
301 d = fold_build2 (MINUS_EXPR, niter_type,
302 fold_convert (niter_type, TYPE_MAX_VALUE (type)),
303 fold_convert (niter_type, iv0->base));
304 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
306 else
307 diff = fold_build2 (MINUS_EXPR, niter_type, step,
308 build_int_cst (niter_type, 1));
309 bound = fold_build2 (MINUS_EXPR, type,
310 TYPE_MAX_VALUE (type), fold_convert (type, diff));
311 assumption = fold_build2 (LE_EXPR, boolean_type_node,
312 iv1->base, bound);
314 else
316 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
317 if (iv1->no_overflow)
318 return true;
320 if (TREE_CODE (iv1->base) == INTEGER_CST)
322 d = fold_build2 (MINUS_EXPR, niter_type,
323 fold_convert (niter_type, iv1->base),
324 fold_convert (niter_type, TYPE_MIN_VALUE (type)));
325 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
327 else
328 diff = fold_build2 (MINUS_EXPR, niter_type, step,
329 build_int_cst (niter_type, 1));
330 bound = fold_build2 (PLUS_EXPR, type,
331 TYPE_MIN_VALUE (type), fold_convert (type, diff));
332 assumption = fold_build2 (GE_EXPR, boolean_type_node,
333 iv0->base, bound);
336 if (zero_p (assumption))
337 return false;
338 if (!nonzero_p (assumption))
339 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
340 niter->assumptions, assumption);
342 iv0->no_overflow = true;
343 iv1->no_overflow = true;
344 return true;
347 /* Add an assumption to NITER that a loop whose ending condition
348 is IV0 < IV1 rolls. TYPE is the type of the control iv. */
350 static void
351 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1,
352 struct tree_niter_desc *niter)
354 tree assumption = boolean_true_node, bound, diff;
355 tree mbz, mbzl, mbzr;
357 if (nonzero_p (iv0->step))
359 diff = fold_build2 (MINUS_EXPR, type,
360 iv0->step, build_int_cst (type, 1));
362 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
363 0 address never belongs to any object, we can assume this for
364 pointers. */
365 if (!POINTER_TYPE_P (type))
367 bound = fold_build2 (PLUS_EXPR, type,
368 TYPE_MIN_VALUE (type), diff);
369 assumption = fold_build2 (GE_EXPR, boolean_type_node,
370 iv0->base, bound);
373 /* And then we can compute iv0->base - diff, and compare it with
374 iv1->base. */
375 mbzl = fold_build2 (MINUS_EXPR, type, iv0->base, diff);
376 mbzr = iv1->base;
378 else
380 diff = fold_build2 (PLUS_EXPR, type,
381 iv1->step, build_int_cst (type, 1));
383 if (!POINTER_TYPE_P (type))
385 bound = fold_build2 (PLUS_EXPR, type,
386 TYPE_MAX_VALUE (type), diff);
387 assumption = fold_build2 (LE_EXPR, boolean_type_node,
388 iv1->base, bound);
391 mbzl = iv0->base;
392 mbzr = fold_build2 (MINUS_EXPR, type, iv1->base, diff);
395 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr);
397 if (!nonzero_p (assumption))
398 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
399 niter->assumptions, assumption);
400 if (!zero_p (mbz))
401 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
402 niter->may_be_zero, mbz);
405 /* Determines number of iterations of loop whose ending condition
406 is IV0 < IV1. TYPE is the type of the iv. The number of
407 iterations is stored to NITER. */
409 static bool
410 number_of_iterations_lt (tree type, affine_iv *iv0, affine_iv *iv1,
411 struct tree_niter_desc *niter,
412 bool never_infinite ATTRIBUTE_UNUSED)
414 tree niter_type = unsigned_type_for (type);
415 tree delta, step, s;
417 if (nonzero_p (iv0->step))
419 niter->control = *iv0;
420 niter->cmp = LT_EXPR;
421 niter->bound = iv1->base;
423 else
425 niter->control = *iv1;
426 niter->cmp = GT_EXPR;
427 niter->bound = iv0->base;
430 delta = fold_build2 (MINUS_EXPR, niter_type,
431 fold_convert (niter_type, iv1->base),
432 fold_convert (niter_type, iv0->base));
434 /* First handle the special case that the step is +-1. */
435 if ((iv0->step && integer_onep (iv0->step)
436 && zero_p (iv1->step))
437 || (iv1->step && integer_all_onesp (iv1->step)
438 && zero_p (iv0->step)))
440 /* for (i = iv0->base; i < iv1->base; i++)
444 for (i = iv1->base; i > iv0->base; i--).
446 In both cases # of iterations is iv1->base - iv0->base, assuming that
447 iv1->base >= iv0->base. */
448 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node,
449 iv1->base, iv0->base);
450 niter->niter = delta;
451 return true;
454 if (nonzero_p (iv0->step))
455 step = fold_convert (niter_type, iv0->step);
456 else
457 step = fold_convert (niter_type,
458 fold_build1 (NEGATE_EXPR, type, iv1->step));
460 /* If we can determine the final value of the control iv exactly, we can
461 transform the condition to != comparison. In particular, this will be
462 the case if DELTA is constant. */
463 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step))
465 affine_iv zps;
467 zps.base = build_int_cst (niter_type, 0);
468 zps.step = step;
469 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
470 zps does not overflow. */
471 zps.no_overflow = true;
473 return number_of_iterations_ne (type, &zps, delta, niter, true);
476 /* Make sure that the control iv does not overflow. */
477 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step))
478 return false;
480 /* We determine the number of iterations as (delta + step - 1) / step. For
481 this to work, we must know that iv1->base >= iv0->base - step + 1,
482 otherwise the loop does not roll. */
483 assert_loop_rolls_lt (type, iv0, iv1, niter);
485 s = fold_build2 (MINUS_EXPR, niter_type,
486 step, build_int_cst (niter_type, 1));
487 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s);
488 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step);
489 return true;
492 /* Determines number of iterations of loop whose ending condition
493 is IV0 <= IV1. TYPE is the type of the iv. The number of
494 iterations is stored to NITER. NEVER_INFINITE is true if
495 we know that this condition must eventually become false (we derived this
496 earlier, and possibly set NITER->assumptions to make sure this
497 is the case). */
499 static bool
500 number_of_iterations_le (tree type, affine_iv *iv0, affine_iv *iv1,
501 struct tree_niter_desc *niter, bool never_infinite)
503 tree assumption;
505 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
506 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
507 value of the type. This we must know anyway, since if it is
508 equal to this value, the loop rolls forever. */
510 if (!never_infinite)
512 if (nonzero_p (iv0->step))
513 assumption = fold_build2 (NE_EXPR, boolean_type_node,
514 iv1->base, TYPE_MAX_VALUE (type));
515 else
516 assumption = fold_build2 (NE_EXPR, boolean_type_node,
517 iv0->base, TYPE_MIN_VALUE (type));
519 if (zero_p (assumption))
520 return false;
521 if (!nonzero_p (assumption))
522 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
523 niter->assumptions, assumption);
526 if (nonzero_p (iv0->step))
527 iv1->base = fold_build2 (PLUS_EXPR, type,
528 iv1->base, build_int_cst (type, 1));
529 else
530 iv0->base = fold_build2 (MINUS_EXPR, type,
531 iv0->base, build_int_cst (type, 1));
532 return number_of_iterations_lt (type, iv0, iv1, niter, never_infinite);
535 /* Determine the number of iterations according to condition (for staying
536 inside loop) which compares two induction variables using comparison
537 operator CODE. The induction variable on left side of the comparison
538 is IV0, the right-hand side is IV1. Both induction variables must have
539 type TYPE, which must be an integer or pointer type. The steps of the
540 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
542 ONLY_EXIT is true if we are sure this is the only way the loop could be
543 exited (including possibly non-returning function calls, exceptions, etc.)
