2017-03-06 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / tree-ssa-loop-niter.c
blobd5eaa33e3ee2495db5f9d12c89f028234ee97e0d
1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004-2017 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "rtl.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "tree-pass.h"
28 #include "ssa.h"
29 #include "gimple-pretty-print.h"
30 #include "diagnostic-core.h"
31 #include "stor-layout.h"
32 #include "fold-const.h"
33 #include "calls.h"
34 #include "intl.h"
35 #include "gimplify.h"
36 #include "gimple-iterator.h"
37 #include "tree-cfg.h"
38 #include "tree-ssa-loop-ivopts.h"
39 #include "tree-ssa-loop-niter.h"
40 #include "tree-ssa-loop.h"
41 #include "cfgloop.h"
42 #include "tree-chrec.h"
43 #include "tree-scalar-evolution.h"
44 #include "params.h"
47 /* The maximum number of dominator BBs we search for conditions
48 of loop header copies we use for simplifying a conditional
49 expression. */
50 #define MAX_DOMINATORS_TO_WALK 8
54 Analysis of number of iterations of an affine exit test.
58 /* Bounds on some value, BELOW <= X <= UP. */
60 struct bounds
62 mpz_t below, up;
66 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
68 static void
69 split_to_var_and_offset (tree expr, tree *var, mpz_t offset)
71 tree type = TREE_TYPE (expr);
72 tree op0, op1;
73 bool negate = false;
75 *var = expr;
76 mpz_set_ui (offset, 0);
78 switch (TREE_CODE (expr))
80 case MINUS_EXPR:
81 negate = true;
82 /* Fallthru. */
84 case PLUS_EXPR:
85 case POINTER_PLUS_EXPR:
86 op0 = TREE_OPERAND (expr, 0);
87 op1 = TREE_OPERAND (expr, 1);
89 if (TREE_CODE (op1) != INTEGER_CST)
90 break;
92 *var = op0;
93 /* Always sign extend the offset. */
94 wi::to_mpz (op1, offset, SIGNED);
95 if (negate)
96 mpz_neg (offset, offset);
97 break;
99 case INTEGER_CST:
100 *var = build_int_cst_type (type, 0);
101 wi::to_mpz (expr, offset, TYPE_SIGN (type));
102 break;
104 default:
105 break;
109 /* From condition C0 CMP C1 derives information regarding the value range
110 of VAR, which is of TYPE. Results are stored in to BELOW and UP. */
112 static void
113 refine_value_range_using_guard (tree type, tree var,
114 tree c0, enum tree_code cmp, tree c1,
115 mpz_t below, mpz_t up)
117 tree varc0, varc1, ctype;
118 mpz_t offc0, offc1;
119 mpz_t mint, maxt, minc1, maxc1;
120 wide_int minv, maxv;
121 bool no_wrap = nowrap_type_p (type);
122 bool c0_ok, c1_ok;
123 signop sgn = TYPE_SIGN (type);
125 switch (cmp)
127 case LT_EXPR:
128 case LE_EXPR:
129 case GT_EXPR:
130 case GE_EXPR:
131 STRIP_SIGN_NOPS (c0);
132 STRIP_SIGN_NOPS (c1);
133 ctype = TREE_TYPE (c0);
134 if (!useless_type_conversion_p (ctype, type))
135 return;
137 break;
139 case EQ_EXPR:
140 /* We could derive quite precise information from EQ_EXPR, however,
141 such a guard is unlikely to appear, so we do not bother with
142 handling it. */
143 return;
145 case NE_EXPR:
146 /* NE_EXPR comparisons do not contain much of useful information,
147 except for cases of comparing with bounds. */
148 if (TREE_CODE (c1) != INTEGER_CST
149 || !INTEGRAL_TYPE_P (type))
150 return;
152 /* Ensure that the condition speaks about an expression in the same
153 type as X and Y. */
154 ctype = TREE_TYPE (c0);
155 if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type))
156 return;
157 c0 = fold_convert (type, c0);
158 c1 = fold_convert (type, c1);
160 if (operand_equal_p (var, c0, 0))
162 mpz_t valc1;
164 /* Case of comparing VAR with its below/up bounds. */
165 mpz_init (valc1);
166 wi::to_mpz (c1, valc1, TYPE_SIGN (type));
167 if (mpz_cmp (valc1, below) == 0)
168 cmp = GT_EXPR;
169 if (mpz_cmp (valc1, up) == 0)
170 cmp = LT_EXPR;
172 mpz_clear (valc1);
174 else
176 /* Case of comparing with the bounds of the type. */
177 wide_int min = wi::min_value (type);
178 wide_int max = wi::max_value (type);
180 if (wi::eq_p (c1, min))
181 cmp = GT_EXPR;
182 if (wi::eq_p (c1, max))
183 cmp = LT_EXPR;
186 /* Quick return if no useful information. */
187 if (cmp == NE_EXPR)
188 return;
190 break;
192 default:
193 return;
196 mpz_init (offc0);
197 mpz_init (offc1);
198 split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0);
199 split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1);
201 /* We are only interested in comparisons of expressions based on VAR. */
202 if (operand_equal_p (var, varc1, 0))
204 std::swap (varc0, varc1);
205 mpz_swap (offc0, offc1);
206 cmp = swap_tree_comparison (cmp);
208 else if (!operand_equal_p (var, varc0, 0))
210 mpz_clear (offc0);
211 mpz_clear (offc1);
212 return;
215 mpz_init (mint);
216 mpz_init (maxt);
217 get_type_static_bounds (type, mint, maxt);
218 mpz_init (minc1);
219 mpz_init (maxc1);
220 /* Setup range information for varc1. */
221 if (integer_zerop (varc1))
223 wi::to_mpz (integer_zero_node, minc1, TYPE_SIGN (type));
224 wi::to_mpz (integer_zero_node, maxc1, TYPE_SIGN (type));
226 else if (TREE_CODE (varc1) == SSA_NAME
227 && INTEGRAL_TYPE_P (type)
228 && get_range_info (varc1, &minv, &maxv) == VR_RANGE)
230 gcc_assert (wi::le_p (minv, maxv, sgn));
231 wi::to_mpz (minv, minc1, sgn);
232 wi::to_mpz (maxv, maxc1, sgn);
234 else
236 mpz_set (minc1, mint);
237 mpz_set (maxc1, maxt);
240 /* Compute valid range information for varc1 + offc1. Note nothing
241 useful can be derived if it overflows or underflows. Overflow or
242 underflow could happen when:
244 offc1 > 0 && varc1 + offc1 > MAX_VAL (type)
245 offc1 < 0 && varc1 + offc1 < MIN_VAL (type). */
246 mpz_add (minc1, minc1, offc1);
247 mpz_add (maxc1, maxc1, offc1);
248 c1_ok = (no_wrap
249 || mpz_sgn (offc1) == 0
250 || (mpz_sgn (offc1) < 0 && mpz_cmp (minc1, mint) >= 0)
251 || (mpz_sgn (offc1) > 0 && mpz_cmp (maxc1, maxt) <= 0));
252 if (!c1_ok)
253 goto end;
255 if (mpz_cmp (minc1, mint) < 0)
256 mpz_set (minc1, mint);
257 if (mpz_cmp (maxc1, maxt) > 0)
258 mpz_set (maxc1, maxt);
260 if (cmp == LT_EXPR)
262 cmp = LE_EXPR;
263 mpz_sub_ui (maxc1, maxc1, 1);
265 if (cmp == GT_EXPR)
267 cmp = GE_EXPR;
268 mpz_add_ui (minc1, minc1, 1);
271 /* Compute range information for varc0. If there is no overflow,
272 the condition implied that
274 (varc0) cmp (varc1 + offc1 - offc0)
276 We can possibly improve the upper bound of varc0 if cmp is LE_EXPR,
277 or the below bound if cmp is GE_EXPR.
279 To prove there is no overflow/underflow, we need to check below
280 four cases:
281 1) cmp == LE_EXPR && offc0 > 0
283 (varc0 + offc0) doesn't overflow
284 && (varc1 + offc1 - offc0) doesn't underflow
286 2) cmp == LE_EXPR && offc0 < 0
288 (varc0 + offc0) doesn't underflow
289 && (varc1 + offc1 - offc0) doesn't overfloe
291 In this case, (varc0 + offc0) will never underflow if we can
292 prove (varc1 + offc1 - offc0) doesn't overflow.
294 3) cmp == GE_EXPR && offc0 < 0
296 (varc0 + offc0) doesn't underflow
297 && (varc1 + offc1 - offc0) doesn't overflow
299 4) cmp == GE_EXPR && offc0 > 0
301 (varc0 + offc0) doesn't overflow
302 && (varc1 + offc1 - offc0) doesn't underflow
304 In this case, (varc0 + offc0) will never overflow if we can
305 prove (varc1 + offc1 - offc0) doesn't underflow.
307 Note we only handle case 2 and 4 in below code. */
309 mpz_sub (minc1, minc1, offc0);
310 mpz_sub (maxc1, maxc1, offc0);
311 c0_ok = (no_wrap
312 || mpz_sgn (offc0) == 0
313 || (cmp == LE_EXPR
314 && mpz_sgn (offc0) < 0 && mpz_cmp (maxc1, maxt) <= 0)
315 || (cmp == GE_EXPR
316 && mpz_sgn (offc0) > 0 && mpz_cmp (minc1, mint) >= 0));
317 if (!c0_ok)
318 goto end;
320 if (cmp == LE_EXPR)
322 if (mpz_cmp (up, maxc1) > 0)
323 mpz_set (up, maxc1);
325 else
327 if (mpz_cmp (below, minc1) < 0)
328 mpz_set (below, minc1);
331 end:
332 mpz_clear (mint);
333 mpz_clear (maxt);
334 mpz_clear (minc1);
335 mpz_clear (maxc1);
336 mpz_clear (offc0);
337 mpz_clear (offc1);
340 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
341 in TYPE to MIN and MAX. */
343 static void
344 determine_value_range (struct loop *loop, tree type, tree var, mpz_t off,
345 mpz_t min, mpz_t max)
347 int cnt = 0;
348 mpz_t minm, maxm;
349 basic_block bb;
350 wide_int minv, maxv;
351 enum value_range_type rtype = VR_VARYING;
353 /* If the expression is a constant, we know its value exactly. */
354 if (integer_zerop (var))
356 mpz_set (min, off);
357 mpz_set (max, off);
358 return;
361 get_type_static_bounds (type, min, max);
363 /* See if we have some range info from VRP. */
364 if (TREE_CODE (var) == SSA_NAME && INTEGRAL_TYPE_P (type))
366 edge e = loop_preheader_edge (loop);
367 signop sgn = TYPE_SIGN (type);
368 gphi_iterator gsi;
370 /* Either for VAR itself... */
371 rtype = get_range_info (var, &minv, &maxv);
372 /* Or for PHI results in loop->header where VAR is used as
373 PHI argument from the loop preheader edge. */
374 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi))
376 gphi *phi = gsi.phi ();
377 wide_int minc, maxc;
378 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var
379 && (get_range_info (gimple_phi_result (phi), &minc, &maxc)
380 == VR_RANGE))
382 if (rtype != VR_RANGE)
384 rtype = VR_RANGE;
385 minv = minc;
386 maxv = maxc;
388 else
390 minv = wi::max (minv, minc, sgn);
391 maxv = wi::min (maxv, maxc, sgn);
392 /* If the PHI result range are inconsistent with
393 the VAR range, give up on looking at the PHI
394 results. This can happen if VR_UNDEFINED is
395 involved. */
396 if (wi::gt_p (minv, maxv, sgn))
398 rtype = get_range_info (var, &minv, &maxv);
399 break;
404 mpz_init (minm);
405 mpz_init (maxm);
406 if (rtype != VR_RANGE)
408 mpz_set (minm, min);
409 mpz_set (maxm, max);
411 else
413 gcc_assert (wi::le_p (minv, maxv, sgn));
414 wi::to_mpz (minv, minm, sgn);
415 wi::to_mpz (maxv, maxm, sgn);
417 /* Now walk the dominators of the loop header and use the entry
418 guards to refine the estimates. */
419 for (bb = loop->header;
420 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK;
421 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
423 edge e;
424 tree c0, c1;
425 gimple *cond;
426 enum tree_code cmp;
428 if (!single_pred_p (bb))
429 continue;
430 e = single_pred_edge (bb);
432 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
433 continue;
435 cond = last_stmt (e->src);
436 c0 = gimple_cond_lhs (cond);
437 cmp = gimple_cond_code (cond);
438 c1 = gimple_cond_rhs (cond);
440 if (e->flags & EDGE_FALSE_VALUE)
441 cmp = invert_tree_comparison (cmp, false);
443 refine_value_range_using_guard (type, var, c0, cmp, c1, minm, maxm);
444 ++cnt;
447 mpz_add (minm, minm, off);
448 mpz_add (maxm, maxm, off);
449 /* If the computation may not wrap or off is zero, then this
450 is always fine. If off is negative and minv + off isn't
451 smaller than type's minimum, or off is positive and
452 maxv + off isn't bigger than type's maximum, use the more
453 precise range too. */
454 if (nowrap_type_p (type)
455 || mpz_sgn (off) == 0
456 || (mpz_sgn (off) < 0 && mpz_cmp (minm, min) >= 0)
457 || (mpz_sgn (off) > 0 && mpz_cmp (maxm, max) <= 0))
459 mpz_set (min, minm);
460 mpz_set (max, maxm);
461 mpz_clear (minm);
462 mpz_clear (maxm);
463 return;
465 mpz_clear (minm);
466 mpz_clear (maxm);
469 /* If the computation may wrap, we know nothing about the value, except for
470 the range of the type. */
471 if (!nowrap_type_p (type))
472 return;
474 /* Since the addition of OFF does not wrap, if OFF is positive, then we may
475 add it to MIN, otherwise to MAX. */
476 if (mpz_sgn (off) < 0)
477 mpz_add (max, max, off);
478 else
479 mpz_add (min, min, off);
482 /* Stores the bounds on the difference of the values of the expressions
483 (var + X) and (var + Y), computed in TYPE, to BNDS. */
485 static void
486 bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y,
487 bounds *bnds)
489 int rel = mpz_cmp (x, y);
490 bool may_wrap = !nowrap_type_p (type);
491 mpz_t m;
493 /* If X == Y, then the expressions are always equal.
494 If X > Y, there are the following possibilities:
495 a) neither of var + X and var + Y overflow or underflow, or both of
496 them do. Then their difference is X - Y.
497 b) var + X overflows, and var + Y does not. Then the values of the
498 expressions are var + X - M and var + Y, where M is the range of
499 the type, and their difference is X - Y - M.
500 c) var + Y underflows and var + X does not. Their difference again
501 is M - X + Y.
502 Therefore, if the arithmetics in type does not overflow, then the
503 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y)
504 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or
505 (X - Y, X - Y + M). */
507 if (rel == 0)
509 mpz_set_ui (bnds->below, 0);
510 mpz_set_ui (bnds->up, 0);
511 return;
514 mpz_init (m);
515 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), m, UNSIGNED);
516 mpz_add_ui (m, m, 1);
517 mpz_sub (bnds->up, x, y);
518 mpz_set (bnds->below, bnds->up);
520 if (may_wrap)
522 if (rel > 0)
523 mpz_sub (bnds->below, bnds->below, m);
524 else
525 mpz_add (bnds->up, bnds->up, m);
528 mpz_clear (m);
531 /* From condition C0 CMP C1 derives information regarding the
532 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE,
533 and stores it to BNDS. */
535 static void
536 refine_bounds_using_guard (tree type, tree varx, mpz_t offx,
537 tree vary, mpz_t offy,
538 tree c0, enum tree_code cmp, tree c1,
539 bounds *bnds)
541 tree varc0, varc1, ctype;
542 mpz_t offc0, offc1, loffx, loffy, bnd;
543 bool lbound = false;
544 bool no_wrap = nowrap_type_p (type);
545 bool x_ok, y_ok;
547 switch (cmp)
549 case LT_EXPR:
550 case LE_EXPR:
551 case GT_EXPR:
552 case GE_EXPR:
553 STRIP_SIGN_NOPS (c0);
554 STRIP_SIGN_NOPS (c1);
555 ctype = TREE_TYPE (c0);
556 if (!useless_type_conversion_p (ctype, type))
557 return;
559 break;
561 case EQ_EXPR:
562 /* We could derive quite precise information from EQ_EXPR, however, such
563 a guard is unlikely to appear, so we do not bother with handling
564 it. */
565 return;
567 case NE_EXPR:
568 /* NE_EXPR comparisons do not contain much of useful information, except for
569 special case of comparing with the bounds of the type. */
570 if (TREE_CODE (c1) != INTEGER_CST
571 || !INTEGRAL_TYPE_P (type))
572 return;
574 /* Ensure that the condition speaks about an expression in the same type
575 as X and Y. */
576 ctype = TREE_TYPE (c0);
577 if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type))
578 return;
579 c0 = fold_convert (type, c0);
580 c1 = fold_convert (type, c1);
582 if (TYPE_MIN_VALUE (type)
583 && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0))
585 cmp = GT_EXPR;
586 break;
588 if (TYPE_MAX_VALUE (type)
589 && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0))
591 cmp = LT_EXPR;
592 break;
595 return;
596 default:
597 return;
600 mpz_init (offc0);
601 mpz_init (offc1);
602 split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0);
603 split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1);
605 /* We are only interested in comparisons of expressions based on VARX and
606 VARY. TODO -- we might also be able to derive some bounds from
607 expressions containing just one of the variables. */
609 if (operand_equal_p (varx, varc1, 0))
611 std::swap (varc0, varc1);
612 mpz_swap (offc0, offc1);
613 cmp = swap_tree_comparison (cmp);
616 if (!operand_equal_p (varx, varc0, 0)
617 || !operand_equal_p (vary, varc1, 0))
618 goto end;
620 mpz_init_set (loffx, offx);
621 mpz_init_set (loffy, offy);
623 if (cmp == GT_EXPR || cmp == GE_EXPR)
625 std::swap (varx, vary);
626 mpz_swap (offc0, offc1);
627 mpz_swap (loffx, loffy);
628 cmp = swap_tree_comparison (cmp);
629 lbound = true;
632 /* If there is no overflow, the condition implies that
634 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
636 The overflows and underflows may complicate things a bit; each
637 overflow decreases the appropriate offset by M, and underflow
638 increases it by M. The above inequality would not necessarily be
639 true if
641 -- VARX + OFFX underflows and VARX + OFFC0 does not, or
642 VARX + OFFC0 overflows, but VARX + OFFX does not.
643 This may only happen if OFFX < OFFC0.
644 -- VARY + OFFY overflows and VARY + OFFC1 does not, or
645 VARY + OFFC1 underflows and VARY + OFFY does not.
646 This may only happen if OFFY > OFFC1. */
648 if (no_wrap)
650 x_ok = true;
651 y_ok = true;
653 else
655 x_ok = (integer_zerop (varx)
656 || mpz_cmp (loffx, offc0) >= 0);
657 y_ok = (integer_zerop (vary)
658 || mpz_cmp (loffy, offc1) <= 0);
661 if (x_ok && y_ok)
663 mpz_init (bnd);
664 mpz_sub (bnd, loffx, loffy);
665 mpz_add (bnd, bnd, offc1);
666 mpz_sub (bnd, bnd, offc0);
668 if (cmp == LT_EXPR)
669 mpz_sub_ui (bnd, bnd, 1);
671 if (lbound)
673 mpz_neg (bnd, bnd);
674 if (mpz_cmp (bnds->below, bnd) < 0)
675 mpz_set (bnds->below, bnd);
677 else
679 if (mpz_cmp (bnd, bnds->up) < 0)
680 mpz_set (bnds->up, bnd);
682 mpz_clear (bnd);
685 mpz_clear (loffx);
686 mpz_clear (loffy);
687 end:
688 mpz_clear (offc0);
689 mpz_clear (offc1);
692 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS.
