1 /* Functions to determine/estimate number of iterations of a loop.
2 Copyright (C) 2004-2013 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"
26 #include "basic-block.h"
27 #include "gimple-pretty-print.h"
29 #include "tree-flow.h"
33 #include "tree-chrec.h"
34 #include "tree-scalar-evolution.h"
35 #include "tree-data-ref.h"
38 #include "diagnostic-core.h"
39 #include "tree-inline.h"
41 #define SWAP(X, Y) do { affine_iv *tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
43 /* The maximum number of dominator BBs we search for conditions
44 of loop header copies we use for simplifying a conditional
46 #define MAX_DOMINATORS_TO_WALK 8
50 Analysis of number of iterations of an affine exit test.
54 /* Bounds on some value, BELOW <= X <= UP. */
62 /* Splits expression EXPR to a variable part VAR and constant OFFSET. */
65 split_to_var_and_offset (tree expr
, tree
*var
, mpz_t offset
)
67 tree type
= TREE_TYPE (expr
);
73 mpz_set_ui (offset
, 0);
75 switch (TREE_CODE (expr
))
82 case POINTER_PLUS_EXPR
:
83 op0
= TREE_OPERAND (expr
, 0);
84 op1
= TREE_OPERAND (expr
, 1);
86 if (TREE_CODE (op1
) != INTEGER_CST
)
90 /* Always sign extend the offset. */
91 off
= tree_to_double_int (op1
);
92 off
= off
.sext (TYPE_PRECISION (type
));
93 mpz_set_double_int (offset
, off
, false);
95 mpz_neg (offset
, offset
);
99 *var
= build_int_cst_type (type
, 0);
100 off
= tree_to_double_int (expr
);
101 mpz_set_double_int (offset
, off
, TYPE_UNSIGNED (type
));
109 /* Stores estimate on the minimum/maximum value of the expression VAR + OFF
110 in TYPE to MIN and MAX. */
113 determine_value_range (tree type
, tree var
, mpz_t off
,
114 mpz_t min
, mpz_t max
)
116 /* If the expression is a constant, we know its value exactly. */
117 if (integer_zerop (var
))
124 /* If the computation may wrap, we know nothing about the value, except for
125 the range of the type. */
126 get_type_static_bounds (type
, min
, max
);
127 if (!nowrap_type_p (type
))
130 /* Since the addition of OFF does not wrap, if OFF is positive, then we may
131 add it to MIN, otherwise to MAX. */
132 if (mpz_sgn (off
) < 0)
133 mpz_add (max
, max
, off
);
135 mpz_add (min
, min
, off
);
138 /* Stores the bounds on the difference of the values of the expressions
139 (var + X) and (var + Y), computed in TYPE, to BNDS. */
142 bound_difference_of_offsetted_base (tree type
, mpz_t x
, mpz_t y
,
145 int rel
= mpz_cmp (x
, y
);
146 bool may_wrap
= !nowrap_type_p (type
);
149 /* If X == Y, then the expressions are always equal.
150 If X > Y, there are the following possibilities:
151 a) neither of var + X and var + Y overflow or underflow, or both of
152 them do. Then their difference is X - Y.
153 b) var + X overflows, and var + Y does not. Then the values of the
154 expressions are var + X - M and var + Y, where M is the range of
155 the type, and their difference is X - Y - M.