544 -- in this case we can use the information whether the control induction
545 variables can overflow or not in a more efficient way.
547 The results (number of iterations and assumptions as described in
548 comments at struct tree_niter_desc in tree-flow.h) are stored to NITER.
549 Returns false if it fails to determine number of iterations, true if it
550 was determined (possibly with some assumptions). */
552 static bool
553 number_of_iterations_cond (tree type, affine_iv *iv0, enum tree_code code,
554 affine_iv *iv1, struct tree_niter_desc *niter,
555 bool only_exit)
557 bool never_infinite;
559 /* The meaning of these assumptions is this:
560 if !assumptions
561 then the rest of information does not have to be valid
562 if may_be_zero then the loop does not roll, even if
563 niter != 0. */
564 niter->assumptions = boolean_true_node;
565 niter->may_be_zero = boolean_false_node;
566 niter->niter = NULL_TREE;
567 niter->additional_info = boolean_true_node;
569 niter->bound = NULL_TREE;
570 niter->cmp = ERROR_MARK;
572 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
573 the control variable is on lhs. */
574 if (code == GE_EXPR || code == GT_EXPR
575 || (code == NE_EXPR && zero_p (iv0->step)))
577 SWAP (iv0, iv1);
578 code = swap_tree_comparison (code);
581 if (!only_exit)
583 /* If this is not the only possible exit from the loop, the information
584 that the induction variables cannot overflow as derived from
585 signedness analysis cannot be relied upon. We use them e.g. in the
586 following way: given loop for (i = 0; i <= n; i++), if i is
587 signed, it cannot overflow, thus this loop is equivalent to
588 for (i = 0; i < n + 1; i++); however, if n == MAX, but the loop
589 is exited in some other way before i overflows, this transformation
590 is incorrect (the new loop exits immediately). */
591 iv0->no_overflow = false;
592 iv1->no_overflow = false;
595 if (POINTER_TYPE_P (type))
597 /* Comparison of pointers is undefined unless both iv0 and iv1 point
598 to the same object. If they do, the control variable cannot wrap
599 (as wrap around the bounds of memory will never return a pointer
600 that would be guaranteed to point to the same object, even if we
601 avoid undefined behavior by casting to size_t and back). The
602 restrictions on pointer arithmetics and comparisons of pointers
603 ensure that using the no-overflow assumptions is correct in this
604 case even if ONLY_EXIT is false. */
605 iv0->no_overflow = true;
606 iv1->no_overflow = true;
609 /* If the control induction variable does not overflow, the loop obviously
610 cannot be infinite. */
611 if (!zero_p (iv0->step) && iv0->no_overflow)
612 never_infinite = true;
613 else if (!zero_p (iv1->step) && iv1->no_overflow)
614 never_infinite = true;
615 else
616 never_infinite = false;
618 /* We can handle the case when neither of the sides of the comparison is
619 invariant, provided that the test is NE_EXPR. This rarely occurs in
620 practice, but it is simple enough to manage. */
621 if (!zero_p (iv0->step) && !zero_p (iv1->step))
623 if (code != NE_EXPR)
624 return false;
626 iv0->step = fold_binary_to_constant (MINUS_EXPR, type,
627 iv0->step, iv1->step);
628 iv0->no_overflow = false;
629 iv1->step = NULL_TREE;
630 iv1->no_overflow = true;
633 /* If the result of the comparison is a constant, the loop is weird. More
634 precise handling would be possible, but the situation is not common enough
635 to waste time on it. */
636 if (zero_p (iv0->step) && zero_p (iv1->step))
637 return false;
639 /* Ignore loops of while (i-- < 10) type. */
640 if (code != NE_EXPR)
642 if (iv0->step && tree_int_cst_sign_bit (iv0->step))
643 return false;
645 if (!zero_p (iv1->step) && !tree_int_cst_sign_bit (iv1->step))
646 return false;
649 /* If the loop exits immediately, there is nothing to do. */
650 if (zero_p (fold_build2 (code, boolean_type_node, iv0->base, iv1->base)))
652 niter->niter = build_int_cst (unsigned_type_for (type), 0);
653 return true;
656 /* OK, now we know we have a senseful loop. Handle several cases, depending
657 on what comparison operator is used. */
658 switch (code)
660 case NE_EXPR:
661 gcc_assert (zero_p (iv1->step));
662 return number_of_iterations_ne (type, iv0, iv1->base, niter, never_infinite);
663 case LT_EXPR:
664 return number_of_iterations_lt (type, iv0, iv1, niter, never_infinite);
665 case LE_EXPR:
666 return number_of_iterations_le (type, iv0, iv1, niter, never_infinite);
667 default:
668 gcc_unreachable ();
672 /* Substitute NEW for OLD in EXPR and fold the result. */
674 static tree
675 simplify_replace_tree (tree expr, tree old, tree new)
677 unsigned i, n;
678 tree ret = NULL_TREE, e, se;
680 if (!expr)
681 return NULL_TREE;
683 if (expr == old
684 || operand_equal_p (expr, old, 0))
685 return unshare_expr (new);
687 if (!EXPR_P (expr))
688 return expr;
690 n = TREE_CODE_LENGTH (TREE_CODE (expr));
691 for (i = 0; i < n; i++)
693 e = TREE_OPERAND (expr, i);
694 se = simplify_replace_tree (e, old, new);
695 if (e == se)
696 continue;
698 if (!ret)
699 ret = copy_node (expr);
701 TREE_OPERAND (ret, i) = se;
704 return (ret ? fold (ret) : expr);
707 /* Expand definitions of ssa names in EXPR as long as they are simple
708 enough, and return the new expression. */
710 tree
711 expand_simple_operations (tree expr)
713 unsigned i, n;
714 tree ret = NULL_TREE, e, ee, stmt;
715 enum tree_code code;
717 if (expr == NULL_TREE)
718 return expr;
720 if (is_gimple_min_invariant (expr))
721 return expr;
723 code = TREE_CODE (expr);
724 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code)))
726 n = TREE_CODE_LENGTH (code);
727 for (i = 0; i < n; i++)
729 e = TREE_OPERAND (expr, i);
730 ee = expand_simple_operations (e);
731 if (e == ee)
732 continue;
734 if (!ret)
735 ret = copy_node (expr);
737 TREE_OPERAND (ret, i) = ee;
740 return (ret ? fold (ret) : expr);
743 if (TREE_CODE (expr) != SSA_NAME)
744 return expr;
746 stmt = SSA_NAME_DEF_STMT (expr);
747 if (TREE_CODE (stmt) != MODIFY_EXPR)
748 return expr;
750 e = TREE_OPERAND (stmt, 1);
751 if (/* Casts are simple. */
752 TREE_CODE (e) != NOP_EXPR
753 && TREE_CODE (e) != CONVERT_EXPR
754 /* Copies are simple. */
755 && TREE_CODE (e) != SSA_NAME
756 /* Assignments of invariants are simple. */
757 && !is_gimple_min_invariant (e)
758 /* And increments and decrements by a constant are simple. */
759 && !((TREE_CODE (e) == PLUS_EXPR
760 || TREE_CODE (e) == MINUS_EXPR)
761 && is_gimple_min_invariant (TREE_OPERAND (e, 1))))
762 return expr;