693 The subtraction is considered to be performed in arbitrary precision,
694 without overflows.
696 We do not attempt to be too clever regarding the value ranges of X and
697 Y; most of the time, they are just integers or ssa names offsetted by
698 integer. However, we try to use the information contained in the
699 comparisons before the loop (usually created by loop header copying). */
701 static void
702 bound_difference (struct loop *loop, tree x, tree y, bounds *bnds)
704 tree type = TREE_TYPE (x);
705 tree varx, vary;
706 mpz_t offx, offy;
707 mpz_t minx, maxx, miny, maxy;
708 int cnt = 0;
709 edge e;
710 basic_block bb;
711 tree c0, c1;
712 gimple *cond;
713 enum tree_code cmp;
715 /* Get rid of unnecessary casts, but preserve the value of
716 the expressions. */
717 STRIP_SIGN_NOPS (x);
718 STRIP_SIGN_NOPS (y);
720 mpz_init (bnds->below);
721 mpz_init (bnds->up);
722 mpz_init (offx);
723 mpz_init (offy);
724 split_to_var_and_offset (x, &varx, offx);
725 split_to_var_and_offset (y, &vary, offy);
727 if (!integer_zerop (varx)
728 && operand_equal_p (varx, vary, 0))
730 /* Special case VARX == VARY -- we just need to compare the
731 offsets. The matters are a bit more complicated in the
732 case addition of offsets may wrap. */
733 bound_difference_of_offsetted_base (type, offx, offy, bnds);
735 else
737 /* Otherwise, use the value ranges to determine the initial
738 estimates on below and up. */
739 mpz_init (minx);
740 mpz_init (maxx);
741 mpz_init (miny);
742 mpz_init (maxy);
743 determine_value_range (loop, type, varx, offx, minx, maxx);
744 determine_value_range (loop, type, vary, offy, miny, maxy);
746 mpz_sub (bnds->below, minx, maxy);
747 mpz_sub (bnds->up, maxx, miny);
748 mpz_clear (minx);
749 mpz_clear (maxx);
750 mpz_clear (miny);
751 mpz_clear (maxy);
754 /* If both X and Y are constants, we cannot get any more precise. */
755 if (integer_zerop (varx) && integer_zerop (vary))
756 goto end;
758 /* Now walk the dominators of the loop header and use the entry
759 guards to refine the estimates. */
760 for (bb = loop->header;
761 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK;
762 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
764 if (!single_pred_p (bb))
765 continue;
766 e = single_pred_edge (bb);
768 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
769 continue;
771 cond = last_stmt (e->src);
772 c0 = gimple_cond_lhs (cond);
773 cmp = gimple_cond_code (cond);
774 c1 = gimple_cond_rhs (cond);
776 if (e->flags & EDGE_FALSE_VALUE)
777 cmp = invert_tree_comparison (cmp, false);
779 refine_bounds_using_guard (type, varx, offx, vary, offy,
780 c0, cmp, c1, bnds);
781 ++cnt;
784 end:
785 mpz_clear (offx);
786 mpz_clear (offy);
789 /* Update the bounds in BNDS that restrict the value of X to the bounds
790 that restrict the value of X + DELTA. X can be obtained as a
791 difference of two values in TYPE. */
793 static void
794 bounds_add (bounds *bnds, const widest_int &delta, tree type)
796 mpz_t mdelta, max;
798 mpz_init (mdelta);
799 wi::to_mpz (delta, mdelta, SIGNED);
801 mpz_init (max);
802 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED);
804 mpz_add (bnds->up, bnds->up, mdelta);
805 mpz_add (bnds->below, bnds->below, mdelta);
807 if (mpz_cmp (bnds->up, max) > 0)
808 mpz_set (bnds->up, max);
810 mpz_neg (max, max);
811 if (mpz_cmp (bnds->below, max) < 0)
812 mpz_set (bnds->below, max);
814 mpz_clear (mdelta);
815 mpz_clear (max);
818 /* Update the bounds in BNDS that restrict the value of X to the bounds
819 that restrict the value of -X. */
821 static void
822 bounds_negate (bounds *bnds)
824 mpz_t tmp;
826 mpz_init_set (tmp, bnds->up);
827 mpz_neg (bnds->up, bnds->below);
828 mpz_neg (bnds->below, tmp);
829 mpz_clear (tmp);
832 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
834 static tree
835 inverse (tree x, tree mask)
837 tree type = TREE_TYPE (x);
838 tree rslt;
839 unsigned ctr = tree_floor_log2 (mask);
841 if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT)
843 unsigned HOST_WIDE_INT ix;
844 unsigned HOST_WIDE_INT imask;
845 unsigned HOST_WIDE_INT irslt = 1;
847 gcc_assert (cst_and_fits_in_hwi (x));
848 gcc_assert (cst_and_fits_in_hwi (mask));
850 ix = int_cst_value (x);
851 imask = int_cst_value (mask);
853 for (; ctr; ctr--)
855 irslt *= ix;
856 ix *= ix;
858 irslt &= imask;
860 rslt = build_int_cst_type (type, irslt);
862 else
864 rslt = build_int_cst (type, 1);
865 for (; ctr; ctr--)
867 rslt = int_const_binop (MULT_EXPR, rslt, x);
868 x = int_const_binop (MULT_EXPR, x, x);
870 rslt = int_const_binop (BIT_AND_EXPR, rslt, mask);
873 return rslt;
876 /* Derives the upper bound BND on the number of executions of loop with exit
877 condition S * i <> C. If NO_OVERFLOW is true, then the control variable of
878 the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed
879 that the loop ends through this exit, i.e., the induction variable ever
880 reaches the value of C.
882 The value C is equal to final - base, where final and base are the final and
883 initial value of the actual induction variable in the analysed loop. BNDS
884 bounds the value of this difference when computed in signed type with
885 unbounded range, while the computation of C is performed in an unsigned
886 type with the range matching the range of the type of the induction variable.
887 In particular, BNDS.up contains an upper bound on C in the following cases:
888 -- if the iv must reach its final value without overflow, i.e., if
889 NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or
890 -- if final >= base, which we know to hold when BNDS.below >= 0. */
892 static void
893 number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s,
894 bounds *bnds, bool exit_must_be_taken)
896 widest_int max;
897 mpz_t d;
898 tree type = TREE_TYPE (c);
899 bool bnds_u_valid = ((no_overflow && exit_must_be_taken)
900 || mpz_sgn (bnds->below) >= 0);
902 if (integer_onep (s)
903 || (TREE_CODE (c) == INTEGER_CST
904 && TREE_CODE (s) == INTEGER_CST
905 && wi::mod_trunc (c, s, TYPE_SIGN (type)) == 0)
906 || (TYPE_OVERFLOW_UNDEFINED (type)
907 && multiple_of_p (type, c, s)))
909 /* If C is an exact multiple of S, then its value will be reached before
910 the induction variable overflows (unless the loop is exited in some
911 other way before). Note that the actual induction variable in the
912 loop (which ranges from base to final instead of from 0 to C) may
913 overflow, in which case BNDS.up will not be giving a correct upper
914 bound on C; thus, BNDS_U_VALID had to be computed in advance. */
915 no_overflow = true;
916 exit_must_be_taken = true;
919 /* If the induction variable can overflow, the number of iterations is at
920 most the period of the control variable (or infinite, but in that case
921 the whole # of iterations analysis will fail). */
922 if (!no_overflow)
924 max = wi::mask <widest_int> (TYPE_PRECISION (type) - wi::ctz (s), false);
925 wi::to_mpz (max, bnd, UNSIGNED);
926 return;
929 /* Now we know that the induction variable does not overflow, so the loop
930 iterates at most (range of type / S) times. */
931 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), bnd, UNSIGNED);
933 /* If the induction variable is guaranteed to reach the value of C before
934 overflow, ... */
935 if (exit_must_be_taken)
937 /* ... then we can strengthen this to C / S, and possibly we can use
938 the upper bound on C given by BNDS. */
939 if (TREE_CODE (c) == INTEGER_CST)
940 wi::to_mpz (c, bnd, UNSIGNED);
941 else if (bnds_u_valid)
942 mpz_set (bnd, bnds->up);
945 mpz_init (d);
946 wi::to_mpz (s, d, UNSIGNED);
947 mpz_fdiv_q (bnd, bnd, d);
948 mpz_clear (d);
951 /* Determines number of iterations of loop whose ending condition
952 is IV <> FINAL. TYPE is the type of the iv. The number of
953 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
954 we know that the exit must be taken eventually, i.e., that the IV
955 ever reaches the value FINAL (we derived this earlier, and possibly set
956 NITER->assumptions to make sure this is the case). BNDS contains the
957 bounds on the difference FINAL - IV->base. */
959 static bool
960 number_of_iterations_ne (struct loop *loop, tree type, affine_iv *iv,
961 tree final, struct tree_niter_desc *niter,
962 bool exit_must_be_taken, bounds *bnds)
964 tree niter_type = unsigned_type_for (type);
965 tree s, c, d, bits, assumption, tmp, bound;
966 mpz_t max;
968 niter->control = *iv;
969 niter->bound = final;
970 niter->cmp = NE_EXPR;
972 /* Rearrange the terms so that we get inequality S * i <> C, with S
973 positive. Also cast everything to the unsigned type. If IV does
974 not overflow, BNDS bounds the value of C. Also, this is the
975 case if the computation |FINAL - IV->base| does not overflow, i.e.,
976 if BNDS->below in the result is nonnegative. */
977 if (tree_int_cst_sign_bit (iv->step))
979 s = fold_convert (niter_type,
980 fold_build1 (NEGATE_EXPR, type, iv->step));
981 c = fold_build2 (MINUS_EXPR, niter_type,
982 fold_convert (niter_type, iv->base),
983 fold_convert (niter_type, final));
984 bounds_negate (bnds);
986 else
988 s = fold_convert (niter_type, iv->step);
989 c = fold_build2 (MINUS_EXPR, niter_type,
990 fold_convert (niter_type, final),
991 fold_convert (niter_type, iv->base));
994 mpz_init (max);
995 number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds,
996 exit_must_be_taken);
997 niter->max = widest_int::from (wi::from_mpz (niter_type, max, false),
998 TYPE_SIGN (niter_type));
999 mpz_clear (max);
1001 /* Compute no-overflow information for the control iv. This can be
1002 proven when below two conditions are satisfied:
1004 1) IV evaluates toward FINAL at beginning, i.e:
1005 base <= FINAL ; step > 0
1006 base >= FINAL ; step < 0
1008 2) |FINAL - base| is an exact multiple of step.
1010 Unfortunately, it's hard to prove above conditions after pass loop-ch
1011 because loop with exit condition (IV != FINAL) usually will be guarded
1012 by initial-condition (IV.base - IV.step != FINAL). In this case, we
1013 can alternatively try to prove below conditions:
1015 1') IV evaluates toward FINAL at beginning, i.e:
1016 new_base = base - step < FINAL ; step > 0
1017 && base - step doesn't underflow
1018 new_base = base - step > FINAL ; step < 0
1019 && base - step doesn't overflow
1021 2') |FINAL - new_base| is an exact multiple of step.
1023 Please refer to PR34114 as an example of loop-ch's impact, also refer
1024 to PR72817 as an example why condition 2') is necessary.
1026 Note, for NE_EXPR, base equals to FINAL is a special case, in
1027 which the loop exits immediately, and the iv does not overflow. */
1028 if (!niter->control.no_overflow
1029 && (integer_onep (s) || multiple_of_p (type, c, s)))
1031 tree t, cond, new_c, relaxed_cond = boolean_false_node;
1033 if (tree_int_cst_sign_bit (iv->step))
1035 cond = fold_build2 (GE_EXPR, boolean_type_node, iv->base, final);
1036 if (TREE_CODE (type) == INTEGER_TYPE)
1038 /* Only when base - step doesn't overflow. */
1039 t = TYPE_MAX_VALUE (type);
1040 t = fold_build2 (PLUS_EXPR, type, t, iv->step);
1041 t = fold_build2 (GE_EXPR, boolean_type_node, t, iv->base);
1042 if (integer_nonzerop (t))
1044 t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step);
1045 new_c = fold_build2 (MINUS_EXPR, niter_type,
1046 fold_convert (niter_type, t),
1047 fold_convert (niter_type, final));
1048 if (multiple_of_p (type, new_c, s))
1049 relaxed_cond = fold_build2 (GT_EXPR, boolean_type_node,
1050 t, final);
1054 else
1056 cond = fold_build2 (LE_EXPR, boolean_type_node, iv->base, final);
1057 if (TREE_CODE (type) == INTEGER_TYPE)
1059 /* Only when base - step doesn't underflow. */
1060 t = TYPE_MIN_VALUE (type);
1061 t = fold_build2 (PLUS_EXPR, type, t, iv->step);
1062 t = fold_build2 (LE_EXPR, boolean_type_node, t, iv->base);
1063 if (integer_nonzerop (t))
1065 t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step);
1066 new_c = fold_build2 (MINUS_EXPR, niter_type,
1067 fold_convert (niter_type, final),
1068 fold_convert (niter_type, t));
1069 if (multiple_of_p (type, new_c, s))
1070 relaxed_cond = fold_build2 (LT_EXPR, boolean_type_node,
1071 t, final);
1076 t = simplify_using_initial_conditions (loop, cond);
1077 if (!t || !integer_onep (t))
1078 t = simplify_using_initial_conditions (loop, relaxed_cond);
1080 if (t && integer_onep (t))
1081 niter->control.no_overflow = true;
1084 /* First the trivial cases -- when the step is 1. */
1085 if (integer_onep (s))
1087 niter->niter = c;
1088 return true;
1090 if (niter->control.no_overflow && multiple_of_p (type, c, s))
1092 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, c, s);
1093 return true;
1096 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
1097 is infinite. Otherwise, the number of iterations is
1098 (inverse(s/d) * (c/d)) mod (size of mode/d). */
1099 bits = num_ending_zeros (s);
1100 bound = build_low_bits_mask (niter_type,
1101 (TYPE_PRECISION (niter_type)
1102 - tree_to_uhwi (bits)));
1104 d = fold_binary_to_constant (LSHIFT_EXPR, niter_type,
1105 build_int_cst (niter_type, 1), bits);
1106 s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits);
1108 if (!exit_must_be_taken)
1110 /* If we cannot assume that the exit is taken eventually, record the
1111 assumptions for divisibility of c. */
1112 assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d);
1113 assumption = fold_build2 (EQ_EXPR, boolean_type_node,
1114 assumption, build_int_cst (niter_type, 0));
1115 if (!integer_nonzerop (assumption))
1116 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1117 niter->assumptions, assumption);
1120 c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d);
1121 tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound));
1122 niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound);
1123 return true;
1126 /* Checks whether we can determine the final value of the control variable
1127 of the loop with ending condition IV0 < IV1 (computed in TYPE).
1128 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
1129 of the step. The assumptions necessary to ensure that the computation
1130 of the final value does not overflow are recorded in NITER. If we
1131 find the final value, we adjust DELTA and return TRUE. Otherwise
1132 we return false. BNDS bounds the value of IV1->base - IV0->base,
1133 and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is
1134 true if we know that the exit must be taken eventually. */
1136 static bool
1137 number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1,
1138 struct tree_niter_desc *niter,
1139 tree *delta, tree step,
1140 bool exit_must_be_taken, bounds *bnds)
1142 tree niter_type = TREE_TYPE (step);
1143 tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step);
1144 tree tmod;
1145 tree assumption = boolean_true_node, bound;
1146 tree type1 = (POINTER_TYPE_P (type)) ? sizetype : type;
1148 if (TREE_CODE (mod) != INTEGER_CST)
1149 return false;
1150 if (integer_nonzerop (mod))
1151 mod = fold_build2 (MINUS_EXPR, niter_type, step, mod);
1152 tmod = fold_convert (type1, mod);
1154 /* If the induction variable does not overflow and the exit is taken,
1155 then the computation of the final value does not overflow. There
1156 are three cases:
1157 1) The case if the new final value is equal to the current one.