156 c) var + Y underflows and var + X does not. Their difference again
158 Therefore, if the arithmetics in type does not overflow, then the
159 bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y)
160 Similarly, if X < Y, the bounds are either (X - Y, X - Y) or
161 (X - Y, X - Y + M). */
165 mpz_set_ui (bnds
->below
, 0);
166 mpz_set_ui (bnds
->up
, 0);
171 mpz_set_double_int (m
, double_int::mask (TYPE_PRECISION (type
)), true);
172 mpz_add_ui (m
, m
, 1);
173 mpz_sub (bnds
->up
, x
, y
);
174 mpz_set (bnds
->below
, bnds
->up
);
179 mpz_sub (bnds
->below
, bnds
->below
, m
);
181 mpz_add (bnds
->up
, bnds
->up
, m
);
187 /* From condition C0 CMP C1 derives information regarding the
188 difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE,
189 and stores it to BNDS. */
192 refine_bounds_using_guard (tree type
, tree varx
, mpz_t offx
,
193 tree vary
, mpz_t offy
,
194 tree c0
, enum tree_code cmp
, tree c1
,
197 tree varc0
, varc1
, tmp
, ctype
;
198 mpz_t offc0
, offc1
, loffx
, loffy
, bnd
;
200 bool no_wrap
= nowrap_type_p (type
);
209 STRIP_SIGN_NOPS (c0
);
210 STRIP_SIGN_NOPS (c1
);
211 ctype
= TREE_TYPE (c0
);
212 if (!useless_type_conversion_p (ctype
, type
))
218 /* We could derive quite precise information from EQ_EXPR, however, such
219 a guard is unlikely to appear, so we do not bother with handling
224 /* NE_EXPR comparisons do not contain much of useful information, except for
225 special case of comparing with the bounds of the type. */
226 if (TREE_CODE (c1
) != INTEGER_CST
227 || !INTEGRAL_TYPE_P (type
))
230 /* Ensure that the condition speaks about an expression in the same type
232 ctype
= TREE_TYPE (c0
);
233 if (TYPE_PRECISION (ctype
) != TYPE_PRECISION (type
))
235 c0
= fold_convert (type
, c0
);
236 c1
= fold_convert (type
, c1
);
238 if (TYPE_MIN_VALUE (type
)
239 && operand_equal_p (c1
, TYPE_MIN_VALUE (type
), 0))
244 if (TYPE_MAX_VALUE (type
)
245 && operand_equal_p (c1
, TYPE_MAX_VALUE (type
), 0))
258 split_to_var_and_offset (expand_simple_operations (c0
), &varc0
, offc0
);
259 split_to_var_and_offset (expand_simple_operations (c1
), &varc1
, offc1
);
261 /* We are only interested in comparisons of expressions based on VARX and
262 VARY. TODO -- we might also be able to derive some bounds from
263 expressions containing just one of the variables. */
265 if (operand_equal_p (varx
, varc1
, 0))
267 tmp
= varc0
; varc0
= varc1
; varc1
= tmp
;
268 mpz_swap (offc0
, offc1
);
269 cmp
= swap_tree_comparison (cmp
);
272 if (!operand_equal_p (varx
, varc0
, 0)
273 || !operand_equal_p (vary
, varc1
, 0))
276 mpz_init_set (loffx
, offx
);
277 mpz_init_set (loffy
, offy
);
279 if (cmp
== GT_EXPR
|| cmp
== GE_EXPR
)
281 tmp
= varx
; varx
= vary
; vary
= tmp
;
282 mpz_swap (offc0
, offc1
);
283 mpz_swap (loffx
, loffy
);
284 cmp
= swap_tree_comparison (cmp
);
288 /* If there is no overflow, the condition implies that
290 (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0).
292 The overflows and underflows may complicate things a bit; each
293 overflow decreases the appropriate offset by M, and underflow
294 increases it by M. The above inequality would not necessarily be
297 -- VARX + OFFX underflows and VARX + OFFC0 does not, or
298 VARX + OFFC0 overflows, but VARX + OFFX does not.
299 This may only happen if OFFX < OFFC0.
300 -- VARY + OFFY overflows and VARY + OFFC1 does not, or
301 VARY + OFFC1 underflows and VARY + OFFY does not.
302 This may only happen if OFFY > OFFC1. */
311 x_ok
= (integer_zerop (varx
)
312 || mpz_cmp (loffx
, offc0
) >= 0);
313 y_ok
= (integer_zerop (vary
)
314 || mpz_cmp (loffy
, offc1
) <= 0);
320 mpz_sub (bnd
, loffx
, loffy
);
321 mpz_add (bnd
, bnd
, offc1
);
322 mpz_sub (bnd
, bnd
, offc0
);
325 mpz_sub_ui (bnd
, bnd
, 1);
330 if (mpz_cmp (bnds
->below
, bnd
) < 0)
331 mpz_set (bnds
->below
, bnd
);
335 if (mpz_cmp (bnd
, bnds
->up
) < 0)
336 mpz_set (bnds
->up
, bnd
);
348 /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS.