764 return expand_simple_operations (e);
767 /* Tries to simplify EXPR using the condition COND. Returns the simplified
768 expression (or EXPR unchanged, if no simplification was possible). */
770 static tree
771 tree_simplify_using_condition_1 (tree cond, tree expr)
773 bool changed;
774 tree e, te, e0, e1, e2, notcond;
775 enum tree_code code = TREE_CODE (expr);
777 if (code == INTEGER_CST)
778 return expr;
780 if (code == TRUTH_OR_EXPR
781 || code == TRUTH_AND_EXPR
782 || code == COND_EXPR)
784 changed = false;
786 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0));
787 if (TREE_OPERAND (expr, 0) != e0)
788 changed = true;
790 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1));
791 if (TREE_OPERAND (expr, 1) != e1)
792 changed = true;
794 if (code == COND_EXPR)
796 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2));
797 if (TREE_OPERAND (expr, 2) != e2)
798 changed = true;
800 else
801 e2 = NULL_TREE;
803 if (changed)
805 if (code == COND_EXPR)
806 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
807 else
808 expr = fold_build2 (code, boolean_type_node, e0, e1);
811 return expr;
814 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
815 propagation, and vice versa. Fold does not handle this, since it is
816 considered too expensive. */
817 if (TREE_CODE (cond) == EQ_EXPR)
819 e0 = TREE_OPERAND (cond, 0);
820 e1 = TREE_OPERAND (cond, 1);
822 /* We know that e0 == e1. Check whether we cannot simplify expr
823 using this fact. */
824 e = simplify_replace_tree (expr, e0, e1);
825 if (zero_p (e) || nonzero_p (e))
826 return e;
828 e = simplify_replace_tree (expr, e1, e0);
829 if (zero_p (e) || nonzero_p (e))
830 return e;
832 if (TREE_CODE (expr) == EQ_EXPR)
834 e0 = TREE_OPERAND (expr, 0);
835 e1 = TREE_OPERAND (expr, 1);
837 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
838 e = simplify_replace_tree (cond, e0, e1);
839 if (zero_p (e))
840 return e;
841 e = simplify_replace_tree (cond, e1, e0);
842 if (zero_p (e))
843 return e;
845 if (TREE_CODE (expr) == NE_EXPR)
847 e0 = TREE_OPERAND (expr, 0);
848 e1 = TREE_OPERAND (expr, 1);
850 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
851 e = simplify_replace_tree (cond, e0, e1);
852 if (zero_p (e))
853 return boolean_true_node;
854 e = simplify_replace_tree (cond, e1, e0);
855 if (zero_p (e))
856 return boolean_true_node;
859 te = expand_simple_operations (expr);
861 /* Check whether COND ==> EXPR. */
862 notcond = invert_truthvalue (cond);
863 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, te);
864 if (nonzero_p (e))
865 return e;
867 /* Check whether COND ==> not EXPR. */
868 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, te);
869 if (e && zero_p (e))
870 return e;
872 return expr;
875 /* Tries to simplify EXPR using the condition COND. Returns the simplified
876 expression (or EXPR unchanged, if no simplification was possible).
877 Wrapper around tree_simplify_using_condition_1 that ensures that chains
878 of simple operations in definitions of ssa names in COND are expanded,
879 so that things like casts or incrementing the value of the bound before
880 the loop do not cause us to fail. */
882 static tree
883 tree_simplify_using_condition (tree cond, tree expr)
885 cond = expand_simple_operations (cond);
887 return tree_simplify_using_condition_1 (cond, expr);
890 /* Tries to simplify EXPR using the conditions on entry to LOOP.
891 Record the conditions used for simplification to CONDS_USED.
892 Returns the simplified expression (or EXPR unchanged, if no
893 simplification was possible).*/
895 static tree
896 simplify_using_initial_conditions (struct loop *loop, tree expr,
897 tree *conds_used)
899 edge e;
900 basic_block bb;
901 tree exp, cond;
903 if (TREE_CODE (expr) == INTEGER_CST)
904 return expr;
906 for (bb = loop->header;
907 bb != ENTRY_BLOCK_PTR;
908 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
910 if (!single_pred_p (bb))
911 continue;
912 e = single_pred_edge (bb);
914 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
915 continue;
917 cond = COND_EXPR_COND (last_stmt (e->src));
918 if (e->flags & EDGE_FALSE_VALUE)
919 cond = invert_truthvalue (cond);
920 exp = tree_simplify_using_condition (cond, expr);
922 if (exp != expr)
923 *conds_used = fold_build2 (TRUTH_AND_EXPR,
924 boolean_type_node,
925 *conds_used,
926 cond);
928 expr = exp;
931 return expr;
934 /* Tries to simplify EXPR using the evolutions of the loop invariants
935 in the superloops of LOOP. Returns the simplified expression
936 (or EXPR unchanged, if no simplification was possible). */
938 static tree
939 simplify_using_outer_evolutions (struct loop *loop, tree expr)
941 enum tree_code code = TREE_CODE (expr);
942 bool changed;
943 tree e, e0, e1, e2;
945 if (is_gimple_min_invariant (expr))
946 return expr;
948 if (code == TRUTH_OR_EXPR
949 || code == TRUTH_AND_EXPR
950 || code == COND_EXPR)
952 changed = false;
954 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0));
955 if (TREE_OPERAND (expr, 0) != e0)
956 changed = true;
958 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1));
959 if (TREE_OPERAND (expr, 1) != e1)
960 changed = true;
962 if (code == COND_EXPR)
964 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2));
965 if (TREE_OPERAND (expr, 2) != e2)
966 changed = true;
968 else
969 e2 = NULL_TREE;
971 if (changed)
973 if (code == COND_EXPR)
974 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
975 else
976 expr = fold_build2 (code, boolean_type_node, e0, e1);
979 return expr;
982 e = instantiate_parameters (loop, expr);
983 if (is_gimple_min_invariant (e))
984 return e;
986 return expr;
989 /* Returns true if EXIT is the only possible exit from LOOP. */
991 static bool
992 loop_only_exit_p (struct loop *loop, edge exit)
994 basic_block *body;
995 block_stmt_iterator bsi;
996 unsigned i;
997 tree call;
999 if (exit != loop->single_exit)
1000 return false;
1002 body = get_loop_body (loop);
1003 for (i = 0; i < loop->num_nodes; i++)
1005 for (bsi = bsi_start (body[0]); !bsi_end_p (bsi); bsi_next (&bsi))
1007 call = get_call_expr_in (bsi_stmt (bsi));
1008 if (call && TREE_SIDE_EFFECTS (call))
1010 free (body);
1011 return false;
1016 free (body);
1017 return true;
1020 /* Stores description of number of iterations of LOOP derived from
1021 EXIT (an exit edge of the LOOP) in NITER. Returns true if some
1022 useful information could be derived (and fields of NITER has
1023 meaning described in comments at struct tree_niter_desc
1024 declaration), false otherwise. If WARN is true and