1158 2) Induction varaible has pointer type, as the code cannot rely
1159 on the object to that the pointer points being placed at the
1160 end of the address space (and more pragmatically,
1161 TYPE_{MIN,MAX}_VALUE is not defined for pointers).
1162 3) EXIT_MUST_BE_TAKEN is true, note it implies that the induction
1163 variable does not overflow. */
1164 if (!integer_zerop (mod) && !POINTER_TYPE_P (type) && !exit_must_be_taken)
1166 if (integer_nonzerop (iv0->step))
1168 /* The final value of the iv is iv1->base + MOD, assuming
1169 that this computation does not overflow, and that
1170 iv0->base <= iv1->base + MOD. */
1171 bound = fold_build2 (MINUS_EXPR, type1,
1172 TYPE_MAX_VALUE (type1), tmod);
1173 assumption = fold_build2 (LE_EXPR, boolean_type_node,
1174 iv1->base, bound);
1176 else
1178 /* The final value of the iv is iv0->base - MOD, assuming
1179 that this computation does not overflow, and that
1180 iv0->base - MOD <= iv1->base. */
1181 bound = fold_build2 (PLUS_EXPR, type1,
1182 TYPE_MIN_VALUE (type1), tmod);
1183 assumption = fold_build2 (GE_EXPR, boolean_type_node,
1184 iv0->base, bound);
1186 if (integer_zerop (assumption))
1187 return false;
1188 else if (!integer_nonzerop (assumption))
1189 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1190 niter->assumptions, assumption);
1193 /* Since we are transforming LT to NE and DELTA is constant, there
1194 is no need to compute may_be_zero because this loop must roll. */
1196 bounds_add (bnds, wi::to_widest (mod), type);
1197 *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod);
1198 return true;
1201 /* Add assertions to NITER that ensure that the control variable of the loop
1202 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
1203 are TYPE. Returns false if we can prove that there is an overflow, true
1204 otherwise. STEP is the absolute value of the step. */
1206 static bool
1207 assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1,
1208 struct tree_niter_desc *niter, tree step)
1210 tree bound, d, assumption, diff;
1211 tree niter_type = TREE_TYPE (step);
1213 if (integer_nonzerop (iv0->step))
1215 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
1216 if (iv0->no_overflow)
1217 return true;
1219 /* If iv0->base is a constant, we can determine the last value before
1220 overflow precisely; otherwise we conservatively assume
1221 MAX - STEP + 1. */
1223 if (TREE_CODE (iv0->base) == INTEGER_CST)
1225 d = fold_build2 (MINUS_EXPR, niter_type,
1226 fold_convert (niter_type, TYPE_MAX_VALUE (type)),
1227 fold_convert (niter_type, iv0->base));
1228 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
1230 else
1231 diff = fold_build2 (MINUS_EXPR, niter_type, step,
1232 build_int_cst (niter_type, 1));
1233 bound = fold_build2 (MINUS_EXPR, type,
1234 TYPE_MAX_VALUE (type), fold_convert (type, diff));
1235 assumption = fold_build2 (LE_EXPR, boolean_type_node,
1236 iv1->base, bound);
1238 else
1240 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
1241 if (iv1->no_overflow)
1242 return true;
1244 if (TREE_CODE (iv1->base) == INTEGER_CST)
1246 d = fold_build2 (MINUS_EXPR, niter_type,
1247 fold_convert (niter_type, iv1->base),
1248 fold_convert (niter_type, TYPE_MIN_VALUE (type)));
1249 diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step);
1251 else
1252 diff = fold_build2 (MINUS_EXPR, niter_type, step,
1253 build_int_cst (niter_type, 1));
1254 bound = fold_build2 (PLUS_EXPR, type,
1255 TYPE_MIN_VALUE (type), fold_convert (type, diff));
1256 assumption = fold_build2 (GE_EXPR, boolean_type_node,
1257 iv0->base, bound);
1260 if (integer_zerop (assumption))
1261 return false;
1262 if (!integer_nonzerop (assumption))
1263 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1264 niter->assumptions, assumption);
1266 iv0->no_overflow = true;
1267 iv1->no_overflow = true;
1268 return true;
1271 /* Add an assumption to NITER that a loop whose ending condition
1272 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS
1273 bounds the value of IV1->base - IV0->base. */
1275 static void
1276 assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1,
1277 struct tree_niter_desc *niter, bounds *bnds)
1279 tree assumption = boolean_true_node, bound, diff;
1280 tree mbz, mbzl, mbzr, type1;
1281 bool rolls_p, no_overflow_p;
1282 widest_int dstep;
1283 mpz_t mstep, max;
1285 /* We are going to compute the number of iterations as
1286 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned
1287 variant of TYPE. This formula only works if
1289 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
1291 (where MAX is the maximum value of the unsigned variant of TYPE, and
1292 the computations in this formula are performed in full precision,
1293 i.e., without overflows).
1295 Usually, for loops with exit condition iv0->base + step * i < iv1->base,
1296 we have a condition of the form iv0->base - step < iv1->base before the loop,
1297 and for loops iv0->base < iv1->base - step * i the condition
1298 iv0->base < iv1->base + step, due to loop header copying, which enable us
1299 to prove the lower bound.
1301 The upper bound is more complicated. Unless the expressions for initial
1302 and final value themselves contain enough information, we usually cannot
1303 derive it from the context. */
1305 /* First check whether the answer does not follow from the bounds we gathered
1306 before. */
1307 if (integer_nonzerop (iv0->step))
1308 dstep = wi::to_widest (iv0->step);
1309 else
1311 dstep = wi::sext (wi::to_widest (iv1->step), TYPE_PRECISION (type));
1312 dstep = -dstep;
1315 mpz_init (mstep);
1316 wi::to_mpz (dstep, mstep, UNSIGNED);
1317 mpz_neg (mstep, mstep);
1318 mpz_add_ui (mstep, mstep, 1);
1320 rolls_p = mpz_cmp (mstep, bnds->below) <= 0;
1322 mpz_init (max);
1323 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED);
1324 mpz_add (max, max, mstep);
1325 no_overflow_p = (mpz_cmp (bnds->up, max) <= 0
1326 /* For pointers, only values lying inside a single object
1327 can be compared or manipulated by pointer arithmetics.
1328 Gcc in general does not allow or handle objects larger
1329 than half of the address space, hence the upper bound
1330 is satisfied for pointers. */
1331 || POINTER_TYPE_P (type));
1332 mpz_clear (mstep);
1333 mpz_clear (max);
1335 if (rolls_p && no_overflow_p)
1336 return;
1338 type1 = type;
1339 if (POINTER_TYPE_P (type))
1340 type1 = sizetype;
1342 /* Now the hard part; we must formulate the assumption(s) as expressions, and
1343 we must be careful not to introduce overflow. */
1345 if (integer_nonzerop (iv0->step))
1347 diff = fold_build2 (MINUS_EXPR, type1,
1348 iv0->step, build_int_cst (type1, 1));
1350 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
1351 0 address never belongs to any object, we can assume this for
1352 pointers. */
1353 if (!POINTER_TYPE_P (type))
1355 bound = fold_build2 (PLUS_EXPR, type1,
1356 TYPE_MIN_VALUE (type), diff);
1357 assumption = fold_build2 (GE_EXPR, boolean_type_node,
1358 iv0->base, bound);
1361 /* And then we can compute iv0->base - diff, and compare it with
1362 iv1->base. */
1363 mbzl = fold_build2 (MINUS_EXPR, type1,
1364 fold_convert (type1, iv0->base), diff);
1365 mbzr = fold_convert (type1, iv1->base);
1367 else
1369 diff = fold_build2 (PLUS_EXPR, type1,
1370 iv1->step, build_int_cst (type1, 1));
1372 if (!POINTER_TYPE_P (type))
1374 bound = fold_build2 (PLUS_EXPR, type1,
1375 TYPE_MAX_VALUE (type), diff);
1376 assumption = fold_build2 (LE_EXPR, boolean_type_node,
1377 iv1->base, bound);
1380 mbzl = fold_convert (type1, iv0->base);
1381 mbzr = fold_build2 (MINUS_EXPR, type1,
1382 fold_convert (type1, iv1->base), diff);
1385 if (!integer_nonzerop (assumption))
1386 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1387 niter->assumptions, assumption);
1388 if (!rolls_p)
1390 mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr);
1391 niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
1392 niter->may_be_zero, mbz);
1396 /* Determines number of iterations of loop whose ending condition
1397 is IV0 < IV1. TYPE is the type of the iv. The number of
1398 iterations is stored to NITER. BNDS bounds the difference
1399 IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know
1400 that the exit must be taken eventually. */
1402 static bool
1403 number_of_iterations_lt (struct loop *loop, tree type, affine_iv *iv0,
1404 affine_iv *iv1, struct tree_niter_desc *niter,
1405 bool exit_must_be_taken, bounds *bnds)
1407 tree niter_type = unsigned_type_for (type);
1408 tree delta, step, s;
1409 mpz_t mstep, tmp;
1411 if (integer_nonzerop (iv0->step))
1413 niter->control = *iv0;
1414 niter->cmp = LT_EXPR;
1415 niter->bound = iv1->base;
1417 else
1419 niter->control = *iv1;
1420 niter->cmp = GT_EXPR;
1421 niter->bound = iv0->base;
1424 delta = fold_build2 (MINUS_EXPR, niter_type,
1425 fold_convert (niter_type, iv1->base),
1426 fold_convert (niter_type, iv0->base));
1428 /* First handle the special case that the step is +-1. */
1429 if ((integer_onep (iv0->step) && integer_zerop (iv1->step))
1430 || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step)))
1432 /* for (i = iv0->base; i < iv1->base; i++)
1436 for (i = iv1->base; i > iv0->base; i--).
1438 In both cases # of iterations is iv1->base - iv0->base, assuming that
1439 iv1->base >= iv0->base.
1441 First try to derive a lower bound on the value of
1442 iv1->base - iv0->base, computed in full precision. If the difference
1443 is nonnegative, we are done, otherwise we must record the
1444 condition. */
1446 if (mpz_sgn (bnds->below) < 0)
1447 niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node,
1448 iv1->base, iv0->base);
1449 niter->niter = delta;
1450 niter->max = widest_int::from (wi::from_mpz (niter_type, bnds->up, false),
1451 TYPE_SIGN (niter_type));
1452 niter->control.no_overflow = true;
1453 return true;
1456 if (integer_nonzerop (iv0->step))
1457 step = fold_convert (niter_type, iv0->step);
1458 else
1459 step = fold_convert (niter_type,
1460 fold_build1 (NEGATE_EXPR, type, iv1->step));
1462 /* If we can determine the final value of the control iv exactly, we can
1463 transform the condition to != comparison. In particular, this will be
1464 the case if DELTA is constant. */
1465 if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step,
1466 exit_must_be_taken, bnds))
1468 affine_iv zps;
1470 zps.base = build_int_cst (niter_type, 0);
1471 zps.step = step;
1472 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1473 zps does not overflow. */
1474 zps.no_overflow = true;
1476 return number_of_iterations_ne (loop, type, &zps,
1477 delta, niter, true, bnds);
1480 /* Make sure that the control iv does not overflow. */
1481 if (!assert_no_overflow_lt (type, iv0, iv1, niter, step))
1482 return false;
1484 /* We determine the number of iterations as (delta + step - 1) / step. For
1485 this to work, we must know that iv1->base >= iv0->base - step + 1,
1486 otherwise the loop does not roll. */
1487 assert_loop_rolls_lt (type, iv0, iv1, niter, bnds);
1489 s = fold_build2 (MINUS_EXPR, niter_type,
1490 step, build_int_cst (niter_type, 1));
1491 delta = fold_build2 (PLUS_EXPR, niter_type, delta, s);
1492 niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step);
1494 mpz_init (mstep);
1495 mpz_init (tmp);
1496 wi::to_mpz (step, mstep, UNSIGNED);
1497 mpz_add (tmp, bnds->up, mstep);
1498 mpz_sub_ui (tmp, tmp, 1);
1499 mpz_fdiv_q (tmp, tmp, mstep);
1500 niter->max = widest_int::from (wi::from_mpz (niter_type, tmp, false),
1501 TYPE_SIGN (niter_type));
1502 mpz_clear (mstep);
1503 mpz_clear (tmp);
1505 return true;
1508 /* Determines number of iterations of loop whose ending condition
1509 is IV0 <= IV1. TYPE is the type of the iv. The number of
1510 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
1511 we know that this condition must eventually become false (we derived this
1512 earlier, and possibly set NITER->assumptions to make sure this
1513 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1515 static bool
1516 number_of_iterations_le (struct loop *loop, tree type, affine_iv *iv0,
1517 affine_iv *iv1, struct tree_niter_desc *niter,
1518 bool exit_must_be_taken, bounds *bnds)
1520 tree assumption;
1521 tree type1 = type;
1522 if (POINTER_TYPE_P (type))
1523 type1 = sizetype;
1525 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1526 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1527 value of the type. This we must know anyway, since if it is
1528 equal to this value, the loop rolls forever. We do not check
1529 this condition for pointer type ivs, as the code cannot rely on
1530 the object to that the pointer points being placed at the end of
1531 the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is
1532 not defined for pointers). */
1534 if (!exit_must_be_taken && !POINTER_TYPE_P (type))
1536 if (integer_nonzerop (iv0->step))
1537 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1538 iv1->base, TYPE_MAX_VALUE (type));
1539 else
1540 assumption = fold_build2 (NE_EXPR, boolean_type_node,
1541 iv0->base, TYPE_MIN_VALUE (type));
1543 if (integer_zerop (assumption))
1544 return false;
1545 if (!integer_nonzerop (assumption))
1546 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1547 niter->assumptions, assumption);
1550 if (integer_nonzerop (iv0->step))
1552 if (POINTER_TYPE_P (type))
1553 iv1->base = fold_build_pointer_plus_hwi (iv1->base, 1);
1554 else
1555 iv1->base = fold_build2 (PLUS_EXPR, type1, iv1->base,
1556 build_int_cst (type1, 1));
1558 else if (POINTER_TYPE_P (type))
1559 iv0->base = fold_build_pointer_plus_hwi (iv0->base, -1);
1560 else
1561 iv0->base = fold_build2 (MINUS_EXPR, type1,
1562 iv0->base, build_int_cst (type1, 1));
1564 bounds_add (bnds, 1, type1);
1566 return number_of_iterations_lt (loop, type, iv0, iv1, niter, exit_must_be_taken,
1567 bnds);
1570 /* Dumps description of affine induction variable IV to FILE. */
1572 static void
1573 dump_affine_iv (FILE *file, affine_iv *iv)
1575 if (!integer_zerop (iv->step))
1576 fprintf (file, "[");
1578 print_generic_expr (dump_file, iv->base, TDF_SLIM);
1580 if (!integer_zerop (iv->step))
1582 fprintf (file, ", + , ");
1583 print_generic_expr (dump_file, iv->step, TDF_SLIM);
1584 fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : "");
1588 /* Determine the number of iterations according to condition (for staying
1589 inside loop) which compares two induction variables using comparison
1590 operator CODE. The induction variable on left side of the comparison
1591 is IV0, the right-hand side is IV1. Both induction variables must have
1592 type TYPE, which must be an integer or pointer type. The steps of the
1593 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1595 LOOP is the loop whose number of iterations we are determining.
1597 ONLY_EXIT is true if we are sure this is the only way the loop could be
1598 exited (including possibly non-returning function calls, exceptions, etc.)
1599 -- in this case we can use the information whether the control induction
1600 variables can overflow or not in a more efficient way.
1602 if EVERY_ITERATION is true, we know the test is executed on every iteration.
1604 The results (number of iterations and assumptions as described in
1605 comments at struct tree_niter_desc in tree-ssa-loop.h) are stored to NITER.
1606 Returns false if it fails to determine number of iterations, true if it
1607 was determined (possibly with some assumptions). */
1609 static bool
1610 number_of_iterations_cond (struct loop *loop,
1611 tree type, affine_iv *iv0, enum tree_code code,
1612 affine_iv *iv1, struct tree_niter_desc *niter,
1613 bool only_exit, bool every_iteration)
1615 bool exit_must_be_taken = false, ret;
1616 bounds bnds;
1618 /* If the test is not executed every iteration, wrapping may make the test
1619 to pass again.
