1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004-2021 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
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
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/>. */
22 #include "coretypes.h"
27 #include "tree-pass.h"
29 #include "gimple-pretty-print.h"
30 #include "diagnostic-core.h"
31 #include "stor-layout.h"
32 #include "fold-const.h"
36 #include "gimple-iterator.h"
38 #include "tree-ssa-loop-ivopts.h"
39 #include "tree-ssa-loop-niter.h"
40 #include "tree-ssa-loop.h"
42 #include "tree-chrec.h"
43 #include "tree-scalar-evolution.h"
45 #include "gimple-range.h"
48 /* The maximum number of dominator BBs we search for conditions
49 of loop header copies we use for simplifying a conditional
51 #define MAX_DOMINATORS_TO_WALK 8
55 Analysis of number of iterations of an affine exit test.
59 /* Bounds on some value, BELOW <= X <= UP. */
66 static bool number_of_iterations_popcount (loop_p loop
, edge exit
,
68 class tree_niter_desc
*niter
);
71 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
74 split_to_var_and_offset (tree expr
, tree
*var
, mpz_t offset
)
76 tree type
= TREE_TYPE (expr
);
81 mpz_set_ui (offset
, 0);
83 switch (TREE_CODE (expr
))
90 case POINTER_PLUS_EXPR
:
91 op0
= TREE_OPERAND (expr
, 0);
92 op1
= TREE_OPERAND (expr
, 1);
94 if (TREE_CODE (op1
) != INTEGER_CST
)
98 /* Always sign extend the offset. */
99 wi::to_mpz (wi::to_wide (op1
), offset
, SIGNED
);
101 mpz_neg (offset
, offset
);
105 *var
= build_int_cst_type (type
, 0);
106 wi::to_mpz (wi::to_wide (expr
), offset
, TYPE_SIGN (type
));
114 /* From condition C0 CMP C1 derives information regarding the value range
115 of VAR, which is of TYPE. Results are stored in to BELOW and UP. */
118 refine_value_range_using_guard (tree type
, tree var
,
119 tree c0
, enum tree_code cmp
, tree c1
,
120 mpz_t below
, mpz_t up
)
122 tree varc0
, varc1
, ctype
;
124 mpz_t mint
, maxt
, minc1
, maxc1
;
125 bool no_wrap
= nowrap_type_p (type
);
127 signop sgn
= TYPE_SIGN (type
);
135 STRIP_SIGN_NOPS (c0
);
136 STRIP_SIGN_NOPS (c1
);
137 ctype
= TREE_TYPE (c0
);
138 if (!useless_type_conversion_p (ctype
, type
))
144 /* We could derive quite precise information from EQ_EXPR, however,
145 such a guard is unlikely to appear, so we do not bother with
150 /* NE_EXPR comparisons do not contain much of useful information,
151 except for cases of comparing with bounds. */
152 if (TREE_CODE (c1
) != INTEGER_CST
153 || !INTEGRAL_TYPE_P (type
))
156 /* Ensure that the condition speaks about an expression in the same
158 ctype
= TREE_TYPE (c0
);
159 if (TYPE_PRECISION (ctype
) != TYPE_PRECISION (type
))
161 c0
= fold_convert (type
, c0
);
162 c1
= fold_convert (type
, c1
);
164 if (operand_equal_p (var
, c0
, 0))
168 /* Case of comparing VAR with its below/up bounds. */
170 wi::to_mpz (wi::to_wide (c1
), valc1
, TYPE_SIGN (type
));
171 if (mpz_cmp (valc1
, below
) == 0)
173 if (mpz_cmp (valc1
, up
) == 0)
180 /* Case of comparing with the bounds of the type. */
181 wide_int min
= wi::min_value (type
);
182 wide_int max
= wi::max_value (type
);
184 if (wi::to_wide (c1
) == min
)
186 if (wi::to_wide (c1
) == max
)
190 /* Quick return if no useful information. */
202 split_to_var_and_offset (expand_simple_operations (c0
), &varc0
, offc0
);
203 split_to_var_and_offset (expand_simple_operations (c1
), &varc1
, offc1
);
205 /* We are only interested in comparisons of expressions based on VAR. */
206 if (operand_equal_p (var
, varc1
, 0))
208 std::swap (varc0
, varc1
);
209 mpz_swap (offc0
, offc1
);
210 cmp
= swap_tree_comparison (cmp
);
212 else if (!operand_equal_p (var
, varc0
, 0))
221 get_type_static_bounds (type
, mint
, maxt
);
225 /* Setup range information for varc1. */
226 if (integer_zerop (varc1
))
228 wi::to_mpz (0, minc1
, TYPE_SIGN (type
));
229 wi::to_mpz (0, maxc1
, TYPE_SIGN (type
));
231 else if (TREE_CODE (varc1
) == SSA_NAME
232 && INTEGRAL_TYPE_P (type
)
233 && get_range_query (cfun
)->range_of_expr (r
, varc1
)
234 && r
.kind () == VR_RANGE
)
236 gcc_assert (wi::le_p (r
.lower_bound (), r
.upper_bound (), sgn
));
237 wi::to_mpz (r
.lower_bound (), minc1
, sgn
);
238 wi::to_mpz (r
.upper_bound (), maxc1
, sgn
);
242 mpz_set (minc1
, mint
);
243 mpz_set (maxc1
, maxt
);
246 /* Compute valid range information for varc1 + offc1. Note nothing
247 useful can be derived if it overflows or underflows. Overflow or
248 underflow could happen when:
250 offc1 > 0 && varc1 + offc1 > MAX_VAL (type)
251 offc1 < 0 && varc1 + offc1 < MIN_VAL (type). */
252 mpz_add (minc1
, minc1
, offc1
);
253 mpz_add (maxc1
, maxc1
, offc1
);
255 || mpz_sgn (offc1
) == 0
256 || (mpz_sgn (offc1
) < 0 && mpz_cmp (minc1
, mint
) >= 0)
257 || (mpz_sgn (offc1
) > 0 && mpz_cmp (maxc1
, maxt
) <= 0));
261 if (mpz_cmp (minc1
, mint
) < 0)
262 mpz_set (minc1
, mint
);
263 if (mpz_cmp (maxc1
, maxt
) > 0)
264 mpz_set (maxc1
, maxt
);
269 mpz_sub_ui (maxc1
, maxc1
, 1);
274 mpz_add_ui (minc1
, minc1
, 1);
277 /* Compute range information for varc0. If there is no overflow,
278 the condition implied that
280 (varc0) cmp (varc1 + offc1 - offc0)
282 We can possibly improve the upper bound of varc0 if cmp is LE_EXPR,
283 or the below bound if cmp is GE_EXPR.
285 To prove there is no overflow/underflow, we need to check below
287 1) cmp == LE_EXPR && offc0 > 0
289 (varc0 + offc0) doesn't overflow
290 && (varc1 + offc1 - offc0) doesn't underflow
292 2) cmp == LE_EXPR && offc0 < 0
294 (varc0 + offc0) doesn't underflow
295 && (varc1 + offc1 - offc0) doesn't overfloe
297 In this case, (varc0 + offc0) will never underflow if we can
298 prove (varc1 + offc1 - offc0) doesn't overflow.
300 3) cmp == GE_EXPR && offc0 < 0
302 (varc0 + offc0) doesn't underflow
303 && (varc1 + offc1 - offc0) doesn't overflow
305 4) cmp == GE_EXPR && offc0 > 0
307 (varc0 + offc0) doesn't overflow
308 && (varc1 + offc1 - offc0) doesn't underflow
310 In this case, (varc0 + offc0) will never overflow if we can
311 prove (varc1 + offc1 - offc0) doesn't underflow.
313 Note we only handle case 2 and 4 in below code. */
315 mpz_sub (minc1
, minc1
, offc0
);
316 mpz_sub (maxc1
, maxc1
, offc0
);
318 || mpz_sgn (offc0
) == 0
320 && mpz_sgn (offc0
) < 0 && mpz_cmp (maxc1
, maxt
) <= 0)
322 && mpz_sgn (offc0
) > 0 && mpz_cmp (minc1
, mint
) >= 0));
328 if (mpz_cmp (up
, maxc1
) > 0)
333 if (mpz_cmp (below
, minc1
) < 0)
334 mpz_set (below
, minc1
);
346 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
347 in TYPE to MIN and MAX. */
350 determine_value_range (class loop
*loop
, tree type
, tree var
, mpz_t off
,
351 mpz_t min
, mpz_t max
)
357 enum value_range_kind rtype
= VR_VARYING
;
359 /* If the expression is a constant, we know its value exactly. */
360 if (integer_zerop (var
))
367 get_type_static_bounds (type
, min
, max
);
369 /* See if we have some range info from VRP. */
370 if (TREE_CODE (var
) == SSA_NAME
&& INTEGRAL_TYPE_P (type
))
372 edge e
= loop_preheader_edge (loop
);
373 signop sgn
= TYPE_SIGN (type
);
376 /* Either for VAR itself... */
377 value_range var_range
;
378 get_range_query (cfun
)->range_of_expr (var_range
, var
);
379 rtype
= var_range
.kind ();
380 if (!var_range
.undefined_p ())
382 minv
= var_range
.lower_bound ();
383 maxv
= var_range
.upper_bound ();
386 /* Or for PHI results in loop->header where VAR is used as
387 PHI argument from the loop preheader edge. */
388 for (gsi
= gsi_start_phis (loop
->header
); !gsi_end_p (gsi
); gsi_next (&gsi
))
390 gphi
*phi
= gsi
.phi ();
391 value_range phi_range
;
392 if (PHI_ARG_DEF_FROM_EDGE (phi
, e
) == var
393 && get_range_query (cfun
)->range_of_expr (phi_range
,
394 gimple_phi_result (phi
))
395 && phi_range
.kind () == VR_RANGE
)
397 if (rtype
!= VR_RANGE
)
400 minv
= phi_range
.lower_bound ();
401 maxv
= phi_range
.upper_bound ();
405 minv
= wi::max (minv
, phi_range
.lower_bound (), sgn
);
406 maxv
= wi::min (maxv
, phi_range
.upper_bound (), sgn
);
407 /* If the PHI result range are inconsistent with
408 the VAR range, give up on looking at the PHI
409 results. This can happen if VR_UNDEFINED is
411 if (wi::gt_p (minv
, maxv
, sgn
))
414 get_range_query (cfun
)->range_of_expr (vr
, var
);
416 if (!vr
.undefined_p ())
418 minv
= vr
.lower_bound ();
419 maxv
= vr
.upper_bound ();
428 if (rtype
!= VR_RANGE
)
435 gcc_assert (wi::le_p (minv
, maxv
, sgn
));
436 wi::to_mpz (minv
, minm
, sgn
);
437 wi::to_mpz (maxv
, maxm
, sgn
);
439 /* Now walk the dominators of the loop header and use the entry
440 guards to refine the estimates. */
441 for (bb
= loop
->header
;
442 bb
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
) && cnt
< MAX_DOMINATORS_TO_WALK
;
443 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
450 if (!single_pred_p (bb
))
452 e
= single_pred_edge (bb
);
454 if (!(e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)))
457 cond
= last_stmt (e
->src
);
458 c0
= gimple_cond_lhs (cond
);
459 cmp
= gimple_cond_code (cond
);
460 c1
= gimple_cond_rhs (cond
);
462 if (e
->flags
& EDGE_FALSE_VALUE
)
463 cmp
= invert_tree_comparison (cmp
, false);
465 refine_value_range_using_guard (type
, var
, c0
, cmp
, c1
, minm
, maxm
);
469 mpz_add (minm
, minm
, off
);
470 mpz_add (maxm
, maxm
, off
);
471 /* If the computation may not wrap or off is zero, then this
472 is always fine. If off is negative and minv + off isn't
473 smaller than type's minimum, or off is positive and
474 maxv + off isn't bigger than type's maximum, use the more
475 precise range too. */
476 if (nowrap_type_p (type
)
477 || mpz_sgn (off
) == 0
478 || (mpz_sgn (off
) < 0 && mpz_cmp (minm
, min
) >= 0)
479 || (mpz_sgn (off
) > 0 && mpz_cmp (maxm
, max
) <= 0))
491 /* If the computation may wrap, we know nothing about the value, except for
492 the range of the type. */
493 if (!nowrap_type_p (type
))
496 /* Since the addition of OFF does not wrap, if OFF is positive, then we may
497 add it to MIN, otherwise to MAX. */
498 if (mpz_sgn (off
) < 0)
499 mpz_add (max
, max
, off
);
501 mpz_add (min
, min
, off
);
504 /* Stores the bounds on the difference of the values of the expressions
505 (var + X) and (var + Y), computed in TYPE, to BNDS. */
508 bound_difference_of_offsetted_base (tree type
, mpz_t x
, mpz_t y
,
511 int rel
= mpz_cmp (x
, y
);
512 bool may_wrap
= !nowrap_type_p (type
);
515 /* If X == Y, then the expressions are always equal.
516 If X > Y, there are the following possibilities:
517 a) neither of var + X and var + Y overflow or underflow, or both of
518 them do. Then their difference is X - Y.
519 b) var + X overflows, and var + Y does not. Then the values of the
520 expressions are var + X - M and var + Y, where M is the range of
521 the type, and their difference is X - Y - M.
522 c) var + Y underflows and var + X does not. Their difference again
524 Therefore, if the arithmetics in type does not overflow, then the
525 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y)
526 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or
527 (X - Y, X - Y + M). */
531 mpz_set_ui (bnds
->below
, 0);
532 mpz_set_ui (bnds
->up
, 0);
537 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type
)), m
, UNSIGNED
);
538 mpz_add_ui (m
, m
, 1);
539 mpz_sub (bnds
->up
, x
, y
);
540 mpz_set (bnds
->below
, bnds
->up
);
545 mpz_sub (bnds
->below
, bnds
->below
, m
);
547 mpz_add (bnds
->up
, bnds
->up
, m
);
553 /* From condition C0 CMP C1 derives information regarding the
554 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE,
555 and stores it to BNDS. */
558 refine_bounds_using_guard (tree type
, tree varx
, mpz_t offx
,
559 tree vary
, mpz_t offy
,
560 tree c0
, enum tree_code cmp
, tree c1
,
563 tree varc0
, varc1
, ctype
;
564 mpz_t offc0
, offc1
, loffx
, loffy
, bnd
;
566 bool no_wrap
= nowrap_type_p (type
);
575 STRIP_SIGN_NOPS (c0
);
576 STRIP_SIGN_NOPS (c1
);
577 ctype
= TREE_TYPE (c0
);
578 if (!useless_type_conversion_p (ctype
, type
))
584 /* We could derive quite precise information from EQ_EXPR, however, such
585 a guard is unlikely to appear, so we do not bother with handling
590 /* NE_EXPR comparisons do not contain much of useful information, except for
591 special case of comparing with the bounds of the type. */
592 if (TREE_CODE (c1
) != INTEGER_CST
593 || !INTEGRAL_TYPE_P (type
))
596 /* Ensure that the condition speaks about an expression in the same type
598 ctype
= TREE_TYPE (c0
);
599 if (TYPE_PRECISION (ctype
) != TYPE_PRECISION (type
))
601 c0
= fold_convert (type
, c0
);
602 c1
= fold_convert (type
, c1
);
604 if (TYPE_MIN_VALUE (type
)
605 && operand_equal_p (c1
, TYPE_MIN_VALUE (type
), 0))
610 if (TYPE_MAX_VALUE (type
)
611 && operand_equal_p (c1
, TYPE_MAX_VALUE (type
), 0))
624 split_to_var_and_offset (expand_simple_operations (c0
), &varc0
, offc0
);
625 split_to_var_and_offset (expand_simple_operations (c1
), &varc1
, offc1
);
627 /* We are only interested in comparisons of expressions based on VARX and
628 VARY. TODO -- we might also be able to derive some bounds from
629 expressions containing just one of the variables. */
631 if (operand_equal_p (varx
, varc1
, 0))
633 std::swap (varc0
, varc1
);
634 mpz_swap (offc0
, offc1
);
635 cmp
= swap_tree_comparison (cmp
);
638 if (!operand_equal_p (varx
, varc0
, 0)
639 || !operand_equal_p (vary
, varc1
, 0))
642 mpz_init_set (loffx
, offx
);
643 mpz_init_set (loffy
, offy
);
645 if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
647 std::swap (varx
, vary
);
648 mpz_swap (offc0
, offc1
);
649 mpz_swap (loffx
, loffy
);
650 cmp
= swap_tree_comparison (cmp
);
654 /* If there is no overflow, the condition implies that
656 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
658 The overflows and underflows may complicate things a bit; each
659 overflow decreases the appropriate offset by M, and underflow
660 increases it by M. The above inequality would not necessarily be
663 -- VARX + OFFX underflows and VARX + OFFC0 does not, or
664 VARX + OFFC0 overflows, but VARX + OFFX does not.
665 This may only happen if OFFX < OFFC0.
666 -- VARY + OFFY overflows and VARY + OFFC1 does not, or
667 VARY + OFFC1 underflows and VARY + OFFY does not.
668 This may only happen if OFFY > OFFC1. */
677 x_ok
= (integer_zerop (varx
)
678 || mpz_cmp (loffx
, offc0
) >= 0);
679 y_ok
= (integer_zerop (vary
)
680 || mpz_cmp (loffy
, offc1
) <= 0);
686 mpz_sub (bnd
, loffx
, loffy
);
687 mpz_add (bnd
, bnd
, offc1
);
688 mpz_sub (bnd
, bnd
, offc0
);
691 mpz_sub_ui (bnd
, bnd
, 1);
696 if (mpz_cmp (bnds
->below
, bnd
) < 0)
697 mpz_set (bnds
->below
, bnd
);
701 if (mpz_cmp (bnd
, bnds
->up
) < 0)
702 mpz_set (bnds
->up
, bnd
);
714 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS.