1025 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use
1026 potentially unsafe assumptions. */
1028 bool
1029 number_of_iterations_exit (struct loop *loop, edge exit,
1030 struct tree_niter_desc *niter,
1031 bool warn)
1033 tree stmt, cond, type;
1034 tree op0, op1;
1035 enum tree_code code;
1036 affine_iv iv0, iv1;
1038 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src))
1039 return false;
1041 niter->assumptions = boolean_false_node;
1042 stmt = last_stmt (exit->src);
1043 if (!stmt || TREE_CODE (stmt) != COND_EXPR)
1044 return false;
1046 /* We want the condition for staying inside loop. */
1047 cond = COND_EXPR_COND (stmt);
1048 if (exit->flags & EDGE_TRUE_VALUE)
1049 cond = invert_truthvalue (cond);
1051 code = TREE_CODE (cond);
1052 switch (code)
1054 case GT_EXPR:
1055 case GE_EXPR:
1056 case NE_EXPR:
1057 case LT_EXPR:
1058 case LE_EXPR:
1059 break;
1061 default:
1062 return false;
1065 op0 = TREE_OPERAND (cond, 0);
1066 op1 = TREE_OPERAND (cond, 1);
1067 type = TREE_TYPE (op0);
1069 if (TREE_CODE (type) != INTEGER_TYPE
1070 && !POINTER_TYPE_P (type))
1071 return false;
1073 if (!simple_iv (loop, stmt, op0, &iv0, false))
1074 return false;
1075 if (!simple_iv (loop, stmt, op1, &iv1, false))
1076 return false;
1078 iv0.base = expand_simple_operations (iv0.base);
1079 iv1.base = expand_simple_operations (iv1.base);
1080 if (!number_of_iterations_cond (type, &iv0, code, &iv1, niter,
1081 loop_only_exit_p (loop, exit)))
1082 return false;
1084 if (optimize >= 3)
1086 niter->assumptions = simplify_using_outer_evolutions (loop,
1087 niter->assumptions);
1088 niter->may_be_zero = simplify_using_outer_evolutions (loop,
1089 niter->may_be_zero);
1090 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
1093 niter->additional_info = boolean_true_node;
1094 niter->assumptions
1095 = simplify_using_initial_conditions (loop,
1096 niter->assumptions,
1097 &niter->additional_info);
1098 niter->may_be_zero
1099 = simplify_using_initial_conditions (loop,
1100 niter->may_be_zero,
1101 &niter->additional_info);
1103 if (integer_onep (niter->assumptions))
1104 return true;
1106 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
1107 But if we can prove that there is overflow or some other source of weird
1108 behavior, ignore the loop even with -funsafe-loop-optimizations. */
1109 if (integer_zerop (niter->assumptions))
1110 return false;
1112 if (flag_unsafe_loop_optimizations)
1113 niter->assumptions = boolean_true_node;
1115 if (warn)
1117 const char *wording;
1118 location_t loc = EXPR_LOCATION (stmt);
1120 /* We can provide a more specific warning if one of the operator is
1121 constant and the other advances by +1 or -1. */
1122 if (!zero_p (iv1.step)
1123 ? (zero_p (iv0.step)
1124 && (integer_onep (iv1.step) || integer_all_onesp (iv1.step)))
1125 : (iv0.step
1126 && (integer_onep (iv0.step) || integer_all_onesp (iv0.step))))
1127 wording =
1128 flag_unsafe_loop_optimizations
1129 ? N_("assuming that the loop is not infinite")
1130 : N_("cannot optimize possibly infinite loops");
1131 else
1132 wording =
1133 flag_unsafe_loop_optimizations
1134 ? N_("assuming that the loop counter does not overflow")
1135 : N_("cannot optimize loop, the loop counter may overflow");
1137 if (LOCATION_LINE (loc) > 0)
1138 warning (OPT_Wunsafe_loop_optimizations, "%H%s", &loc, gettext (wording));
1139 else
1140 warning (OPT_Wunsafe_loop_optimizations, "%s", gettext (wording));
1143 return flag_unsafe_loop_optimizations;
1146 /* Try to determine the number of iterations of LOOP. If we succeed,
1147 expression giving number of iterations is returned and *EXIT is
1148 set to the edge from that the information is obtained. Otherwise
1149 chrec_dont_know is returned. */
1151 tree
1152 find_loop_niter (struct loop *loop, edge *exit)
1154 unsigned n_exits, i;
1155 edge *exits = get_loop_exit_edges (loop, &n_exits);
1156 edge ex;
1157 tree niter = NULL_TREE, aniter;
1158 struct tree_niter_desc desc;
1160 *exit = NULL;
1161 for (i = 0; i < n_exits; i++)
1163 ex = exits[i];
1164 if (!just_once_each_iteration_p (loop, ex->src))
1165 continue;
1167 if (!number_of_iterations_exit (loop, ex, &desc, false))
1168 continue;
1170 if (nonzero_p (desc.may_be_zero))
1172 /* We exit in the first iteration through this exit.
1173 We won't find anything better. */
1174 niter = build_int_cst (unsigned_type_node, 0);
1175 *exit = ex;
1176 break;
1179 if (!zero_p (desc.may_be_zero))
1180 continue;
1182 aniter = desc.niter;
1184 if (!niter)
1186 /* Nothing recorded yet. */
1187 niter = aniter;
1188 *exit = ex;
1189 continue;
1192 /* Prefer constants, the lower the better. */
1193 if (TREE_CODE (aniter) != INTEGER_CST)
1194 continue;
1196 if (TREE_CODE (niter) != INTEGER_CST)
1198 niter = aniter;
1199 *exit = ex;
1200 continue;
1203 if (tree_int_cst_lt (aniter, niter))
1205 niter = aniter;
1206 *exit = ex;
1207 continue;
1210 free (exits);
1212 return niter ? niter : chrec_dont_know;
1217 Analysis of a number of iterations of a loop by a brute-force evaluation.
1221 /* Bound on the number of iterations we try to evaluate. */
1223 #define MAX_ITERATIONS_TO_TRACK \
1224 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
1226 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
1227 result by a chain of operations such that all but exactly one of their
1228 operands are constants. */
1230 static tree
1231 chain_of_csts_start (struct loop *loop, tree x)
1233 tree stmt = SSA_NAME_DEF_STMT (x);
1234 tree use;
1235 basic_block bb = bb_for_stmt (stmt);
1237 if (!bb
1238 || !flow_bb_inside_loop_p (loop, bb))
1239 return NULL_TREE;
1241 if (TREE_CODE (stmt) == PHI_NODE)
1243 if (bb == loop->header)
1244 return stmt;
1246 return NULL_TREE;
1249 if (TREE_CODE (stmt) != MODIFY_EXPR)
1250 return NULL_TREE;
1252 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
1253 return NULL_TREE;
1254 if (SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF) == NULL_DEF_OPERAND_P)
1255 return NULL_TREE;
1257 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
1258 if (use == NULL_USE_OPERAND_P)
1259 return NULL_TREE;
1261 return chain_of_csts_start (loop, use);
1264 /* Determines whether the expression X is derived from a result of a phi node
1265 in header of LOOP such that
1267 * the derivation of X consists only from operations with constants
1268 * the initial value of the phi node is constant
1269 * the value of the phi node in the next iteration can be derived from the
1270 value in the current iteration by a chain of operations with constants.