1620 TODO: the overflow case can be still used as unreliable estimate of upper
1621 bound. But we have no API to pass it down to number of iterations code
1622 and, at present, it will not use it anyway. */
1623 if (!every_iteration
1624 && (!iv0->no_overflow || !iv1->no_overflow
1625 || code == NE_EXPR || code == EQ_EXPR))
1626 return false;
1628 /* The meaning of these assumptions is this:
1629 if !assumptions
1630 then the rest of information does not have to be valid
1631 if may_be_zero then the loop does not roll, even if
1632 niter != 0. */
1633 niter->assumptions = boolean_true_node;
1634 niter->may_be_zero = boolean_false_node;
1635 niter->niter = NULL_TREE;
1636 niter->max = 0;
1637 niter->bound = NULL_TREE;
1638 niter->cmp = ERROR_MARK;
1640 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1641 the control variable is on lhs. */
1642 if (code == GE_EXPR || code == GT_EXPR
1643 || (code == NE_EXPR && integer_zerop (iv0->step)))
1645 std::swap (iv0, iv1);
1646 code = swap_tree_comparison (code);
1649 if (POINTER_TYPE_P (type))
1651 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1652 to the same object. If they do, the control variable cannot wrap
1653 (as wrap around the bounds of memory will never return a pointer
1654 that would be guaranteed to point to the same object, even if we
1655 avoid undefined behavior by casting to size_t and back). */
1656 iv0->no_overflow = true;
1657 iv1->no_overflow = true;
1660 /* If the control induction variable does not overflow and the only exit
1661 from the loop is the one that we analyze, we know it must be taken
1662 eventually. */
1663 if (only_exit)
1665 if (!integer_zerop (iv0->step) && iv0->no_overflow)
1666 exit_must_be_taken = true;
1667 else if (!integer_zerop (iv1->step) && iv1->no_overflow)
1668 exit_must_be_taken = true;
1671 /* We can handle the case when neither of the sides of the comparison is
1672 invariant, provided that the test is NE_EXPR. This rarely occurs in
1673 practice, but it is simple enough to manage. */
1674 if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step))
1676 tree step_type = POINTER_TYPE_P (type) ? sizetype : type;
1677 if (code != NE_EXPR)
1678 return false;
1680 iv0->step = fold_binary_to_constant (MINUS_EXPR, step_type,
1681 iv0->step, iv1->step);
1682 iv0->no_overflow = false;
1683 iv1->step = build_int_cst (step_type, 0);
1684 iv1->no_overflow = true;
1687 /* If the result of the comparison is a constant, the loop is weird. More
1688 precise handling would be possible, but the situation is not common enough
1689 to waste time on it. */
1690 if (integer_zerop (iv0->step) && integer_zerop (iv1->step))
1691 return false;
1693 /* Ignore loops of while (i-- < 10) type. */
1694 if (code != NE_EXPR)
1696 if (iv0->step && tree_int_cst_sign_bit (iv0->step))
1697 return false;
1699 if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step))
1700 return false;
1703 /* If the loop exits immediately, there is nothing to do. */
1704 tree tem = fold_binary (code, boolean_type_node, iv0->base, iv1->base);
1705 if (tem && integer_zerop (tem))
1707 niter->niter = build_int_cst (unsigned_type_for (type), 0);
1708 niter->max = 0;
1709 return true;
1712 /* OK, now we know we have a senseful loop. Handle several cases, depending
1713 on what comparison operator is used. */
1714 bound_difference (loop, iv1->base, iv0->base, &bnds);
1716 if (dump_file && (dump_flags & TDF_DETAILS))
1718 fprintf (dump_file,
1719 "Analyzing # of iterations of loop %d\n", loop->num);
1721 fprintf (dump_file, " exit condition ");
1722 dump_affine_iv (dump_file, iv0);
1723 fprintf (dump_file, " %s ",
1724 code == NE_EXPR ? "!="
1725 : code == LT_EXPR ? "<"
1726 : "<=");
1727 dump_affine_iv (dump_file, iv1);
1728 fprintf (dump_file, "\n");
1730 fprintf (dump_file, " bounds on difference of bases: ");
1731 mpz_out_str (dump_file, 10, bnds.below);
1732 fprintf (dump_file, " ... ");
1733 mpz_out_str (dump_file, 10, bnds.up);
1734 fprintf (dump_file, "\n");
1737 switch (code)
1739 case NE_EXPR:
1740 gcc_assert (integer_zerop (iv1->step));
1741 ret = number_of_iterations_ne (loop, type, iv0, iv1->base, niter,
1742 exit_must_be_taken, &bnds);
1743 break;
1745 case LT_EXPR:
1746 ret = number_of_iterations_lt (loop, type, iv0, iv1, niter,
1747 exit_must_be_taken, &bnds);
1748 break;
1750 case LE_EXPR:
1751 ret = number_of_iterations_le (loop, type, iv0, iv1, niter,
1752 exit_must_be_taken, &bnds);
1753 break;
1755 default:
1756 gcc_unreachable ();
1759 mpz_clear (bnds.up);
1760 mpz_clear (bnds.below);
1762 if (dump_file && (dump_flags & TDF_DETAILS))
1764 if (ret)
1766 fprintf (dump_file, " result:\n");
1767 if (!integer_nonzerop (niter->assumptions))
1769 fprintf (dump_file, " under assumptions ");
1770 print_generic_expr (dump_file, niter->assumptions, TDF_SLIM);
1771 fprintf (dump_file, "\n");
1774 if (!integer_zerop (niter->may_be_zero))
1776 fprintf (dump_file, " zero if ");
1777 print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM);
1778 fprintf (dump_file, "\n");
1781 fprintf (dump_file, " # of iterations ");
1782 print_generic_expr (dump_file, niter->niter, TDF_SLIM);
1783 fprintf (dump_file, ", bounded by ");
1784 print_decu (niter->max, dump_file);
1785 fprintf (dump_file, "\n");
1787 else
1788 fprintf (dump_file, " failed\n\n");
1790 return ret;
1793 /* Substitute NEW for OLD in EXPR and fold the result. */
1795 static tree
1796 simplify_replace_tree (tree expr, tree old, tree new_tree)
1798 unsigned i, n;
1799 tree ret = NULL_TREE, e, se;
1801 if (!expr)
1802 return NULL_TREE;
1804 /* Do not bother to replace constants. */
1805 if (CONSTANT_CLASS_P (old))
1806 return expr;
1808 if (expr == old
1809 || operand_equal_p (expr, old, 0))
1810 return unshare_expr (new_tree);
1812 if (!EXPR_P (expr))
1813 return expr;
1815 n = TREE_OPERAND_LENGTH (expr);
1816 for (i = 0; i < n; i++)
1818 e = TREE_OPERAND (expr, i);
1819 se = simplify_replace_tree (e, old, new_tree);
1820 if (e == se)
1821 continue;
1823 if (!ret)
1824 ret = copy_node (expr);
1826 TREE_OPERAND (ret, i) = se;
1829 return (ret ? fold (ret) : expr);
1832 /* Expand definitions of ssa names in EXPR as long as they are simple
1833 enough, and return the new expression. If STOP is specified, stop
1834 expanding if EXPR equals to it. */
1836 tree
1837 expand_simple_operations (tree expr, tree stop)
1839 unsigned i, n;
1840 tree ret = NULL_TREE, e, ee, e1;
1841 enum tree_code code;
1842 gimple *stmt;
1844 if (expr == NULL_TREE)
1845 return expr;
1847 if (is_gimple_min_invariant (expr))
1848 return expr;
1850 code = TREE_CODE (expr);
1851 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code)))
1853 n = TREE_OPERAND_LENGTH (expr);
1854 for (i = 0; i < n; i++)
1856 e = TREE_OPERAND (expr, i);
1857 ee = expand_simple_operations (e, stop);
1858 if (e == ee)
1859 continue;
1861 if (!ret)
1862 ret = copy_node (expr);
1864 TREE_OPERAND (ret, i) = ee;
1867 if (!ret)
1868 return expr;
1870 fold_defer_overflow_warnings ();
1871 ret = fold (ret);
1872 fold_undefer_and_ignore_overflow_warnings ();
1873 return ret;
1876 /* Stop if it's not ssa name or the one we don't want to expand. */
1877 if (TREE_CODE (expr) != SSA_NAME || expr == stop)
1878 return expr;
1880 stmt = SSA_NAME_DEF_STMT (expr);
1881 if (gimple_code (stmt) == GIMPLE_PHI)
1883 basic_block src, dest;
1885 if (gimple_phi_num_args (stmt) != 1)
1886 return expr;
1887 e = PHI_ARG_DEF (stmt, 0);
1889 /* Avoid propagating through loop exit phi nodes, which
1890 could break loop-closed SSA form restrictions. */
1891 dest = gimple_bb (stmt);
1892 src = single_pred (dest);
1893 if (TREE_CODE (e) == SSA_NAME
1894 && src->loop_father != dest->loop_father)
1895 return expr;
1897 return expand_simple_operations (e, stop);
1899 if (gimple_code (stmt) != GIMPLE_ASSIGN)
1900 return expr;
1902 /* Avoid expanding to expressions that contain SSA names that need
1903 to take part in abnormal coalescing. */
1904 ssa_op_iter iter;
1905 FOR_EACH_SSA_TREE_OPERAND (e, stmt, iter, SSA_OP_USE)
1906 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e))
1907 return expr;
1909 e = gimple_assign_rhs1 (stmt);
1910 code = gimple_assign_rhs_code (stmt);
1911 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1913 if (is_gimple_min_invariant (e))
1914 return e;
1916 if (code == SSA_NAME)
1917 return expand_simple_operations (e, stop);
1919 return expr;
1922 switch (code)
1924 CASE_CONVERT:
1925 /* Casts are simple. */
1926 ee = expand_simple_operations (e, stop);
1927 return fold_build1 (code, TREE_TYPE (expr), ee);
1929 case PLUS_EXPR:
1930 case MINUS_EXPR:
1931 if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (expr))
1932 && TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr)))
1933 return expr;
1934 /* Fallthru. */
1935 case POINTER_PLUS_EXPR:
1936 /* And increments and decrements by a constant are simple. */
1937 e1 = gimple_assign_rhs2 (stmt);
1938 if (!is_gimple_min_invariant (e1))
1939 return expr;
1941 ee = expand_simple_operations (e, stop);
1942 return fold_build2 (code, TREE_TYPE (expr), ee, e1);
1944 default:
1945 return expr;
1949 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1950 expression (or EXPR unchanged, if no simplification was possible). */
1952 static tree
1953 tree_simplify_using_condition_1 (tree cond, tree expr)
1955 bool changed;
1956 tree e, e0, e1, e2, notcond;
1957 enum tree_code code = TREE_CODE (expr);
1959 if (code == INTEGER_CST)
1960 return expr;
1962 if (code == TRUTH_OR_EXPR
1963 || code == TRUTH_AND_EXPR
1964 || code == COND_EXPR)
1966 changed = false;
1968 e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0));
1969 if (TREE_OPERAND (expr, 0) != e0)
1970 changed = true;
1972 e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1));
1973 if (TREE_OPERAND (expr, 1) != e1)
1974 changed = true;
1976 if (code == COND_EXPR)
1978 e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2));
1979 if (TREE_OPERAND (expr, 2) != e2)
1980 changed = true;
1982 else
1983 e2 = NULL_TREE;
1985 if (changed)
1987 if (code == COND_EXPR)
1988 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
1989 else
1990 expr = fold_build2 (code, boolean_type_node, e0, e1);
1993 return expr;
1996 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
1997 propagation, and vice versa. Fold does not handle this, since it is
1998 considered too expensive. */
1999 if (TREE_CODE (cond) == EQ_EXPR)
2001 e0 = TREE_OPERAND (cond, 0);
2002 e1 = TREE_OPERAND (cond, 1);
2004 /* We know that e0 == e1. Check whether we cannot simplify expr
2005 using this fact. */
2006 e = simplify_replace_tree (expr, e0, e1);
2007 if (integer_zerop (e) || integer_nonzerop (e))
2008 return e;
2010 e = simplify_replace_tree (expr, e1, e0);
2011 if (integer_zerop (e) || integer_nonzerop (e))
2012 return e;
2014 if (TREE_CODE (expr) == EQ_EXPR)
2016 e0 = TREE_OPERAND (expr, 0);
2017 e1 = TREE_OPERAND (expr, 1);
2019 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
2020 e = simplify_replace_tree (cond, e0, e1);
2021 if (integer_zerop (e))
2022 return e;
2023 e = simplify_replace_tree (cond, e1, e0);
2024 if (integer_zerop (e))
2025 return e;
2027 if (TREE_CODE (expr) == NE_EXPR)
2029 e0 = TREE_OPERAND (expr, 0);
2030 e1 = TREE_OPERAND (expr, 1);
2032 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
2033 e = simplify_replace_tree (cond, e0, e1);
2034 if (integer_zerop (e))
2035 return boolean_true_node;
2036 e = simplify_replace_tree (cond, e1, e0);
2037 if (integer_zerop (e))
2038 return boolean_true_node;
2041 /* Check whether COND ==> EXPR. */
2042 notcond = invert_truthvalue (cond);
2043 e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, expr);
2044 if (e && integer_nonzerop (e))
2045 return e;
2047 /* Check whether COND ==> not EXPR. */
2048 e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, expr);
2049 if (e && integer_zerop (e))
2050 return e;
2052 return expr;
2055 /* Tries to simplify EXPR using the condition COND. Returns the simplified
2056 expression (or EXPR unchanged, if no simplification was possible).
2057 Wrapper around tree_simplify_using_condition_1 that ensures that chains
2058 of simple operations in definitions of ssa names in COND are expanded,
2059 so that things like casts or incrementing the value of the bound before
2060 the loop do not cause us to fail. */
2062 static tree
2063 tree_simplify_using_condition (tree cond, tree expr)
2065 cond = expand_simple_operations (cond);
2067 return tree_simplify_using_condition_1 (cond, expr);
2070 /* Tries to simplify EXPR using the conditions on entry to LOOP.
2071 Returns the simplified expression (or EXPR unchanged, if no
2072 simplification was possible). */
2074 tree
2075 simplify_using_initial_conditions (struct loop *loop, tree expr)
2077 edge e;
2078 basic_block bb;
2079 gimple *stmt;
2080 tree cond, expanded, backup;
2081 int cnt = 0;
2083 if (TREE_CODE (expr) == INTEGER_CST)
2084 return expr;
2086 backup = expanded = expand_simple_operations (expr);
2088 /* Limit walking the dominators to avoid quadraticness in
2089 the number of BBs times the number of loops in degenerate
2090 cases. */
2091 for (bb = loop->header;
2092 bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK;
2093 bb = get_immediate_dominator (CDI_DOMINATORS, bb))
2095 if (!single_pred_p (bb))
2096 continue;
2097 e = single_pred_edge (bb);
2099 if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
2100 continue;
2102 stmt = last_stmt (e->src);
2103 cond = fold_build2 (gimple_cond_code (stmt),
2104 boolean_type_node,
2105 gimple_cond_lhs (stmt),
2106 gimple_cond_rhs (stmt));
2107 if (e->flags & EDGE_FALSE_VALUE)
2108 cond = invert_truthvalue (cond);
2109 expanded = tree_simplify_using_condition (cond, expanded);
2110 /* Break if EXPR is simplified to const values. */
2111 if (expanded
2112 && (integer_zerop (expanded) || integer_nonzerop (expanded)))
2113 return expanded;
2115 ++cnt;
2118 /* Return the original expression if no simplification is done. */
2119 return operand_equal_p (backup, expanded, 0) ? expr : expanded;
2122 /* Tries to simplify EXPR using the evolutions of the loop invariants
2123 in the superloops of LOOP. Returns the simplified expression
2124 (or EXPR unchanged, if no simplification was possible). */
2126 static tree
2127 simplify_using_outer_evolutions (struct loop *loop, tree expr)
2129 enum tree_code code = TREE_CODE (expr);
2130 bool changed;
2131 tree e, e0, e1, e2;
2133 if (is_gimple_min_invariant (expr))
2134 return expr;
2136 if (code == TRUTH_OR_EXPR
2137 || code == TRUTH_AND_EXPR
2138 || code == COND_EXPR)
2140 changed = false;
2142 e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0));
2143 if (TREE_OPERAND (expr, 0) != e0)
2144 changed = true;
2146 e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1));
2147 if (TREE_OPERAND (expr, 1) != e1)
2148 changed = true;
2150 if (code == COND_EXPR)
2152 e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2));
2153 if (TREE_OPERAND (expr, 2) != e2)
2154 changed = true;
2156 else
2157 e2 = NULL_TREE;
2159 if (changed)
2161 if (code == COND_EXPR)
2162 expr = fold_build3 (code, boolean_type_node, e0, e1, e2);
2163 else
2164 expr = fold_build2 (code, boolean_type_node, e0, e1);
2167 return expr;
2170 e = instantiate_parameters (loop, expr);
2171 if (is_gimple_min_invariant (e))
2172 return e;
2174 return expr;
2177 /* Returns true if EXIT is the only possible exit from LOOP. */
2179 bool
2180 loop_only_exit_p (const struct loop *loop, const_edge exit)
2182 basic_block *body;
2183 gimple_stmt_iterator bsi;
2184 unsigned i;
2186 if (exit != single_exit (loop))
2187 return false;
2189 body = get_loop_body (loop);
2190 for (i = 0; i < loop->num_nodes; i++)
2192 for (bsi = gsi_start_bb (body[i]); !gsi_end_p (bsi); gsi_next (&bsi))
2193 if (stmt_can_terminate_bb_p (gsi_stmt (bsi)))
2195 free (body);
2196 return true;
2200 free (body);
2201 return true;
2204 /* Stores description of number of iterations of LOOP derived from
2205 EXIT (an exit edge of the LOOP) in NITER. Returns true if some useful
2206 information could be derived (and fields of NITER have meaning described
2207 in comments at struct tree_niter_desc declaration), false otherwise.
2208 When EVERY_ITERATION is true, only tests that are known to be executed
2209 every iteration are considered (i.e. only test that alone bounds the loop).