715 The subtraction is considered to be performed in arbitrary precision,
718 We do not attempt to be too clever regarding the value ranges of X and
719 Y; most of the time, they are just integers or ssa names offsetted by
720 integer. However, we try to use the information contained in the
721 comparisons before the loop (usually created by loop header copying). */
724 bound_difference (class loop
*loop
, tree x
, tree y
, bounds
*bnds
)
726 tree type
= TREE_TYPE (x
);
729 mpz_t minx
, maxx
, miny
, maxy
;
737 /* Get rid of unnecessary casts, but preserve the value of
742 mpz_init (bnds
->below
);
746 split_to_var_and_offset (x
, &varx
, offx
);
747 split_to_var_and_offset (y
, &vary
, offy
);
749 if (!integer_zerop (varx
)
750 && operand_equal_p (varx
, vary
, 0))
752 /* Special case VARX == VARY -- we just need to compare the
753 offsets. The matters are a bit more complicated in the
754 case addition of offsets may wrap. */
755 bound_difference_of_offsetted_base (type
, offx
, offy
, bnds
);
759 /* Otherwise, use the value ranges to determine the initial
760 estimates on below and up. */
765 determine_value_range (loop
, type
, varx
, offx
, minx
, maxx
);
766 determine_value_range (loop
, type
, vary
, offy
, miny
, maxy
);
768 mpz_sub (bnds
->below
, minx
, maxy
);
769 mpz_sub (bnds
->up
, maxx
, miny
);
776 /* If both X and Y are constants, we cannot get any more precise. */
777 if (integer_zerop (varx
) && integer_zerop (vary
))
780 /* Now walk the dominators of the loop header and use the entry
781 guards to refine the estimates. */
782 for (bb
= loop
->header
;
783 bb
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
) && cnt
< MAX_DOMINATORS_TO_WALK
;
784 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
786 if (!single_pred_p (bb
))
788 e
= single_pred_edge (bb
);
790 if (!(e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)))
793 cond
= last_stmt (e
->src
);
794 c0
= gimple_cond_lhs (cond
);
795 cmp
= gimple_cond_code (cond
);
796 c1
= gimple_cond_rhs (cond
);
798 if (e
->flags
& EDGE_FALSE_VALUE
)
799 cmp
= invert_tree_comparison (cmp
, false);
801 refine_bounds_using_guard (type
, varx
, offx
, vary
, offy
,
811 /* Update the bounds in BNDS that restrict the value of X to the bounds
812 that restrict the value of X + DELTA. X can be obtained as a
813 difference of two values in TYPE. */
816 bounds_add (bounds
*bnds
, const widest_int
&delta
, tree type
)
821 wi::to_mpz (delta
, mdelta
, SIGNED
);
824 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type
)), max
, UNSIGNED
);
826 mpz_add (bnds
->up
, bnds
->up
, mdelta
);
827 mpz_add (bnds
->below
, bnds
->below
, mdelta
);
829 if (mpz_cmp (bnds
->up
, max
) > 0)
830 mpz_set (bnds
->up
, max
);
833 if (mpz_cmp (bnds
->below
, max
) < 0)
834 mpz_set (bnds
->below
, max
);
840 /* Update the bounds in BNDS that restrict the value of X to the bounds
841 that restrict the value of -X. */
844 bounds_negate (bounds
*bnds
)
848 mpz_init_set (tmp
, bnds
->up
);
849 mpz_neg (bnds
->up
, bnds
->below
);
850 mpz_neg (bnds
->below
, tmp
);
854 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
857 inverse (tree x
, tree mask
)
859 tree type
= TREE_TYPE (x
);
861 unsigned ctr
= tree_floor_log2 (mask
);
863 if (TYPE_PRECISION (type
) <= HOST_BITS_PER_WIDE_INT
)
865 unsigned HOST_WIDE_INT ix
;
866 unsigned HOST_WIDE_INT imask
;
867 unsigned HOST_WIDE_INT irslt
= 1;
869 gcc_assert (cst_and_fits_in_hwi (x
));
870 gcc_assert (cst_and_fits_in_hwi (mask
));
872 ix
= int_cst_value (x
);
873 imask
= int_cst_value (mask
);
882 rslt
= build_int_cst_type (type
, irslt
);
886 rslt
= build_int_cst (type
, 1);
889 rslt
= int_const_binop (MULT_EXPR
, rslt
, x
);
890 x
= int_const_binop (MULT_EXPR
, x
, x
);
892 rslt
= int_const_binop (BIT_AND_EXPR
, rslt
, mask
);
898 /* Derives the upper bound BND on the number of executions of loop with exit
899 condition S * i <> C. If NO_OVERFLOW is true, then the control variable of
900 the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed
901 that the loop ends through this exit, i.e., the induction variable ever
902 reaches the value of C.
904 The value C is equal to final - base, where final and base are the final and
905 initial value of the actual induction variable in the analysed loop. BNDS
906 bounds the value of this difference when computed in signed type with
907 unbounded range, while the computation of C is performed in an unsigned
908 type with the range matching the range of the type of the induction variable.
909 In particular, BNDS.up contains an upper bound on C in the following cases:
910 -- if the iv must reach its final value without overflow, i.e., if
911 NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or
912 -- if final >= base, which we know to hold when BNDS.below >= 0. */
915 number_of_iterations_ne_max (mpz_t bnd
, bool no_overflow
, tree c
, tree s
,
916 bounds
*bnds
, bool exit_must_be_taken
)
920 tree type
= TREE_TYPE (c
);
921 bool bnds_u_valid
= ((no_overflow
&& exit_must_be_taken
)
922 || mpz_sgn (bnds
->below
) >= 0);
925 || (TREE_CODE (c
) == INTEGER_CST
926 && TREE_CODE (s
) == INTEGER_CST
927 && wi::mod_trunc (wi::to_wide (c
), wi::to_wide (s
),
928 TYPE_SIGN (type
)) == 0)
929 || (TYPE_OVERFLOW_UNDEFINED (type
)
930 && multiple_of_p (type
, c
, s
)))
932 /* If C is an exact multiple of S, then its value will be reached before
933 the induction variable overflows (unless the loop is exited in some
934 other way before). Note that the actual induction variable in the
935 loop (which ranges from base to final instead of from 0 to C) may
936 overflow, in which case BNDS.up will not be giving a correct upper
937 bound on C; thus, BNDS_U_VALID had to be computed in advance. */
939 exit_must_be_taken
= true;
942 /* If the induction variable can overflow, the number of iterations is at
943 most the period of the control variable (or infinite, but in that case
944 the whole # of iterations analysis will fail). */
947 max
= wi::mask
<widest_int
> (TYPE_PRECISION (type
)
948 - wi::ctz (wi::to_wide (s
)), false);
949 wi::to_mpz (max
, bnd
, UNSIGNED
);
953 /* Now we know that the induction variable does not overflow, so the loop
954 iterates at most (range of type / S) times. */
955 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type
)), bnd
, UNSIGNED
);
957 /* If the induction variable is guaranteed to reach the value of C before
959 if (exit_must_be_taken
)
961 /* ... then we can strengthen this to C / S, and possibly we can use
962 the upper bound on C given by BNDS. */
963 if (TREE_CODE (c
) == INTEGER_CST
)
964 wi::to_mpz (wi::to_wide (c
), bnd
, UNSIGNED
);
965 else if (bnds_u_valid
)
966 mpz_set (bnd
, bnds
->up
);
970 wi::to_mpz (wi::to_wide (s
), d
, UNSIGNED
);
971 mpz_fdiv_q (bnd
, bnd
, d
);
975 /* Determines number of iterations of loop whose ending condition
976 is IV <> FINAL. TYPE is the type of the iv. The number of
977 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
978 we know that the exit must be taken eventually, i.e., that the IV
979 ever reaches the value FINAL (we derived this earlier, and possibly set
980 NITER->assumptions to make sure this is the case). BNDS contains the
981 bounds on the difference FINAL - IV->base. */
984 number_of_iterations_ne (class loop
*loop
, tree type
, affine_iv
*iv
,
985 tree final
, class tree_niter_desc
*niter
,
986 bool exit_must_be_taken
, bounds
*bnds
)
988 tree niter_type
= unsigned_type_for (type
);
989 tree s
, c
, d
, bits
, assumption
, tmp
, bound
;
992 niter
->control
= *iv
;
993 niter
->bound
= final
;
994 niter
->cmp
= NE_EXPR
;
996 /* Rearrange the terms so that we get inequality S * i <> C, with S
997 positive. Also cast everything to the unsigned type. If IV does
998 not overflow, BNDS bounds the value of C. Also, this is the
999 case if the computation |FINAL - IV->base| does not overflow, i.e.,
1000 if BNDS->below in the result is nonnegative. */
1001 if (tree_int_cst_sign_bit (iv
->step
))
1003 s
= fold_convert (niter_type
,
1004 fold_build1 (NEGATE_EXPR
, type
, iv
->step
));
1005 c
= fold_build2 (MINUS_EXPR
, niter_type
,
1006 fold_convert (niter_type
, iv
->base
),
1007 fold_convert (niter_type
, final
));
1008 bounds_negate (bnds
);
1012 s
= fold_convert (niter_type
, iv
->step
);
1013 c
= fold_build2 (MINUS_EXPR
, niter_type
,
1014 fold_convert (niter_type
, final
),
1015 fold_convert (niter_type
, iv
->base
));
1019 number_of_iterations_ne_max (max
, iv
->no_overflow
, c
, s
, bnds
,
1020 exit_must_be_taken
);
1021 niter
->max
= widest_int::from (wi::from_mpz (niter_type
, max
, false),
1022 TYPE_SIGN (niter_type
));
1025 /* Compute no-overflow information for the control iv. This can be
1026 proven when below two conditions are satisfied:
1028 1) IV evaluates toward FINAL at beginning, i.e:
1029 base <= FINAL ; step > 0
1030 base >= FINAL ; step < 0
1032 2) |FINAL - base| is an exact multiple of step.
1034 Unfortunately, it's hard to prove above conditions after pass loop-ch
1035 because loop with exit condition (IV != FINAL) usually will be guarded
1036 by initial-condition (IV.base - IV.step != FINAL). In this case, we
1037 can alternatively try to prove below conditions:
1039 1') IV evaluates toward FINAL at beginning, i.e:
1040 new_base = base - step < FINAL ; step > 0
1041 && base - step doesn't underflow
1042 new_base = base - step > FINAL ; step < 0
1043 && base - step doesn't overflow
1045 2') |FINAL - new_base| is an exact multiple of step.
1047 Please refer to PR34114 as an example of loop-ch's impact, also refer
1048 to PR72817 as an example why condition 2') is necessary.
1050 Note, for NE_EXPR, base equals to FINAL is a special case, in
1051 which the loop exits immediately, and the iv does not overflow. */
1052 if (!niter
->control
.no_overflow
1053 && (integer_onep (s
) || multiple_of_p (type
, c
, s
)))
1055 tree t
, cond
, new_c
, relaxed_cond
= boolean_false_node
;
1057 if (tree_int_cst_sign_bit (iv
->step
))
1059 cond
= fold_build2 (GE_EXPR
, boolean_type_node
, iv
->base
, final
);
1060 if (TREE_CODE (type
) == INTEGER_TYPE
)
1062 /* Only when base - step doesn't overflow. */
1063 t
= TYPE_MAX_VALUE (type
);
1064 t
= fold_build2 (PLUS_EXPR
, type
, t
, iv
->step
);
1065 t
= fold_build2 (GE_EXPR
, boolean_type_node
, t
, iv
->base
);
1066 if (integer_nonzerop (t
))
1068 t
= fold_build2 (MINUS_EXPR
, type
, iv
->base
, iv
->step
);
1069 new_c
= fold_build2 (MINUS_EXPR
, niter_type
,
1070 fold_convert (niter_type
, t
),
1071 fold_convert (niter_type
, final
));
1072 if (multiple_of_p (type
, new_c
, s
))
1073 relaxed_cond
= fold_build2 (GT_EXPR
, boolean_type_node
,
1080 cond
= fold_build2 (LE_EXPR
, boolean_type_node
, iv
->base
, final
);
1081 if (TREE_CODE (type
) == INTEGER_TYPE
)
1083 /* Only when base - step doesn't underflow. */
1084 t
= TYPE_MIN_VALUE (type
);
1085 t
= fold_build2 (PLUS_EXPR
, type
, t
, iv
->step
);
1086 t
= fold_build2 (LE_EXPR
, boolean_type_node
, t
, iv
->base
);
1087 if (integer_nonzerop (t
))
1089 t
= fold_build2 (MINUS_EXPR
, type
, iv
->base
, iv
->step
);
1090 new_c
= fold_build2 (MINUS_EXPR
, niter_type
,
1091 fold_convert (niter_type
, final
),
1092 fold_convert (niter_type
, t
));
1093 if (multiple_of_p (type
, new_c
, s
))
1094 relaxed_cond
= fold_build2 (LT_EXPR
, boolean_type_node
,
1100 t
= simplify_using_initial_conditions (loop
, cond
);
1101 if (!t
|| !integer_onep (t
))
1102 t
= simplify_using_initial_conditions (loop
, relaxed_cond
);
1104 if (t
&& integer_onep (t
))
1105 niter
->control
.no_overflow
= true;
1108 /* First the trivial cases -- when the step is 1. */
1109 if (integer_onep (s
))
1114 if (niter
->control
.no_overflow
&& multiple_of_p (type
, c
, s
))
1116 niter
->niter
= fold_build2 (FLOOR_DIV_EXPR
, niter_type
, c
, s
);
1120 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
1121 is infinite. Otherwise, the number of iterations is
1122 (inverse(s/d) * (c/d)) mod (size of mode/d). */
1123 bits
= num_ending_zeros (s
);
1124 bound
= build_low_bits_mask (niter_type
,
1125 (TYPE_PRECISION (niter_type
)
1126 - tree_to_uhwi (bits
)));
1128 d
= fold_binary_to_constant (LSHIFT_EXPR
, niter_type
,
1129 build_int_cst (niter_type
, 1), bits
);
1130 s
= fold_binary_to_constant (RSHIFT_EXPR
, niter_type
, s
, bits
);
1132 if (!exit_must_be_taken
)
1134 /* If we cannot assume that the exit is taken eventually, record the
1135 assumptions for divisibility of c. */
1136 assumption
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, c
, d
);
1137 assumption
= fold_build2 (EQ_EXPR
, boolean_type_node
,
1138 assumption
, build_int_cst (niter_type
, 0));
1139 if (!integer_nonzerop (assumption
))
1140 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1141 niter
->assumptions
, assumption
);
1144 c
= fold_build2 (EXACT_DIV_EXPR
, niter_type
, c
, d
);
1145 if (integer_onep (s
))
1151 tmp
= fold_build2 (MULT_EXPR
, niter_type
, c
, inverse (s
, bound
));
1152 niter
->niter
= fold_build2 (BIT_AND_EXPR
, niter_type
, tmp
, bound
);
1157 /* Checks whether we can determine the final value of the control variable
1158 of the loop with ending condition IV0 < IV1 (computed in TYPE).
1159 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
1160 of the step. The assumptions necessary to ensure that the computation
1161 of the final value does not overflow are recorded in NITER. If we
1162 find the final value, we adjust DELTA and return TRUE. Otherwise
1163 we return false. BNDS bounds the value of IV1->base - IV0->base,
1164 and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is
1165 true if we know that the exit must be taken eventually. */
1168 number_of_iterations_lt_to_ne (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
1169 class tree_niter_desc
*niter
,
1170 tree
*delta
, tree step
,
1171 bool exit_must_be_taken
, bounds
*bnds
)
1173 tree niter_type
= TREE_TYPE (step
);
1174 tree mod
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, *delta
, step
);
1177 tree assumption
= boolean_true_node
, bound
, noloop
;
1178 bool ret
= false, fv_comp_no_overflow
;
1180 if (POINTER_TYPE_P (type
))
1183 if (TREE_CODE (mod
) != INTEGER_CST
)
1185 if (integer_nonzerop (mod
))
1186 mod
= fold_build2 (MINUS_EXPR
, niter_type
, step
, mod
);
1187 tmod
= fold_convert (type1
, mod
);
1190 wi::to_mpz (wi::to_wide (mod
), mmod
, UNSIGNED
);
1191 mpz_neg (mmod
, mmod
);
1193 /* If the induction variable does not overflow and the exit is taken,
1194 then the computation of the final value does not overflow. This is
1195 also obviously the case if the new final value is equal to the
1196 current one. Finally, we postulate this for pointer type variables,
1197 as the code cannot rely on the object to that the pointer points being
1198 placed at the end of the address space (and more pragmatically,
1199 TYPE_{MIN,MAX}_VALUE is not defined for pointers). */
1200 if (integer_zerop (mod
) || POINTER_TYPE_P (type
))
1201 fv_comp_no_overflow
= true;
1202 else if (!exit_must_be_taken
)
1203 fv_comp_no_overflow
= false;
1205 fv_comp_no_overflow
=
1206 (iv0
->no_overflow
&& integer_nonzerop (iv0
->step
))
1207 || (iv1
->no_overflow
&& integer_nonzerop (iv1
->step
));
1209 if (integer_nonzerop (iv0
->step
))
1211 /* The final value of the iv is iv1->base + MOD, assuming that this
1212 computation does not overflow, and that
1213 iv0->base <= iv1->base + MOD. */
1214 if (!fv_comp_no_overflow
)
1216 bound
= fold_build2 (MINUS_EXPR
, type1
,
1217 TYPE_MAX_VALUE (type1
), tmod
);
1218 assumption
= fold_build2 (LE_EXPR
, boolean_type_node
,
1220 if (integer_zerop (assumption
))
1223 if (mpz_cmp (mmod
, bnds
->below
) < 0)
1224 noloop
= boolean_false_node
;
1225 else if (POINTER_TYPE_P (type
))
1226 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
1228 fold_build_pointer_plus (iv1
->base
, tmod
));
1230 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
1232 fold_build2 (PLUS_EXPR
, type1
,
1237 /* The final value of the iv is iv0->base - MOD, assuming that this
1238 computation does not overflow, and that
1239 iv0->base - MOD <= iv1->base. */
1240 if (!fv_comp_no_overflow
)
1242 bound
= fold_build2 (PLUS_EXPR
, type1
,
1243 TYPE_MIN_VALUE (type1
), tmod
);
1244 assumption
= fold_build2 (GE_EXPR
, boolean_type_node
,
1246 if (integer_zerop (assumption
))
1249 if (mpz_cmp (mmod
, bnds
->below
) < 0)
1250 noloop
= boolean_false_node
;
1251 else if (POINTER_TYPE_P (type
))
1252 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
1253 fold_build_pointer_plus (iv0
->base
,
1254 fold_build1 (NEGATE_EXPR
,
1258 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
1259 fold_build2 (MINUS_EXPR
, type1
,
1264 if (!integer_nonzerop (assumption
))
1265 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1268 if (!integer_zerop (noloop
))
1269 niter
->may_be_zero
= fold_build2 (TRUTH_OR_EXPR
, boolean_type_node
,
1272 bounds_add (bnds
, wi::to_widest (mod
), type
);
1273 *delta
= fold_build2 (PLUS_EXPR
, niter_type
, *delta
, mod
);
1281 /* Add assertions to NITER that ensure that the control variable of the loop
1282 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
1283 are TYPE. Returns false if we can prove that there is an overflow, true
1284 otherwise. STEP is the absolute value of the step. */
1287 assert_no_overflow_lt (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
1288 class tree_niter_desc
*niter
, tree step
)
1290 tree bound
, d
, assumption
, diff
;
1291 tree niter_type
= TREE_TYPE (step
);
1293 if (integer_nonzerop (iv0
->step
))
1295 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
1296 if (iv0
->no_overflow
)
1299 /* If iv0->base is a constant, we can determine the last value before
1300 overflow precisely; otherwise we conservatively assume
1303 if (TREE_CODE (iv0
->base
) == INTEGER_CST
)
1305 d
= fold_build2 (MINUS_EXPR
, niter_type
,
1306 fold_convert (niter_type
, TYPE_MAX_VALUE (type
)),
1307 fold_convert (niter_type
, iv0
->base
));
1308 diff
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, d
, step
);
1311 diff
= fold_build2 (MINUS_EXPR
, niter_type
, step
,
1312 build_int_cst (niter_type
, 1));
1313 bound
= fold_build2 (MINUS_EXPR
, type
,
1314 TYPE_MAX_VALUE (type
), fold_convert (type
, diff
));
1315 assumption
= fold_build2 (LE_EXPR
, boolean_type_node
,
1320 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
1321 if (iv1
->no_overflow
)
1324 if (TREE_CODE (iv1
->base
) == INTEGER_CST
)
1326 d
= fold_build2 (MINUS_EXPR
, niter_type
,
1327 fold_convert (niter_type
, iv1
->base
),
1328 fold_convert (niter_type
, TYPE_MIN_VALUE (type
)));
1329 diff
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, d
, step
);
1332 diff
= fold_build2 (MINUS_EXPR
, niter_type
, step
,
1333 build_int_cst (niter_type
, 1));
1334 bound
= fold_build2 (PLUS_EXPR
, type
,
1335 TYPE_MIN_VALUE (type
), fold_convert (type
, diff
));
1336 assumption
= fold_build2 (GE_EXPR
, boolean_type_node
,
1340 if (integer_zerop (assumption
))
1342 if (!integer_nonzerop (assumption
))
1343 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1344 niter
->assumptions
, assumption
);
1346 iv0
->no_overflow
= true;
1347 iv1
->no_overflow
= true;
1351 /* Add an assumption to NITER that a loop whose ending condition
1352 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS
1353 bounds the value of IV1->base - IV0->base. */
1356 assert_loop_rolls_lt (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
1357 class tree_niter_desc
*niter
, bounds
*bnds
)
1359 tree assumption
= boolean_true_node
, bound
, diff
;
1360 tree mbz
, mbzl
, mbzr
, type1
;
1361 bool rolls_p
, no_overflow_p
;
1365 /* We are going to compute the number of iterations as
1366 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned
1367 variant of TYPE. This formula only works if
1369 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
1371 (where MAX is the maximum value of the unsigned variant of TYPE, and
1372 the computations in this formula are performed in full precision,
1373 i.e., without overflows).