349 The subtraction is considered to be performed in arbitrary precision,
352 We do not attempt to be too clever regarding the value ranges of X and
353 Y; most of the time, they are just integers or ssa names offsetted by
354 integer. However, we try to use the information contained in the
355 comparisons before the loop (usually created by loop header copying). */
358 bound_difference (struct loop
*loop
, tree x
, tree y
, bounds
*bnds
)
360 tree type
= TREE_TYPE (x
);
363 mpz_t minx
, maxx
, miny
, maxy
;
371 /* Get rid of unnecessary casts, but preserve the value of
376 mpz_init (bnds
->below
);
380 split_to_var_and_offset (x
, &varx
, offx
);
381 split_to_var_and_offset (y
, &vary
, offy
);
383 if (!integer_zerop (varx
)
384 && operand_equal_p (varx
, vary
, 0))
386 /* Special case VARX == VARY -- we just need to compare the
387 offsets. The matters are a bit more complicated in the
388 case addition of offsets may wrap. */
389 bound_difference_of_offsetted_base (type
, offx
, offy
, bnds
);
393 /* Otherwise, use the value ranges to determine the initial
394 estimates on below and up. */
399 determine_value_range (type
, varx
, offx
, minx
, maxx
);
400 determine_value_range (type
, vary
, offy
, miny
, maxy
);
402 mpz_sub (bnds
->below
, minx
, maxy
);
403 mpz_sub (bnds
->up
, maxx
, miny
);
410 /* If both X and Y are constants, we cannot get any more precise. */
411 if (integer_zerop (varx
) && integer_zerop (vary
))
414 /* Now walk the dominators of the loop header and use the entry
415 guards to refine the estimates. */
416 for (bb
= loop
->header
;
417 bb
!= ENTRY_BLOCK_PTR
&& cnt
< MAX_DOMINATORS_TO_WALK
;
418 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
420 if (!single_pred_p (bb
))
422 e
= single_pred_edge (bb
);
424 if (!(e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)))
427 cond
= last_stmt (e
->src
);
428 c0
= gimple_cond_lhs (cond
);
429 cmp
= gimple_cond_code (cond
);
430 c1
= gimple_cond_rhs (cond
);
432 if (e
->flags
& EDGE_FALSE_VALUE
)
433 cmp
= invert_tree_comparison (cmp
, false);
435 refine_bounds_using_guard (type
, varx
, offx
, vary
, offy
,
445 /* Update the bounds in BNDS that restrict the value of X to the bounds
446 that restrict the value of X + DELTA. X can be obtained as a
447 difference of two values in TYPE. */
450 bounds_add (bounds
*bnds
, double_int delta
, tree type
)
455 mpz_set_double_int (mdelta
, delta
, false);
458 mpz_set_double_int (max
, double_int::mask (TYPE_PRECISION (type
)), true);
460 mpz_add (bnds
->up
, bnds
->up
, mdelta
);
461 mpz_add (bnds
->below
, bnds
->below
, mdelta
);
463 if (mpz_cmp (bnds
->up
, max
) > 0)
464 mpz_set (bnds
->up
, max
);
467 if (mpz_cmp (bnds
->below
, max
) < 0)
468 mpz_set (bnds
->below
, max
);
474 /* Update the bounds in BNDS that restrict the value of X to the bounds
475 that restrict the value of -X. */
478 bounds_negate (bounds
*bnds
)
482 mpz_init_set (tmp
, bnds
->up
);
483 mpz_neg (bnds
->up
, bnds
->below
);
484 mpz_neg (bnds
->below
, tmp
);
488 /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */
491 inverse (tree x
, tree mask
)
493 tree type
= TREE_TYPE (x
);
495 unsigned ctr
= tree_floor_log2 (mask
);
497 if (TYPE_PRECISION (type
) <= HOST_BITS_PER_WIDE_INT
)
499 unsigned HOST_WIDE_INT ix
;
500 unsigned HOST_WIDE_INT imask
;
501 unsigned HOST_WIDE_INT irslt
= 1;
503 gcc_assert (cst_and_fits_in_hwi (x
));
504 gcc_assert (cst_and_fits_in_hwi (mask
));
506 ix
= int_cst_value (x
);
507 imask
= int_cst_value (mask
);
516 rslt
= build_int_cst_type (type
, irslt
);
520 rslt
= build_int_cst (type
, 1);
523 rslt
= int_const_binop (MULT_EXPR
, rslt
, x
);
524 x
= int_const_binop (MULT_EXPR
, x
, x
);
526 rslt
= int_const_binop (BIT_AND_EXPR
, rslt
, mask
);
532 /* Derives the upper bound BND on the number of executions of loop with exit
533 condition S * i <> C. If NO_OVERFLOW is true, then the control variable of
534 the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed
535 that the loop ends through this exit, i.e., the induction variable ever
536 reaches the value of C.