2210 If AT_STMT is not NULL, this function stores LOOP's condition statement in
2211 it when returning true. */
2213 bool
2214 number_of_iterations_exit_assumptions (struct loop *loop, edge exit,
2215 struct tree_niter_desc *niter,
2216 gcond **at_stmt, bool every_iteration)
2218 gimple *last;
2219 gcond *stmt;
2220 tree type;
2221 tree op0, op1;
2222 enum tree_code code;
2223 affine_iv iv0, iv1;
2224 bool safe;
2226 /* Nothing to analyze if the loop is known to be infinite. */
2227 if (loop_constraint_set_p (loop, LOOP_C_INFINITE))
2228 return false;
2230 safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src);
2232 if (every_iteration && !safe)
2233 return false;
2235 niter->assumptions = boolean_false_node;
2236 niter->control.base = NULL_TREE;
2237 niter->control.step = NULL_TREE;
2238 niter->control.no_overflow = false;
2239 last = last_stmt (exit->src);
2240 if (!last)
2241 return false;
2242 stmt = dyn_cast <gcond *> (last);
2243 if (!stmt)
2244 return false;
2246 /* We want the condition for staying inside loop. */
2247 code = gimple_cond_code (stmt);
2248 if (exit->flags & EDGE_TRUE_VALUE)
2249 code = invert_tree_comparison (code, false);
2251 switch (code)
2253 case GT_EXPR:
2254 case GE_EXPR:
2255 case LT_EXPR:
2256 case LE_EXPR:
2257 case NE_EXPR:
2258 break;
2260 default:
2261 return false;
2264 op0 = gimple_cond_lhs (stmt);
2265 op1 = gimple_cond_rhs (stmt);
2266 type = TREE_TYPE (op0);
2268 if (TREE_CODE (type) != INTEGER_TYPE
2269 && !POINTER_TYPE_P (type))
2270 return false;
2272 tree iv0_niters = NULL_TREE;
2273 if (!simple_iv_with_niters (loop, loop_containing_stmt (stmt),
2274 op0, &iv0, &iv0_niters, false))
2275 return false;
2276 tree iv1_niters = NULL_TREE;
2277 if (!simple_iv_with_niters (loop, loop_containing_stmt (stmt),
2278 op1, &iv1, &iv1_niters, false))
2279 return false;
2280 /* Give up on complicated case. */
2281 if (iv0_niters && iv1_niters)
2282 return false;
2284 /* We don't want to see undefined signed overflow warnings while
2285 computing the number of iterations. */
2286 fold_defer_overflow_warnings ();
2288 iv0.base = expand_simple_operations (iv0.base);
2289 iv1.base = expand_simple_operations (iv1.base);
2290 if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter,
2291 loop_only_exit_p (loop, exit), safe))
2293 fold_undefer_and_ignore_overflow_warnings ();
2294 return false;
2297 /* Incorporate additional assumption implied by control iv. */
2298 tree iv_niters = iv0_niters ? iv0_niters : iv1_niters;
2299 if (iv_niters)
2301 tree assumption = fold_build2 (LE_EXPR, boolean_type_node, niter->niter,
2302 fold_convert (TREE_TYPE (niter->niter),
2303 iv_niters));
2305 if (!integer_nonzerop (assumption))
2306 niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
2307 niter->assumptions, assumption);
2309 /* Refine upper bound if possible. */
2310 if (TREE_CODE (iv_niters) == INTEGER_CST
2311 && niter->max > wi::to_widest (iv_niters))
2312 niter->max = wi::to_widest (iv_niters);
2315 /* There is no assumptions if the loop is known to be finite. */
2316 if (!integer_zerop (niter->assumptions)
2317 && loop_constraint_set_p (loop, LOOP_C_FINITE))
2318 niter->assumptions = boolean_true_node;
2320 if (optimize >= 3)
2322 niter->assumptions = simplify_using_outer_evolutions (loop,
2323 niter->assumptions);
2324 niter->may_be_zero = simplify_using_outer_evolutions (loop,
2325 niter->may_be_zero);
2326 niter->niter = simplify_using_outer_evolutions (loop, niter->niter);
2329 niter->assumptions
2330 = simplify_using_initial_conditions (loop,
2331 niter->assumptions);
2332 niter->may_be_zero
2333 = simplify_using_initial_conditions (loop,
2334 niter->may_be_zero);
2336 fold_undefer_and_ignore_overflow_warnings ();
2338 /* If NITER has simplified into a constant, update MAX. */
2339 if (TREE_CODE (niter->niter) == INTEGER_CST)
2340 niter->max = wi::to_widest (niter->niter);
2342 if (at_stmt)
2343 *at_stmt = stmt;
2345 return (!integer_zerop (niter->assumptions));
2348 /* Like number_of_iterations_exit_assumptions, but return TRUE only if
2349 the niter information holds unconditionally. */
2351 bool
2352 number_of_iterations_exit (struct loop *loop, edge exit,
2353 struct tree_niter_desc *niter,
2354 bool warn, bool every_iteration)
2356 gcond *stmt;
2357 if (!number_of_iterations_exit_assumptions (loop, exit, niter,
2358 &stmt, every_iteration))
2359 return false;
2361 if (integer_nonzerop (niter->assumptions))
2362 return true;
2364 if (warn)
2365 warning_at (gimple_location_safe (stmt),
2366 OPT_Wunsafe_loop_optimizations,
2367 "missed loop optimization, the loop counter may overflow");
2369 return false;
2372 /* Try to determine the number of iterations of LOOP. If we succeed,
2373 expression giving number of iterations is returned and *EXIT is
2374 set to the edge from that the information is obtained. Otherwise
2375 chrec_dont_know is returned. */
2377 tree
2378 find_loop_niter (struct loop *loop, edge *exit)
2380 unsigned i;
2381 vec<edge> exits = get_loop_exit_edges (loop);
2382 edge ex;
2383 tree niter = NULL_TREE, aniter;
2384 struct tree_niter_desc desc;
2386 *exit = NULL;
2387 FOR_EACH_VEC_ELT (exits, i, ex)
2389 if (!number_of_iterations_exit (loop, ex, &desc, false))
2390 continue;
2392 if (integer_nonzerop (desc.may_be_zero))
2394 /* We exit in the first iteration through this exit.
2395 We won't find anything better. */
2396 niter = build_int_cst (unsigned_type_node, 0);
2397 *exit = ex;
2398 break;
2401 if (!integer_zerop (desc.may_be_zero))
2402 continue;
2404 aniter = desc.niter;
2406 if (!niter)
2408 /* Nothing recorded yet. */
2409 niter = aniter;
2410 *exit = ex;
2411 continue;
2414 /* Prefer constants, the lower the better. */
2415 if (TREE_CODE (aniter) != INTEGER_CST)
2416 continue;
2418 if (TREE_CODE (niter) != INTEGER_CST)
2420 niter = aniter;
2421 *exit = ex;
2422 continue;
2425 if (tree_int_cst_lt (aniter, niter))
2427 niter = aniter;
2428 *exit = ex;
2429 continue;
2432 exits.release ();
2434 return niter ? niter : chrec_dont_know;
2437 /* Return true if loop is known to have bounded number of iterations. */
2439 bool
2440 finite_loop_p (struct loop *loop)
2442 widest_int nit;
2443 int flags;
2445 flags = flags_from_decl_or_type (current_function_decl);
2446 if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE))
2448 if (dump_file && (dump_flags & TDF_DETAILS))
2449 fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n",
2450 loop->num);
2451 return true;
2454 if (loop->any_upper_bound
2455 || max_loop_iterations (loop, &nit))
2457 if (dump_file && (dump_flags & TDF_DETAILS))
2458 fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n",
2459 loop->num);
2460 return true;
2462 return false;
2467 Analysis of a number of iterations of a loop by a brute-force evaluation.
2471 /* Bound on the number of iterations we try to evaluate. */
2473 #define MAX_ITERATIONS_TO_TRACK \
2474 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
2476 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
2477 result by a chain of operations such that all but exactly one of their
2478 operands are constants. */
2480 static gphi *
2481 chain_of_csts_start (struct loop *loop, tree x)
2483 gimple *stmt = SSA_NAME_DEF_STMT (x);
2484 tree use;
2485 basic_block bb = gimple_bb (stmt);
2486 enum tree_code code;
2488 if (!bb
2489 || !flow_bb_inside_loop_p (loop, bb))
2490 return NULL;
2492 if (gimple_code (stmt) == GIMPLE_PHI)
2494 if (bb == loop->header)
2495 return as_a <gphi *> (stmt);
2497 return NULL;
2500 if (gimple_code (stmt) != GIMPLE_ASSIGN
2501 || gimple_assign_rhs_class (stmt) == GIMPLE_TERNARY_RHS)
2502 return NULL;
2504 code = gimple_assign_rhs_code (stmt);
2505 if (gimple_references_memory_p (stmt)
2506 || TREE_CODE_CLASS (code) == tcc_reference
2507 || (code == ADDR_EXPR
2508 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
2509 return NULL;
2511 use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
2512 if (use == NULL_TREE)
2513 return NULL;
2515 return chain_of_csts_start (loop, use);
2518 /* Determines whether the expression X is derived from a result of a phi node
2519 in header of LOOP such that
2521 * the derivation of X consists only from operations with constants
2522 * the initial value of the phi node is constant
2523 * the value of the phi node in the next iteration can be derived from the
2524 value in the current iteration by a chain of operations with constants.
2526 If such phi node exists, it is returned, otherwise NULL is returned. */
2528 static gphi *
2529 get_base_for (struct loop *loop, tree x)
2531 gphi *phi;
2532 tree init, next;
2534 if (is_gimple_min_invariant (x))
2535 return NULL;
2537 phi = chain_of_csts_start (loop, x);
2538 if (!phi)
2539 return NULL;
2541 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2542 next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2544 if (TREE_CODE (next) != SSA_NAME)
2545 return NULL;
2547 if (!is_gimple_min_invariant (init))
2548 return NULL;
2550 if (chain_of_csts_start (loop, next) != phi)
2551 return NULL;
2553 return phi;
2556 /* Given an expression X, then
2558 * if X is NULL_TREE, we return the constant BASE.
2559 * otherwise X is a SSA name, whose value in the considered loop is derived
2560 by a chain of operations with constant from a result of a phi node in
2561 the header of the loop. Then we return value of X when the value of the
2562 result of this phi node is given by the constant BASE. */
2564 static tree
2565 get_val_for (tree x, tree base)
2567 gimple *stmt;
2569 gcc_checking_assert (is_gimple_min_invariant (base));
2571 if (!x)
2572 return base;
2574 stmt = SSA_NAME_DEF_STMT (x);
2575 if (gimple_code (stmt) == GIMPLE_PHI)
2576 return base;
2578 gcc_checking_assert (is_gimple_assign (stmt));
2580 /* STMT must be either an assignment of a single SSA name or an
2581 expression involving an SSA name and a constant. Try to fold that
2582 expression using the value for the SSA name. */
2583 if (gimple_assign_ssa_name_copy_p (stmt))
2584 return get_val_for (gimple_assign_rhs1 (stmt), base);
2585 else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS
2586 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
2588 return fold_build1 (gimple_assign_rhs_code (stmt),
2589 gimple_expr_type (stmt),
2590 get_val_for (gimple_assign_rhs1 (stmt), base));
2592 else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS)
2594 tree rhs1 = gimple_assign_rhs1 (stmt);
2595 tree rhs2 = gimple_assign_rhs2 (stmt);
2596 if (TREE_CODE (rhs1) == SSA_NAME)
2597 rhs1 = get_val_for (rhs1, base);
2598 else if (TREE_CODE (rhs2) == SSA_NAME)
2599 rhs2 = get_val_for (rhs2, base);
2600 else
2601 gcc_unreachable ();
2602 return fold_build2 (gimple_assign_rhs_code (stmt),
2603 gimple_expr_type (stmt), rhs1, rhs2);
2605 else
2606 gcc_unreachable ();
2610 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2611 by brute force -- i.e. by determining the value of the operands of the
2612 condition at EXIT in first few iterations of the loop (assuming that
2613 these values are constant) and determining the first one in that the
2614 condition is not satisfied. Returns the constant giving the number
2615 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2617 tree
2618 loop_niter_by_eval (struct loop *loop, edge exit)
2620 tree acnd;
2621 tree op[2], val[2], next[2], aval[2];
2622 gphi *phi;
2623 gimple *cond;
2624 unsigned i, j;
2625 enum tree_code cmp;
2627 cond = last_stmt (exit->src);
2628 if (!cond || gimple_code (cond) != GIMPLE_COND)
2629 return chrec_dont_know;
2631 cmp = gimple_cond_code (cond);
2632 if (exit->flags & EDGE_TRUE_VALUE)
2633 cmp = invert_tree_comparison (cmp, false);
2635 switch (cmp)
2637 case EQ_EXPR:
2638 case NE_EXPR:
2639 case GT_EXPR:
2640 case GE_EXPR:
2641 case LT_EXPR:
2642 case LE_EXPR:
2643 op[0] = gimple_cond_lhs (cond);
2644 op[1] = gimple_cond_rhs (cond);
2645 break;
2647 default:
2648 return chrec_dont_know;
2651 for (j = 0; j < 2; j++)
2653 if (is_gimple_min_invariant (op[j]))
2655 val[j] = op[j];
2656 next[j] = NULL_TREE;
2657 op[j] = NULL_TREE;
2659 else
2661 phi = get_base_for (loop, op[j]);
2662 if (!phi)
2663 return chrec_dont_know;
2664 val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
2665 next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop));
2669 /* Don't issue signed overflow warnings. */
2670 fold_defer_overflow_warnings ();
2672 for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++)
2674 for (j = 0; j < 2; j++)
2675 aval[j] = get_val_for (op[j], val[j]);
2677 acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]);
2678 if (acnd && integer_zerop (acnd))
2680 fold_undefer_and_ignore_overflow_warnings ();
2681 if (dump_file && (dump_flags & TDF_DETAILS))
2682 fprintf (dump_file,
2683 "Proved that loop %d iterates %d times using brute force.\n",
2684 loop->num, i);
2685 return build_int_cst (unsigned_type_node, i);
2688 for (j = 0; j < 2; j++)
2690 val[j] = get_val_for (next[j], val[j]);
2691 if (!is_gimple_min_invariant (val[j]))
2693 fold_undefer_and_ignore_overflow_warnings ();
2694 return chrec_dont_know;
2699 fold_undefer_and_ignore_overflow_warnings ();
2701 return chrec_dont_know;
2704 /* Finds the exit of the LOOP by that the loop exits after a constant
2705 number of iterations and stores the exit edge to *EXIT. The constant
2706 giving the number of iterations of LOOP is returned. The number of
2707 iterations is determined using loop_niter_by_eval (i.e. by brute force
2708 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2709 determines the number of iterations, chrec_dont_know is returned. */
2711 tree
2712 find_loop_niter_by_eval (struct loop *loop, edge *exit)
2714 unsigned i;
2715 vec<edge> exits = get_loop_exit_edges (loop);
2716 edge ex;
2717 tree niter = NULL_TREE, aniter;
2719 *exit = NULL;
2721 /* Loops with multiple exits are expensive to handle and less important. */
2722 if (!flag_expensive_optimizations
2723 && exits.length () > 1)
2725 exits.release ();
2726 return chrec_dont_know;
2729 FOR_EACH_VEC_ELT (exits, i, ex)
2731 if (!just_once_each_iteration_p (loop, ex->src))
2732 continue;
2734 aniter = loop_niter_by_eval (loop, ex);
2735 if (chrec_contains_undetermined (aniter))
2736 continue;
2738 if (niter
2739 && !tree_int_cst_lt (aniter, niter))
2740 continue;
2742 niter = aniter;
2743 *exit = ex;
2745 exits.release ();
2747 return niter ? niter : chrec_dont_know;
2752 Analysis of upper bounds on number of iterations of a loop.
2756 static widest_int derive_constant_upper_bound_ops (tree, tree,
2757 enum tree_code, tree);
2759 /* Returns a constant upper bound on the value of the right-hand side of
2760 an assignment statement STMT. */
2762 static widest_int
2763 derive_constant_upper_bound_assign (gimple *stmt)
2765 enum tree_code code = gimple_assign_rhs_code (stmt);
2766 tree op0 = gimple_assign_rhs1 (stmt);
2767 tree op1 = gimple_assign_rhs2 (stmt);
2769 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)),
2770 op0, code, op1);
2773 /* Returns a constant upper bound on the value of expression VAL. VAL
2774 is considered to be unsigned. If its type is signed, its value must
2775 be nonnegative. */
2777 static widest_int
2778 derive_constant_upper_bound (tree val)
2780 enum tree_code code;
2781 tree op0, op1, op2;
2783 extract_ops_from_tree (val, &code, &op0, &op1, &op2);
2784 return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1);
2787 /* Returns a constant upper bound on the value of expression OP0 CODE OP1,
2788 whose type is TYPE. The expression is considered to be unsigned. If
2789 its type is signed, its value must be nonnegative. */
2791 static widest_int
2792 derive_constant_upper_bound_ops (tree type, tree op0,
2793 enum tree_code code, tree op1)
2795 tree subtype, maxt;
2796 widest_int bnd, max, cst;
2797 gimple *stmt;
2799 if (INTEGRAL_TYPE_P (type))
2800 maxt = TYPE_MAX_VALUE (type);
2801 else
2802 maxt = upper_bound_in_type (type, type);
2804 max = wi::to_widest (maxt);
2806 switch (code)
2808 case INTEGER_CST:
2809 return wi::to_widest (op0);
2811 CASE_CONVERT:
2812 subtype = TREE_TYPE (op0);
2813 if (!TYPE_UNSIGNED (subtype)
2814 /* If TYPE is also signed, the fact that VAL is nonnegative implies
2815 that OP0 is nonnegative. */
2816 && TYPE_UNSIGNED (type)
2817 && !tree_expr_nonnegative_p (op0))
2819 /* If we cannot prove that the casted expression is nonnegative,
2820 we cannot establish more useful upper bound than the precision
2821 of the type gives us. */
2822 return max;
2825 /* We now know that op0 is an nonnegative value. Try deriving an upper
2826 bound for it. */
2827 bnd = derive_constant_upper_bound (op0);
2829 /* If the bound does not fit in TYPE, max. value of TYPE could be
2830 attained. */
2831 if (wi::ltu_p (max, bnd))
2832 return max;
2834 return bnd;
2836 case PLUS_EXPR:
2837 case POINTER_PLUS_EXPR:
2838 case MINUS_EXPR:
2839 if (TREE_CODE (op1) != INTEGER_CST
2840 || !tree_expr_nonnegative_p (op0))
2841 return max;
2843 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
2844 choose the most logical way how to treat this constant regardless
2845 of the signedness of the type. */
2846 cst = wi::sext (wi::to_widest (op1), TYPE_PRECISION (type));
2847 if (code != MINUS_EXPR)
2848 cst = -cst;
2850 bnd = derive_constant_upper_bound (op0);
2852 if (wi::neg_p (cst))
2854 cst = -cst;
2855 /* Avoid CST == 0x80000... */
2856 if (wi::neg_p (cst))
2857 return max;
2859 /* OP0 + CST. We need to check that
2860 BND <= MAX (type) - CST. */
2862 widest_int mmax = max - cst;
2863 if (wi::leu_p (bnd, mmax))
2864 return max;
2866 return bnd + cst;
2868 else
2870 /* OP0 - CST, where CST >= 0.