1375 Usually, for loops with exit condition iv0->base + step * i < iv1->base,
1376 we have a condition of the form iv0->base - step < iv1->base before the loop,
1377 and for loops iv0->base < iv1->base - step * i the condition
1378 iv0->base < iv1->base + step, due to loop header copying, which enable us
1379 to prove the lower bound.
1381 The upper bound is more complicated. Unless the expressions for initial
1382 and final value themselves contain enough information, we usually cannot
1383 derive it from the context. */
1385 /* First check whether the answer does not follow from the bounds we gathered
1387 if (integer_nonzerop (iv0
->step
))
1388 dstep
= wi::to_widest (iv0
->step
);
1391 dstep
= wi::sext (wi::to_widest (iv1
->step
), TYPE_PRECISION (type
));
1396 wi::to_mpz (dstep
, mstep
, UNSIGNED
);
1397 mpz_neg (mstep
, mstep
);
1398 mpz_add_ui (mstep
, mstep
, 1);
1400 rolls_p
= mpz_cmp (mstep
, bnds
->below
) <= 0;
1403 wi::to_mpz (wi::minus_one (TYPE_PRECISION (type
)), max
, UNSIGNED
);
1404 mpz_add (max
, max
, mstep
);
1405 no_overflow_p
= (mpz_cmp (bnds
->up
, max
) <= 0
1406 /* For pointers, only values lying inside a single object
1407 can be compared or manipulated by pointer arithmetics.
1408 Gcc in general does not allow or handle objects larger
1409 than half of the address space, hence the upper bound
1410 is satisfied for pointers. */
1411 || POINTER_TYPE_P (type
));
1415 if (rolls_p
&& no_overflow_p
)
1419 if (POINTER_TYPE_P (type
))
1422 /* Now the hard part; we must formulate the assumption(s) as expressions, and
1423 we must be careful not to introduce overflow. */
1425 if (integer_nonzerop (iv0
->step
))
1427 diff
= fold_build2 (MINUS_EXPR
, type1
,
1428 iv0
->step
, build_int_cst (type1
, 1));
1430 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
1431 0 address never belongs to any object, we can assume this for
1433 if (!POINTER_TYPE_P (type
))
1435 bound
= fold_build2 (PLUS_EXPR
, type1
,
1436 TYPE_MIN_VALUE (type
), diff
);
1437 assumption
= fold_build2 (GE_EXPR
, boolean_type_node
,
1441 /* And then we can compute iv0->base - diff, and compare it with
1443 mbzl
= fold_build2 (MINUS_EXPR
, type1
,
1444 fold_convert (type1
, iv0
->base
), diff
);
1445 mbzr
= fold_convert (type1
, iv1
->base
);
1449 diff
= fold_build2 (PLUS_EXPR
, type1
,
1450 iv1
->step
, build_int_cst (type1
, 1));
1452 if (!POINTER_TYPE_P (type
))
1454 bound
= fold_build2 (PLUS_EXPR
, type1
,
1455 TYPE_MAX_VALUE (type
), diff
);
1456 assumption
= fold_build2 (LE_EXPR
, boolean_type_node
,
1460 mbzl
= fold_convert (type1
, iv0
->base
);
1461 mbzr
= fold_build2 (MINUS_EXPR
, type1
,
1462 fold_convert (type1
, iv1
->base
), diff
);
1465 if (!integer_nonzerop (assumption
))
1466 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1467 niter
->assumptions
, assumption
);
1470 mbz
= fold_build2 (GT_EXPR
, boolean_type_node
, mbzl
, mbzr
);
1471 niter
->may_be_zero
= fold_build2 (TRUTH_OR_EXPR
, boolean_type_node
,
1472 niter
->may_be_zero
, mbz
);
1476 /* Determines number of iterations of loop whose ending condition
1477 is IV0 < IV1 which likes: {base, -C} < n, or n < {base, C}.
1478 The number of iterations is stored to NITER. */
1481 number_of_iterations_until_wrap (class loop
*loop
, tree type
, affine_iv
*iv0
,
1482 affine_iv
*iv1
, class tree_niter_desc
*niter
)
1484 tree niter_type
= unsigned_type_for (type
);
1485 tree step
, num
, assumptions
, may_be_zero
, span
;
1486 wide_int high
, low
, max
, min
;
1488 may_be_zero
= fold_build2 (LE_EXPR
, boolean_type_node
, iv1
->base
, iv0
->base
);
1489 if (integer_onep (may_be_zero
))
1492 int prec
= TYPE_PRECISION (type
);
1493 signop sgn
= TYPE_SIGN (type
);
1494 min
= wi::min_value (prec
, sgn
);
1495 max
= wi::max_value (prec
, sgn
);
1497 /* n < {base, C}. */
1498 if (integer_zerop (iv0
->step
) && !tree_int_cst_sign_bit (iv1
->step
))
1501 /* MIN + C - 1 <= n. */
1502 tree last
= wide_int_to_tree (type
, min
+ wi::to_wide (step
) - 1);
1503 assumptions
= fold_build2 (LE_EXPR
, boolean_type_node
, last
, iv0
->base
);
1504 if (integer_zerop (assumptions
))
1507 num
= fold_build2 (MINUS_EXPR
, niter_type
, wide_int_to_tree (type
, max
),
1510 /* When base has the form iv + 1, if we know iv >= n, then iv + 1 < n
1511 only when iv + 1 overflows, i.e. when iv == TYPE_VALUE_MAX. */
1513 && integer_onep (step
)
1514 && TREE_CODE (iv1
->base
) == PLUS_EXPR
1515 && integer_onep (TREE_OPERAND (iv1
->base
, 1)))
1517 tree cond
= fold_build2 (GE_EXPR
, boolean_type_node
,
1518 TREE_OPERAND (iv1
->base
, 0), iv0
->base
);
1519 cond
= simplify_using_initial_conditions (loop
, cond
);
1520 if (integer_onep (cond
))
1521 may_be_zero
= fold_build2 (EQ_EXPR
, boolean_type_node
,
1522 TREE_OPERAND (iv1
->base
, 0),
1523 TYPE_MAX_VALUE (type
));
1527 if (TREE_CODE (iv1
->base
) == INTEGER_CST
)
1528 low
= wi::to_wide (iv1
->base
) - 1;
1529 else if (TREE_CODE (iv0
->base
) == INTEGER_CST
)
1530 low
= wi::to_wide (iv0
->base
);
1534 /* {base, -C} < n. */
1535 else if (tree_int_cst_sign_bit (iv0
->step
) && integer_zerop (iv1
->step
))
1537 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (iv0
->step
), iv0
->step
);
1538 /* MAX - C + 1 >= n. */
1539 tree last
= wide_int_to_tree (type
, max
- wi::to_wide (step
) + 1);
1540 assumptions
= fold_build2 (GE_EXPR
, boolean_type_node
, last
, iv1
->base
);
1541 if (integer_zerop (assumptions
))
1544 num
= fold_build2 (MINUS_EXPR
, niter_type
, iv0
->base
,
1545 wide_int_to_tree (type
, min
));
1547 if (TREE_CODE (iv0
->base
) == INTEGER_CST
)
1548 high
= wi::to_wide (iv0
->base
) + 1;
1549 else if (TREE_CODE (iv1
->base
) == INTEGER_CST
)
1550 high
= wi::to_wide (iv1
->base
);
1557 /* (delta + step - 1) / step */
1558 step
= fold_convert (niter_type
, step
);
1559 num
= fold_convert (niter_type
, num
);
1560 num
= fold_build2 (PLUS_EXPR
, niter_type
, num
, step
);
1561 niter
->niter
= fold_build2 (FLOOR_DIV_EXPR
, niter_type
, num
, step
);
1563 widest_int delta
, s
;
1564 delta
= widest_int::from (high
, sgn
) - widest_int::from (low
, sgn
);
1565 s
= wi::to_widest (step
);
1566 delta
= delta
+ s
- 1;
1567 niter
->max
= wi::udiv_floor (delta
, s
);
1569 niter
->may_be_zero
= may_be_zero
;
1571 if (!integer_nonzerop (assumptions
))
1572 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1573 niter
->assumptions
, assumptions
);
1575 niter
->control
.no_overflow
= false;
1577 /* Update bound and exit condition as:
1578 bound = niter * STEP + (IVbase - STEP).
1579 { IVbase - STEP, +, STEP } != bound
1580 Here, biasing IVbase by 1 step makes 'bound' be the value before wrap.
1582 niter
->control
.base
= fold_build2 (MINUS_EXPR
, niter_type
,
1583 niter
->control
.base
, niter
->control
.step
);
1584 span
= fold_build2 (MULT_EXPR
, niter_type
, niter
->niter
,
1585 fold_convert (niter_type
, niter
->control
.step
));
1586 niter
->bound
= fold_build2 (PLUS_EXPR
, niter_type
, span
,
1587 fold_convert (niter_type
, niter
->control
.base
));
1588 niter
->bound
= fold_convert (type
, niter
->bound
);
1589 niter
->cmp
= NE_EXPR
;
1594 /* Determines number of iterations of loop whose ending condition
1595 is IV0 < IV1. TYPE is the type of the iv. The number of
1596 iterations is stored to NITER. BNDS bounds the difference
1597 IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know
1598 that the exit must be taken eventually. */
1601 number_of_iterations_lt (class loop
*loop
, tree type
, affine_iv
*iv0
,
1602 affine_iv
*iv1
, class tree_niter_desc
*niter
,
1603 bool exit_must_be_taken
, bounds
*bnds
)
1605 tree niter_type
= unsigned_type_for (type
);
1606 tree delta
, step
, s
;
1609 if (integer_nonzerop (iv0
->step
))
1611 niter
->control
= *iv0
;
1612 niter
->cmp
= LT_EXPR
;
1613 niter
->bound
= iv1
->base
;
1617 niter
->control
= *iv1
;
1618 niter
->cmp
= GT_EXPR
;
1619 niter
->bound
= iv0
->base
;
1622 /* {base, -C} < n, or n < {base, C} */
1623 if (tree_int_cst_sign_bit (iv0
->step
)
1624 || (!integer_zerop (iv1
->step
) && !tree_int_cst_sign_bit (iv1
->step
)))
1625 return number_of_iterations_until_wrap (loop
, type
, iv0
, iv1
, niter
);
1627 delta
= fold_build2 (MINUS_EXPR
, niter_type
,
1628 fold_convert (niter_type
, iv1
->base
),
1629 fold_convert (niter_type
, iv0
->base
));
1631 /* First handle the special case that the step is +-1. */
1632 if ((integer_onep (iv0
->step
) && integer_zerop (iv1
->step
))
1633 || (integer_all_onesp (iv1
->step
) && integer_zerop (iv0
->step
)))
1635 /* for (i = iv0->base; i < iv1->base; i++)
1639 for (i = iv1->base; i > iv0->base; i--).
1641 In both cases # of iterations is iv1->base - iv0->base, assuming that
1642 iv1->base >= iv0->base.
1644 First try to derive a lower bound on the value of
1645 iv1->base - iv0->base, computed in full precision. If the difference
1646 is nonnegative, we are done, otherwise we must record the
1649 if (mpz_sgn (bnds
->below
) < 0)
1650 niter
->may_be_zero
= fold_build2 (LT_EXPR
, boolean_type_node
,
1651 iv1
->base
, iv0
->base
);
1652 niter
->niter
= delta
;
1653 niter
->max
= widest_int::from (wi::from_mpz (niter_type
, bnds
->up
, false),
1654 TYPE_SIGN (niter_type
));
1655 niter
->control
.no_overflow
= true;
1659 if (integer_nonzerop (iv0
->step
))
1660 step
= fold_convert (niter_type
, iv0
->step
);
1662 step
= fold_convert (niter_type
,
1663 fold_build1 (NEGATE_EXPR
, type
, iv1
->step
));
1665 /* If we can determine the final value of the control iv exactly, we can
1666 transform the condition to != comparison. In particular, this will be
1667 the case if DELTA is constant. */
1668 if (number_of_iterations_lt_to_ne (type
, iv0
, iv1
, niter
, &delta
, step
,
1669 exit_must_be_taken
, bnds
))
1673 zps
.base
= build_int_cst (niter_type
, 0);
1675 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1676 zps does not overflow. */
1677 zps
.no_overflow
= true;
1679 return number_of_iterations_ne (loop
, type
, &zps
,
1680 delta
, niter
, true, bnds
);
1683 /* Make sure that the control iv does not overflow. */
1684 if (!assert_no_overflow_lt (type
, iv0
, iv1
, niter
, step
))
1687 /* We determine the number of iterations as (delta + step - 1) / step. For
1688 this to work, we must know that iv1->base >= iv0->base - step + 1,
1689 otherwise the loop does not roll. */
1690 assert_loop_rolls_lt (type
, iv0
, iv1
, niter
, bnds
);
1692 s
= fold_build2 (MINUS_EXPR
, niter_type
,
1693 step
, build_int_cst (niter_type
, 1));
1694 delta
= fold_build2 (PLUS_EXPR
, niter_type
, delta
, s
);
1695 niter
->niter
= fold_build2 (FLOOR_DIV_EXPR
, niter_type
, delta
, step
);
1699 wi::to_mpz (wi::to_wide (step
), mstep
, UNSIGNED
);
1700 mpz_add (tmp
, bnds
->up
, mstep
);
1701 mpz_sub_ui (tmp
, tmp
, 1);
1702 mpz_fdiv_q (tmp
, tmp
, mstep
);
1703 niter
->max
= widest_int::from (wi::from_mpz (niter_type
, tmp
, false),
1704 TYPE_SIGN (niter_type
));
1711 /* Determines number of iterations of loop whose ending condition
1712 is IV0 <= IV1. TYPE is the type of the iv. The number of
1713 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
1714 we know that this condition must eventually become false (we derived this
1715 earlier, and possibly set NITER->assumptions to make sure this
1716 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1719 number_of_iterations_le (class loop
*loop
, tree type
, affine_iv
*iv0
,
1720 affine_iv
*iv1
, class tree_niter_desc
*niter
,
1721 bool exit_must_be_taken
, bounds
*bnds
)
1725 if (POINTER_TYPE_P (type
))
1728 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1729 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1730 value of the type. This we must know anyway, since if it is
1731 equal to this value, the loop rolls forever. We do not check
1732 this condition for pointer type ivs, as the code cannot rely on
1733 the object to that the pointer points being placed at the end of
1734 the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is
1735 not defined for pointers). */
1737 if (!exit_must_be_taken
&& !POINTER_TYPE_P (type
))
1739 if (integer_nonzerop (iv0
->step
))
1740 assumption
= fold_build2 (NE_EXPR
, boolean_type_node
,
1741 iv1
->base
, TYPE_MAX_VALUE (type
));
1743 assumption
= fold_build2 (NE_EXPR
, boolean_type_node
,
1744 iv0
->base
, TYPE_MIN_VALUE (type
));
1746 if (integer_zerop (assumption
))
1748 if (!integer_nonzerop (assumption
))
1749 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1750 niter
->assumptions
, assumption
);
1753 if (integer_nonzerop (iv0
->step
))
1755 if (POINTER_TYPE_P (type
))
1756 iv1
->base
= fold_build_pointer_plus_hwi (iv1
->base
, 1);
1758 iv1
->base
= fold_build2 (PLUS_EXPR
, type1
, iv1
->base
,
1759 build_int_cst (type1
, 1));
1761 else if (POINTER_TYPE_P (type
))
1762 iv0
->base
= fold_build_pointer_plus_hwi (iv0
->base
, -1);
1764 iv0
->base
= fold_build2 (MINUS_EXPR
, type1
,
1765 iv0
->base
, build_int_cst (type1
, 1));
1767 bounds_add (bnds
, 1, type1
);
1769 return number_of_iterations_lt (loop
, type
, iv0
, iv1
, niter
, exit_must_be_taken
,
1773 /* Dumps description of affine induction variable IV to FILE. */
1776 dump_affine_iv (FILE *file
, affine_iv
*iv
)
1778 if (!integer_zerop (iv
->step
))
1779 fprintf (file
, "[");
1781 print_generic_expr (dump_file
, iv
->base
, TDF_SLIM
);
1783 if (!integer_zerop (iv
->step
))
1785 fprintf (file
, ", + , ");
1786 print_generic_expr (dump_file
, iv
->step
, TDF_SLIM
);
1787 fprintf (file
, "]%s", iv
->no_overflow
? "(no_overflow)" : "");
1791 /* Determine the number of iterations according to condition (for staying
1792 inside loop) which compares two induction variables using comparison
1793 operator CODE. The induction variable on left side of the comparison
1794 is IV0, the right-hand side is IV1. Both induction variables must have
1795 type TYPE, which must be an integer or pointer type. The steps of the
1796 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1798 LOOP is the loop whose number of iterations we are determining.
1800 ONLY_EXIT is true if we are sure this is the only way the loop could be
1801 exited (including possibly non-returning function calls, exceptions, etc.)
1802 -- in this case we can use the information whether the control induction
1803 variables can overflow or not in a more efficient way.
1805 if EVERY_ITERATION is true, we know the test is executed on every iteration.
1807 The results (number of iterations and assumptions as described in
1808 comments at class tree_niter_desc in tree-ssa-loop.h) are stored to NITER.
1809 Returns false if it fails to determine number of iterations, true if it
1810 was determined (possibly with some assumptions). */
1813 number_of_iterations_cond (class loop
*loop
,
1814 tree type
, affine_iv
*iv0
, enum tree_code code
,
1815 affine_iv
*iv1
, class tree_niter_desc
*niter
,
1816 bool only_exit
, bool every_iteration
)
1818 bool exit_must_be_taken
= false, ret
;
1821 /* If the test is not executed every iteration, wrapping may make the test
1823 TODO: the overflow case can be still used as unreliable estimate of upper
1824 bound. But we have no API to pass it down to number of iterations code
1825 and, at present, it will not use it anyway. */
1826 if (!every_iteration
1827 && (!iv0
->no_overflow
|| !iv1
->no_overflow
1828 || code
== NE_EXPR
|| code
== EQ_EXPR
))
1831 /* The meaning of these assumptions is this:
1833 then the rest of information does not have to be valid
1834 if may_be_zero then the loop does not roll, even if
1836 niter
->assumptions
= boolean_true_node
;
1837 niter
->may_be_zero
= boolean_false_node
;
1838 niter
->niter
= NULL_TREE
;
1840 niter
->bound
= NULL_TREE
;
1841 niter
->cmp
= ERROR_MARK
;
1843 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1844 the control variable is on lhs. */
1845 if (code
== GE_EXPR
|| code
== GT_EXPR
1846 || (code
== NE_EXPR
&& integer_zerop (iv0
->step
)))
1848 std::swap (iv0
, iv1
);
1849 code
= swap_tree_comparison (code
);
1852 if (POINTER_TYPE_P (type
))
1854 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1855 to the same object. If they do, the control variable cannot wrap
1856 (as wrap around the bounds of memory will never return a pointer
1857 that would be guaranteed to point to the same object, even if we
1858 avoid undefined behavior by casting to size_t and back). */
1859 iv0
->no_overflow
= true;
1860 iv1
->no_overflow
= true;
1863 /* If the control induction variable does not overflow and the only exit
1864 from the loop is the one that we analyze, we know it must be taken
1868 if (!integer_zerop (iv0
->step
) && iv0
->no_overflow
)
1869 exit_must_be_taken
= true;
1870 else if (!integer_zerop (iv1
->step
) && iv1
->no_overflow
)
1871 exit_must_be_taken
= true;
1874 /* We can handle cases which neither of the sides of the comparison is
1877 {iv0.base, iv0.step} cmp_code {iv1.base, iv1.step}
1879 {iv0.base, iv0.step - iv1.step} cmp_code {iv1.base, 0}
1881 provided that either below condition is satisfied:
1883 a) the test is NE_EXPR;
1884 b) iv0.step - iv1.step is integer and iv0/iv1 don't overflow.