1272 If such phi node exists, it is returned. If X is a constant, X is returned
1273 unchanged. Otherwise NULL_TREE is returned. */
1275 static tree
1276 get_base_for (struct loop *loop, tree x)
1278 tree phi, init, next;
1280 if (is_gimple_min_invariant (x))
1281 return x;
1283 phi = chain_of_csts_start (loop, x);
1284 if (!phi)
1285 return NULL_TREE;
1287 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
1288 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
1290 if (TREE_CODE (next) != SSA_NAME)
1291 return NULL_TREE;
1293 if (!is_gimple_min_invariant (init))
1294 return NULL_TREE;
1296 if (chain_of_csts_start (loop, next) != phi)
1297 return NULL_TREE;
1299 return phi;
1302 /* Given an expression X, then
1304 * if BASE is NULL_TREE, X must be a constant and we return X.
1305 * otherwise X is a SSA name, whose value in the considered loop is derived
1306 by a chain of operations with constant from a result of a phi node in
1307 the header of the loop. Then we return value of X when the value of the
1308 result of this phi node is given by the constant BASE. */
1310 static tree
1311 get_val_for (tree x, tree base)
1313 tree stmt, nx, val;
1314 use_operand_p op;
1315 ssa_op_iter iter;
1317 if (!x)
1318 return base;
1320 stmt = SSA_NAME_DEF_STMT (x);
1321 if (TREE_CODE (stmt) == PHI_NODE)
1322 return base;
1324 FOR_EACH_SSA_USE_OPERAND (op, stmt, iter, SSA_OP_USE)
1326 nx = USE_FROM_PTR (op);
1327 val = get_val_for (nx, base);
1328 SET_USE (op, val);
1329 val = fold (TREE_OPERAND (stmt, 1));
1330 SET_USE (op, nx);
1331 /* only iterate loop once. */
1332 return val;
1335 /* Should never reach here. */
1336 gcc_unreachable();
1339 /* Tries to count the number of iterations of LOOP till it exits by EXIT
1340 by brute force -- i.e. by determining the value of the operands of the
1341 condition at EXIT in first few iterations of the loop (assuming that
1342 these values are constant) and determining the first one in that the
1343 condition is not satisfied. Returns the constant giving the number
1344 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
1346 tree
1347 loop_niter_by_eval (struct loop *loop, edge exit)
1349 tree cond, cnd, acnd;
1350 tree op[2], val[2], next[2], aval[2], phi[2];
1351 unsigned i, j;
1352 enum tree_code cmp;
1354 cond = last_stmt (exit->src);
1355 if (!cond || TREE_CODE (cond) != COND_EXPR)
1356 return chrec_dont_know;
1358 cnd = COND_EXPR_COND (cond);
1359 if (exit->flags & EDGE_TRUE_VALUE)
1360 cnd = invert_truthvalue (cnd);
1362 cmp = TREE_CODE (cnd);
1363 switch (cmp)
1365 case EQ_EXPR:
1366 case NE_EXPR:
1367 case GT_EXPR:
1368 case GE_EXPR:
1369 case LT_EXPR:
1370 case LE_EXPR:
1371 for (j = 0; j < 2; j++)
1372 op[j] = TREE_OPERAND (cnd, j);
1373 break;
1375 default:
1376 return chrec_dont_know;
1379 for (j = 0; j < 2; j++)
1381 phi[j] = get_base_for (loop, op[j]);
1382 if (!phi[j])
1383 return chrec_dont_know;
1386 for (j = 0; j < 2; j++)
1388 if (TREE_CODE (phi[j]) == PHI_NODE)
1390 val[j] = PHI_ARG_DEF_FROM_EDGE (phi[j], loop_preheader_edge (loop));
1391 next[j] = PHI_ARG_DEF_FROM_EDGE (phi[j], loop_latch_edge (loop));
1393 else
1395 val[j] = phi[j];
1396 next[j] = NULL_TREE;
1397 op[j] = NULL_TREE;
1401 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
1403 for (j = 0; j < 2; j++)
1404 aval[j] = get_val_for (op[j], val[j]);
1406 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
1407 if (acnd && zero_p (acnd))
1409 if (dump_file && (dump_flags & TDF_DETAILS))
1410 fprintf (dump_file,
1411 "Proved that loop %d iterates %d times using brute force.\n",
1412 loop->num, i);
1413 return build_int_cst (unsigned_type_node, i);
1416 for (j = 0; j < 2; j++)
1417 val[j] = get_val_for (next[j], val[j]);
1420 return chrec_dont_know;
1423 /* Finds the exit of the LOOP by that the loop exits after a constant
1424 number of iterations and stores the exit edge to *EXIT. The constant
1425 giving the number of iterations of LOOP is returned. The number of
1426 iterations is determined using loop_niter_by_eval (i.e. by brute force
1427 evaluation). If we are unable to find the exit for that loop_niter_by_eval
1428 determines the number of iterations, chrec_dont_know is returned. */
1430 tree
1431 find_loop_niter_by_eval (struct loop *loop, edge *exit)
1433 unsigned n_exits, i;
1434 edge *exits = get_loop_exit_edges (loop, &n_exits);
1435 edge ex;
1436 tree niter = NULL_TREE, aniter;
1438 *exit = NULL;
1439 for (i = 0; i < n_exits; i++)
1441 ex = exits[i];
1442 if (!just_once_each_iteration_p (loop, ex->src))
1443 continue;
1445 aniter = loop_niter_by_eval (loop, ex);
1446 if (chrec_contains_undetermined (aniter))
1447 continue;
1449 if (niter
1450 && !tree_int_cst_lt (aniter, niter))
1451 continue;
1453 niter = aniter;
1454 *exit = ex;
1456 free (exits);
1458 return niter ? niter : chrec_dont_know;
1463 Analysis of upper bounds on number of iterations of a loop.
1467 /* Returns a constant upper bound on the value of expression VAL. The
1468 condition ADDITIONAL must be satisfied (for example, if VAL is
1469 "(unsigned) n" and ADDITIONAL is "n > 0", then we can derive that
1470 VAL is at most (unsigned) MAX_INT).
1472 TODO -- actually do something nontrivial here. */
1474 static tree
1475 derive_constant_upper_bound (tree val, tree additional ATTRIBUTE_UNUSED)
1477 tree type = TREE_TYPE (val);
1478 tree unsigned_type = unsigned_type_for (type);
1480 if (TREE_CODE (val) != INTEGER_CST)
1481 val = upper_bound_in_type (type, type);
1482 return fold_convert (unsigned_type, val);
1485 /* Records that AT_STMT is executed at most BOUND times in LOOP. The
1486 additional condition ADDITIONAL is recorded with the bound. */
1488 void
1489 record_estimate (struct loop *loop, tree bound, tree additional, tree at_stmt)
1491 struct nb_iter_bound *elt = xmalloc (sizeof (struct nb_iter_bound));
1492 tree c_bound = derive_constant_upper_bound (bound, additional);
1494 if (dump_file && (dump_flags & TDF_DETAILS))
1496 fprintf (dump_file, "Statements after ");
1497 print_generic_expr (dump_file, at_stmt, TDF_SLIM);
1498 fprintf (dump_file, " are executed at most ");
1499 print_generic_expr (dump_file, bound, TDF_SLIM);
1500 fprintf (dump_file, " (bounded by ");
1501 print_generic_expr (dump_file, c_bound, TDF_SLIM);
1502 fprintf (dump_file, ") times in loop %d.\n", loop->num);
1505 elt->bound = c_bound;
1506 elt->at_stmt = at_stmt;
1507 elt->next = loop->bounds;
1508 loop->bounds = elt;
1511 /* Initialize LOOP->ESTIMATED_NB_ITERATIONS with the lowest safe
1512 approximation of the number of iterations for LOOP. */
1514 static void
1515 compute_estimated_nb_iterations (struct loop *loop)
1517 struct nb_iter_bound *bound;
1519 for (bound = loop->bounds; bound; bound = bound->next)
1521 if (TREE_CODE (bound->bound) != INTEGER_CST)
1522 continue;
1524 /* Update only when there is no previous estimation, or when the current
1525 estimation is smaller. */
1526 if (chrec_contains_undetermined (loop->estimated_nb_iterations)
1527 || tree_int_cst_lt (bound->bound, loop->estimated_nb_iterations))
1528 loop->estimated_nb_iterations = bound->bound;
1532 /* The following analyzers are extracting informations on the bounds
1533 of LOOP from the following undefined behaviors:
1535 - data references should not access elements over the statically
1536 allocated size,
1538 - signed variables should not overflow when flag_wrapv is not set.