2872 If TYPE is signed, we have already verified that OP0 >= 0, and we
2873 know that the result is nonnegative. This implies that
2874 VAL <= BND - CST.
2876 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2877 otherwise the operation underflows.
2880 /* This should only happen if the type is unsigned; however, for
2881 buggy programs that use overflowing signed arithmetics even with
2882 -fno-wrapv, this condition may also be true for signed values. */
2883 if (wi::ltu_p (bnd, cst))
2884 return max;
2886 if (TYPE_UNSIGNED (type))
2888 tree tem = fold_binary (GE_EXPR, boolean_type_node, op0,
2889 wide_int_to_tree (type, cst));
2890 if (!tem || integer_nonzerop (tem))
2891 return max;
2894 bnd -= cst;
2897 return bnd;
2899 case FLOOR_DIV_EXPR:
2900 case EXACT_DIV_EXPR:
2901 if (TREE_CODE (op1) != INTEGER_CST
2902 || tree_int_cst_sign_bit (op1))
2903 return max;
2905 bnd = derive_constant_upper_bound (op0);
2906 return wi::udiv_floor (bnd, wi::to_widest (op1));
2908 case BIT_AND_EXPR:
2909 if (TREE_CODE (op1) != INTEGER_CST
2910 || tree_int_cst_sign_bit (op1))
2911 return max;
2912 return wi::to_widest (op1);
2914 case SSA_NAME:
2915 stmt = SSA_NAME_DEF_STMT (op0);
2916 if (gimple_code (stmt) != GIMPLE_ASSIGN
2917 || gimple_assign_lhs (stmt) != op0)
2918 return max;
2919 return derive_constant_upper_bound_assign (stmt);
2921 default:
2922 return max;
2926 /* Emit a -Waggressive-loop-optimizations warning if needed. */
2928 static void
2929 do_warn_aggressive_loop_optimizations (struct loop *loop,
2930 widest_int i_bound, gimple *stmt)
2932 /* Don't warn if the loop doesn't have known constant bound. */
2933 if (!loop->nb_iterations
2934 || TREE_CODE (loop->nb_iterations) != INTEGER_CST
2935 || !warn_aggressive_loop_optimizations
2936 /* To avoid warning multiple times for the same loop,
2937 only start warning when we preserve loops. */
2938 || (cfun->curr_properties & PROP_loops) == 0
2939 /* Only warn once per loop. */
2940 || loop->warned_aggressive_loop_optimizations
2941 /* Only warn if undefined behavior gives us lower estimate than the
2942 known constant bound. */
2943 || wi::cmpu (i_bound, wi::to_widest (loop->nb_iterations)) >= 0
2944 /* And undefined behavior happens unconditionally. */
2945 || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt)))
2946 return;
2948 edge e = single_exit (loop);
2949 if (e == NULL)
2950 return;
2952 gimple *estmt = last_stmt (e->src);
2953 char buf[WIDE_INT_PRINT_BUFFER_SIZE];
2954 print_dec (i_bound, buf, TYPE_UNSIGNED (TREE_TYPE (loop->nb_iterations))
2955 ? UNSIGNED : SIGNED);
2956 if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations,
2957 "iteration %s invokes undefined behavior", buf))
2958 inform (gimple_location (estmt), "within this loop");
2959 loop->warned_aggressive_loop_optimizations = true;
2962 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
2963 is true if the loop is exited immediately after STMT, and this exit
2964 is taken at last when the STMT is executed BOUND + 1 times.
2965 REALISTIC is true if BOUND is expected to be close to the real number
2966 of iterations. UPPER is true if we are sure the loop iterates at most
2967 BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */
2969 static void
2970 record_estimate (struct loop *loop, tree bound, const widest_int &i_bound,
2971 gimple *at_stmt, bool is_exit, bool realistic, bool upper)
2973 widest_int delta;
2975 if (dump_file && (dump_flags & TDF_DETAILS))
2977 fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : "");
2978 print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM);
2979 fprintf (dump_file, " is %sexecuted at most ",
2980 upper ? "" : "probably ");
2981 print_generic_expr (dump_file, bound, TDF_SLIM);
2982 fprintf (dump_file, " (bounded by ");
2983 print_decu (i_bound, dump_file);
2984 fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num);
2987 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
2988 real number of iterations. */
2989 if (TREE_CODE (bound) != INTEGER_CST)
2990 realistic = false;
2991 else
2992 gcc_checking_assert (i_bound == wi::to_widest (bound));
2994 /* If we have a guaranteed upper bound, record it in the appropriate
2995 list, unless this is an !is_exit bound (i.e. undefined behavior in
2996 at_stmt) in a loop with known constant number of iterations. */
2997 if (upper
2998 && (is_exit
2999 || loop->nb_iterations == NULL_TREE
3000 || TREE_CODE (loop->nb_iterations) != INTEGER_CST))
3002 struct nb_iter_bound *elt = ggc_alloc<nb_iter_bound> ();
3004 elt->bound = i_bound;
3005 elt->stmt = at_stmt;
3006 elt->is_exit = is_exit;
3007 elt->next = loop->bounds;
3008 loop->bounds = elt;
3011 /* If statement is executed on every path to the loop latch, we can directly
3012 infer the upper bound on the # of iterations of the loop. */
3013 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt)))
3014 upper = false;
3016 /* Update the number of iteration estimates according to the bound.
3017 If at_stmt is an exit then the loop latch is executed at most BOUND times,
3018 otherwise it can be executed BOUND + 1 times. We will lower the estimate
3019 later if such statement must be executed on last iteration */
3020 if (is_exit)
3021 delta = 0;
3022 else
3023 delta = 1;
3024 widest_int new_i_bound = i_bound + delta;
3026 /* If an overflow occurred, ignore the result. */
3027 if (wi::ltu_p (new_i_bound, delta))
3028 return;
3030 if (upper && !is_exit)
3031 do_warn_aggressive_loop_optimizations (loop, new_i_bound, at_stmt);
3032 record_niter_bound (loop, new_i_bound, realistic, upper);
3035 /* Records the control iv analyzed in NITER for LOOP if the iv is valid
3036 and doesn't overflow. */
3038 static void
3039 record_control_iv (struct loop *loop, struct tree_niter_desc *niter)
3041 struct control_iv *iv;
3043 if (!niter->control.base || !niter->control.step)
3044 return;
3046 if (!integer_onep (niter->assumptions) || !niter->control.no_overflow)
3047 return;
3049 iv = ggc_alloc<control_iv> ();
3050 iv->base = niter->control.base;
3051 iv->step = niter->control.step;
3052 iv->next = loop->control_ivs;
3053 loop->control_ivs = iv;
3055 return;
3058 /* This function returns TRUE if below conditions are satisfied:
3059 1) VAR is SSA variable.
3060 2) VAR is an IV:{base, step} in its defining loop.
3061 3) IV doesn't overflow.
3062 4) Both base and step are integer constants.
3063 5) Base is the MIN/MAX value depends on IS_MIN.
3064 Store value of base to INIT correspondingly. */
3066 static bool
3067 get_cst_init_from_scev (tree var, wide_int *init, bool is_min)
3069 if (TREE_CODE (var) != SSA_NAME)
3070 return false;
3072 gimple *def_stmt = SSA_NAME_DEF_STMT (var);
3073 struct loop *loop = loop_containing_stmt (def_stmt);
3075 if (loop == NULL)
3076 return false;
3078 affine_iv iv;
3079 if (!simple_iv (loop, loop, var, &iv, false))
3080 return false;
3082 if (!iv.no_overflow)
3083 return false;
3085 if (TREE_CODE (iv.base) != INTEGER_CST || TREE_CODE (iv.step) != INTEGER_CST)
3086 return false;
3088 if (is_min == tree_int_cst_sign_bit (iv.step))
3089 return false;
3091 *init = iv.base;
3092 return true;
3095 /* Record the estimate on number of iterations of LOOP based on the fact that
3096 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
3097 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
3098 estimated number of iterations is expected to be close to the real one.
3099 UPPER is true if we are sure the induction variable does not wrap. */
3101 static void
3102 record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple *stmt,
3103 tree low, tree high, bool realistic, bool upper)
3105 tree niter_bound, extreme, delta;
3106 tree type = TREE_TYPE (base), unsigned_type;
3107 tree orig_base = base;
3109 if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step))
3110 return;
3112 if (dump_file && (dump_flags & TDF_DETAILS))
3114 fprintf (dump_file, "Induction variable (");
3115 print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM);
3116 fprintf (dump_file, ") ");
3117 print_generic_expr (dump_file, base, TDF_SLIM);
3118 fprintf (dump_file, " + ");
3119 print_generic_expr (dump_file, step, TDF_SLIM);
3120 fprintf (dump_file, " * iteration does not wrap in statement ");
3121 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
3122 fprintf (dump_file, " in loop %d.\n", loop->num);
3125 unsigned_type = unsigned_type_for (type);
3126 base = fold_convert (unsigned_type, base);
3127 step = fold_convert (unsigned_type, step);
3129 if (tree_int_cst_sign_bit (step))
3131 wide_int min, max;
3132 extreme = fold_convert (unsigned_type, low);
3133 if (TREE_CODE (orig_base) == SSA_NAME
3134 && TREE_CODE (high) == INTEGER_CST
3135 && INTEGRAL_TYPE_P (TREE_TYPE (orig_base))
3136 && (get_range_info (orig_base, &min, &max) == VR_RANGE
3137 || get_cst_init_from_scev (orig_base, &max, false))
3138 && wi::gts_p (high, max))
3139 base = wide_int_to_tree (unsigned_type, max);
3140 else if (TREE_CODE (base) != INTEGER_CST
3141 && dominated_by_p (CDI_DOMINATORS,
3142 loop->latch, gimple_bb (stmt)))
3143 base = fold_convert (unsigned_type, high);
3144 delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme);
3145 step = fold_build1 (NEGATE_EXPR, unsigned_type, step);
3147 else
3149 wide_int min, max;
3150 extreme = fold_convert (unsigned_type, high);
3151 if (TREE_CODE (orig_base) == SSA_NAME
3152 && TREE_CODE (low) == INTEGER_CST
3153 && INTEGRAL_TYPE_P (TREE_TYPE (orig_base))
3154 && (get_range_info (orig_base, &min, &max) == VR_RANGE
3155 || get_cst_init_from_scev (orig_base, &min, true))
3156 && wi::gts_p (min, low))
3157 base = wide_int_to_tree (unsigned_type, min);
3158 else if (TREE_CODE (base) != INTEGER_CST
3159 && dominated_by_p (CDI_DOMINATORS,
3160 loop->latch, gimple_bb (stmt)))
3161 base = fold_convert (unsigned_type, low);
3162 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);
3165 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
3166 would get out of the range. */
3167 niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step);
3168 widest_int max = derive_constant_upper_bound (niter_bound);
3169 record_estimate (loop, niter_bound, max, stmt, false, realistic, upper);
3172 /* Determine information about number of iterations a LOOP from the index
3173 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
3174 guaranteed to be executed in every iteration of LOOP. Callback for
3175 for_each_index. */
3177 struct ilb_data
3179 struct loop *loop;
3180 gimple *stmt;
3183 static bool
3184 idx_infer_loop_bounds (tree base, tree *idx, void *dta)
3186 struct ilb_data *data = (struct ilb_data *) dta;
3187 tree ev, init, step;
3188 tree low, high, type, next;
3189 bool sign, upper = true, at_end = false;
3190 struct loop *loop = data->loop;
3192 if (TREE_CODE (base) != ARRAY_REF)
3193 return true;
3195 /* For arrays at the end of the structure, we are not guaranteed that they
3196 do not really extend over their declared size. However, for arrays of
3197 size greater than one, this is unlikely to be intended. */
3198 if (array_at_struct_end_p (base))
3200 at_end = true;
3201 upper = false;
3204 struct loop *dloop = loop_containing_stmt (data->stmt);
3205 if (!dloop)
3206 return true;
3208 ev = analyze_scalar_evolution (dloop, *idx);
3209 ev = instantiate_parameters (loop, ev);
3210 init = initial_condition (ev);
3211 step = evolution_part_in_loop_num (ev, loop->num);
3213 if (!init
3214 || !step
3215 || TREE_CODE (step) != INTEGER_CST
3216 || integer_zerop (step)
3217 || tree_contains_chrecs (init, NULL)
3218 || chrec_contains_symbols_defined_in_loop (init, loop->num))
3219 return true;
3221 low = array_ref_low_bound (base);
3222 high = array_ref_up_bound (base);
3224 /* The case of nonconstant bounds could be handled, but it would be
3225 complicated. */
3226 if (TREE_CODE (low) != INTEGER_CST
3227 || !high
3228 || TREE_CODE (high) != INTEGER_CST)
3229 return true;
3230 sign = tree_int_cst_sign_bit (step);
3231 type = TREE_TYPE (step);
3233 /* The array of length 1 at the end of a structure most likely extends
3234 beyond its bounds. */
3235 if (at_end
3236 && operand_equal_p (low, high, 0))
3237 return true;
3239 /* In case the relevant bound of the array does not fit in type, or
3240 it does, but bound + step (in type) still belongs into the range of the
3241 array, the index may wrap and still stay within the range of the array
3242 (consider e.g. if the array is indexed by the full range of
3243 unsigned char).
3245 To make things simpler, we require both bounds to fit into type, although
3246 there are cases where this would not be strictly necessary. */
3247 if (!int_fits_type_p (high, type)
3248 || !int_fits_type_p (low, type))
3249 return true;
3250 low = fold_convert (type, low);
3251 high = fold_convert (type, high);
3253 if (sign)
3254 next = fold_binary (PLUS_EXPR, type, low, step);
3255 else
3256 next = fold_binary (PLUS_EXPR, type, high, step);
3258 if (tree_int_cst_compare (low, next) <= 0
3259 && tree_int_cst_compare (next, high) <= 0)
3260 return true;
3262 /* If access is not executed on every iteration, we must ensure that overlow
3263 may not make the access valid later. */
3264 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt))
3265 && scev_probably_wraps_p (NULL_TREE,
3266 initial_condition_in_loop_num (ev, loop->num),
3267 step, data->stmt, loop, true))
3268 upper = false;
3270 record_nonwrapping_iv (loop, init, step, data->stmt, low, high, false, upper);
3271 return true;
3274 /* Determine information about number of iterations a LOOP from the bounds
3275 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
3276 STMT is guaranteed to be executed in every iteration of LOOP.*/
3278 static void
3279 infer_loop_bounds_from_ref (struct loop *loop, gimple *stmt, tree ref)
3281 struct ilb_data data;
3283 data.loop = loop;
3284 data.stmt = stmt;
3285 for_each_index (&ref, idx_infer_loop_bounds, &data);
3288 /* Determine information about number of iterations of a LOOP from the way
3289 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
3290 executed in every iteration of LOOP. */
3292 static void
3293 infer_loop_bounds_from_array (struct loop *loop, gimple *stmt)
3295 if (is_gimple_assign (stmt))
3297 tree op0 = gimple_assign_lhs (stmt);
3298 tree op1 = gimple_assign_rhs1 (stmt);
3300 /* For each memory access, analyze its access function
3301 and record a bound on the loop iteration domain. */
3302 if (REFERENCE_CLASS_P (op0))
3303 infer_loop_bounds_from_ref (loop, stmt, op0);
3305 if (REFERENCE_CLASS_P (op1))
3306 infer_loop_bounds_from_ref (loop, stmt, op1);
3308 else if (is_gimple_call (stmt))
3310 tree arg, lhs;
3311 unsigned i, n = gimple_call_num_args (stmt);
3313 lhs = gimple_call_lhs (stmt);
3314 if (lhs && REFERENCE_CLASS_P (lhs))
3315 infer_loop_bounds_from_ref (loop, stmt, lhs);
3317 for (i = 0; i < n; i++)
3319 arg = gimple_call_arg (stmt, i);
3320 if (REFERENCE_CLASS_P (arg))
3321 infer_loop_bounds_from_ref (loop, stmt, arg);
3326 /* Determine information about number of iterations of a LOOP from the fact
3327 that pointer arithmetics in STMT does not overflow. */
3329 static void
3330 infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple *stmt)
3332 tree def, base, step, scev, type, low, high;
3333 tree var, ptr;
3335 if (!is_gimple_assign (stmt)
3336 || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR)
3337 return;
3339 def = gimple_assign_lhs (stmt);
3340 if (TREE_CODE (def) != SSA_NAME)
3341 return;
3343 type = TREE_TYPE (def);
3344 if (!nowrap_type_p (type))
3345 return;
3347 ptr = gimple_assign_rhs1 (stmt);
3348 if (!expr_invariant_in_loop_p (loop, ptr))
3349 return;
3351 var = gimple_assign_rhs2 (stmt);
3352 if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var)))
3353 return;
3355 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
3356 if (chrec_contains_undetermined (scev))
3357 return;
3359 base = initial_condition_in_loop_num (scev, loop->num);
3360 step = evolution_part_in_loop_num (scev, loop->num);
3362 if (!base || !step
3363 || TREE_CODE (step) != INTEGER_CST
3364 || tree_contains_chrecs (base, NULL)
3365 || chrec_contains_symbols_defined_in_loop (base, loop->num))
3366 return;
3368 low = lower_bound_in_type (type, type);
3369 high = upper_bound_in_type (type, type);
3371 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot
3372 produce a NULL pointer. The contrary would mean NULL points to an object,
3373 while NULL is supposed to compare unequal with the address of all objects.