1886 This rarely occurs in practice, but it is simple enough to manage. */
1887 if (!integer_zerop (iv0
->step
) && !integer_zerop (iv1
->step
))
1889 tree step_type
= POINTER_TYPE_P (type
) ? sizetype
: type
;
1890 tree step
= fold_binary_to_constant (MINUS_EXPR
, step_type
,
1891 iv0
->step
, iv1
->step
);
1893 /* No need to check sign of the new step since below code takes care
1896 && (TREE_CODE (step
) != INTEGER_CST
1897 || !iv0
->no_overflow
|| !iv1
->no_overflow
))
1901 if (!POINTER_TYPE_P (type
))
1902 iv0
->no_overflow
= false;
1904 iv1
->step
= build_int_cst (step_type
, 0);
1905 iv1
->no_overflow
= true;
1908 /* If the result of the comparison is a constant, the loop is weird. More
1909 precise handling would be possible, but the situation is not common enough
1910 to waste time on it. */
1911 if (integer_zerop (iv0
->step
) && integer_zerop (iv1
->step
))
1914 /* If the loop exits immediately, there is nothing to do. */
1915 tree tem
= fold_binary (code
, boolean_type_node
, iv0
->base
, iv1
->base
);
1916 if (tem
&& integer_zerop (tem
))
1918 if (!every_iteration
)
1920 niter
->niter
= build_int_cst (unsigned_type_for (type
), 0);
1925 /* OK, now we know we have a senseful loop. Handle several cases, depending
1926 on what comparison operator is used. */
1927 bound_difference (loop
, iv1
->base
, iv0
->base
, &bnds
);
1929 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1932 "Analyzing # of iterations of loop %d\n", loop
->num
);
1934 fprintf (dump_file
, " exit condition ");
1935 dump_affine_iv (dump_file
, iv0
);
1936 fprintf (dump_file
, " %s ",
1937 code
== NE_EXPR
? "!="
1938 : code
== LT_EXPR
? "<"
1940 dump_affine_iv (dump_file
, iv1
);
1941 fprintf (dump_file
, "\n");
1943 fprintf (dump_file
, " bounds on difference of bases: ");
1944 mpz_out_str (dump_file
, 10, bnds
.below
);
1945 fprintf (dump_file
, " ... ");
1946 mpz_out_str (dump_file
, 10, bnds
.up
);
1947 fprintf (dump_file
, "\n");
1953 gcc_assert (integer_zerop (iv1
->step
));
1954 ret
= number_of_iterations_ne (loop
, type
, iv0
, iv1
->base
, niter
,
1955 exit_must_be_taken
, &bnds
);
1959 ret
= number_of_iterations_lt (loop
, type
, iv0
, iv1
, niter
,
1960 exit_must_be_taken
, &bnds
);
1964 ret
= number_of_iterations_le (loop
, type
, iv0
, iv1
, niter
,
1965 exit_must_be_taken
, &bnds
);
1972 mpz_clear (bnds
.up
);
1973 mpz_clear (bnds
.below
);
1975 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1979 fprintf (dump_file
, " result:\n");
1980 if (!integer_nonzerop (niter
->assumptions
))
1982 fprintf (dump_file
, " under assumptions ");
1983 print_generic_expr (dump_file
, niter
->assumptions
, TDF_SLIM
);
1984 fprintf (dump_file
, "\n");
1987 if (!integer_zerop (niter
->may_be_zero
))
1989 fprintf (dump_file
, " zero if ");
1990 print_generic_expr (dump_file
, niter
->may_be_zero
, TDF_SLIM
);
1991 fprintf (dump_file
, "\n");
1994 fprintf (dump_file
, " # of iterations ");
1995 print_generic_expr (dump_file
, niter
->niter
, TDF_SLIM
);
1996 fprintf (dump_file
, ", bounded by ");
1997 print_decu (niter
->max
, dump_file
);
1998 fprintf (dump_file
, "\n");
2001 fprintf (dump_file
, " failed\n\n");
2006 /* Substitute NEW_TREE for OLD in EXPR and fold the result.
2007 If VALUEIZE is non-NULL then OLD and NEW_TREE are ignored and instead
2008 all SSA names are replaced with the result of calling the VALUEIZE
2009 function with the SSA name as argument. */
2012 simplify_replace_tree (tree expr
, tree old
, tree new_tree
,
2013 tree (*valueize
) (tree
, void*), void *context
,
2017 tree ret
= NULL_TREE
, e
, se
;
2022 /* Do not bother to replace constants. */
2023 if (CONSTANT_CLASS_P (expr
))
2028 if (TREE_CODE (expr
) == SSA_NAME
)
2030 new_tree
= valueize (expr
, context
);
2031 if (new_tree
!= expr
)
2035 else if (expr
== old
2036 || operand_equal_p (expr
, old
, 0))
2037 return unshare_expr (new_tree
);
2042 n
= TREE_OPERAND_LENGTH (expr
);
2043 for (i
= 0; i
< n
; i
++)
2045 e
= TREE_OPERAND (expr
, i
);
2046 se
= simplify_replace_tree (e
, old
, new_tree
, valueize
, context
, do_fold
);
2051 ret
= copy_node (expr
);
2053 TREE_OPERAND (ret
, i
) = se
;
2056 return (ret
? (do_fold
? fold (ret
) : ret
) : expr
);
2059 /* Expand definitions of ssa names in EXPR as long as they are simple
2060 enough, and return the new expression. If STOP is specified, stop
2061 expanding if EXPR equals to it. */
2064 expand_simple_operations (tree expr
, tree stop
, hash_map
<tree
, tree
> &cache
)
2067 tree ret
= NULL_TREE
, e
, ee
, e1
;
2068 enum tree_code code
;
2071 if (expr
== NULL_TREE
)
2074 if (is_gimple_min_invariant (expr
))
2077 code
= TREE_CODE (expr
);
2078 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code
)))
2080 n
= TREE_OPERAND_LENGTH (expr
);
2081 for (i
= 0; i
< n
; i
++)
2083 e
= TREE_OPERAND (expr
, i
);
2084 /* SCEV analysis feeds us with a proper expression
2085 graph matching the SSA graph. Avoid turning it
2086 into a tree here, thus handle tree sharing
2088 ??? The SSA walk below still turns the SSA graph
2089 into a tree but until we find a testcase do not
2090 introduce additional tree sharing here. */
2092 tree
&cee
= cache
.get_or_insert (e
, &existed_p
);
2098 ee
= expand_simple_operations (e
, stop
, cache
);
2100 *cache
.get (e
) = ee
;
2106 ret
= copy_node (expr
);
2108 TREE_OPERAND (ret
, i
) = ee
;
2114 fold_defer_overflow_warnings ();
2116 fold_undefer_and_ignore_overflow_warnings ();
2120 /* Stop if it's not ssa name or the one we don't want to expand. */
2121 if (TREE_CODE (expr
) != SSA_NAME
|| expr
== stop
)
2124 stmt
= SSA_NAME_DEF_STMT (expr
);
2125 if (gimple_code (stmt
) == GIMPLE_PHI
)
2127 basic_block src
, dest
;
2129 if (gimple_phi_num_args (stmt
) != 1)
2131 e
= PHI_ARG_DEF (stmt
, 0);
2133 /* Avoid propagating through loop exit phi nodes, which
2134 could break loop-closed SSA form restrictions. */
2135 dest
= gimple_bb (stmt
);
2136 src
= single_pred (dest
);
2137 if (TREE_CODE (e
) == SSA_NAME
2138 && src
->loop_father
!= dest
->loop_father
)
2141 return expand_simple_operations (e
, stop
, cache
);
2143 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
2146 /* Avoid expanding to expressions that contain SSA names that need
2147 to take part in abnormal coalescing. */
2149 FOR_EACH_SSA_TREE_OPERAND (e
, stmt
, iter
, SSA_OP_USE
)
2150 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e
))
2153 e
= gimple_assign_rhs1 (stmt
);
2154 code
= gimple_assign_rhs_code (stmt
);
2155 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
2157 if (is_gimple_min_invariant (e
))
2160 if (code
== SSA_NAME
)
2161 return expand_simple_operations (e
, stop
, cache
);
2162 else if (code
== ADDR_EXPR
)
2165 tree base
= get_addr_base_and_unit_offset (TREE_OPERAND (e
, 0),
2168 && TREE_CODE (base
) == MEM_REF
)
2170 ee
= expand_simple_operations (TREE_OPERAND (base
, 0), stop
,
2172 return fold_build2 (POINTER_PLUS_EXPR
, TREE_TYPE (expr
), ee
,
2173 wide_int_to_tree (sizetype
,
2174 mem_ref_offset (base
)
2185 /* Casts are simple. */
2186 ee
= expand_simple_operations (e
, stop
, cache
);
2187 return fold_build1 (code
, TREE_TYPE (expr
), ee
);
2191 if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (expr
))
2192 && TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr
)))
2195 case POINTER_PLUS_EXPR
:
2196 /* And increments and decrements by a constant are simple. */
2197 e1
= gimple_assign_rhs2 (stmt
);
2198 if (!is_gimple_min_invariant (e1
))
2201 ee
= expand_simple_operations (e
, stop
, cache
);
2202 return fold_build2 (code
, TREE_TYPE (expr
), ee
, e1
);
2210 expand_simple_operations (tree expr
, tree stop
)
2212 hash_map
<tree
, tree
> cache
;
2213 return expand_simple_operations (expr
, stop
, cache
);
2216 /* Tries to simplify EXPR using the condition COND. Returns the simplified
2217 expression (or EXPR unchanged, if no simplification was possible). */
2220 tree_simplify_using_condition_1 (tree cond
, tree expr
)
2223 tree e
, e0
, e1
, e2
, notcond
;
2224 enum tree_code code
= TREE_CODE (expr
);
2226 if (code
== INTEGER_CST
)
2229 if (code
== TRUTH_OR_EXPR
2230 || code
== TRUTH_AND_EXPR
2231 || code
== COND_EXPR
)
2235 e0
= tree_simplify_using_condition_1 (cond
, TREE_OPERAND (expr
, 0));
2236 if (TREE_OPERAND (expr
, 0) != e0
)
2239 e1
= tree_simplify_using_condition_1 (cond
, TREE_OPERAND (expr
, 1));
2240 if (TREE_OPERAND (expr
, 1) != e1
)
2243 if (code
== COND_EXPR
)
2245 e2
= tree_simplify_using_condition_1 (cond
, TREE_OPERAND (expr
, 2));
2246 if (TREE_OPERAND (expr
, 2) != e2
)
2254 if (code
== COND_EXPR
)
2255 expr
= fold_build3 (code
, boolean_type_node
, e0
, e1
, e2
);
2257 expr
= fold_build2 (code
, boolean_type_node
, e0
, e1
);
2263 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
2264 propagation, and vice versa. Fold does not handle this, since it is
2265 considered too expensive. */
2266 if (TREE_CODE (cond
) == EQ_EXPR
)
2268 e0
= TREE_OPERAND (cond
, 0);
2269 e1
= TREE_OPERAND (cond
, 1);
2271 /* We know that e0 == e1. Check whether we cannot simplify expr
2273 e
= simplify_replace_tree (expr
, e0
, e1
);
2274 if (integer_zerop (e
) || integer_nonzerop (e
))
2277 e
= simplify_replace_tree (expr
, e1
, e0
);
2278 if (integer_zerop (e
) || integer_nonzerop (e
))
2281 if (TREE_CODE (expr
) == EQ_EXPR
)
2283 e0
= TREE_OPERAND (expr
, 0);
2284 e1
= TREE_OPERAND (expr
, 1);
2286 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
2287 e
= simplify_replace_tree (cond
, e0
, e1
);
2288 if (integer_zerop (e
))
2290 e
= simplify_replace_tree (cond
, e1
, e0
);
2291 if (integer_zerop (e
))
2294 if (TREE_CODE (expr
) == NE_EXPR
)
2296 e0
= TREE_OPERAND (expr
, 0);
2297 e1
= TREE_OPERAND (expr
, 1);
2299 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
2300 e
= simplify_replace_tree (cond
, e0
, e1
);
2301 if (integer_zerop (e
))
2302 return boolean_true_node
;
2303 e
= simplify_replace_tree (cond
, e1
, e0
);
2304 if (integer_zerop (e
))
2305 return boolean_true_node
;
2308 /* Check whether COND ==> EXPR. */
2309 notcond
= invert_truthvalue (cond
);
2310 e
= fold_binary (TRUTH_OR_EXPR
, boolean_type_node
, notcond
, expr
);
2311 if (e
&& integer_nonzerop (e
))
2314 /* Check whether COND ==> not EXPR. */
2315 e
= fold_binary (TRUTH_AND_EXPR
, boolean_type_node
, cond
, expr
);
2316 if (e
&& integer_zerop (e
))
2322 /* Tries to simplify EXPR using the condition COND. Returns the simplified
2323 expression (or EXPR unchanged, if no simplification was possible).
2324 Wrapper around tree_simplify_using_condition_1 that ensures that chains
2325 of simple operations in definitions of ssa names in COND are expanded,
2326 so that things like casts or incrementing the value of the bound before
2327 the loop do not cause us to fail. */
2330 tree_simplify_using_condition (tree cond
, tree expr
)
2332 cond
= expand_simple_operations (cond
);
2334 return tree_simplify_using_condition_1 (cond
, expr
);
2337 /* Tries to simplify EXPR using the conditions on entry to LOOP.
2338 Returns the simplified expression (or EXPR unchanged, if no
2339 simplification was possible). */
2342 simplify_using_initial_conditions (class loop
*loop
, tree expr
)
2347 tree cond
, expanded
, backup
;
2350 if (TREE_CODE (expr
) == INTEGER_CST
)
2353 backup
= expanded
= expand_simple_operations (expr
);
2355 /* Limit walking the dominators to avoid quadraticness in
2356 the number of BBs times the number of loops in degenerate
2358 for (bb
= loop
->header
;
2359 bb
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
) && cnt
< MAX_DOMINATORS_TO_WALK
;
2360 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
2362 if (!single_pred_p (bb
))
2364 e
= single_pred_edge (bb
);
2366 if (!(e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)))
2369 stmt
= last_stmt (e
->src
);
2370 cond
= fold_build2 (gimple_cond_code (stmt
),
2372 gimple_cond_lhs (stmt
),
2373 gimple_cond_rhs (stmt
));
2374 if (e
->flags
& EDGE_FALSE_VALUE
)
2375 cond
= invert_truthvalue (cond
);
2376 expanded
= tree_simplify_using_condition (cond
, expanded
);
2377 /* Break if EXPR is simplified to const values. */
2379 && (integer_zerop (expanded
) || integer_nonzerop (expanded
)))
2385 /* Return the original expression if no simplification is done. */
2386 return operand_equal_p (backup
, expanded
, 0) ? expr
: expanded
;
2389 /* Tries to simplify EXPR using the evolutions of the loop invariants
2390 in the superloops of LOOP. Returns the simplified expression
2391 (or EXPR unchanged, if no simplification was possible). */
2394 simplify_using_outer_evolutions (class loop
*loop
, tree expr
)
2396 enum tree_code code
= TREE_CODE (expr
);
2400 if (is_gimple_min_invariant (expr
))
2403 if (code
== TRUTH_OR_EXPR
2404 || code
== TRUTH_AND_EXPR
2405 || code
== COND_EXPR
)
2409 e0
= simplify_using_outer_evolutions (loop
, TREE_OPERAND (expr
, 0));
2410 if (TREE_OPERAND (expr
, 0) != e0
)
2413 e1
= simplify_using_outer_evolutions (loop
, TREE_OPERAND (expr
, 1));
2414 if (TREE_OPERAND (expr
, 1) != e1
)
2417 if (code
== COND_EXPR
)
2419 e2
= simplify_using_outer_evolutions (loop
, TREE_OPERAND (expr
, 2));
2420 if (TREE_OPERAND (expr
, 2) != e2
)
2428 if (code
== COND_EXPR
)
2429 expr
= fold_build3 (code
, boolean_type_node
, e0
, e1
, e2
);
2431 expr
= fold_build2 (code
, boolean_type_node
, e0
, e1
);
2437 e
= instantiate_parameters (loop
, expr
);
2438 if (is_gimple_min_invariant (e
))
2444 /* Returns true if EXIT is the only possible exit from LOOP. */
2447 loop_only_exit_p (const class loop
*loop
, basic_block
*body
, const_edge exit
)
2449 gimple_stmt_iterator bsi
;
2452 if (exit
!= single_exit (loop
))
2455 for (i
= 0; i
< loop
->num_nodes
; i
++)
2456 for (bsi
= gsi_start_bb (body
[i
]); !gsi_end_p (bsi
); gsi_next (&bsi
))
2457 if (stmt_can_terminate_bb_p (gsi_stmt (bsi
)))
2463 /* Stores description of number of iterations of LOOP derived from
2464 EXIT (an exit edge of the LOOP) in NITER. Returns true if some useful
2465 information could be derived (and fields of NITER have meaning described
2466 in comments at class tree_niter_desc declaration), false otherwise.
2467 When EVERY_ITERATION is true, only tests that are known to be executed
2468 every iteration are considered (i.e. only test that alone bounds the loop).