538 The value C is equal to final - base, where final and base are the final and
539 initial value of the actual induction variable in the analysed loop. BNDS
540 bounds the value of this difference when computed in signed type with
541 unbounded range, while the computation of C is performed in an unsigned
542 type with the range matching the range of the type of the induction variable.
543 In particular, BNDS.up contains an upper bound on C in the following cases:
544 -- if the iv must reach its final value without overflow, i.e., if
545 NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or
546 -- if final >= base, which we know to hold when BNDS.below >= 0. */
549 number_of_iterations_ne_max (mpz_t bnd
, bool no_overflow
, tree c
, tree s
,
550 bounds
*bnds
, bool exit_must_be_taken
)
554 bool bnds_u_valid
= ((no_overflow
&& exit_must_be_taken
)
555 || mpz_sgn (bnds
->below
) >= 0);
557 if (multiple_of_p (TREE_TYPE (c
), c
, s
))
559 /* If C is an exact multiple of S, then its value will be reached before
560 the induction variable overflows (unless the loop is exited in some
561 other way before). Note that the actual induction variable in the
562 loop (which ranges from base to final instead of from 0 to C) may
563 overflow, in which case BNDS.up will not be giving a correct upper
564 bound on C; thus, BNDS_U_VALID had to be computed in advance. */
566 exit_must_be_taken
= true;
569 /* If the induction variable can overflow, the number of iterations is at
570 most the period of the control variable (or infinite, but in that case
571 the whole # of iterations analysis will fail). */
574 max
= double_int::mask (TYPE_PRECISION (TREE_TYPE (c
))
575 - tree_low_cst (num_ending_zeros (s
), 1));
576 mpz_set_double_int (bnd
, max
, true);
580 /* Now we know that the induction variable does not overflow, so the loop
581 iterates at most (range of type / S) times. */
582 mpz_set_double_int (bnd
, double_int::mask (TYPE_PRECISION (TREE_TYPE (c
))),
585 /* If the induction variable is guaranteed to reach the value of C before
587 if (exit_must_be_taken
)
589 /* ... then we can strengthen this to C / S, and possibly we can use
590 the upper bound on C given by BNDS. */
591 if (TREE_CODE (c
) == INTEGER_CST
)
592 mpz_set_double_int (bnd
, tree_to_double_int (c
), true);
593 else if (bnds_u_valid
)
594 mpz_set (bnd
, bnds
->up
);
598 mpz_set_double_int (d
, tree_to_double_int (s
), true);
599 mpz_fdiv_q (bnd
, bnd
, d
);
603 /* Determines number of iterations of loop whose ending condition
604 is IV <> FINAL. TYPE is the type of the iv. The number of
605 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
606 we know that the exit must be taken eventually, i.e., that the IV
607 ever reaches the value FINAL (we derived this earlier, and possibly set
608 NITER->assumptions to make sure this is the case). BNDS contains the
609 bounds on the difference FINAL - IV->base. */
612 number_of_iterations_ne (tree type
, affine_iv
*iv
, tree final
,
613 struct tree_niter_desc
*niter
, bool exit_must_be_taken
,
616 tree niter_type
= unsigned_type_for (type
);
617 tree s
, c
, d
, bits
, assumption
, tmp
, bound
;
620 niter
->control
= *iv
;
621 niter
->bound
= final
;
622 niter
->cmp
= NE_EXPR
;
624 /* Rearrange the terms so that we get inequality S * i <> C, with S
625 positive. Also cast everything to the unsigned type. If IV does
626 not overflow, BNDS bounds the value of C. Also, this is the
627 case if the computation |FINAL - IV->base| does not overflow, i.e.,
628 if BNDS->below in the result is nonnegative. */
629 if (tree_int_cst_sign_bit (iv
->step
))
631 s
= fold_convert (niter_type
,
632 fold_build1 (NEGATE_EXPR
, type
, iv
->step
));
633 c
= fold_build2 (MINUS_EXPR
, niter_type
,
634 fold_convert (niter_type
, iv
->base
),
635 fold_convert (niter_type
, final
));
636 bounds_negate (bnds
);
640 s
= fold_convert (niter_type
, iv
->step
);
641 c
= fold_build2 (MINUS_EXPR
, niter_type
,
642 fold_convert (niter_type
, final
),
643 fold_convert (niter_type
, iv
->base
));
647 number_of_iterations_ne_max (max
, iv
->no_overflow
, c
, s
, bnds
,