3374 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a
3375 NULL pointer since that would mean wrapping, which we assume here not to
3376 happen. So, we can exclude NULL from the valid range of pointer
3377 arithmetic. */
3378 if (flag_delete_null_pointer_checks && int_cst_value (low) == 0)
3379 low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type)));
3381 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
3384 /* Determine information about number of iterations of a LOOP from the fact
3385 that signed arithmetics in STMT does not overflow. */
3387 static void
3388 infer_loop_bounds_from_signedness (struct loop *loop, gimple *stmt)
3390 tree def, base, step, scev, type, low, high;
3392 if (gimple_code (stmt) != GIMPLE_ASSIGN)
3393 return;
3395 def = gimple_assign_lhs (stmt);
3397 if (TREE_CODE (def) != SSA_NAME)
3398 return;
3400 type = TREE_TYPE (def);
3401 if (!INTEGRAL_TYPE_P (type)
3402 || !TYPE_OVERFLOW_UNDEFINED (type))
3403 return;
3405 scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def));
3406 if (chrec_contains_undetermined (scev))
3407 return;
3409 base = initial_condition_in_loop_num (scev, loop->num);
3410 step = evolution_part_in_loop_num (scev, loop->num);
3412 if (!base || !step
3413 || TREE_CODE (step) != INTEGER_CST
3414 || tree_contains_chrecs (base, NULL)
3415 || chrec_contains_symbols_defined_in_loop (base, loop->num))
3416 return;
3418 low = lower_bound_in_type (type, type);
3419 high = upper_bound_in_type (type, type);
3421 record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true);
3424 /* The following analyzers are extracting informations on the bounds
3425 of LOOP from the following undefined behaviors:
3427 - data references should not access elements over the statically
3428 allocated size,
3430 - signed variables should not overflow when flag_wrapv is not set.
3433 static void
3434 infer_loop_bounds_from_undefined (struct loop *loop)
3436 unsigned i;
3437 basic_block *bbs;
3438 gimple_stmt_iterator bsi;
3439 basic_block bb;
3440 bool reliable;
3442 bbs = get_loop_body (loop);
3444 for (i = 0; i < loop->num_nodes; i++)
3446 bb = bbs[i];
3448 /* If BB is not executed in each iteration of the loop, we cannot
3449 use the operations in it to infer reliable upper bound on the
3450 # of iterations of the loop. However, we can use it as a guess.
3451 Reliable guesses come only from array bounds. */
3452 reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb);
3454 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
3456 gimple *stmt = gsi_stmt (bsi);
3458 infer_loop_bounds_from_array (loop, stmt);
3460 if (reliable)
3462 infer_loop_bounds_from_signedness (loop, stmt);
3463 infer_loop_bounds_from_pointer_arith (loop, stmt);
3469 free (bbs);
3472 /* Compare wide ints, callback for qsort. */
3474 static int
3475 wide_int_cmp (const void *p1, const void *p2)
3477 const widest_int *d1 = (const widest_int *) p1;
3478 const widest_int *d2 = (const widest_int *) p2;
3479 return wi::cmpu (*d1, *d2);
3482 /* Return index of BOUND in BOUNDS array sorted in increasing order.
3483 Lookup by binary search. */
3485 static int
3486 bound_index (vec<widest_int> bounds, const widest_int &bound)
3488 unsigned int end = bounds.length ();
3489 unsigned int begin = 0;
3491 /* Find a matching index by means of a binary search. */
3492 while (begin != end)
3494 unsigned int middle = (begin + end) / 2;
3495 widest_int index = bounds[middle];
3497 if (index == bound)
3498 return middle;
3499 else if (wi::ltu_p (index, bound))
3500 begin = middle + 1;
3501 else
3502 end = middle;
3504 gcc_unreachable ();
3507 /* We recorded loop bounds only for statements dominating loop latch (and thus
3508 executed each loop iteration). If there are any bounds on statements not
3509 dominating the loop latch we can improve the estimate by walking the loop
3510 body and seeing if every path from loop header to loop latch contains
3511 some bounded statement. */
3513 static void
3514 discover_iteration_bound_by_body_walk (struct loop *loop)
3516 struct nb_iter_bound *elt;
3517 auto_vec<widest_int> bounds;
3518 vec<vec<basic_block> > queues = vNULL;
3519 vec<basic_block> queue = vNULL;
3520 ptrdiff_t queue_index;
3521 ptrdiff_t latch_index = 0;
3523 /* Discover what bounds may interest us. */
3524 for (elt = loop->bounds; elt; elt = elt->next)
3526 widest_int bound = elt->bound;
3528 /* Exit terminates loop at given iteration, while non-exits produce undefined
3529 effect on the next iteration. */
3530 if (!elt->is_exit)
3532 bound += 1;
3533 /* If an overflow occurred, ignore the result. */
3534 if (bound == 0)
3535 continue;
3538 if (!loop->any_upper_bound
3539 || wi::ltu_p (bound, loop->nb_iterations_upper_bound))
3540 bounds.safe_push (bound);
3543 /* Exit early if there is nothing to do. */
3544 if (!bounds.exists ())
3545 return;
3547 if (dump_file && (dump_flags & TDF_DETAILS))
3548 fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n");
3550 /* Sort the bounds in decreasing order. */
3551 bounds.qsort (wide_int_cmp);
3553 /* For every basic block record the lowest bound that is guaranteed to
3554 terminate the loop. */
3556 hash_map<basic_block, ptrdiff_t> bb_bounds;
3557 for (elt = loop->bounds; elt; elt = elt->next)
3559 widest_int bound = elt->bound;
3560 if (!elt->is_exit)
3562 bound += 1;
3563 /* If an overflow occurred, ignore the result. */
3564 if (bound == 0)
3565 continue;
3568 if (!loop->any_upper_bound
3569 || wi::ltu_p (bound, loop->nb_iterations_upper_bound))
3571 ptrdiff_t index = bound_index (bounds, bound);
3572 ptrdiff_t *entry = bb_bounds.get (gimple_bb (elt->stmt));
3573 if (!entry)
3574 bb_bounds.put (gimple_bb (elt->stmt), index);
3575 else if ((ptrdiff_t)*entry > index)
3576 *entry = index;
3580 hash_map<basic_block, ptrdiff_t> block_priority;
3582 /* Perform shortest path discovery loop->header ... loop->latch.
3584 The "distance" is given by the smallest loop bound of basic block
3585 present in the path and we look for path with largest smallest bound
3586 on it.
3588 To avoid the need for fibonacci heap on double ints we simply compress
3589 double ints into indexes to BOUNDS array and then represent the queue
3590 as arrays of queues for every index.
3591 Index of BOUNDS.length() means that the execution of given BB has
3592 no bounds determined.
3594 VISITED is a pointer map translating basic block into smallest index
3595 it was inserted into the priority queue with. */
3596 latch_index = -1;
3598 /* Start walk in loop header with index set to infinite bound. */
3599 queue_index = bounds.length ();
3600 queues.safe_grow_cleared (queue_index + 1);
3601 queue.safe_push (loop->header);
3602 queues[queue_index] = queue;
3603 block_priority.put (loop->header, queue_index);
3605 for (; queue_index >= 0; queue_index--)
3607 if (latch_index < queue_index)
3609 while (queues[queue_index].length ())
3611 basic_block bb;
3612 ptrdiff_t bound_index = queue_index;
3613 edge e;
3614 edge_iterator ei;
3616 queue = queues[queue_index];
3617 bb = queue.pop ();
3619 /* OK, we later inserted the BB with lower priority, skip it. */
3620 if (*block_priority.get (bb) > queue_index)
3621 continue;
3623 /* See if we can improve the bound. */
3624 ptrdiff_t *entry = bb_bounds.get (bb);
3625 if (entry && *entry < bound_index)
3626 bound_index = *entry;
3628 /* Insert succesors into the queue, watch for latch edge
3629 and record greatest index we saw. */
3630 FOR_EACH_EDGE (e, ei, bb->succs)
3632 bool insert = false;
3634 if (loop_exit_edge_p (loop, e))
3635 continue;
3637 if (e == loop_latch_edge (loop)
3638 && latch_index < bound_index)
3639 latch_index = bound_index;
3640 else if (!(entry = block_priority.get (e->dest)))
3642 insert = true;
3643 block_priority.put (e->dest, bound_index);
3645 else if (*entry < bound_index)
3647 insert = true;
3648 *entry = bound_index;
3651 if (insert)
3652 queues[bound_index].safe_push (e->dest);
3656 queues[queue_index].release ();
3659 gcc_assert (latch_index >= 0);
3660 if ((unsigned)latch_index < bounds.length ())
3662 if (dump_file && (dump_flags & TDF_DETAILS))
3664 fprintf (dump_file, "Found better loop bound ");
3665 print_decu (bounds[latch_index], dump_file);
3666 fprintf (dump_file, "\n");
3668 record_niter_bound (loop, bounds[latch_index], false, true);
3671 queues.release ();
3674 /* See if every path cross the loop goes through a statement that is known
3675 to not execute at the last iteration. In that case we can decrese iteration
3676 count by 1. */
3678 static void
3679 maybe_lower_iteration_bound (struct loop *loop)
3681 hash_set<gimple *> *not_executed_last_iteration = NULL;
3682 struct nb_iter_bound *elt;
3683 bool found_exit = false;
3684 auto_vec<basic_block> queue;
3685 bitmap visited;
3687 /* Collect all statements with interesting (i.e. lower than
3688 nb_iterations_upper_bound) bound on them.
3690 TODO: Due to the way record_estimate choose estimates to store, the bounds
3691 will be always nb_iterations_upper_bound-1. We can change this to record
3692 also statements not dominating the loop latch and update the walk bellow
3693 to the shortest path algorthm. */
3694 for (elt = loop->bounds; elt; elt = elt->next)
3696 if (!elt->is_exit
3697 && wi::ltu_p (elt->bound, loop->nb_iterations_upper_bound))
3699 if (!not_executed_last_iteration)
3700 not_executed_last_iteration = new hash_set<gimple *>;
3701 not_executed_last_iteration->add (elt->stmt);
3704 if (!not_executed_last_iteration)
3705 return;
3707 /* Start DFS walk in the loop header and see if we can reach the
3708 loop latch or any of the exits (including statements with side
3709 effects that may terminate the loop otherwise) without visiting
3710 any of the statements known to have undefined effect on the last
3711 iteration. */
3712 queue.safe_push (loop->header);
3713 visited = BITMAP_ALLOC (NULL);
3714 bitmap_set_bit (visited, loop->header->index);
3715 found_exit = false;
3719 basic_block bb = queue.pop ();
3720 gimple_stmt_iterator gsi;
3721 bool stmt_found = false;
3723 /* Loop for possible exits and statements bounding the execution. */
3724 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
3726 gimple *stmt = gsi_stmt (gsi);
3727 if (not_executed_last_iteration->contains (stmt))
3729 stmt_found = true;
3730 break;
3732 if (gimple_has_side_effects (stmt))
3734 found_exit = true;
3735 break;
3738 if (found_exit)
3739 break;
3741 /* If no bounding statement is found, continue the walk. */
3742 if (!stmt_found)
3744 edge e;
3745 edge_iterator ei;
3747 FOR_EACH_EDGE (e, ei, bb->succs)
3749 if (loop_exit_edge_p (loop, e)
3750 || e == loop_latch_edge (loop))
3752 found_exit = true;
3753 break;
3755 if (bitmap_set_bit (visited, e->dest->index))
3756 queue.safe_push (e->dest);
3760 while (queue.length () && !found_exit);
3762 /* If every path through the loop reach bounding statement before exit,
3763 then we know the last iteration of the loop will have undefined effect
3764 and we can decrease number of iterations. */
3766 if (!found_exit)
3768 if (dump_file && (dump_flags & TDF_DETAILS))
3769 fprintf (dump_file, "Reducing loop iteration estimate by 1; "
3770 "undefined statement must be executed at the last iteration.\n");
3771 record_niter_bound (loop, loop->nb_iterations_upper_bound - 1,
3772 false, true);
3775 BITMAP_FREE (visited);
3776 delete not_executed_last_iteration;
3779 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P
3780 is true also use estimates derived from undefined behavior. */
3782 static void
3783 estimate_numbers_of_iterations_loop (struct loop *loop)
3785 vec<edge> exits;
3786 tree niter, type;
3787 unsigned i;
3788 struct tree_niter_desc niter_desc;
3789 edge ex;
3790 widest_int bound;
3791 edge likely_exit;
3793 /* Give up if we already have tried to compute an estimation. */
3794 if (loop->estimate_state != EST_NOT_COMPUTED)
3795 return;
3797 loop->estimate_state = EST_AVAILABLE;
3799 /* If we have a measured profile, use it to estimate the number of
3800 iterations. Normally this is recorded by branch_prob right after
3801 reading the profile. In case we however found a new loop, record the
3802 information here.
3804 Explicitly check for profile status so we do not report
3805 wrong prediction hitrates for guessed loop iterations heuristics.
3806 Do not recompute already recorded bounds - we ought to be better on
3807 updating iteration bounds than updating profile in general and thus
3808 recomputing iteration bounds later in the compilation process will just
3809 introduce random roundoff errors. */
3810 if (!loop->any_estimate
3811 && loop->header->count != 0
3812 && profile_status_for_fn (cfun) >= PROFILE_READ)
3814 gcov_type nit = expected_loop_iterations_unbounded (loop);
3815 bound = gcov_type_to_wide_int (nit);
3816 record_niter_bound (loop, bound, true, false);
3819 /* Ensure that loop->nb_iterations is computed if possible. If it turns out
3820 to be constant, we avoid undefined behavior implied bounds and instead
3821 diagnose those loops with -Waggressive-loop-optimizations. */
3822 number_of_latch_executions (loop);
3824 exits = get_loop_exit_edges (loop);
3825 likely_exit = single_likely_exit (loop);
3826 FOR_EACH_VEC_ELT (exits, i, ex)
3828 if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false))
3829 continue;
3831 niter = niter_desc.niter;
3832 type = TREE_TYPE (niter);
3833 if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST)
3834 niter = build3 (COND_EXPR, type, niter_desc.may_be_zero,
3835 build_int_cst (type, 0),
3836 niter);
3837 record_estimate (loop, niter, niter_desc.max,
3838 last_stmt (ex->src),
3839 true, ex == likely_exit, true);
3840 record_control_iv (loop, &niter_desc);
3842 exits.release ();
3844 if (flag_aggressive_loop_optimizations)
3845 infer_loop_bounds_from_undefined (loop);
3847 discover_iteration_bound_by_body_walk (loop);
3849 maybe_lower_iteration_bound (loop);
3851 /* If we know the exact number of iterations of this loop, try to
3852 not break code with undefined behavior by not recording smaller
3853 maximum number of iterations. */
3854 if (loop->nb_iterations
3855 && TREE_CODE (loop->nb_iterations) == INTEGER_CST)
3857 loop->any_upper_bound = true;
3858 loop->nb_iterations_upper_bound = wi::to_widest (loop->nb_iterations);
3862 /* Sets NIT to the estimated number of executions of the latch of the
3863 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
3864 large as the number of iterations. If we have no reliable estimate,
3865 the function returns false, otherwise returns true. */
3867 bool
3868 estimated_loop_iterations (struct loop *loop, widest_int *nit)
3870 /* When SCEV information is available, try to update loop iterations
3871 estimate. Otherwise just return whatever we recorded earlier. */
3872 if (scev_initialized_p ())
3873 estimate_numbers_of_iterations_loop (loop);
3875 return (get_estimated_loop_iterations (loop, nit));
3878 /* Similar to estimated_loop_iterations, but returns the estimate only
3879 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3880 on the number of iterations of LOOP could not be derived, returns -1. */
3882 HOST_WIDE_INT
3883 estimated_loop_iterations_int (struct loop *loop)
3885 widest_int nit;
3886 HOST_WIDE_INT hwi_nit;
3888 if (!estimated_loop_iterations (loop, &nit))
3889 return -1;
3891 if (!wi::fits_shwi_p (nit))
3892 return -1;
3893 hwi_nit = nit.to_shwi ();
3895 return hwi_nit < 0 ? -1 : hwi_nit;
3899 /* Sets NIT to an upper bound for the maximum number of executions of the
3900 latch of the LOOP. If we have no reliable estimate, the function returns
3901 false, otherwise returns true. */
3903 bool
3904 max_loop_iterations (struct loop *loop, widest_int *nit)
3906 /* When SCEV information is available, try to update loop iterations
3907 estimate. Otherwise just return whatever we recorded earlier. */
3908 if (scev_initialized_p ())
3909 estimate_numbers_of_iterations_loop (loop);
3911 return get_max_loop_iterations (loop, nit);
3914 /* Similar to max_loop_iterations, but returns the estimate only
3915 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3916 on the number of iterations of LOOP could not be derived, returns -1. */
3918 HOST_WIDE_INT
3919 max_loop_iterations_int (struct loop *loop)
3921 widest_int nit;
3922 HOST_WIDE_INT hwi_nit;
3924 if (!max_loop_iterations (loop, &nit))
3925 return -1;
3927 if (!wi::fits_shwi_p (nit))
3928 return -1;
3929 hwi_nit = nit.to_shwi ();
3931 return hwi_nit < 0 ? -1 : hwi_nit;
3934 /* Sets NIT to an likely upper bound for the maximum number of executions of the
3935 latch of the LOOP. If we have no reliable estimate, the function returns
3936 false, otherwise returns true. */
3938 bool
3939 likely_max_loop_iterations (struct loop *loop, widest_int *nit)
3941 /* When SCEV information is available, try to update loop iterations
3942 estimate. Otherwise just return whatever we recorded earlier. */
3943 if (scev_initialized_p ())
3944 estimate_numbers_of_iterations_loop (loop);
3946 return get_likely_max_loop_iterations (loop, nit);
3949 /* Similar to max_loop_iterations, but returns the estimate only
3950 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3951 on the number of iterations of LOOP could not be derived, returns -1. */
3953 HOST_WIDE_INT
3954 likely_max_loop_iterations_int (struct loop *loop)
3956 widest_int nit;
3957 HOST_WIDE_INT hwi_nit;
3959 if (!likely_max_loop_iterations (loop, &nit))
3960 return -1;
3962 if (!wi::fits_shwi_p (nit))
3963 return -1;
3964 hwi_nit = nit.to_shwi ();
3966 return hwi_nit < 0 ? -1 : hwi_nit;
3969 /* Returns an estimate for the number of executions of statements
3970 in the LOOP. For statements before the loop exit, this exceeds
3971 the number of execution of the latch by one. */
3973 HOST_WIDE_INT
3974 estimated_stmt_executions_int (struct loop *loop)
3976 HOST_WIDE_INT nit = estimated_loop_iterations_int (loop);
3977 HOST_WIDE_INT snit;
3979 if (nit == -1)
3980 return -1;
3982 snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1);
3984 /* If the computation overflows, return -1. */
3985 return snit < 0 ? -1 : snit;
3988 /* Sets NIT to the maximum number of executions of the latch of the
3989 LOOP, plus one. If we have no reliable estimate, the function returns
3990 false, otherwise returns true. */
3992 bool
3993 max_stmt_executions (struct loop *loop, widest_int *nit)
3995 widest_int nit_minus_one;
3997 if (!max_loop_iterations (loop, nit))
3998 return false;
4000 nit_minus_one = *nit;
4002 *nit += 1;
4004 return wi::gtu_p (*nit, nit_minus_one);
4007 /* Sets NIT to the estimated maximum number of executions of the latch of the
4008 LOOP, plus one. If we have no likely estimate, the function returns
4009 false, otherwise returns true. */
4011 bool
4012 likely_max_stmt_executions (struct loop *loop, widest_int *nit)
4014 widest_int nit_minus_one;
4016 if (!likely_max_loop_iterations (loop, nit))
4017 return false;
4019 nit_minus_one = *nit;
4021 *nit += 1;
4023 return wi::gtu_p (*nit, nit_minus_one);
4026 /* Sets NIT to the estimated number of executions of the latch of the
4027 LOOP, plus one. If we have no reliable estimate, the function returns
4028 false, otherwise returns true. */
4030 bool
4031 estimated_stmt_executions (struct loop *loop, widest_int *nit)
4033 widest_int nit_minus_one;
4035 if (!estimated_loop_iterations (loop, nit))
4036 return false;
4038 nit_minus_one = *nit;
4040 *nit += 1;
4042 return wi::gtu_p (*nit, nit_minus_one);
4045 /* Records estimates on numbers of iterations of loops. */
4047 void
4048 estimate_numbers_of_iterations (void)
4050 struct loop *loop;
4052 /* We don't want to issue signed overflow warnings while getting
4053 loop iteration estimates. */
4054 fold_defer_overflow_warnings ();
4056 FOR_EACH_LOOP (loop, 0)
4058 estimate_numbers_of_iterations_loop (loop);
4061 fold_undefer_and_ignore_overflow_warnings ();
4064 /* Returns true if statement S1 dominates statement S2. */
4066 bool
4067 stmt_dominates_stmt_p (gimple *s1, gimple *s2)
4069 basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
4071 if (!bb1
4072 || s1 == s2)
4073 return true;
4075 if (bb1 == bb2)
4077 gimple_stmt_iterator bsi;
4079 if (gimple_code (s2) == GIMPLE_PHI)
4080 return false;
4082 if (gimple_code (s1) == GIMPLE_PHI)
4083 return true;
4085 for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi))
4086 if (gsi_stmt (bsi) == s1)
4087 return true;
4089 return false;
4092 return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
4095 /* Returns true when we can prove that the number of executions of
4096 STMT in the loop is at most NITER, according to the bound on
4097 the number of executions of the statement NITER_BOUND->stmt recorded in
4098 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT.