2469 If AT_STMT is not NULL, this function stores LOOP's condition statement in
2470 it when returning true. */
2473 number_of_iterations_exit_assumptions (class loop
*loop
, edge exit
,
2474 class tree_niter_desc
*niter
,
2475 gcond
**at_stmt
, bool every_iteration
,
2482 enum tree_code code
;
2486 /* The condition at a fake exit (if it exists) does not control its
2488 if (exit
->flags
& EDGE_FAKE
)
2491 /* Nothing to analyze if the loop is known to be infinite. */
2492 if (loop_constraint_set_p (loop
, LOOP_C_INFINITE
))
2495 safe
= dominated_by_p (CDI_DOMINATORS
, loop
->latch
, exit
->src
);
2497 if (every_iteration
&& !safe
)
2500 niter
->assumptions
= boolean_false_node
;
2501 niter
->control
.base
= NULL_TREE
;
2502 niter
->control
.step
= NULL_TREE
;
2503 niter
->control
.no_overflow
= false;
2504 last
= last_stmt (exit
->src
);
2507 stmt
= dyn_cast
<gcond
*> (last
);
2511 /* We want the condition for staying inside loop. */
2512 code
= gimple_cond_code (stmt
);
2513 if (exit
->flags
& EDGE_TRUE_VALUE
)
2514 code
= invert_tree_comparison (code
, false);
2529 op0
= gimple_cond_lhs (stmt
);
2530 op1
= gimple_cond_rhs (stmt
);
2531 type
= TREE_TYPE (op0
);
2533 if (TREE_CODE (type
) != INTEGER_TYPE
2534 && !POINTER_TYPE_P (type
))
2537 tree iv0_niters
= NULL_TREE
;
2538 if (!simple_iv_with_niters (loop
, loop_containing_stmt (stmt
),
2539 op0
, &iv0
, safe
? &iv0_niters
: NULL
, false))
2540 return number_of_iterations_popcount (loop
, exit
, code
, niter
);
2541 tree iv1_niters
= NULL_TREE
;
2542 if (!simple_iv_with_niters (loop
, loop_containing_stmt (stmt
),
2543 op1
, &iv1
, safe
? &iv1_niters
: NULL
, false))
2545 /* Give up on complicated case. */
2546 if (iv0_niters
&& iv1_niters
)
2549 /* We don't want to see undefined signed overflow warnings while
2550 computing the number of iterations. */
2551 fold_defer_overflow_warnings ();
2553 iv0
.base
= expand_simple_operations (iv0
.base
);
2554 iv1
.base
= expand_simple_operations (iv1
.base
);
2555 bool body_from_caller
= true;
2558 body
= get_loop_body (loop
);
2559 body_from_caller
= false;
2561 bool only_exit_p
= loop_only_exit_p (loop
, body
, exit
);
2562 if (!body_from_caller
)
2564 if (!number_of_iterations_cond (loop
, type
, &iv0
, code
, &iv1
, niter
,
2567 fold_undefer_and_ignore_overflow_warnings ();
2571 /* Incorporate additional assumption implied by control iv. */
2572 tree iv_niters
= iv0_niters
? iv0_niters
: iv1_niters
;
2575 tree assumption
= fold_build2 (LE_EXPR
, boolean_type_node
, niter
->niter
,
2576 fold_convert (TREE_TYPE (niter
->niter
),
2579 if (!integer_nonzerop (assumption
))
2580 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
2581 niter
->assumptions
, assumption
);
2583 /* Refine upper bound if possible. */
2584 if (TREE_CODE (iv_niters
) == INTEGER_CST
2585 && niter
->max
> wi::to_widest (iv_niters
))
2586 niter
->max
= wi::to_widest (iv_niters
);
2589 /* There is no assumptions if the loop is known to be finite. */
2590 if (!integer_zerop (niter
->assumptions
)
2591 && loop_constraint_set_p (loop
, LOOP_C_FINITE
))
2592 niter
->assumptions
= boolean_true_node
;
2596 niter
->assumptions
= simplify_using_outer_evolutions (loop
,
2597 niter
->assumptions
);
2598 niter
->may_be_zero
= simplify_using_outer_evolutions (loop
,
2599 niter
->may_be_zero
);
2600 niter
->niter
= simplify_using_outer_evolutions (loop
, niter
->niter
);
2604 = simplify_using_initial_conditions (loop
,
2605 niter
->assumptions
);
2607 = simplify_using_initial_conditions (loop
,
2608 niter
->may_be_zero
);
2610 fold_undefer_and_ignore_overflow_warnings ();
2612 /* If NITER has simplified into a constant, update MAX. */
2613 if (TREE_CODE (niter
->niter
) == INTEGER_CST
)
2614 niter
->max
= wi::to_widest (niter
->niter
);
2619 return (!integer_zerop (niter
->assumptions
));
2623 /* Utility function to check if OP is defined by a stmt
2624 that is a val - 1. */
2627 ssa_defined_by_minus_one_stmt_p (tree op
, tree val
)
2630 return (TREE_CODE (op
) == SSA_NAME
2631 && (stmt
= SSA_NAME_DEF_STMT (op
))
2632 && is_gimple_assign (stmt
)
2633 && (gimple_assign_rhs_code (stmt
) == PLUS_EXPR
)
2634 && val
== gimple_assign_rhs1 (stmt
)
2635 && integer_minus_onep (gimple_assign_rhs2 (stmt
)));
2639 /* See if LOOP is a popcout implementation, determine NITER for the loop
2650 b_11 = PHI <b_5(D)(2), b_6(3)>
2658 OR we match copy-header version:
2665 b_11 = PHI <b_5(2), b_6(3)>
2675 If popcount pattern, update NITER accordingly.
2676 i.e., set NITER to __builtin_popcount (b)
2677 return true if we did, false otherwise.
2682 number_of_iterations_popcount (loop_p loop
, edge exit
,
2683 enum tree_code code
,
2684 class tree_niter_desc
*niter
)
2690 tree fn
= NULL_TREE
;
2692 /* Check loop terminating branch is like
2694 gimple
*stmt
= last_stmt (exit
->src
);
2696 || gimple_code (stmt
) != GIMPLE_COND
2698 || !integer_zerop (gimple_cond_rhs (stmt
))
2699 || TREE_CODE (gimple_cond_lhs (stmt
)) != SSA_NAME
)
2702 gimple
*and_stmt
= SSA_NAME_DEF_STMT (gimple_cond_lhs (stmt
));
2704 /* Depending on copy-header is performed, feeding PHI stmts might be in
2705 the loop header or loop latch, handle this. */
2706 if (gimple_code (and_stmt
) == GIMPLE_PHI
2707 && gimple_bb (and_stmt
) == loop
->header
2708 && gimple_phi_num_args (and_stmt
) == 2
2709 && (TREE_CODE (gimple_phi_arg_def (and_stmt
,
2710 loop_latch_edge (loop
)->dest_idx
))
2713 /* SSA used in exit condition is defined by PHI stmt
2714 b_11 = PHI <b_5(D)(2), b_6(3)>
2715 from the PHI stmt, get the and_stmt
2717 tree t
= gimple_phi_arg_def (and_stmt
, loop_latch_edge (loop
)->dest_idx
);
2718 and_stmt
= SSA_NAME_DEF_STMT (t
);
2722 /* Make sure it is indeed an and stmt (b_6 = _1 & b_11). */
2723 if (!is_gimple_assign (and_stmt
)
2724 || gimple_assign_rhs_code (and_stmt
) != BIT_AND_EXPR
)
2727 tree b_11
= gimple_assign_rhs1 (and_stmt
);
2728 tree _1
= gimple_assign_rhs2 (and_stmt
);
2730 /* Check that _1 is defined by _b11 + -1 (_1 = b_11 + -1).
2731 Also make sure that b_11 is the same in and_stmt and _1 defining stmt.
2732 Also canonicalize if _1 and _b11 are revrsed. */
2733 if (ssa_defined_by_minus_one_stmt_p (b_11
, _1
))
2734 std::swap (b_11
, _1
);
2735 else if (ssa_defined_by_minus_one_stmt_p (_1
, b_11
))
2739 /* Check the recurrence:
2740 ... = PHI <b_5(2), b_6(3)>. */
2741 gimple
*phi
= SSA_NAME_DEF_STMT (b_11
);
2742 if (gimple_code (phi
) != GIMPLE_PHI
2743 || (gimple_bb (phi
) != loop_latch_edge (loop
)->dest
)
2744 || (gimple_assign_lhs (and_stmt
)
2745 != gimple_phi_arg_def (phi
, loop_latch_edge (loop
)->dest_idx
)))
2748 /* We found a match. Get the corresponding popcount builtin. */
2749 tree src
= gimple_phi_arg_def (phi
, loop_preheader_edge (loop
)->dest_idx
);
2750 if (TYPE_PRECISION (TREE_TYPE (src
)) <= TYPE_PRECISION (integer_type_node
))
2751 fn
= builtin_decl_implicit (BUILT_IN_POPCOUNT
);
2752 else if (TYPE_PRECISION (TREE_TYPE (src
))
2753 == TYPE_PRECISION (long_integer_type_node
))
2754 fn
= builtin_decl_implicit (BUILT_IN_POPCOUNTL
);
2755 else if (TYPE_PRECISION (TREE_TYPE (src
))
2756 == TYPE_PRECISION (long_long_integer_type_node
)
2757 || (TYPE_PRECISION (TREE_TYPE (src
))
2758 == 2 * TYPE_PRECISION (long_long_integer_type_node
)))
2759 fn
= builtin_decl_implicit (BUILT_IN_POPCOUNTLL
);
2764 /* Update NITER params accordingly */
2765 tree utype
= unsigned_type_for (TREE_TYPE (src
));
2766 src
= fold_convert (utype
, src
);
2767 if (TYPE_PRECISION (TREE_TYPE (src
)) < TYPE_PRECISION (integer_type_node
))
2768 src
= fold_convert (unsigned_type_node
, src
);
2770 if (TYPE_PRECISION (TREE_TYPE (src
))
2771 == 2 * TYPE_PRECISION (long_long_integer_type_node
))
2773 int prec
= TYPE_PRECISION (long_long_integer_type_node
);
2774 tree src1
= fold_convert (long_long_unsigned_type_node
,
2775 fold_build2 (RSHIFT_EXPR
, TREE_TYPE (src
),
2777 build_int_cst (integer_type_node
,
2779 tree src2
= fold_convert (long_long_unsigned_type_node
, src
);
2780 call
= build_call_expr (fn
, 1, src1
);
2781 call
= fold_build2 (PLUS_EXPR
, TREE_TYPE (call
), call
,
2782 build_call_expr (fn
, 1, src2
));
2783 call
= fold_convert (utype
, call
);
2786 call
= fold_convert (utype
, build_call_expr (fn
, 1, src
));
2788 iter
= fold_build2 (MINUS_EXPR
, utype
, call
, build_int_cst (utype
, 1));
2792 if (TREE_CODE (call
) == INTEGER_CST
)
2793 max
= tree_to_uhwi (call
);
2795 max
= TYPE_PRECISION (TREE_TYPE (src
));
2799 niter
->niter
= iter
;
2800 niter
->assumptions
= boolean_true_node
;
2804 tree may_be_zero
= fold_build2 (EQ_EXPR
, boolean_type_node
, src
,
2805 build_zero_cst (TREE_TYPE (src
)));
2807 = simplify_using_initial_conditions (loop
, may_be_zero
);
2810 niter
->may_be_zero
= boolean_false_node
;
2813 niter
->bound
= NULL_TREE
;
2814 niter
->cmp
= ERROR_MARK
;
2819 /* Like number_of_iterations_exit_assumptions, but return TRUE only if
2820 the niter information holds unconditionally. */
2823 number_of_iterations_exit (class loop
*loop
, edge exit
,
2824 class tree_niter_desc
*niter
,
2825 bool warn
, bool every_iteration
,
2829 if (!number_of_iterations_exit_assumptions (loop
, exit
, niter
,
2830 &stmt
, every_iteration
, body
))
2833 if (integer_nonzerop (niter
->assumptions
))
2836 if (warn
&& dump_enabled_p ())
2837 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, stmt
,
2838 "missed loop optimization: niters analysis ends up "
2839 "with assumptions.\n");
2844 /* Try to determine the number of iterations of LOOP. If we succeed,
2845 expression giving number of iterations is returned and *EXIT is
2846 set to the edge from that the information is obtained. Otherwise
2847 chrec_dont_know is returned. */
2850 find_loop_niter (class loop
*loop
, edge
*exit
)
2853 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
2855 tree niter
= NULL_TREE
, aniter
;
2856 class tree_niter_desc desc
;
2859 FOR_EACH_VEC_ELT (exits
, i
, ex
)
2861 if (!number_of_iterations_exit (loop
, ex
, &desc
, false))
2864 if (integer_nonzerop (desc
.may_be_zero
))
2866 /* We exit in the first iteration through this exit.
2867 We won't find anything better. */
2868 niter
= build_int_cst (unsigned_type_node
, 0);
2873 if (!integer_zerop (desc
.may_be_zero
))
2876 aniter
= desc
.niter
;
2880 /* Nothing recorded yet. */
2886 /* Prefer constants, the lower the better. */
2887 if (TREE_CODE (aniter
) != INTEGER_CST
)
2890 if (TREE_CODE (niter
) != INTEGER_CST
)
2897 if (tree_int_cst_lt (aniter
, niter
))
2905 return niter
? niter
: chrec_dont_know
;
2908 /* Return true if loop is known to have bounded number of iterations. */
2911 finite_loop_p (class loop
*loop
)
2916 flags
= flags_from_decl_or_type (current_function_decl
);
2917 if ((flags
& (ECF_CONST
|ECF_PURE
)) && !(flags
& ECF_LOOPING_CONST_OR_PURE
))
2919 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2920 fprintf (dump_file
, "Found loop %i to be finite: it is within pure or const function.\n",
2925 if (loop
->any_upper_bound
2926 || max_loop_iterations (loop
, &nit
))
2928 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2929 fprintf (dump_file
, "Found loop %i to be finite: upper bound found.\n",
2937 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
2940 /* If the loop has a normal exit, we can assume it will terminate. */
2941 FOR_EACH_VEC_ELT (exits
, i
, ex
)
2942 if (!(ex
->flags
& (EDGE_EH
| EDGE_ABNORMAL
| EDGE_FAKE
)))
2945 fprintf (dump_file
, "Assume loop %i to be finite: it has an exit "
2946 "and -ffinite-loops is on.\n", loop
->num
);
2956 Analysis of a number of iterations of a loop by a brute-force evaluation.
2960 /* Bound on the number of iterations we try to evaluate. */
2962 #define MAX_ITERATIONS_TO_TRACK \
2963 ((unsigned) param_max_iterations_to_track)
2965 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
2966 result by a chain of operations such that all but exactly one of their
2967 operands are constants. */
2970 chain_of_csts_start (class loop
*loop
, tree x
)
2972 gimple
*stmt
= SSA_NAME_DEF_STMT (x
);
2974 basic_block bb
= gimple_bb (stmt
);
2975 enum tree_code code
;
2978 || !flow_bb_inside_loop_p (loop
, bb
))
2981 if (gimple_code (stmt
) == GIMPLE_PHI
)
2983 if (bb
== loop
->header
)
2984 return as_a
<gphi
*> (stmt
);
2989 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2990 || gimple_assign_rhs_class (stmt
) == GIMPLE_TERNARY_RHS
)
2993 code
= gimple_assign_rhs_code (stmt
);
2994 if (gimple_references_memory_p (stmt
)
2995 || TREE_CODE_CLASS (code
) == tcc_reference
2996 || (code
== ADDR_EXPR
2997 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt
))))
3000 use
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_USE
);
3001 if (use
== NULL_TREE
)
3004 return chain_of_csts_start (loop
, use
);
3007 /* Determines whether the expression X is derived from a result of a phi node
3008 in header of LOOP such that
3010 * the derivation of X consists only from operations with constants
3011 * the initial value of the phi node is constant
3012 * the value of the phi node in the next iteration can be derived from the
3013 value in the current iteration by a chain of operations with constants,
3014 or is also a constant
3016 If such phi node exists, it is returned, otherwise NULL is returned. */
3019 get_base_for (class loop
*loop
, tree x
)
3024 if (is_gimple_min_invariant (x
))
3027 phi
= chain_of_csts_start (loop
, x
);
3031 init
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
3032 next
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (loop
));
3034 if (!is_gimple_min_invariant (init
))
3037 if (TREE_CODE (next
) == SSA_NAME
3038 && chain_of_csts_start (loop
, next
) != phi
)
3044 /* Given an expression X, then
3046 * if X is NULL_TREE, we return the constant BASE.
3047 * if X is a constant, we return the constant X.
3048 * otherwise X is a SSA name, whose value in the considered loop is derived
3049 by a chain of operations with constant from a result of a phi node in
3050 the header of the loop. Then we return value of X when the value of the
3051 result of this phi node is given by the constant BASE. */
3054 get_val_for (tree x
, tree base
)
3058 gcc_checking_assert (is_gimple_min_invariant (base
));
3062 else if (is_gimple_min_invariant (x
))
3065 stmt
= SSA_NAME_DEF_STMT (x
);
3066 if (gimple_code (stmt
) == GIMPLE_PHI
)
3069 gcc_checking_assert (is_gimple_assign (stmt
));
3071 /* STMT must be either an assignment of a single SSA name or an
3072 expression involving an SSA name and a constant. Try to fold that
3073 expression using the value for the SSA name. */
3074 if (gimple_assign_ssa_name_copy_p (stmt
))
3075 return get_val_for (gimple_assign_rhs1 (stmt
), base
);
3076 else if (gimple_assign_rhs_class (stmt
) == GIMPLE_UNARY_RHS
3077 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
)
3078 return fold_build1 (gimple_assign_rhs_code (stmt
),
3079 TREE_TYPE (gimple_assign_lhs (stmt
)),
3080 get_val_for (gimple_assign_rhs1 (stmt
), base
));
3081 else if (gimple_assign_rhs_class (stmt
) == GIMPLE_BINARY_RHS
)
3083 tree rhs1
= gimple_assign_rhs1 (stmt
);
3084 tree rhs2
= gimple_assign_rhs2 (stmt
);
3085 if (TREE_CODE (rhs1
) == SSA_NAME
)
3086 rhs1
= get_val_for (rhs1
, base
);
3087 else if (TREE_CODE (rhs2
) == SSA_NAME
)
3088 rhs2
= get_val_for (rhs2
, base
);
3091 return fold_build2 (gimple_assign_rhs_code (stmt
),
3092 TREE_TYPE (gimple_assign_lhs (stmt
)), rhs1
, rhs2
);
3099 /* Tries to count the number of iterations of LOOP till it exits by EXIT
3100 by brute force -- i.e. by determining the value of the operands of the
3101 condition at EXIT in first few iterations of the loop (assuming that
3102 these values are constant) and determining the first one in that the
3103 condition is not satisfied. Returns the constant giving the number
3104 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
3107 loop_niter_by_eval (class loop
*loop
, edge exit
)
3110 tree op
[2], val
[2], next
[2], aval
[2];
3116 cond
= last_stmt (exit
->src
);
3117 if (!cond
|| gimple_code (cond
) != GIMPLE_COND
)
3118 return chrec_dont_know
;
3120 cmp
= gimple_cond_code (cond
);
3121 if (exit
->flags
& EDGE_TRUE_VALUE
)
3122 cmp
= invert_tree_comparison (cmp
, false);
3132 op
[0] = gimple_cond_lhs (cond
);
3133 op
[1] = gimple_cond_rhs (cond
);
3137 return chrec_dont_know
;
3140 for (j
= 0; j
< 2; j
++)
3142 if (is_gimple_min_invariant (op
[j
]))
3145 next
[j
] = NULL_TREE
;
3150 phi
= get_base_for (loop
, op
[j
]);
3152 return chrec_dont_know
;
3153 val
[j
] = PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
3154 next
[j
] = PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (loop
));
3158 /* Don't issue signed overflow warnings. */
3159 fold_defer_overflow_warnings ();
3161 for (i
= 0; i
< MAX_ITERATIONS_TO_TRACK
; i
++)
3163 for (j
= 0; j
< 2; j
++)
3164 aval
[j
] = get_val_for (op
[j
], val
[j
]);
3166 acnd
= fold_binary (cmp
, boolean_type_node
, aval
[0], aval
[1]);
3167 if (acnd
&& integer_zerop (acnd
))
3169 fold_undefer_and_ignore_overflow_warnings ();
3170 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3172 "Proved that loop %d iterates %d times using brute force.\n",
3174 return build_int_cst (unsigned_type_node
, i
);
3177 for (j
= 0; j
< 2; j
++)
3180 val
[j
] = get_val_for (next
[j
], val
[j
]);
3181 if (!is_gimple_min_invariant (val
[j
]))
3183 fold_undefer_and_ignore_overflow_warnings ();
3184 return chrec_dont_know
;
3188 /* If the next iteration would use the same base values
3189 as the current one, there is no point looping further,
3190 all following iterations will be the same as this one. */
3191 if (val
[0] == aval
[0] && val
[1] == aval
[1])
3195 fold_undefer_and_ignore_overflow_warnings ();
3197 return chrec_dont_know
;
3200 /* Finds the exit of the LOOP by that the loop exits after a constant
3201 number of iterations and stores the exit edge to *EXIT. The constant
3202 giving the number of iterations of LOOP is returned. The number of
3203 iterations is determined using loop_niter_by_eval (i.e. by brute force
3204 evaluation). If we are unable to find the exit for that loop_niter_by_eval
3205 determines the number of iterations, chrec_dont_know is returned. */
3208 find_loop_niter_by_eval (class loop
*loop
, edge
*exit
)
3211 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
3213 tree niter
= NULL_TREE
, aniter
;
3217 /* Loops with multiple exits are expensive to handle and less important. */
3218 if (!flag_expensive_optimizations
3219 && exits
.length () > 1)
3220 return chrec_dont_know
;
3222 FOR_EACH_VEC_ELT (exits
, i
, ex
)
3224 if (!just_once_each_iteration_p (loop
, ex
->src
))
3227 aniter
= loop_niter_by_eval (loop
, ex
);
3228 if (chrec_contains_undetermined (aniter
))
3232 && !tree_int_cst_lt (aniter
, niter
))
3239 return niter
? niter
: chrec_dont_know
;