1541 static void
1542 infer_loop_bounds_from_undefined (struct loop *loop)
1544 unsigned i;
1545 basic_block bb, *bbs;
1546 block_stmt_iterator bsi;
1548 bbs = get_loop_body (loop);
1550 for (i = 0; i < loop->num_nodes; i++)
1552 bb = bbs[i];
1554 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
1556 tree stmt = bsi_stmt (bsi);
1558 switch (TREE_CODE (stmt))
1560 case MODIFY_EXPR:
1562 tree op0 = TREE_OPERAND (stmt, 0);
1563 tree op1 = TREE_OPERAND (stmt, 1);
1565 /* For each array access, analyze its access function
1566 and record a bound on the loop iteration domain. */
1567 if (TREE_CODE (op1) == ARRAY_REF
1568 && !array_ref_contains_indirect_ref (op1))
1569 estimate_iters_using_array (stmt, op1);
1571 if (TREE_CODE (op0) == ARRAY_REF
1572 && !array_ref_contains_indirect_ref (op0))
1573 estimate_iters_using_array (stmt, op0);
1575 /* For each signed type variable in LOOP, analyze its
1576 scalar evolution and record a bound of the loop
1577 based on the type's ranges. */
1578 else if (!flag_wrapv && TREE_CODE (op0) == SSA_NAME)
1580 tree init, step, diff, estimation;
1581 tree scev = instantiate_parameters
1582 (loop, analyze_scalar_evolution (loop, op0));
1583 tree type = chrec_type (scev);
1584 tree utype;
1586 if (chrec_contains_undetermined (scev)
1587 || TYPE_UNSIGNED (type))
1588 break;
1590 init = initial_condition_in_loop_num (scev, loop->num);
1591 step = evolution_part_in_loop_num (scev, loop->num);
1593 if (init == NULL_TREE
1594 || step == NULL_TREE
1595 || TREE_CODE (init) != INTEGER_CST
1596 || TREE_CODE (step) != INTEGER_CST
1597 || TYPE_MIN_VALUE (type) == NULL_TREE
1598 || TYPE_MAX_VALUE (type) == NULL_TREE)
1599 break;
1601 utype = unsigned_type_for (type);
1602 if (tree_int_cst_lt (step, integer_zero_node))
1603 diff = fold_build2 (MINUS_EXPR, utype, init,
1604 TYPE_MIN_VALUE (type));
1605 else
1606 diff = fold_build2 (MINUS_EXPR, utype,
1607 TYPE_MAX_VALUE (type), init);
1609 if (!integer_zerop (step))
1611 estimation = fold_build2 (CEIL_DIV_EXPR, utype, diff,
1612 step);
1613 record_estimate (loop, estimation, boolean_true_node,
1614 stmt);
1618 break;
1621 case CALL_EXPR:
1623 tree args;
1625 for (args = TREE_OPERAND (stmt, 1); args;
1626 args = TREE_CHAIN (args))
1627 if (TREE_CODE (TREE_VALUE (args)) == ARRAY_REF
1628 && !array_ref_contains_indirect_ref (TREE_VALUE (args)))
1629 estimate_iters_using_array (stmt, TREE_VALUE (args));
1631 break;
1634 default:
1635 break;
1640 compute_estimated_nb_iterations (loop);
1641 free (bbs);
1644 /* Records estimates on numbers of iterations of LOOP. */
1646 static void
1647 estimate_numbers_of_iterations_loop (struct loop *loop)
1649 edge *exits;
1650 tree niter, type;
1651 unsigned i, n_exits;
1652 struct tree_niter_desc niter_desc;
1654 /* Give up if we already have tried to compute an estimation. */
1655 if (loop->estimated_nb_iterations == chrec_dont_know
1656 /* Or when we already have an estimation. */
1657 || (loop->estimated_nb_iterations != NULL_TREE
1658 && TREE_CODE (loop->estimated_nb_iterations) == INTEGER_CST))
1659 return;
1660 else
1661 loop->estimated_nb_iterations = chrec_dont_know;
1663 exits = get_loop_exit_edges (loop, &n_exits);
1664 for (i = 0; i < n_exits; i++)
1666 if (!number_of_iterations_exit (loop, exits[i], &niter_desc, false))
1667 continue;
1669 niter = niter_desc.niter;
1670 type = TREE_TYPE (niter);
1671 if (!zero_p (niter_desc.may_be_zero)
1672 && !nonzero_p (niter_desc.may_be_zero))
1673 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
1674 build_int_cst (type, 0),
1675 niter);
1676 record_estimate (loop, niter,
1677 niter_desc.additional_info,
1678 last_stmt (exits[i]->src));
1680 free (exits);
1682 if (chrec_contains_undetermined (loop->estimated_nb_iterations))
1683 infer_loop_bounds_from_undefined (loop);
1686 /* Records estimates on numbers of iterations of LOOPS. */
1688 void
1689 estimate_numbers_of_iterations (struct loops *loops)
1691 unsigned i;
1692 struct loop *loop;
1694 for (i = 1; i < loops->num; i++)
1696 loop = loops->parray[i];
1697 if (loop)
1698 estimate_numbers_of_iterations_loop (loop);
1702 /* If A > B, returns -1. If A == B, returns 0. If A < B, returns 1.