4100 ??? This code can become quite a CPU hog - we can have many bounds,
4101 and large basic block forcing stmt_dominates_stmt_p to be queried
4102 many times on a large basic blocks, so the whole thing is O(n^2)
4103 for scev_probably_wraps_p invocation (that can be done n times).
4105 It would make more sense (and give better answers) to remember BB
4106 bounds computed by discover_iteration_bound_by_body_walk. */
4108 static bool
4109 n_of_executions_at_most (gimple *stmt,
4110 struct nb_iter_bound *niter_bound,
4111 tree niter)
4113 widest_int bound = niter_bound->bound;
4114 tree nit_type = TREE_TYPE (niter), e;
4115 enum tree_code cmp;
4117 gcc_assert (TYPE_UNSIGNED (nit_type));
4119 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
4120 the number of iterations is small. */
4121 if (!wi::fits_to_tree_p (bound, nit_type))
4122 return false;
4124 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
4125 times. This means that:
4127 -- if NITER_BOUND->is_exit is true, then everything after
4128 it at most NITER_BOUND->bound times.
4130 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
4131 is executed, then NITER_BOUND->stmt is executed as well in the same
4132 iteration then STMT is executed at most NITER_BOUND->bound + 1 times.
4134 If we can determine that NITER_BOUND->stmt is always executed
4135 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times.
4136 We conclude that if both statements belong to the same
4137 basic block and STMT is before NITER_BOUND->stmt and there are no
4138 statements with side effects in between. */
4140 if (niter_bound->is_exit)
4142 if (stmt == niter_bound->stmt
4143 || !stmt_dominates_stmt_p (niter_bound->stmt, stmt))
4144 return false;
4145 cmp = GE_EXPR;
4147 else
4149 if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt))
4151 gimple_stmt_iterator bsi;
4152 if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt)
4153 || gimple_code (stmt) == GIMPLE_PHI
4154 || gimple_code (niter_bound->stmt) == GIMPLE_PHI)
4155 return false;
4157 /* By stmt_dominates_stmt_p we already know that STMT appears
4158 before NITER_BOUND->STMT. Still need to test that the loop
4159 can not be terinated by a side effect in between. */
4160 for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt;
4161 gsi_next (&bsi))
4162 if (gimple_has_side_effects (gsi_stmt (bsi)))
4163 return false;
4164 bound += 1;
4165 if (bound == 0
4166 || !wi::fits_to_tree_p (bound, nit_type))
4167 return false;
4169 cmp = GT_EXPR;
4172 e = fold_binary (cmp, boolean_type_node,
4173 niter, wide_int_to_tree (nit_type, bound));
4174 return e && integer_nonzerop (e);
4177 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
4179 bool
4180 nowrap_type_p (tree type)
4182 if (ANY_INTEGRAL_TYPE_P (type)
4183 && TYPE_OVERFLOW_UNDEFINED (type))
4184 return true;
4186 if (POINTER_TYPE_P (type))
4187 return true;
4189 return false;
4192 /* Return true if we can prove LOOP is exited before evolution of induction
4193 variable {BASE, STEP} overflows with respect to its type bound. */
4195 static bool
4196 loop_exits_before_overflow (tree base, tree step,
4197 gimple *at_stmt, struct loop *loop)
4199 widest_int niter;
4200 struct control_iv *civ;
4201 struct nb_iter_bound *bound;
4202 tree e, delta, step_abs, unsigned_base;
4203 tree type = TREE_TYPE (step);
4204 tree unsigned_type, valid_niter;
4206 /* Don't issue signed overflow warnings. */
4207 fold_defer_overflow_warnings ();
4209 /* Compute the number of iterations before we reach the bound of the
4210 type, and verify that the loop is exited before this occurs. */
4211 unsigned_type = unsigned_type_for (type);
4212 unsigned_base = fold_convert (unsigned_type, base);
4214 if (tree_int_cst_sign_bit (step))
4216 tree extreme = fold_convert (unsigned_type,
4217 lower_bound_in_type (type, type));
4218 delta = fold_build2 (MINUS_EXPR, unsigned_type, unsigned_base, extreme);
4219 step_abs = fold_build1 (NEGATE_EXPR, unsigned_type,
4220 fold_convert (unsigned_type, step));
4222 else
4224 tree extreme = fold_convert (unsigned_type,
4225 upper_bound_in_type (type, type));
4226 delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, unsigned_base);
4227 step_abs = fold_convert (unsigned_type, step);
4230 valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs);
4232 estimate_numbers_of_iterations_loop (loop);
4234 if (max_loop_iterations (loop, &niter)
4235 && wi::fits_to_tree_p (niter, TREE_TYPE (valid_niter))
4236 && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter,
4237 wide_int_to_tree (TREE_TYPE (valid_niter),
4238 niter))) != NULL
4239 && integer_nonzerop (e))
4241 fold_undefer_and_ignore_overflow_warnings ();
4242 return true;
4244 if (at_stmt)
4245 for (bound = loop->bounds; bound; bound = bound->next)
4247 if (n_of_executions_at_most (at_stmt, bound, valid_niter))
4249 fold_undefer_and_ignore_overflow_warnings ();
4250 return true;
4253 fold_undefer_and_ignore_overflow_warnings ();
4255 /* Try to prove loop is exited before {base, step} overflows with the
4256 help of analyzed loop control IV. This is done only for IVs with
4257 constant step because otherwise we don't have the information. */
4258 if (TREE_CODE (step) == INTEGER_CST)
4260 for (civ = loop->control_ivs; civ; civ = civ->next)
4262 enum tree_code code;
4263 tree civ_type = TREE_TYPE (civ->step);
4265 /* Have to consider type difference because operand_equal_p ignores
4266 that for constants. */
4267 if (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (civ_type)
4268 || element_precision (type) != element_precision (civ_type))
4269 continue;
4271 /* Only consider control IV with same step. */
4272 if (!operand_equal_p (step, civ->step, 0))
4273 continue;
4275 /* Done proving if this is a no-overflow control IV. */
4276 if (operand_equal_p (base, civ->base, 0))
4277 return true;
4279 /* Control IV is recorded after expanding simple operations,
4280 Here we expand base and compare it too. */
4281 tree expanded_base = expand_simple_operations (base);
4282 if (operand_equal_p (expanded_base, civ->base, 0))
4283 return true;
4285 /* If this is a before stepping control IV, in other words, we have
4287 {civ_base, step} = {base + step, step}
4289 Because civ {base + step, step} doesn't overflow during loop
4290 iterations, {base, step} will not overflow if we can prove the
4291 operation "base + step" does not overflow. Specifically, we try
4292 to prove below conditions are satisfied:
4294 base <= UPPER_BOUND (type) - step ;;step > 0
4295 base >= LOWER_BOUND (type) - step ;;step < 0
4297 by proving the reverse conditions are false using loop's initial
4298 condition. */
4299 if (POINTER_TYPE_P (TREE_TYPE (base)))
4300 code = POINTER_PLUS_EXPR;
4301 else
4302 code = PLUS_EXPR;
4304 tree stepped = fold_build2 (code, TREE_TYPE (base), base, step);
4305 tree expanded_stepped = fold_build2 (code, TREE_TYPE (base),
4306 expanded_base, step);
4307 if (operand_equal_p (stepped, civ->base, 0)
4308 || operand_equal_p (expanded_stepped, civ->base, 0))
4310 tree extreme;
4312 if (tree_int_cst_sign_bit (step))
4314 code = LT_EXPR;
4315 extreme = lower_bound_in_type (type, type);
4317 else
4319 code = GT_EXPR;
4320 extreme = upper_bound_in_type (type, type);
4322 extreme = fold_build2 (MINUS_EXPR, type, extreme, step);
4323 e = fold_build2 (code, boolean_type_node, base, extreme);
4324 e = simplify_using_initial_conditions (loop, e);
4325 if (integer_zerop (e))
4326 return true;
4331 return false;
4334 /* VAR is scev variable whose evolution part is constant STEP, this function
4335 proves that VAR can't overflow by using value range info. If VAR's value
4336 range is [MIN, MAX], it can be proven by:
4337 MAX + step doesn't overflow ; if step > 0
4339 MIN + step doesn't underflow ; if step < 0.
4341 We can only do this if var is computed in every loop iteration, i.e, var's
4342 definition has to dominate loop latch. Consider below example:
4345 unsigned int i;
4347 <bb 3>:
4349 <bb 4>:
4350 # RANGE [0, 4294967294] NONZERO 65535
4351 # i_21 = PHI <0(3), i_18(9)>
4352 if (i_21 != 0)
4353 goto <bb 6>;
4354 else
4355 goto <bb 8>;
4357 <bb 6>:
4358 # RANGE [0, 65533] NONZERO 65535
4359 _6 = i_21 + 4294967295;
4360 # RANGE [0, 65533] NONZERO 65535
4361 _7 = (long unsigned int) _6;
4362 # RANGE [0, 524264] NONZERO 524280
4363 _8 = _7 * 8;
4364 # PT = nonlocal escaped
4365 _9 = a_14 + _8;
4366 *_9 = 0;
4368 <bb 8>:
4369 # RANGE [1, 65535] NONZERO 65535
4370 i_18 = i_21 + 1;
4371 if (i_18 >= 65535)
4372 goto <bb 10>;
4373 else
4374 goto <bb 9>;
4376 <bb 9>:
4377 goto <bb 4>;
4379 <bb 10>:
4380 return;
4383 VAR _6 doesn't overflow only with pre-condition (i_21 != 0), here we
4384 can't use _6 to prove no-overlfow for _7. In fact, var _7 takes value
4385 sequence (4294967295, 0, 1, ..., 65533) in loop life time, rather than
4386 (4294967295, 4294967296, ...). */
4388 static bool
4389 scev_var_range_cant_overflow (tree var, tree step, struct loop *loop)
4391 tree type;
4392 wide_int minv, maxv, diff, step_wi;
4393 enum value_range_type rtype;
4395 if (TREE_CODE (step) != INTEGER_CST || !INTEGRAL_TYPE_P (TREE_TYPE (var)))
4396 return false;
4398 /* Check if VAR evaluates in every loop iteration. It's not the case
4399 if VAR is default definition or does not dominate loop's latch. */
4400 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var));
4401 if (!def_bb || !dominated_by_p (CDI_DOMINATORS, loop->latch, def_bb))
4402 return false;
4404 rtype = get_range_info (var, &minv, &maxv);
4405 if (rtype != VR_RANGE)
4406 return false;
4408 /* VAR is a scev whose evolution part is STEP and value range info
4409 is [MIN, MAX], we can prove its no-overflowness by conditions:
4411 type_MAX - MAX >= step ; if step > 0
4412 MIN - type_MIN >= |step| ; if step < 0.
4414 Or VAR must take value outside of value range, which is not true. */
4415 step_wi = step;
4416 type = TREE_TYPE (var);
4417 if (tree_int_cst_sign_bit (step))
4419 diff = lower_bound_in_type (type, type);
4420 diff = minv - diff;
4421 step_wi = - step_wi;
4423 else
4425 diff = upper_bound_in_type (type, type);
4426 diff = diff - maxv;
4429 return (wi::geu_p (diff, step_wi));
4432 /* Return false only when the induction variable BASE + STEP * I is
4433 known to not overflow: i.e. when the number of iterations is small
4434 enough with respect to the step and initial condition in order to
4435 keep the evolution confined in TYPEs bounds. Return true when the
4436 iv is known to overflow or when the property is not computable.
4438 USE_OVERFLOW_SEMANTICS is true if this function should assume that
4439 the rules for overflow of the given language apply (e.g., that signed
4440 arithmetics in C does not overflow).
4442 If VAR is a ssa variable, this function also returns false if VAR can
4443 be proven not overflow with value range info. */
4445 bool
4446 scev_probably_wraps_p (tree var, tree base, tree step,
4447 gimple *at_stmt, struct loop *loop,
4448 bool use_overflow_semantics)
4450 /* FIXME: We really need something like
4451 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
4453 We used to test for the following situation that frequently appears
4454 during address arithmetics:
4456 D.1621_13 = (long unsigned intD.4) D.1620_12;
4457 D.1622_14 = D.1621_13 * 8;
4458 D.1623_15 = (doubleD.29 *) D.1622_14;
4460 And derived that the sequence corresponding to D_14
4461 can be proved to not wrap because it is used for computing a
4462 memory access; however, this is not really the case -- for example,
4463 if D_12 = (unsigned char) [254,+,1], then D_14 has values
4464 2032, 2040, 0, 8, ..., but the code is still legal. */
4466 if (chrec_contains_undetermined (base)
4467 || chrec_contains_undetermined (step))
4468 return true;
4470 if (integer_zerop (step))
4471 return false;
4473 /* If we can use the fact that signed and pointer arithmetics does not
4474 wrap, we are done. */
4475 if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base)))
4476 return false;
4478 /* To be able to use estimates on number of iterations of the loop,
4479 we must have an upper bound on the absolute value of the step. */
4480 if (TREE_CODE (step) != INTEGER_CST)
4481 return true;
4483 /* Check if var can be proven not overflow with value range info. */
4484 if (var && TREE_CODE (var) == SSA_NAME
4485 && scev_var_range_cant_overflow (var, step, loop))
4486 return false;
4488 if (loop_exits_before_overflow (base, step, at_stmt, loop))
4489 return false;
4491 /* At this point we still don't have a proof that the iv does not
4492 overflow: give up. */
4493 return true;
4496 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
4498 void
4499 free_numbers_of_iterations_estimates_loop (struct loop *loop)
4501 struct control_iv *civ;
4502 struct nb_iter_bound *bound;
4504 loop->nb_iterations = NULL;
4505 loop->estimate_state = EST_NOT_COMPUTED;
4506 for (bound = loop->bounds; bound;)
4508 struct nb_iter_bound *next = bound->next;
4509 ggc_free (bound);
4510 bound = next;
4512 loop->bounds = NULL;
4514 for (civ = loop->control_ivs; civ;)
4516 struct control_iv *next = civ->next;
4517 ggc_free (civ);
4518 civ = next;
4520 loop->control_ivs = NULL;
4523 /* Frees the information on upper bounds on numbers of iterations of loops. */
4525 void
4526 free_numbers_of_iterations_estimates (function *fn)
4528 struct loop *loop;
4530 FOR_EACH_LOOP_FN (fn, loop, 0)
4532 free_numbers_of_iterations_estimates_loop (loop);
4536 /* Substitute value VAL for ssa name NAME inside expressions held
4537 at LOOP. */
4539 void
4540 substitute_in_loop_info (struct loop *loop, tree name, tree val)
4542 loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val);