649 niter
->max
= mpz_get_double_int (niter_type
, max
, false);
652 /* First the trivial cases -- when the step is 1. */
653 if (integer_onep (s
))
659 /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
660 is infinite. Otherwise, the number of iterations is
661 (inverse(s/d) * (c/d)) mod (size of mode/d). */
662 bits
= num_ending_zeros (s
);
663 bound
= build_low_bits_mask (niter_type
,
664 (TYPE_PRECISION (niter_type
)
665 - tree_low_cst (bits
, 1)));
667 d
= fold_binary_to_constant (LSHIFT_EXPR
, niter_type
,
668 build_int_cst (niter_type
, 1), bits
);
669 s
= fold_binary_to_constant (RSHIFT_EXPR
, niter_type
, s
, bits
);
671 if (!exit_must_be_taken
)
673 /* If we cannot assume that the exit is taken eventually, record the
674 assumptions for divisibility of c. */
675 assumption
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, c
, d
);
676 assumption
= fold_build2 (EQ_EXPR
, boolean_type_node
,
677 assumption
, build_int_cst (niter_type
, 0));
678 if (!integer_nonzerop (assumption
))
679 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
680 niter
->assumptions
, assumption
);
683 c
= fold_build2 (EXACT_DIV_EXPR
, niter_type
, c
, d
);
684 tmp
= fold_build2 (MULT_EXPR
, niter_type
, c
, inverse (s
, bound
));
685 niter
->niter
= fold_build2 (BIT_AND_EXPR
, niter_type
, tmp
, bound
);
689 /* Checks whether we can determine the final value of the control variable
690 of the loop with ending condition IV0 < IV1 (computed in TYPE).
691 DELTA is the difference IV1->base - IV0->base, STEP is the absolute value
692 of the step. The assumptions necessary to ensure that the computation
693 of the final value does not overflow are recorded in NITER. If we
694 find the final value, we adjust DELTA and return TRUE. Otherwise
695 we return false. BNDS bounds the value of IV1->base - IV0->base,
696 and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is
697 true if we know that the exit must be taken eventually. */
700 number_of_iterations_lt_to_ne (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
701 struct tree_niter_desc
*niter
,
702 tree
*delta
, tree step
,
703 bool exit_must_be_taken
, bounds
*bnds
)
705 tree niter_type
= TREE_TYPE (step
);
706 tree mod
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, *delta
, step
);
709 tree assumption
= boolean_true_node
, bound
, noloop
;
710 bool ret
= false, fv_comp_no_overflow
;
712 if (POINTER_TYPE_P (type
))
715 if (TREE_CODE (mod
) != INTEGER_CST
)
717 if (integer_nonzerop (mod
))
718 mod
= fold_build2 (MINUS_EXPR
, niter_type
, step
, mod
);
719 tmod
= fold_convert (type1
, mod
);
722 mpz_set_double_int (mmod
, tree_to_double_int (mod
), true);
723 mpz_neg (mmod
, mmod
);
725 /* If the induction variable does not overflow and the exit is taken,
726 then the computation of the final value does not overflow. This is
727 also obviously the case if the new final value is equal to the
728 current one. Finally, we postulate this for pointer type variables,
729 as the code cannot rely on the object to that the pointer points being
730 placed at the end of the address space (and more pragmatically,
731 TYPE_{MIN,MAX}_VALUE is not defined for pointers). */
732 if (integer_zerop (mod
) || POINTER_TYPE_P (type
))
733 fv_comp_no_overflow
= true;
734 else if (!exit_must_be_taken
)
735 fv_comp_no_overflow
= false;
737 fv_comp_no_overflow
=
738 (iv0
->no_overflow
&& integer_nonzerop (iv0
->step
))
739 || (iv1
->no_overflow
&& integer_nonzerop (iv1
->step
));
741 if (integer_nonzerop (iv0
->step
))
743 /* The final value of the iv is iv1->base + MOD, assuming that this
744 computation does not overflow, and that
745 iv0->base <= iv1->base + MOD. */
746 if (!fv_comp_no_overflow
)
748 bound
= fold_build2 (MINUS_EXPR
, type1
,
749 TYPE_MAX_VALUE (type1
), tmod
);
750 assumption
= fold_build2 (LE_EXPR
, boolean_type_node
,
752 if (integer_zerop (assumption
))
755 if (mpz_cmp (mmod
, bnds
->below
) < 0)
756 noloop
= boolean_false_node
;
757 else if (POINTER_TYPE_P (type
))
758 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
760 fold_build_pointer_plus (iv1
->base
, tmod
));