3244 Analysis of upper bounds on number of iterations of a loop.
3248 static widest_int
derive_constant_upper_bound_ops (tree
, tree
,
3249 enum tree_code
, tree
);
3251 /* Returns a constant upper bound on the value of the right-hand side of
3252 an assignment statement STMT. */
3255 derive_constant_upper_bound_assign (gimple
*stmt
)
3257 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3258 tree op0
= gimple_assign_rhs1 (stmt
);
3259 tree op1
= gimple_assign_rhs2 (stmt
);
3261 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt
)),
3265 /* Returns a constant upper bound on the value of expression VAL. VAL
3266 is considered to be unsigned. If its type is signed, its value must
3270 derive_constant_upper_bound (tree val
)
3272 enum tree_code code
;
3275 extract_ops_from_tree (val
, &code
, &op0
, &op1
, &op2
);
3276 return derive_constant_upper_bound_ops (TREE_TYPE (val
), op0
, code
, op1
);
3279 /* Returns a constant upper bound on the value of expression OP0 CODE OP1,
3280 whose type is TYPE. The expression is considered to be unsigned. If
3281 its type is signed, its value must be nonnegative. */
3284 derive_constant_upper_bound_ops (tree type
, tree op0
,
3285 enum tree_code code
, tree op1
)
3288 widest_int bnd
, max
, cst
;
3291 if (INTEGRAL_TYPE_P (type
))
3292 maxt
= TYPE_MAX_VALUE (type
);
3294 maxt
= upper_bound_in_type (type
, type
);
3296 max
= wi::to_widest (maxt
);
3301 return wi::to_widest (op0
);
3304 subtype
= TREE_TYPE (op0
);
3305 if (!TYPE_UNSIGNED (subtype
)
3306 /* If TYPE is also signed, the fact that VAL is nonnegative implies
3307 that OP0 is nonnegative. */
3308 && TYPE_UNSIGNED (type
)
3309 && !tree_expr_nonnegative_p (op0
))
3311 /* If we cannot prove that the casted expression is nonnegative,
3312 we cannot establish more useful upper bound than the precision
3313 of the type gives us. */
3317 /* We now know that op0 is an nonnegative value. Try deriving an upper
3319 bnd
= derive_constant_upper_bound (op0
);
3321 /* If the bound does not fit in TYPE, max. value of TYPE could be
3323 if (wi::ltu_p (max
, bnd
))
3329 case POINTER_PLUS_EXPR
:
3331 if (TREE_CODE (op1
) != INTEGER_CST
3332 || !tree_expr_nonnegative_p (op0
))
3335 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
3336 choose the most logical way how to treat this constant regardless
3337 of the signedness of the type. */
3338 cst
= wi::sext (wi::to_widest (op1
), TYPE_PRECISION (type
));
3339 if (code
!= MINUS_EXPR
)
3342 bnd
= derive_constant_upper_bound (op0
);
3344 if (wi::neg_p (cst
))
3347 /* Avoid CST == 0x80000... */
3348 if (wi::neg_p (cst
))
3351 /* OP0 + CST. We need to check that
3352 BND <= MAX (type) - CST. */
3354 widest_int mmax
= max
- cst
;
3355 if (wi::leu_p (bnd
, mmax
))
3362 /* OP0 - CST, where CST >= 0.
3364 If TYPE is signed, we have already verified that OP0 >= 0, and we
3365 know that the result is nonnegative. This implies that
3368 If TYPE is unsigned, we must additionally know that OP0 >= CST,
3369 otherwise the operation underflows.
3372 /* This should only happen if the type is unsigned; however, for
3373 buggy programs that use overflowing signed arithmetics even with
3374 -fno-wrapv, this condition may also be true for signed values. */
3375 if (wi::ltu_p (bnd
, cst
))
3378 if (TYPE_UNSIGNED (type
))
3380 tree tem
= fold_binary (GE_EXPR
, boolean_type_node
, op0
,
3381 wide_int_to_tree (type
, cst
));
3382 if (!tem
|| integer_nonzerop (tem
))
3391 case FLOOR_DIV_EXPR
:
3392 case EXACT_DIV_EXPR
:
3393 if (TREE_CODE (op1
) != INTEGER_CST
3394 || tree_int_cst_sign_bit (op1
))
3397 bnd
= derive_constant_upper_bound (op0
);
3398 return wi::udiv_floor (bnd
, wi::to_widest (op1
));
3401 if (TREE_CODE (op1
) != INTEGER_CST
3402 || tree_int_cst_sign_bit (op1
))
3404 return wi::to_widest (op1
);
3407 stmt
= SSA_NAME_DEF_STMT (op0
);
3408 if (gimple_code (stmt
) != GIMPLE_ASSIGN
3409 || gimple_assign_lhs (stmt
) != op0
)
3411 return derive_constant_upper_bound_assign (stmt
);
3418 /* Emit a -Waggressive-loop-optimizations warning if needed. */
3421 do_warn_aggressive_loop_optimizations (class loop
*loop
,
3422 widest_int i_bound
, gimple
*stmt
)
3424 /* Don't warn if the loop doesn't have known constant bound. */
3425 if (!loop
->nb_iterations
3426 || TREE_CODE (loop
->nb_iterations
) != INTEGER_CST
3427 || !warn_aggressive_loop_optimizations
3428 /* To avoid warning multiple times for the same loop,
3429 only start warning when we preserve loops. */
3430 || (cfun
->curr_properties
& PROP_loops
) == 0
3431 /* Only warn once per loop. */
3432 || loop
->warned_aggressive_loop_optimizations
3433 /* Only warn if undefined behavior gives us lower estimate than the
3434 known constant bound. */
3435 || wi::cmpu (i_bound
, wi::to_widest (loop
->nb_iterations
)) >= 0
3436 /* And undefined behavior happens unconditionally. */
3437 || !dominated_by_p (CDI_DOMINATORS
, loop
->latch
, gimple_bb (stmt
)))
3440 edge e
= single_exit (loop
);
3444 gimple
*estmt
= last_stmt (e
->src
);
3445 char buf
[WIDE_INT_PRINT_BUFFER_SIZE
];
3446 print_dec (i_bound
, buf
, TYPE_UNSIGNED (TREE_TYPE (loop
->nb_iterations
))
3447 ? UNSIGNED
: SIGNED
);
3448 auto_diagnostic_group d
;
3449 if (warning_at (gimple_location (stmt
), OPT_Waggressive_loop_optimizations
,
3450 "iteration %s invokes undefined behavior", buf
))
3451 inform (gimple_location (estmt
), "within this loop");
3452 loop
->warned_aggressive_loop_optimizations
= true;
3455 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
3456 is true if the loop is exited immediately after STMT, and this exit
3457 is taken at last when the STMT is executed BOUND + 1 times.
3458 REALISTIC is true if BOUND is expected to be close to the real number
3459 of iterations. UPPER is true if we are sure the loop iterates at most
3460 BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */
3463 record_estimate (class loop
*loop
, tree bound
, const widest_int
&i_bound
,
3464 gimple
*at_stmt
, bool is_exit
, bool realistic
, bool upper
)
3468 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3470 fprintf (dump_file
, "Statement %s", is_exit
? "(exit)" : "");
3471 print_gimple_stmt (dump_file
, at_stmt
, 0, TDF_SLIM
);
3472 fprintf (dump_file
, " is %sexecuted at most ",
3473 upper
? "" : "probably ");
3474 print_generic_expr (dump_file
, bound
, TDF_SLIM
);
3475 fprintf (dump_file
, " (bounded by ");
3476 print_decu (i_bound
, dump_file
);
3477 fprintf (dump_file
, ") + 1 times in loop %d.\n", loop
->num
);
3480 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
3481 real number of iterations. */
3482 if (TREE_CODE (bound
) != INTEGER_CST
)
3485 gcc_checking_assert (i_bound
== wi::to_widest (bound
));
3487 /* If we have a guaranteed upper bound, record it in the appropriate
3488 list, unless this is an !is_exit bound (i.e. undefined behavior in
3489 at_stmt) in a loop with known constant number of iterations. */
3492 || loop
->nb_iterations
== NULL_TREE
3493 || TREE_CODE (loop
->nb_iterations
) != INTEGER_CST
))
3495 class nb_iter_bound
*elt
= ggc_alloc
<nb_iter_bound
> ();
3497 elt
->bound
= i_bound
;
3498 elt
->stmt
= at_stmt
;
3499 elt
->is_exit
= is_exit
;
3500 elt
->next
= loop
->bounds
;
3504 /* If statement is executed on every path to the loop latch, we can directly
3505 infer the upper bound on the # of iterations of the loop. */
3506 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, gimple_bb (at_stmt
)))
3509 /* Update the number of iteration estimates according to the bound.
3510 If at_stmt is an exit then the loop latch is executed at most BOUND times,
3511 otherwise it can be executed BOUND + 1 times. We will lower the estimate
3512 later if such statement must be executed on last iteration */
3517 widest_int new_i_bound
= i_bound
+ delta
;
3519 /* If an overflow occurred, ignore the result. */
3520 if (wi::ltu_p (new_i_bound
, delta
))
3523 if (upper
&& !is_exit
)
3524 do_warn_aggressive_loop_optimizations (loop
, new_i_bound
, at_stmt
);
3525 record_niter_bound (loop
, new_i_bound
, realistic
, upper
);
3528 /* Records the control iv analyzed in NITER for LOOP if the iv is valid
3529 and doesn't overflow. */
3532 record_control_iv (class loop
*loop
, class tree_niter_desc
*niter
)
3534 struct control_iv
*iv
;
3536 if (!niter
->control
.base
|| !niter
->control
.step
)
3539 if (!integer_onep (niter
->assumptions
) || !niter
->control
.no_overflow
)
3542 iv
= ggc_alloc
<control_iv
> ();
3543 iv
->base
= niter
->control
.base
;
3544 iv
->step
= niter
->control
.step
;
3545 iv
->next
= loop
->control_ivs
;
3546 loop
->control_ivs
= iv
;
3551 /* This function returns TRUE if below conditions are satisfied:
3552 1) VAR is SSA variable.
3553 2) VAR is an IV:{base, step} in its defining loop.
3554 3) IV doesn't overflow.
3555 4) Both base and step are integer constants.
3556 5) Base is the MIN/MAX value depends on IS_MIN.
3557 Store value of base to INIT correspondingly. */
3560 get_cst_init_from_scev (tree var
, wide_int
*init
, bool is_min
)
3562 if (TREE_CODE (var
) != SSA_NAME
)
3565 gimple
*def_stmt
= SSA_NAME_DEF_STMT (var
);
3566 class loop
*loop
= loop_containing_stmt (def_stmt
);
3572 if (!simple_iv (loop
, loop
, var
, &iv
, false))
3575 if (!iv
.no_overflow
)
3578 if (TREE_CODE (iv
.base
) != INTEGER_CST
|| TREE_CODE (iv
.step
) != INTEGER_CST
)
3581 if (is_min
== tree_int_cst_sign_bit (iv
.step
))
3584 *init
= wi::to_wide (iv
.base
);
3588 /* Record the estimate on number of iterations of LOOP based on the fact that
3589 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
3590 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
3591 estimated number of iterations is expected to be close to the real one.
3592 UPPER is true if we are sure the induction variable does not wrap. */
3595 record_nonwrapping_iv (class loop
*loop
, tree base
, tree step
, gimple
*stmt
,
3596 tree low
, tree high
, bool realistic
, bool upper
)
3598 tree niter_bound
, extreme
, delta
;
3599 tree type
= TREE_TYPE (base
), unsigned_type
;
3600 tree orig_base
= base
;
3602 if (TREE_CODE (step
) != INTEGER_CST
|| integer_zerop (step
))
3605 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3607 fprintf (dump_file
, "Induction variable (");
3608 print_generic_expr (dump_file
, TREE_TYPE (base
), TDF_SLIM
);
3609 fprintf (dump_file
, ") ");
3610 print_generic_expr (dump_file
, base
, TDF_SLIM
);
3611 fprintf (dump_file
, " + ");
3612 print_generic_expr (dump_file
, step
, TDF_SLIM
);
3613 fprintf (dump_file
, " * iteration does not wrap in statement ");
3614 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
3615 fprintf (dump_file
, " in loop %d.\n", loop
->num
);
3618 unsigned_type
= unsigned_type_for (type
);
3619 base
= fold_convert (unsigned_type
, base
);
3620 step
= fold_convert (unsigned_type
, step
);
3622 if (tree_int_cst_sign_bit (step
))
3625 value_range base_range
;
3626 if (get_range_query (cfun
)->range_of_expr (base_range
, orig_base
)
3627 && !base_range
.undefined_p ())
3628 max
= base_range
.upper_bound ();
3629 extreme
= fold_convert (unsigned_type
, low
);
3630 if (TREE_CODE (orig_base
) == SSA_NAME
3631 && TREE_CODE (high
) == INTEGER_CST
3632 && INTEGRAL_TYPE_P (TREE_TYPE (orig_base
))
3633 && (base_range
.kind () == VR_RANGE
3634 || get_cst_init_from_scev (orig_base
, &max
, false))
3635 && wi::gts_p (wi::to_wide (high
), max
))
3636 base
= wide_int_to_tree (unsigned_type
, max
);
3637 else if (TREE_CODE (base
) != INTEGER_CST
3638 && dominated_by_p (CDI_DOMINATORS
,
3639 loop
->latch
, gimple_bb (stmt
)))
3640 base
= fold_convert (unsigned_type
, high
);
3641 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, base
, extreme
);
3642 step
= fold_build1 (NEGATE_EXPR
, unsigned_type
, step
);
3647 value_range base_range
;
3648 if (get_range_query (cfun
)->range_of_expr (base_range
, orig_base
)
3649 && !base_range
.undefined_p ())
3650 min
= base_range
.lower_bound ();
3651 extreme
= fold_convert (unsigned_type
, high
);
3652 if (TREE_CODE (orig_base
) == SSA_NAME
3653 && TREE_CODE (low
) == INTEGER_CST
3654 && INTEGRAL_TYPE_P (TREE_TYPE (orig_base
))
3655 && (base_range
.kind () == VR_RANGE
3656 || get_cst_init_from_scev (orig_base
, &min
, true))
3657 && wi::gts_p (min
, wi::to_wide (low
)))
3658 base
= wide_int_to_tree (unsigned_type
, min
);
3659 else if (TREE_CODE (base
) != INTEGER_CST
3660 && dominated_by_p (CDI_DOMINATORS
,
3661 loop
->latch
, gimple_bb (stmt
)))
3662 base
= fold_convert (unsigned_type
, low
);
3663 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, extreme
, base
);
3666 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
3667 would get out of the range. */
3668 niter_bound
= fold_build2 (FLOOR_DIV_EXPR
, unsigned_type
, delta
, step
);
3669 widest_int max
= derive_constant_upper_bound (niter_bound
);
3670 record_estimate (loop
, niter_bound
, max
, stmt
, false, realistic
, upper
);
3673 /* Determine information about number of iterations a LOOP from the index
3674 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
3675 guaranteed to be executed in every iteration of LOOP. Callback for
3685 idx_infer_loop_bounds (tree base
, tree
*idx
, void *dta
)
3687 struct ilb_data
*data
= (struct ilb_data
*) dta
;
3688 tree ev
, init
, step
;
3689 tree low
, high
, type
, next
;
3690 bool sign
, upper
= true, at_end
= false;
3691 class loop
*loop
= data
->loop
;
3693 if (TREE_CODE (base
) != ARRAY_REF
)
3696 /* For arrays at the end of the structure, we are not guaranteed that they
3697 do not really extend over their declared size. However, for arrays of
3698 size greater than one, this is unlikely to be intended. */
3699 if (array_at_struct_end_p (base
))
3705 class loop
*dloop
= loop_containing_stmt (data
->stmt
);
3709 ev
= analyze_scalar_evolution (dloop
, *idx
);
3710 ev
= instantiate_parameters (loop
, ev
);
3711 init
= initial_condition (ev
);
3712 step
= evolution_part_in_loop_num (ev
, loop
->num
);
3716 || TREE_CODE (step
) != INTEGER_CST
3717 || integer_zerop (step
)
3718 || tree_contains_chrecs (init
, NULL
)
3719 || chrec_contains_symbols_defined_in_loop (init
, loop
->num
))
3722 low
= array_ref_low_bound (base
);
3723 high
= array_ref_up_bound (base
);
3725 /* The case of nonconstant bounds could be handled, but it would be
3727 if (TREE_CODE (low
) != INTEGER_CST
3729 || TREE_CODE (high
) != INTEGER_CST
)
3731 sign
= tree_int_cst_sign_bit (step
);
3732 type
= TREE_TYPE (step
);
3734 /* The array of length 1 at the end of a structure most likely extends
3735 beyond its bounds. */
3737 && operand_equal_p (low
, high
, 0))
3740 /* In case the relevant bound of the array does not fit in type, or
3741 it does, but bound + step (in type) still belongs into the range of the
3742 array, the index may wrap and still stay within the range of the array
3743 (consider e.g. if the array is indexed by the full range of
3746 To make things simpler, we require both bounds to fit into type, although
3747 there are cases where this would not be strictly necessary. */
3748 if (!int_fits_type_p (high
, type
)
3749 || !int_fits_type_p (low
, type
))
3751 low
= fold_convert (type
, low
);
3752 high
= fold_convert (type
, high
);
3755 next
= fold_binary (PLUS_EXPR
, type
, low
, step
);
3757 next
= fold_binary (PLUS_EXPR
, type
, high
, step
);
3759 if (tree_int_cst_compare (low
, next
) <= 0
3760 && tree_int_cst_compare (next
, high
) <= 0)
3763 /* If access is not executed on every iteration, we must ensure that overlow
3764 may not make the access valid later. */
3765 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, gimple_bb (data
->stmt
))
3766 && scev_probably_wraps_p (NULL_TREE
,
3767 initial_condition_in_loop_num (ev
, loop
->num
),
3768 step
, data
->stmt
, loop
, true))
3771 record_nonwrapping_iv (loop
, init
, step
, data
->stmt
, low
, high
, false, upper
);
3775 /* Determine information about number of iterations a LOOP from the bounds
3776 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
3777 STMT is guaranteed to be executed in every iteration of LOOP.*/
3780 infer_loop_bounds_from_ref (class loop
*loop
, gimple
*stmt
, tree ref
)
3782 struct ilb_data data
;
3786 for_each_index (&ref
, idx_infer_loop_bounds
, &data
);
3789 /* Determine information about number of iterations of a LOOP from the way
3790 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
3791 executed in every iteration of LOOP. */
3794 infer_loop_bounds_from_array (class loop
*loop
, gimple
*stmt
)
3796 if (is_gimple_assign (stmt
))
3798 tree op0
= gimple_assign_lhs (stmt
);
3799 tree op1
= gimple_assign_rhs1 (stmt
);
3801 /* For each memory access, analyze its access function
3802 and record a bound on the loop iteration domain. */
3803 if (REFERENCE_CLASS_P (op0
))
3804 infer_loop_bounds_from_ref (loop
, stmt
, op0
);
3806 if (REFERENCE_CLASS_P (op1
))
3807 infer_loop_bounds_from_ref (loop
, stmt
, op1
);
3809 else if (is_gimple_call (stmt
))
3812 unsigned i
, n
= gimple_call_num_args (stmt
);
3814 lhs
= gimple_call_lhs (stmt
);
3815 if (lhs
&& REFERENCE_CLASS_P (lhs
))
3816 infer_loop_bounds_from_ref (loop
, stmt
, lhs
);
3818 for (i
= 0; i
< n
; i
++)
3820 arg
= gimple_call_arg (stmt
, i
);
3821 if (REFERENCE_CLASS_P (arg
))
3822 infer_loop_bounds_from_ref (loop
, stmt
, arg
);
3827 /* Determine information about number of iterations of a LOOP from the fact
3828 that pointer arithmetics in STMT does not overflow. */
3831 infer_loop_bounds_from_pointer_arith (class loop
*loop
, gimple
*stmt
)
3833 tree def
, base
, step
, scev
, type
, low
, high
;
3836 if (!is_gimple_assign (stmt
)
3837 || gimple_assign_rhs_code (stmt
) != POINTER_PLUS_EXPR
)
3840 def
= gimple_assign_lhs (stmt
);
3841 if (TREE_CODE (def
) != SSA_NAME
)
3844 type
= TREE_TYPE (def
);
3845 if (!nowrap_type_p (type
))
3848 ptr
= gimple_assign_rhs1 (stmt
);
3849 if (!expr_invariant_in_loop_p (loop
, ptr
))
3852 var
= gimple_assign_rhs2 (stmt
);
3853 if (TYPE_PRECISION (type
) != TYPE_PRECISION (TREE_TYPE (var
)))
3856 class loop
*uloop
= loop_containing_stmt (stmt
);
3857 scev
= instantiate_parameters (loop
, analyze_scalar_evolution (uloop
, def
));
3858 if (chrec_contains_undetermined (scev
))
3861 base
= initial_condition_in_loop_num (scev
, loop
->num
);
3862 step
= evolution_part_in_loop_num (scev
, loop
->num
);
3865 || TREE_CODE (step
) != INTEGER_CST
3866 || tree_contains_chrecs (base
, NULL
)
3867 || chrec_contains_symbols_defined_in_loop (base
, loop
->num
))
3870 low
= lower_bound_in_type (type
, type
);
3871 high
= upper_bound_in_type (type
, type
);
3873 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot
3874 produce a NULL pointer. The contrary would mean NULL points to an object,
3875 while NULL is supposed to compare unequal with the address of all objects.