1703 If neither of these relations can be proved, returns 2. */
1705 static int
1706 compare_trees (tree a, tree b)
1708 tree typea = TREE_TYPE (a), typeb = TREE_TYPE (b);
1709 tree type;
1711 if (TYPE_PRECISION (typea) > TYPE_PRECISION (typeb))
1712 type = typea;
1713 else
1714 type = typeb;
1716 a = fold_convert (type, a);
1717 b = fold_convert (type, b);
1719 if (nonzero_p (fold_binary (EQ_EXPR, boolean_type_node, a, b)))
1720 return 0;
1721 if (nonzero_p (fold_binary (LT_EXPR, boolean_type_node, a, b)))
1722 return 1;
1723 if (nonzero_p (fold_binary (GT_EXPR, boolean_type_node, a, b)))
1724 return -1;
1726 return 2;
1729 /* Returns true if statement S1 dominates statement S2. */
1731 static bool
1732 stmt_dominates_stmt_p (tree s1, tree s2)
1734 basic_block bb1 = bb_for_stmt (s1), bb2 = bb_for_stmt (s2);
1736 if (!bb1
1737 || s1 == s2)
1738 return true;
1740 if (bb1 == bb2)
1742 block_stmt_iterator bsi;
1744 for (bsi = bsi_start (bb1); bsi_stmt (bsi) != s2; bsi_next (&bsi))
1745 if (bsi_stmt (bsi) == s1)
1746 return true;
1748 return false;
1751 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
1754 /* Returns true when we can prove that the number of executions of
1755 STMT in the loop is at most NITER, according to the fact
1756 that the statement NITER_BOUND->at_stmt is executed at most
1757 NITER_BOUND->bound times. */
1759 static bool
1760 n_of_executions_at_most (tree stmt,
1761 struct nb_iter_bound *niter_bound,
1762 tree niter)
1764 tree cond;
1765 tree bound = niter_bound->bound;
1766 tree bound_type = TREE_TYPE (bound);
1767 tree nit_type = TREE_TYPE (niter);
1768 enum tree_code cmp;
1770 gcc_assert (TYPE_UNSIGNED (bound_type)
1771 && TYPE_UNSIGNED (nit_type)
1772 && is_gimple_min_invariant (bound));
1773 if (TYPE_PRECISION (nit_type) > TYPE_PRECISION (bound_type))
1774 bound = fold_convert (nit_type, bound);
1775 else
1776 niter = fold_convert (bound_type, niter);
1778 /* After the statement niter_bound->at_stmt we know that anything is
1779 executed at most BOUND times. */
1780 if (stmt && stmt_dominates_stmt_p (niter_bound->at_stmt, stmt))
1781 cmp = GE_EXPR;
1782 /* Before the statement niter_bound->at_stmt we know that anything
1783 is executed at most BOUND + 1 times. */
1784 else
1785 cmp = GT_EXPR;
1787 cond = fold_binary (cmp, boolean_type_node, niter, bound);
1788 return nonzero_p (cond);
1791 /* Checks whether it is correct to count the induction variable BASE +
1792 STEP * I at AT_STMT in a wider type NEW_TYPE, using the bounds on
1793 numbers of iterations of a LOOP. If it is possible, return the
1794 value of step of the induction variable in the NEW_TYPE, otherwise
1795 return NULL_TREE. */
1797 static tree
1798 convert_step_widening (struct loop *loop, tree new_type, tree base, tree step,
1799 tree at_stmt)
1801 struct nb_iter_bound *bound;
1802 tree base_in_new_type, base_plus_step_in_new_type, step_in_new_type;
1803 tree delta, step_abs;
1804 tree unsigned_type, valid_niter;
1806 /* Compute the new step. For example, {(uchar) 100, +, (uchar) 240}
1807 is converted to {(uint) 100, +, (uint) 0xfffffff0} in order to
1808 keep the values of the induction variable unchanged: 100, 84, 68,
1811 Another example is: (uint) {(uchar)100, +, (uchar)3} is converted
1812 to {(uint)100, +, (uint)3}.
1814 Before returning the new step, verify that the number of
1815 iterations is less than DELTA / STEP_ABS (i.e. in the previous
1816 example (256 - 100) / 3) such that the iv does not wrap (in which
1817 case the operations are too difficult to be represented and
1818 handled: the values of the iv should be taken modulo 256 in the
1819 wider type; this is not implemented). */
1820 base_in_new_type = fold_convert (new_type, base);
1821 base_plus_step_in_new_type =
1822 fold_convert (new_type,
1823 fold_build2 (PLUS_EXPR, TREE_TYPE (base), base, step));
1824 step_in_new_type = fold_build2 (MINUS_EXPR, new_type,
1825 base_plus_step_in_new_type,
1826 base_in_new_type);
1828 if (TREE_CODE (step_in_new_type) != INTEGER_CST)
1829 return NULL_TREE;
1831 switch (compare_trees (base_plus_step_in_new_type, base_in_new_type))
1833 case -1:
1835 tree extreme = upper_bound_in_type (new_type, TREE_TYPE (base));
1836 delta = fold_build2 (MINUS_EXPR, new_type, extreme,
1837 base_in_new_type);
1838 step_abs = step_in_new_type;
1839 break;
1842 case 1:
1844 tree extreme = lower_bound_in_type (new_type, TREE_TYPE (base));
1845 delta = fold_build2 (MINUS_EXPR, new_type, base_in_new_type,
1846 extreme);
1847 step_abs = fold_build1 (NEGATE_EXPR, new_type, step_in_new_type);
1848 break;
1851 case 0:
1852 return step_in_new_type;
1854 default:
1855 return NULL_TREE;
1858 unsigned_type = unsigned_type_for (new_type);
1859 delta = fold_convert (unsigned_type, delta);
1860 step_abs = fold_convert (unsigned_type, step_abs);
1861 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type,
1862 delta, step_abs);
1864 estimate_numbers_of_iterations_loop (loop);
1865 for (bound = loop->bounds; bound; bound = bound->next)
1866 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
1867 return step_in_new_type;
1869 /* Fail when the loop has no bound estimations, or when no bound can
1870 be used for verifying the conversion. */
1871 return NULL_TREE;
1874 /* Returns true when VAR is used in pointer arithmetics. DEPTH is
1875 used for limiting the search. */
1877 static bool
1878 used_in_pointer_arithmetic_p (tree var, int depth)
1880 use_operand_p use_p;
1881 imm_use_iterator iter;
1883 if (depth == 0
1884 || TREE_CODE (var) != SSA_NAME
1885 || !has_single_use (var))
1886 return false;
1888 FOR_EACH_IMM_USE_FAST (use_p, iter, var)
1890 tree stmt = USE_STMT (use_p);
1892 if (stmt && TREE_CODE (stmt) == MODIFY_EXPR)
1894 tree rhs = TREE_OPERAND (stmt, 1);
1896 if (TREE_CODE (rhs) == NOP_EXPR
1897 || TREE_CODE (rhs) == CONVERT_EXPR)
1899 if (POINTER_TYPE_P (TREE_TYPE (rhs)))
1900 return true;
1901 return false;
1903 else
1904 return used_in_pointer_arithmetic_p (TREE_OPERAND (stmt, 0),
1905 depth - 1);
1908 return false;
1911 /* Return false only when the induction variable BASE + STEP * I is
1912 known to not overflow: i.e. when the number of iterations is small
1913 enough with respect to the step and initial condition in order to
1914 keep the evolution confined in TYPEs bounds. Return true when the
1915 iv is known to overflow or when the property is not computable.
1917 Initialize INIT_IS_MAX to true when the evolution goes from
1918 INIT_IS_MAX to LOWER_BOUND_IN_TYPE, false in the contrary case.
1919 When this property cannot be determined, UNKNOWN_MAX is set to
1920 true. */
1922 bool
1923 scev_probably_wraps_p (tree type, tree base, tree step,
1924 tree at_stmt, struct loop *loop,
1925 bool *init_is_max, bool *unknown_max)
1927 struct nb_iter_bound *bound;
1928 tree delta, step_abs;
1929 tree unsigned_type, valid_niter;
1930 tree base_plus_step, bpsps;
1931 int cps, cpsps;
1933 /* FIXME: The following code will not be used anymore once
1934 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html is
1935 committed.