762 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
764 fold_build2 (PLUS_EXPR
, type1
,
769 /* The final value of the iv is iv0->base - MOD, assuming that this
770 computation does not overflow, and that
771 iv0->base - MOD <= iv1->base. */
772 if (!fv_comp_no_overflow
)
774 bound
= fold_build2 (PLUS_EXPR
, type1
,
775 TYPE_MIN_VALUE (type1
), tmod
);
776 assumption
= fold_build2 (GE_EXPR
, boolean_type_node
,
778 if (integer_zerop (assumption
))
781 if (mpz_cmp (mmod
, bnds
->below
) < 0)
782 noloop
= boolean_false_node
;
783 else if (POINTER_TYPE_P (type
))
784 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
785 fold_build_pointer_plus (iv0
->base
,
786 fold_build1 (NEGATE_EXPR
,
790 noloop
= fold_build2 (GT_EXPR
, boolean_type_node
,
791 fold_build2 (MINUS_EXPR
, type1
,
796 if (!integer_nonzerop (assumption
))
797 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
800 if (!integer_zerop (noloop
))
801 niter
->may_be_zero
= fold_build2 (TRUTH_OR_EXPR
, boolean_type_node
,
804 bounds_add (bnds
, tree_to_double_int (mod
), type
);
805 *delta
= fold_build2 (PLUS_EXPR
, niter_type
, *delta
, mod
);
813 /* Add assertions to NITER that ensure that the control variable of the loop
814 with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1
815 are TYPE. Returns false if we can prove that there is an overflow, true
816 otherwise. STEP is the absolute value of the step. */
819 assert_no_overflow_lt (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
820 struct tree_niter_desc
*niter
, tree step
)
822 tree bound
, d
, assumption
, diff
;
823 tree niter_type
= TREE_TYPE (step
);
825 if (integer_nonzerop (iv0
->step
))
827 /* for (i = iv0->base; i < iv1->base; i += iv0->step) */
828 if (iv0
->no_overflow
)
831 /* If iv0->base is a constant, we can determine the last value before
832 overflow precisely; otherwise we conservatively assume
835 if (TREE_CODE (iv0
->base
) == INTEGER_CST
)
837 d
= fold_build2 (MINUS_EXPR
, niter_type
,
838 fold_convert (niter_type
, TYPE_MAX_VALUE (type
)),
839 fold_convert (niter_type
, iv0
->base
));
840 diff
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, d
, step
);
843 diff
= fold_build2 (MINUS_EXPR
, niter_type
, step
,
844 build_int_cst (niter_type
, 1));
845 bound
= fold_build2 (MINUS_EXPR
, type
,
846 TYPE_MAX_VALUE (type
), fold_convert (type
, diff
));
847 assumption
= fold_build2 (LE_EXPR
, boolean_type_node
,
852 /* for (i = iv1->base; i > iv0->base; i += iv1->step) */
853 if (iv1
->no_overflow
)
856 if (TREE_CODE (iv1
->base
) == INTEGER_CST
)
858 d
= fold_build2 (MINUS_EXPR
, niter_type
,
859 fold_convert (niter_type
, iv1
->base
),
860 fold_convert (niter_type
, TYPE_MIN_VALUE (type
)));
861 diff
= fold_build2 (FLOOR_MOD_EXPR
, niter_type
, d
, step
);
864 diff
= fold_build2 (MINUS_EXPR
, niter_type
, step
,
865 build_int_cst (niter_type
, 1));
866 bound
= fold_build2 (PLUS_EXPR
, type
,
867 TYPE_MIN_VALUE (type
), fold_convert (type
, diff
));
868 assumption
= fold_build2 (GE_EXPR
, boolean_type_node
,
872 if (integer_zerop (assumption
))
874 if (!integer_nonzerop (assumption
))
875 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
876 niter
->assumptions
, assumption
);
878 iv0
->no_overflow
= true;
879 iv1
->no_overflow
= true;
883 /* Add an assumption to NITER that a loop whose ending condition
884 is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS
885 bounds the value of IV1->base - IV0->base. */
888 assert_loop_rolls_lt (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
889 struct tree_niter_desc
*niter
, bounds
*bnds
)
891 tree assumption
= boolean_true_node
, bound
, diff
;
892 tree mbz
, mbzl
, mbzr
, type1
;
893 bool rolls_p
, no_overflow_p
;
897 /* We are going to compute the number of iterations as
898 (iv1->base - iv0->base + step - 1) / step, computed in the unsigned
899 variant of TYPE. This formula only works if
901 -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1
903 (where MAX is the maximum value of the unsigned variant of TYPE, and
904 the computations in this formula are performed in full precision,
905 i.e., without overflows).