3876 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a
3877 NULL pointer since that would mean wrapping, which we assume here not to
3878 happen. So, we can exclude NULL from the valid range of pointer
3880 if (flag_delete_null_pointer_checks
&& int_cst_value (low
) == 0)
3881 low
= build_int_cstu (TREE_TYPE (low
), TYPE_ALIGN_UNIT (TREE_TYPE (type
)));
3883 record_nonwrapping_iv (loop
, base
, step
, stmt
, low
, high
, false, true);
3886 /* Determine information about number of iterations of a LOOP from the fact
3887 that signed arithmetics in STMT does not overflow. */
3890 infer_loop_bounds_from_signedness (class loop
*loop
, gimple
*stmt
)
3892 tree def
, base
, step
, scev
, type
, low
, high
;
3894 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
3897 def
= gimple_assign_lhs (stmt
);
3899 if (TREE_CODE (def
) != SSA_NAME
)
3902 type
= TREE_TYPE (def
);
3903 if (!INTEGRAL_TYPE_P (type
)
3904 || !TYPE_OVERFLOW_UNDEFINED (type
))
3907 scev
= instantiate_parameters (loop
, analyze_scalar_evolution (loop
, def
));
3908 if (chrec_contains_undetermined (scev
))
3911 base
= initial_condition_in_loop_num (scev
, loop
->num
);
3912 step
= evolution_part_in_loop_num (scev
, loop
->num
);
3915 || TREE_CODE (step
) != INTEGER_CST
3916 || tree_contains_chrecs (base
, NULL
)
3917 || chrec_contains_symbols_defined_in_loop (base
, loop
->num
))
3920 low
= lower_bound_in_type (type
, type
);
3921 high
= upper_bound_in_type (type
, type
);
3923 get_range_query (cfun
)->range_of_expr (r
, def
);
3924 if (r
.kind () == VR_RANGE
)
3926 low
= wide_int_to_tree (type
, r
.lower_bound ());
3927 high
= wide_int_to_tree (type
, r
.upper_bound ());
3930 record_nonwrapping_iv (loop
, base
, step
, stmt
, low
, high
, false, true);
3933 /* The following analyzers are extracting informations on the bounds
3934 of LOOP from the following undefined behaviors:
3936 - data references should not access elements over the statically
3939 - signed variables should not overflow when flag_wrapv is not set.
3943 infer_loop_bounds_from_undefined (class loop
*loop
, basic_block
*bbs
)
3946 gimple_stmt_iterator bsi
;
3950 for (i
= 0; i
< loop
->num_nodes
; i
++)
3954 /* If BB is not executed in each iteration of the loop, we cannot
3955 use the operations in it to infer reliable upper bound on the
3956 # of iterations of the loop. However, we can use it as a guess.
3957 Reliable guesses come only from array bounds. */
3958 reliable
= dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
);
3960 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
3962 gimple
*stmt
= gsi_stmt (bsi
);
3964 infer_loop_bounds_from_array (loop
, stmt
);
3968 infer_loop_bounds_from_signedness (loop
, stmt
);
3969 infer_loop_bounds_from_pointer_arith (loop
, stmt
);
3976 /* Compare wide ints, callback for qsort. */
3979 wide_int_cmp (const void *p1
, const void *p2
)
3981 const widest_int
*d1
= (const widest_int
*) p1
;
3982 const widest_int
*d2
= (const widest_int
*) p2
;
3983 return wi::cmpu (*d1
, *d2
);
3986 /* Return index of BOUND in BOUNDS array sorted in increasing order.
3987 Lookup by binary search. */
3990 bound_index (const vec
<widest_int
> &bounds
, const widest_int
&bound
)
3992 unsigned int end
= bounds
.length ();
3993 unsigned int begin
= 0;
3995 /* Find a matching index by means of a binary search. */
3996 while (begin
!= end
)
3998 unsigned int middle
= (begin
+ end
) / 2;
3999 widest_int index
= bounds
[middle
];
4003 else if (wi::ltu_p (index
, bound
))
4011 /* We recorded loop bounds only for statements dominating loop latch (and thus
4012 executed each loop iteration). If there are any bounds on statements not
4013 dominating the loop latch we can improve the estimate by walking the loop
4014 body and seeing if every path from loop header to loop latch contains
4015 some bounded statement. */
4018 discover_iteration_bound_by_body_walk (class loop
*loop
)
4020 class nb_iter_bound
*elt
;
4021 auto_vec
<widest_int
> bounds
;
4022 vec
<vec
<basic_block
> > queues
= vNULL
;
4023 vec
<basic_block
> queue
= vNULL
;
4024 ptrdiff_t queue_index
;
4025 ptrdiff_t latch_index
= 0;
4027 /* Discover what bounds may interest us. */
4028 for (elt
= loop
->bounds
; elt
; elt
= elt
->next
)
4030 widest_int bound
= elt
->bound
;
4032 /* Exit terminates loop at given iteration, while non-exits produce undefined
4033 effect on the next iteration. */
4037 /* If an overflow occurred, ignore the result. */
4042 if (!loop
->any_upper_bound
4043 || wi::ltu_p (bound
, loop
->nb_iterations_upper_bound
))
4044 bounds
.safe_push (bound
);
4047 /* Exit early if there is nothing to do. */
4048 if (!bounds
.exists ())
4051 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4052 fprintf (dump_file
, " Trying to walk loop body to reduce the bound.\n");
4054 /* Sort the bounds in decreasing order. */
4055 bounds
.qsort (wide_int_cmp
);
4057 /* For every basic block record the lowest bound that is guaranteed to
4058 terminate the loop. */
4060 hash_map
<basic_block
, ptrdiff_t> bb_bounds
;
4061 for (elt
= loop
->bounds
; elt
; elt
= elt
->next
)
4063 widest_int bound
= elt
->bound
;
4067 /* If an overflow occurred, ignore the result. */
4072 if (!loop
->any_upper_bound
4073 || wi::ltu_p (bound
, loop
->nb_iterations_upper_bound
))
4075 ptrdiff_t index
= bound_index (bounds
, bound
);
4076 ptrdiff_t *entry
= bb_bounds
.get (gimple_bb (elt
->stmt
));
4078 bb_bounds
.put (gimple_bb (elt
->stmt
), index
);
4079 else if ((ptrdiff_t)*entry
> index
)
4084 hash_map
<basic_block
, ptrdiff_t> block_priority
;
4086 /* Perform shortest path discovery loop->header ... loop->latch.
4088 The "distance" is given by the smallest loop bound of basic block
4089 present in the path and we look for path with largest smallest bound
4092 To avoid the need for fibonacci heap on double ints we simply compress
4093 double ints into indexes to BOUNDS array and then represent the queue
4094 as arrays of queues for every index.
4095 Index of BOUNDS.length() means that the execution of given BB has
4096 no bounds determined.
4098 VISITED is a pointer map translating basic block into smallest index
4099 it was inserted into the priority queue with. */
4102 /* Start walk in loop header with index set to infinite bound. */
4103 queue_index
= bounds
.length ();
4104 queues
.safe_grow_cleared (queue_index
+ 1, true);
4105 queue
.safe_push (loop
->header
);
4106 queues
[queue_index
] = queue
;
4107 block_priority
.put (loop
->header
, queue_index
);
4109 for (; queue_index
>= 0; queue_index
--)
4111 if (latch_index
< queue_index
)
4113 while (queues
[queue_index
].length ())
4116 ptrdiff_t bound_index
= queue_index
;
4120 queue
= queues
[queue_index
];
4123 /* OK, we later inserted the BB with lower priority, skip it. */
4124 if (*block_priority
.get (bb
) > queue_index
)
4127 /* See if we can improve the bound. */
4128 ptrdiff_t *entry
= bb_bounds
.get (bb
);
4129 if (entry
&& *entry
< bound_index
)
4130 bound_index
= *entry
;
4132 /* Insert succesors into the queue, watch for latch edge
4133 and record greatest index we saw. */
4134 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
4136 bool insert
= false;
4138 if (loop_exit_edge_p (loop
, e
))
4141 if (e
== loop_latch_edge (loop
)
4142 && latch_index
< bound_index
)
4143 latch_index
= bound_index
;
4144 else if (!(entry
= block_priority
.get (e
->dest
)))
4147 block_priority
.put (e
->dest
, bound_index
);
4149 else if (*entry
< bound_index
)
4152 *entry
= bound_index
;
4156 queues
[bound_index
].safe_push (e
->dest
);
4160 queues
[queue_index
].release ();
4163 gcc_assert (latch_index
>= 0);
4164 if ((unsigned)latch_index
< bounds
.length ())
4166 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4168 fprintf (dump_file
, "Found better loop bound ");
4169 print_decu (bounds
[latch_index
], dump_file
);
4170 fprintf (dump_file
, "\n");
4172 record_niter_bound (loop
, bounds
[latch_index
], false, true);
4178 /* See if every path cross the loop goes through a statement that is known
4179 to not execute at the last iteration. In that case we can decrese iteration
4183 maybe_lower_iteration_bound (class loop
*loop
)
4185 hash_set
<gimple
*> *not_executed_last_iteration
= NULL
;
4186 class nb_iter_bound
*elt
;
4187 bool found_exit
= false;
4188 auto_vec
<basic_block
> queue
;
4191 /* Collect all statements with interesting (i.e. lower than
4192 nb_iterations_upper_bound) bound on them.
4194 TODO: Due to the way record_estimate choose estimates to store, the bounds
4195 will be always nb_iterations_upper_bound-1. We can change this to record
4196 also statements not dominating the loop latch and update the walk bellow
4197 to the shortest path algorithm. */
4198 for (elt
= loop
->bounds
; elt
; elt
= elt
->next
)
4201 && wi::ltu_p (elt
->bound
, loop
->nb_iterations_upper_bound
))
4203 if (!not_executed_last_iteration
)
4204 not_executed_last_iteration
= new hash_set
<gimple
*>;
4205 not_executed_last_iteration
->add (elt
->stmt
);
4208 if (!not_executed_last_iteration
)
4211 /* Start DFS walk in the loop header and see if we can reach the
4212 loop latch or any of the exits (including statements with side
4213 effects that may terminate the loop otherwise) without visiting
4214 any of the statements known to have undefined effect on the last
4216 queue
.safe_push (loop
->header
);
4217 visited
= BITMAP_ALLOC (NULL
);
4218 bitmap_set_bit (visited
, loop
->header
->index
);
4223 basic_block bb
= queue
.pop ();
4224 gimple_stmt_iterator gsi
;
4225 bool stmt_found
= false;
4227 /* Loop for possible exits and statements bounding the execution. */
4228 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4230 gimple
*stmt
= gsi_stmt (gsi
);
4231 if (not_executed_last_iteration
->contains (stmt
))
4236 if (gimple_has_side_effects (stmt
))
4245 /* If no bounding statement is found, continue the walk. */
4251 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
4253 if (loop_exit_edge_p (loop
, e
)
4254 || e
== loop_latch_edge (loop
))
4259 if (bitmap_set_bit (visited
, e
->dest
->index
))
4260 queue
.safe_push (e
->dest
);
4264 while (queue
.length () && !found_exit
);
4266 /* If every path through the loop reach bounding statement before exit,
4267 then we know the last iteration of the loop will have undefined effect
4268 and we can decrease number of iterations. */
4272 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4273 fprintf (dump_file
, "Reducing loop iteration estimate by 1; "
4274 "undefined statement must be executed at the last iteration.\n");
4275 record_niter_bound (loop
, loop
->nb_iterations_upper_bound
- 1,
4279 BITMAP_FREE (visited
);
4280 delete not_executed_last_iteration
;
4283 /* Get expected upper bound for number of loop iterations for
4284 BUILT_IN_EXPECT_WITH_PROBABILITY for a condition COND. */
4287 get_upper_bound_based_on_builtin_expr_with_prob (gcond
*cond
)
4292 tree lhs
= gimple_cond_lhs (cond
);
4293 if (TREE_CODE (lhs
) != SSA_NAME
)
4296 gimple
*stmt
= SSA_NAME_DEF_STMT (gimple_cond_lhs (cond
));
4297 gcall
*def
= dyn_cast
<gcall
*> (stmt
);
4301 tree decl
= gimple_call_fndecl (def
);
4303 || !fndecl_built_in_p (decl
, BUILT_IN_EXPECT_WITH_PROBABILITY
)
4304 || gimple_call_num_args (stmt
) != 3)
4307 tree c
= gimple_call_arg (def
, 1);
4308 tree condt
= TREE_TYPE (lhs
);
4309 tree res
= fold_build2 (gimple_cond_code (cond
),
4311 gimple_cond_rhs (cond
));
4312 if (TREE_CODE (res
) != INTEGER_CST
)
4316 tree prob
= gimple_call_arg (def
, 2);
4317 tree t
= TREE_TYPE (prob
);
4319 = build_real_from_int_cst (t
,
4321 if (integer_zerop (res
))
4322 prob
= fold_build2 (MINUS_EXPR
, t
, one
, prob
);
4323 tree r
= fold_build2 (RDIV_EXPR
, t
, one
, prob
);
4324 if (TREE_CODE (r
) != REAL_CST
)
4328 = real_to_integer (TREE_REAL_CST_PTR (r
));
4329 return build_int_cst (condt
, probi
);
4332 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P
4333 is true also use estimates derived from undefined behavior. */
4336 estimate_numbers_of_iterations (class loop
*loop
)
4340 class tree_niter_desc niter_desc
;
4345 /* Give up if we already have tried to compute an estimation. */
4346 if (loop
->estimate_state
!= EST_NOT_COMPUTED
)
4349 loop
->estimate_state
= EST_AVAILABLE
;
4351 /* If we have a measured profile, use it to estimate the number of
4352 iterations. Normally this is recorded by branch_prob right after
4353 reading the profile. In case we however found a new loop, record the
4356 Explicitly check for profile status so we do not report
4357 wrong prediction hitrates for guessed loop iterations heuristics.