1937 If AT_STMT is a cast to unsigned that is later used for
1938 referencing a memory location, it is followed by a pointer
1939 conversion just after. Because pointers do not wrap, the
1940 sequences that reference the memory do not wrap either. In the
1941 following example, sequences corresponding to D_13 and to D_14
1942 can be proved to not wrap because they are used for computing a
1943 memory access:
1945 D.1621_13 = (long unsigned intD.4) D.1620_12;
1946 D.1622_14 = D.1621_13 * 8;
1947 D.1623_15 = (doubleD.29 *) D.1622_14;
1949 if (at_stmt && TREE_CODE (at_stmt) == MODIFY_EXPR)
1951 tree op0 = TREE_OPERAND (at_stmt, 0);
1952 tree op1 = TREE_OPERAND (at_stmt, 1);
1953 tree type_op1 = TREE_TYPE (op1);
1955 if ((TYPE_UNSIGNED (type_op1)
1956 && used_in_pointer_arithmetic_p (op0, 2))
1957 || POINTER_TYPE_P (type_op1))
1959 *unknown_max = true;
1960 return false;
1964 if (chrec_contains_undetermined (base)
1965 || chrec_contains_undetermined (step)
1966 || TREE_CODE (base) == REAL_CST
1967 || TREE_CODE (step) == REAL_CST)
1969 *unknown_max = true;
1970 return true;
1973 *unknown_max = false;
1974 base_plus_step = fold_build2 (PLUS_EXPR, type, base, step);
1975 bpsps = fold_build2 (PLUS_EXPR, type, base_plus_step, step);
1976 cps = compare_trees (base_plus_step, base);
1977 cpsps = compare_trees (bpsps, base_plus_step);
1979 /* Check that the sequence is not wrapping in the first step: it
1980 should have the same monotonicity for the first two steps. See
1981 PR23410. */
1982 if (cps != cpsps)
1983 return true;
1985 switch (cps)
1987 case -1:
1989 tree extreme = upper_bound_in_type (type, TREE_TYPE (base));
1990 delta = fold_build2 (MINUS_EXPR, type, extreme, base);
1991 step_abs = step;
1992 *init_is_max = false;
1993 break;
1996 case 1:
1998 tree extreme = lower_bound_in_type (type, TREE_TYPE (base));
1999 delta = fold_build2 (MINUS_EXPR, type, base, extreme);
2000 step_abs = fold_build1 (NEGATE_EXPR, type, step);
2001 *init_is_max = true;
2002 break;
2005 case 0:
2006 /* This means step is equal to 0. This should not happen. It
2007 could happen in convert step, but not here. Safely answer
2008 don't know as in the default case. */
2010 default:
2011 *unknown_max = true;
2012 return true;
2015 /* If AT_STMT represents a cast operation, we may not be able to
2016 take advantage of the undefinedness of signed type evolutions.
2018 implement-c.texi states: "For conversion to a type of width
2019 N, the value is reduced modulo 2^N to be within range of the
2020 type;"
2022 See PR 21959 for a test case. Essentially, given a cast
2023 operation
2024 unsigned char uc;
2025 signed char sc;
2027 sc = (signed char) uc;
2028 if (sc < 0)
2031 where uc and sc have the scev {0, +, 1}, we would consider uc to
2032 wrap around, but not sc, because it is of a signed type. This
2033 causes VRP to erroneously fold the predicate above because it
2034 thinks that sc cannot be negative. */
2035 if (at_stmt && TREE_CODE (at_stmt) == MODIFY_EXPR)
2037 tree rhs = TREE_OPERAND (at_stmt, 1);
2038 tree outer_t = TREE_TYPE (rhs);
2040 if (!TYPE_UNSIGNED (outer_t)
2041 && (TREE_CODE (rhs) == NOP_EXPR || TREE_CODE (rhs) == CONVERT_EXPR))
2043 tree inner_t = TREE_TYPE (TREE_OPERAND (rhs, 0));
2045 /* If the inner type is unsigned and its size and/or
2046 precision are smaller to that of the outer type, then the
2047 expression may wrap around. */
2048 if (TYPE_UNSIGNED (inner_t)
2049 && (TYPE_SIZE (inner_t) <= TYPE_SIZE (outer_t)
2050 || TYPE_PRECISION (inner_t) <= TYPE_PRECISION (outer_t)))
2052 *unknown_max = true;
2053 return true;
2058 /* After having set INIT_IS_MAX, we can return false: when not using
2059 wrapping arithmetic, signed types don't wrap. */
2060 if (!flag_wrapv && !TYPE_UNSIGNED (type))
2061 return false;
2063 unsigned_type = unsigned_type_for (type);
2064 delta = fold_convert (unsigned_type, delta);
2065 step_abs = fold_convert (unsigned_type, step_abs);
2066 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
2068 estimate_numbers_of_iterations_loop (loop);
2069 for (bound = loop->bounds; bound; bound = bound->next)
2070 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
2071 return false;
2073 /* At this point we still don't have a proof that the iv does not
2074 overflow: give up. */
2075 *unknown_max = true;
2076 return true;
2079 /* Return the conversion to NEW_TYPE of the STEP of an induction
2080 variable BASE + STEP * I at AT_STMT. When it fails, return
2081 NULL_TREE. */
2083 tree
2084 convert_step (struct loop *loop, tree new_type, tree base, tree step,
2085 tree at_stmt)
2087 tree res, base_type;
2089 if (chrec_contains_undetermined (base)
2090 || chrec_contains_undetermined (step))
2091 return NULL_TREE;
2093 base_type = TREE_TYPE (base);
2095 /* When not using wrapping arithmetic, signed types don't wrap. */
2096 if (!flag_wrapv && !TYPE_UNSIGNED (base_type))
2097 goto do_convert_step;
2099 if (TYPE_PRECISION (new_type) > TYPE_PRECISION (base_type))
2100 return convert_step_widening (loop, new_type, base, step, at_stmt);
2102 do_convert_step:
2104 res = fold_convert (new_type, step);
2106 if (TREE_CODE (res) == INTEGER_CST)
2108 TREE_OVERFLOW (res) = 0;
2109 TREE_CONSTANT_OVERFLOW (res) = 0;
2112 return res;
2115 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
2117 void
2118 free_numbers_of_iterations_estimates_loop (struct loop *loop)
2120 struct nb_iter_bound *bound, *next;
2122 loop->nb_iterations = NULL;
2123 loop->estimated_nb_iterations = NULL;
2124 for (bound = loop->bounds; bound; bound = next)
2126 next = bound->next;
2127 free (bound);
2130 loop->bounds = NULL;
2133 /* Frees the information on upper bounds on numbers of iterations of LOOPS. */
2135 void
2136 free_numbers_of_iterations_estimates (struct loops *loops)
2138 unsigned i;
2139 struct loop *loop;
2141 for (i = 1; i < loops->num; i++)
2143 loop = loops->parray[i];
2144 if (loop)
2145 free_numbers_of_iterations_estimates_loop (loop);
2149 /* Substitute value VAL for ssa name NAME inside expressions held
2150 at LOOP. */
2152 void
2153 substitute_in_loop_info (struct loop *loop, tree name, tree val)
2155 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);
2156 loop->estimated_nb_iterations
2157 = simplify_replace_tree (loop->estimated_nb_iterations, name, val);