4358 Do not recompute already recorded bounds - we ought to be better on
4359 updating iteration bounds than updating profile in general and thus
4360 recomputing iteration bounds later in the compilation process will just
4361 introduce random roundoff errors. */
4362 if (!loop
->any_estimate
4363 && loop
->header
->count
.reliable_p ())
4365 gcov_type nit
= expected_loop_iterations_unbounded (loop
);
4366 bound
= gcov_type_to_wide_int (nit
);
4367 record_niter_bound (loop
, bound
, true, false);
4370 /* Ensure that loop->nb_iterations is computed if possible. If it turns out
4371 to be constant, we avoid undefined behavior implied bounds and instead
4372 diagnose those loops with -Waggressive-loop-optimizations. */
4373 number_of_latch_executions (loop
);
4375 basic_block
*body
= get_loop_body (loop
);
4376 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
, body
);
4377 likely_exit
= single_likely_exit (loop
, exits
);
4378 FOR_EACH_VEC_ELT (exits
, i
, ex
)
4380 if (ex
== likely_exit
)
4382 gimple
*stmt
= last_stmt (ex
->src
);
4385 gcond
*cond
= dyn_cast
<gcond
*> (stmt
);
4387 = get_upper_bound_based_on_builtin_expr_with_prob (cond
);
4388 if (niter_bound
!= NULL_TREE
)
4390 widest_int max
= derive_constant_upper_bound (niter_bound
);
4391 record_estimate (loop
, niter_bound
, max
, cond
,
4397 if (!number_of_iterations_exit (loop
, ex
, &niter_desc
,
4398 false, false, body
))
4401 niter
= niter_desc
.niter
;
4402 type
= TREE_TYPE (niter
);
4403 if (TREE_CODE (niter_desc
.may_be_zero
) != INTEGER_CST
)
4404 niter
= build3 (COND_EXPR
, type
, niter_desc
.may_be_zero
,
4405 build_int_cst (type
, 0),
4407 record_estimate (loop
, niter
, niter_desc
.max
,
4408 last_stmt (ex
->src
),
4409 true, ex
== likely_exit
, true);
4410 record_control_iv (loop
, &niter_desc
);
4413 if (flag_aggressive_loop_optimizations
)
4414 infer_loop_bounds_from_undefined (loop
, body
);
4417 discover_iteration_bound_by_body_walk (loop
);
4419 maybe_lower_iteration_bound (loop
);
4421 /* If we know the exact number of iterations of this loop, try to
4422 not break code with undefined behavior by not recording smaller
4423 maximum number of iterations. */
4424 if (loop
->nb_iterations
4425 && TREE_CODE (loop
->nb_iterations
) == INTEGER_CST
)
4427 loop
->any_upper_bound
= true;
4428 loop
->nb_iterations_upper_bound
= wi::to_widest (loop
->nb_iterations
);
4432 /* Sets NIT to the estimated number of executions of the latch of the
4433 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
4434 large as the number of iterations. If we have no reliable estimate,
4435 the function returns false, otherwise returns true. */
4438 estimated_loop_iterations (class loop
*loop
, widest_int
*nit
)
4440 /* When SCEV information is available, try to update loop iterations
4441 estimate. Otherwise just return whatever we recorded earlier. */
4442 if (scev_initialized_p ())
4443 estimate_numbers_of_iterations (loop
);
4445 return (get_estimated_loop_iterations (loop
, nit
));
4448 /* Similar to estimated_loop_iterations, but returns the estimate only
4449 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
4450 on the number of iterations of LOOP could not be derived, returns -1. */
4453 estimated_loop_iterations_int (class loop
*loop
)
4456 HOST_WIDE_INT hwi_nit
;
4458 if (!estimated_loop_iterations (loop
, &nit
))
4461 if (!wi::fits_shwi_p (nit
))
4463 hwi_nit
= nit
.to_shwi ();
4465 return hwi_nit
< 0 ? -1 : hwi_nit
;
4469 /* Sets NIT to an upper bound for the maximum number of executions of the
4470 latch of the LOOP. If we have no reliable estimate, the function returns
4471 false, otherwise returns true. */
4474 max_loop_iterations (class loop
*loop
, widest_int
*nit
)
4476 /* When SCEV information is available, try to update loop iterations
4477 estimate. Otherwise just return whatever we recorded earlier. */
4478 if (scev_initialized_p ())
4479 estimate_numbers_of_iterations (loop
);
4481 return get_max_loop_iterations (loop
, nit
);
4484 /* Similar to max_loop_iterations, but returns the estimate only
4485 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
4486 on the number of iterations of LOOP could not be derived, returns -1. */
4489 max_loop_iterations_int (class loop
*loop
)
4492 HOST_WIDE_INT hwi_nit
;
4494 if (!max_loop_iterations (loop
, &nit
))
4497 if (!wi::fits_shwi_p (nit
))
4499 hwi_nit
= nit
.to_shwi ();
4501 return hwi_nit
< 0 ? -1 : hwi_nit
;
4504 /* Sets NIT to an likely upper bound for the maximum number of executions of the
4505 latch of the LOOP. If we have no reliable estimate, the function returns
4506 false, otherwise returns true. */
4509 likely_max_loop_iterations (class loop
*loop
, widest_int
*nit
)
4511 /* When SCEV information is available, try to update loop iterations
4512 estimate. Otherwise just return whatever we recorded earlier. */
4513 if (scev_initialized_p ())
4514 estimate_numbers_of_iterations (loop
);
4516 return get_likely_max_loop_iterations (loop
, nit
);
4519 /* Similar to max_loop_iterations, but returns the estimate only
4520 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
4521 on the number of iterations of LOOP could not be derived, returns -1. */
4524 likely_max_loop_iterations_int (class loop
*loop
)
4527 HOST_WIDE_INT hwi_nit
;
4529 if (!likely_max_loop_iterations (loop
, &nit
))
4532 if (!wi::fits_shwi_p (nit
))
4534 hwi_nit
= nit
.to_shwi ();
4536 return hwi_nit
< 0 ? -1 : hwi_nit
;
4539 /* Returns an estimate for the number of executions of statements
4540 in the LOOP. For statements before the loop exit, this exceeds
4541 the number of execution of the latch by one. */
4544 estimated_stmt_executions_int (class loop
*loop
)
4546 HOST_WIDE_INT nit
= estimated_loop_iterations_int (loop
);
4552 snit
= (HOST_WIDE_INT
) ((unsigned HOST_WIDE_INT
) nit
+ 1);
4554 /* If the computation overflows, return -1. */
4555 return snit
< 0 ? -1 : snit
;
4558 /* Sets NIT to the maximum number of executions of the latch of the
4559 LOOP, plus one. If we have no reliable estimate, the function returns
4560 false, otherwise returns true. */
4563 max_stmt_executions (class loop
*loop
, widest_int
*nit
)
4565 widest_int nit_minus_one
;
4567 if (!max_loop_iterations (loop
, nit
))
4570 nit_minus_one
= *nit
;
4574 return wi::gtu_p (*nit
, nit_minus_one
);
4577 /* Sets NIT to the estimated maximum number of executions of the latch of the
4578 LOOP, plus one. If we have no likely estimate, the function returns
4579 false, otherwise returns true. */
4582 likely_max_stmt_executions (class loop
*loop
, widest_int
*nit
)
4584 widest_int nit_minus_one
;
4586 if (!likely_max_loop_iterations (loop
, nit
))
4589 nit_minus_one
= *nit
;
4593 return wi::gtu_p (*nit
, nit_minus_one
);
4596 /* Sets NIT to the estimated number of executions of the latch of the
4597 LOOP, plus one. If we have no reliable estimate, the function returns
4598 false, otherwise returns true. */
4601 estimated_stmt_executions (class loop
*loop
, widest_int
*nit
)
4603 widest_int nit_minus_one
;
4605 if (!estimated_loop_iterations (loop
, nit
))
4608 nit_minus_one
= *nit
;
4612 return wi::gtu_p (*nit
, nit_minus_one
);
4615 /* Records estimates on numbers of iterations of loops. */
4618 estimate_numbers_of_iterations (function
*fn
)
4620 /* We don't want to issue signed overflow warnings while getting
4621 loop iteration estimates. */
4622 fold_defer_overflow_warnings ();
4624 for (auto loop
: loops_list (fn
, 0))
4625 estimate_numbers_of_iterations (loop
);
4627 fold_undefer_and_ignore_overflow_warnings ();
4630 /* Returns true if statement S1 dominates statement S2. */
4633 stmt_dominates_stmt_p (gimple
*s1
, gimple
*s2
)
4635 basic_block bb1
= gimple_bb (s1
), bb2
= gimple_bb (s2
);
4643 gimple_stmt_iterator bsi
;
4645 if (gimple_code (s2
) == GIMPLE_PHI
)
4648 if (gimple_code (s1
) == GIMPLE_PHI
)
4651 for (bsi
= gsi_start_bb (bb1
); gsi_stmt (bsi
) != s2
; gsi_next (&bsi
))
4652 if (gsi_stmt (bsi
) == s1
)
4658 return dominated_by_p (CDI_DOMINATORS
, bb2
, bb1
);
4661 /* Returns true when we can prove that the number of executions of
4662 STMT in the loop is at most NITER, according to the bound on
4663 the number of executions of the statement NITER_BOUND->stmt recorded in
4664 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT.
4666 ??? This code can become quite a CPU hog - we can have many bounds,
4667 and large basic block forcing stmt_dominates_stmt_p to be queried
4668 many times on a large basic blocks, so the whole thing is O(n^2)
4669 for scev_probably_wraps_p invocation (that can be done n times).
4671 It would make more sense (and give better answers) to remember BB
4672 bounds computed by discover_iteration_bound_by_body_walk. */
4675 n_of_executions_at_most (gimple
*stmt
,
4676 class nb_iter_bound
*niter_bound
,
4679 widest_int bound
= niter_bound
->bound
;
4680 tree nit_type
= TREE_TYPE (niter
), e
;
4683 gcc_assert (TYPE_UNSIGNED (nit_type
));
4685 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
4686 the number of iterations is small. */
4687 if (!wi::fits_to_tree_p (bound
, nit_type
))
4690 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
4691 times. This means that:
4693 -- if NITER_BOUND->is_exit is true, then everything after
4694 it at most NITER_BOUND->bound times.
4696 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
4697 is executed, then NITER_BOUND->stmt is executed as well in the same
4698 iteration then STMT is executed at most NITER_BOUND->bound + 1 times.
4700 If we can determine that NITER_BOUND->stmt is always executed
4701 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times.
4702 We conclude that if both statements belong to the same
4703 basic block and STMT is before NITER_BOUND->stmt and there are no
4704 statements with side effects in between. */
4706 if (niter_bound
->is_exit
)
4708 if (stmt
== niter_bound
->stmt
4709 || !stmt_dominates_stmt_p (niter_bound
->stmt
, stmt
))
4715 if (!stmt_dominates_stmt_p (niter_bound
->stmt
, stmt
))
4717 gimple_stmt_iterator bsi
;
4718 if (gimple_bb (stmt
) != gimple_bb (niter_bound
->stmt
)
4719 || gimple_code (stmt
) == GIMPLE_PHI
4720 || gimple_code (niter_bound
->stmt
) == GIMPLE_PHI
)
4723 /* By stmt_dominates_stmt_p we already know that STMT appears
4724 before NITER_BOUND->STMT. Still need to test that the loop
4725 cannot be terinated by a side effect in between. */
4726 for (bsi
= gsi_for_stmt (stmt
); gsi_stmt (bsi
) != niter_bound
->stmt
;
4728 if (gimple_has_side_effects (gsi_stmt (bsi
)))
4732 || !wi::fits_to_tree_p (bound
, nit_type
))
4738 e
= fold_binary (cmp
, boolean_type_node
,
4739 niter
, wide_int_to_tree (nit_type
, bound
));
4740 return e
&& integer_nonzerop (e
);
4743 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
4746 nowrap_type_p (tree type
)
4748 if (ANY_INTEGRAL_TYPE_P (type
)
4749 && TYPE_OVERFLOW_UNDEFINED (type
))
4752 if (POINTER_TYPE_P (type
))
4758 /* Return true if we can prove LOOP is exited before evolution of induction
4759 variable {BASE, STEP} overflows with respect to its type bound. */
4762 loop_exits_before_overflow (tree base
, tree step
,
4763 gimple
*at_stmt
, class loop
*loop
)
4766 struct control_iv
*civ
;
4767 class nb_iter_bound
*bound
;
4768 tree e
, delta
, step_abs
, unsigned_base
;
4769 tree type
= TREE_TYPE (step
);
4770 tree unsigned_type
, valid_niter
;
4772 /* Don't issue signed overflow warnings. */
4773 fold_defer_overflow_warnings ();
4775 /* Compute the number of iterations before we reach the bound of the
4776 type, and verify that the loop is exited before this occurs. */
4777 unsigned_type
= unsigned_type_for (type
);
4778 unsigned_base
= fold_convert (unsigned_type
, base
);
4780 if (tree_int_cst_sign_bit (step
))
4782 tree extreme
= fold_convert (unsigned_type
,
4783 lower_bound_in_type (type
, type
));
4784 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, unsigned_base
, extreme
);
4785 step_abs
= fold_build1 (NEGATE_EXPR
, unsigned_type
,
4786 fold_convert (unsigned_type
, step
));
4790 tree extreme
= fold_convert (unsigned_type
,
4791 upper_bound_in_type (type
, type
));
4792 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, extreme
, unsigned_base
);
4793 step_abs
= fold_convert (unsigned_type
, step
);
4796 valid_niter
= fold_build2 (FLOOR_DIV_EXPR
, unsigned_type
, delta
, step_abs
);
4798 estimate_numbers_of_iterations (loop
);
4800 if (max_loop_iterations (loop
, &niter
)
4801 && wi::fits_to_tree_p (niter
, TREE_TYPE (valid_niter
))
4802 && (e
= fold_binary (GT_EXPR
, boolean_type_node
, valid_niter
,
4803 wide_int_to_tree (TREE_TYPE (valid_niter
),
4805 && integer_nonzerop (e
))
4807 fold_undefer_and_ignore_overflow_warnings ();
4811 for (bound
= loop
->bounds
; bound
; bound
= bound
->next
)
4813 if (n_of_executions_at_most (at_stmt
, bound
, valid_niter
))
4815 fold_undefer_and_ignore_overflow_warnings ();
4819 fold_undefer_and_ignore_overflow_warnings ();
4821 /* Try to prove loop is exited before {base, step} overflows with the
4822 help of analyzed loop control IV. This is done only for IVs with
4823 constant step because otherwise we don't have the information. */
4824 if (TREE_CODE (step
) == INTEGER_CST
)
4826 for (civ
= loop
->control_ivs
; civ
; civ
= civ
->next
)
4828 enum tree_code code
;
4829 tree civ_type
= TREE_TYPE (civ
->step
);
4831 /* Have to consider type difference because operand_equal_p ignores
4832 that for constants. */
4833 if (TYPE_UNSIGNED (type
) != TYPE_UNSIGNED (civ_type
)
4834 || element_precision (type
) != element_precision (civ_type
))
4837 /* Only consider control IV with same step. */
4838 if (!operand_equal_p (step
, civ
->step
, 0))
4841 /* Done proving if this is a no-overflow control IV. */
4842 if (operand_equal_p (base
, civ
->base
, 0))
4845 /* Control IV is recorded after expanding simple operations,
4846 Here we expand base and compare it too. */
4847 tree expanded_base
= expand_simple_operations (base
);
4848 if (operand_equal_p (expanded_base
, civ
->base
, 0))
4851 /* If this is a before stepping control IV, in other words, we have
4853 {civ_base, step} = {base + step, step}
4855 Because civ {base + step, step} doesn't overflow during loop
4856 iterations, {base, step} will not overflow if we can prove the
4857 operation "base + step" does not overflow. Specifically, we try
4858 to prove below conditions are satisfied:
4860 base <= UPPER_BOUND (type) - step ;;step > 0
4861 base >= LOWER_BOUND (type) - step ;;step < 0
4863 by proving the reverse conditions are false using loop's initial
4865 if (POINTER_TYPE_P (TREE_TYPE (base
)))
4866 code
= POINTER_PLUS_EXPR
;
4870 tree stepped
= fold_build2 (code
, TREE_TYPE (base
), base
, step
);
4871 tree expanded_stepped
= fold_build2 (code
, TREE_TYPE (base
),
4872 expanded_base
, step
);
4873 if (operand_equal_p (stepped
, civ
->base
, 0)
4874 || operand_equal_p (expanded_stepped
, civ
->base
, 0))
4878 if (tree_int_cst_sign_bit (step
))
4881 extreme
= lower_bound_in_type (type
, type
);
4886 extreme
= upper_bound_in_type (type
, type
);
4888 extreme
= fold_build2 (MINUS_EXPR
, type
, extreme
, step
);
4889 e
= fold_build2 (code
, boolean_type_node
, base
, extreme
);
4890 e
= simplify_using_initial_conditions (loop
, e
);
4891 if (integer_zerop (e
))
4900 /* VAR is scev variable whose evolution part is constant STEP, this function
4901 proves that VAR can't overflow by using value range info. If VAR's value
4902 range is [MIN, MAX], it can be proven by:
4903 MAX + step doesn't overflow ; if step > 0
4905 MIN + step doesn't underflow ; if step < 0.
4907 We can only do this if var is computed in every loop iteration, i.e, var's
4908 definition has to dominate loop latch. Consider below example:
4916 # RANGE [0, 4294967294] NONZERO 65535
4917 # i_21 = PHI <0(3), i_18(9)>
4924 # RANGE [0, 65533] NONZERO 65535
4925 _6 = i_21 + 4294967295;
4926 # RANGE [0, 65533] NONZERO 65535
4927 _7 = (long unsigned int) _6;
4928 # RANGE [0, 524264] NONZERO 524280
4930 # PT = nonlocal escaped
4935 # RANGE [1, 65535] NONZERO 65535
4949 VAR _6 doesn't overflow only with pre-condition (i_21 != 0), here we
4950 can't use _6 to prove no-overlfow for _7. In fact, var _7 takes value
4951 sequence (4294967295, 0, 1, ..., 65533) in loop life time, rather than
4952 (4294967295, 4294967296, ...). */
4955 scev_var_range_cant_overflow (tree var
, tree step
, class loop
*loop
)
4958 wide_int minv
, maxv
, diff
, step_wi
;
4960 if (TREE_CODE (step
) != INTEGER_CST
|| !INTEGRAL_TYPE_P (TREE_TYPE (var
)))
4963 /* Check if VAR evaluates in every loop iteration. It's not the case
4964 if VAR is default definition or does not dominate loop's latch. */
4965 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
4966 if (!def_bb
|| !dominated_by_p (CDI_DOMINATORS
, loop
->latch
, def_bb
))
4970 get_range_query (cfun
)->range_of_expr (r
, var
);
4971 if (r
.kind () != VR_RANGE
)
4974 /* VAR is a scev whose evolution part is STEP and value range info
4975 is [MIN, MAX], we can prove its no-overflowness by conditions:
4977 type_MAX - MAX >= step ; if step > 0
4978 MIN - type_MIN >= |step| ; if step < 0.
4980 Or VAR must take value outside of value range, which is not true. */
4981 step_wi
= wi::to_wide (step
);
4982 type
= TREE_TYPE (var
);
4983 if (tree_int_cst_sign_bit (step
))
4985 diff
= r
.lower_bound () - wi::to_wide (lower_bound_in_type (type
, type
));
4986 step_wi
= - step_wi
;
4989 diff
= wi::to_wide (upper_bound_in_type (type
, type
)) - r
.upper_bound ();
4991 return (wi::geu_p (diff
, step_wi
));
4994 /* Return false only when the induction variable BASE + STEP * I is
4995 known to not overflow: i.e. when the number of iterations is small
4996 enough with respect to the step and initial condition in order to
4997 keep the evolution confined in TYPEs bounds. Return true when the
4998 iv is known to overflow or when the property is not computable.
5000 USE_OVERFLOW_SEMANTICS is true if this function should assume that
5001 the rules for overflow of the given language apply (e.g., that signed
5002 arithmetics in C does not overflow).
5004 If VAR is a ssa variable, this function also returns false if VAR can
5005 be proven not overflow with value range info. */
5008 scev_probably_wraps_p (tree var
, tree base
, tree step
,
5009 gimple
*at_stmt
, class loop
*loop
,
5010 bool use_overflow_semantics
)
5012 /* FIXME: We really need something like
5013 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
5015 We used to test for the following situation that frequently appears
5016 during address arithmetics:
5018 D.1621_13 = (long unsigned intD.4) D.1620_12;
5019 D.1622_14 = D.1621_13 * 8;
5020 D.1623_15 = (doubleD.29 *) D.1622_14;
5022 And derived that the sequence corresponding to D_14
5023 can be proved to not wrap because it is used for computing a
5024 memory access; however, this is not really the case -- for example,
5025 if D_12 = (unsigned char) [254,+,1], then D_14 has values
5026 2032, 2040, 0, 8, ..., but the code is still legal. */
5028 if (chrec_contains_undetermined (base
)
5029 || chrec_contains_undetermined (step
))
5032 if (integer_zerop (step
))
5035 /* If we can use the fact that signed and pointer arithmetics does not
5036 wrap, we are done. */
5037 if (use_overflow_semantics
&& nowrap_type_p (TREE_TYPE (base
)))
5040 /* To be able to use estimates on number of iterations of the loop,
5041 we must have an upper bound on the absolute value of the step. */
5042 if (TREE_CODE (step
) != INTEGER_CST
)
5045 /* Check if var can be proven not overflow with value range info. */
5046 if (var
&& TREE_CODE (var
) == SSA_NAME
5047 && scev_var_range_cant_overflow (var
, step
, loop
))
5050 if (loop_exits_before_overflow (base
, step
, at_stmt
, loop
))
5053 /* At this point we still don't have a proof that the iv does not
5054 overflow: give up. */
5058 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
5061 free_numbers_of_iterations_estimates (class loop
*loop
)
5063 struct control_iv
*civ
;
5064 class nb_iter_bound
*bound
;
5066 loop
->nb_iterations
= NULL
;
5067 loop
->estimate_state
= EST_NOT_COMPUTED
;
5068 for (bound
= loop
->bounds
; bound
;)
5070 class nb_iter_bound
*next
= bound
->next
;
5074 loop
->bounds
= NULL
;
5076 for (civ
= loop
->control_ivs
; civ
;)
5078 struct control_iv
*next
= civ
->next
;
5082 loop
->control_ivs
= NULL
;
5085 /* Frees the information on upper bounds on numbers of iterations of loops. */
5088 free_numbers_of_iterations_estimates (function
*fn
)
5090 for (auto loop
: loops_list (fn
, 0))
5091 free_numbers_of_iterations_estimates (loop
);
5094 /* Substitute value VAL for ssa name NAME inside expressions held
5098 substitute_in_loop_info (class loop
*loop
, tree name
, tree val
)
5100 loop
->nb_iterations
= simplify_replace_tree (loop
->nb_iterations
, name
, val
);