907 Usually, for loops with exit condition iv0->base + step * i < iv1->base,
908 we have a condition of the form iv0->base - step < iv1->base before the loop,
909 and for loops iv0->base < iv1->base - step * i the condition
910 iv0->base < iv1->base + step, due to loop header copying, which enable us
911 to prove the lower bound.
913 The upper bound is more complicated. Unless the expressions for initial
914 and final value themselves contain enough information, we usually cannot
915 derive it from the context. */
917 /* First check whether the answer does not follow from the bounds we gathered
919 if (integer_nonzerop (iv0
->step
))
920 dstep
= tree_to_double_int (iv0
->step
);
923 dstep
= tree_to_double_int (iv1
->step
).sext (TYPE_PRECISION (type
));
928 mpz_set_double_int (mstep
, dstep
, true);
929 mpz_neg (mstep
, mstep
);
930 mpz_add_ui (mstep
, mstep
, 1);
932 rolls_p
= mpz_cmp (mstep
, bnds
->below
) <= 0;
935 mpz_set_double_int (max
, double_int::mask (TYPE_PRECISION (type
)), true);
936 mpz_add (max
, max
, mstep
);
937 no_overflow_p
= (mpz_cmp (bnds
->up
, max
) <= 0
938 /* For pointers, only values lying inside a single object
939 can be compared or manipulated by pointer arithmetics.
940 Gcc in general does not allow or handle objects larger
941 than half of the address space, hence the upper bound
942 is satisfied for pointers. */
943 || POINTER_TYPE_P (type
));
947 if (rolls_p
&& no_overflow_p
)
951 if (POINTER_TYPE_P (type
))
954 /* Now the hard part; we must formulate the assumption(s) as expressions, and
955 we must be careful not to introduce overflow. */
957 if (integer_nonzerop (iv0
->step
))
959 diff
= fold_build2 (MINUS_EXPR
, type1
,
960 iv0
->step
, build_int_cst (type1
, 1));
962 /* We need to know that iv0->base >= MIN + iv0->step - 1. Since
963 0 address never belongs to any object, we can assume this for
965 if (!POINTER_TYPE_P (type
))
967 bound
= fold_build2 (PLUS_EXPR
, type1
,
968 TYPE_MIN_VALUE (type
), diff
);
969 assumption
= fold_build2 (GE_EXPR
, boolean_type_node
,
973 /* And then we can compute iv0->base - diff, and compare it with
975 mbzl
= fold_build2 (MINUS_EXPR
, type1
,
976 fold_convert (type1
, iv0
->base
), diff
);
977 mbzr
= fold_convert (type1
, iv1
->base
);
981 diff
= fold_build2 (PLUS_EXPR
, type1
,
982 iv1
->step
, build_int_cst (type1
, 1));
984 if (!POINTER_TYPE_P (type
))
986 bound
= fold_build2 (PLUS_EXPR
, type1
,
987 TYPE_MAX_VALUE (type
), diff
);
988 assumption
= fold_build2 (LE_EXPR
, boolean_type_node
,
992 mbzl
= fold_convert (type1
, iv0
->base
);
993 mbzr
= fold_build2 (MINUS_EXPR
, type1
,
994 fold_convert (type1
, iv1
->base
), diff
);
997 if (!integer_nonzerop (assumption
))
998 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
999 niter
->assumptions
, assumption
);
1002 mbz
= fold_build2 (GT_EXPR
, boolean_type_node
, mbzl
, mbzr
);
1003 niter
->may_be_zero
= fold_build2 (TRUTH_OR_EXPR
, boolean_type_node
,
1004 niter
->may_be_zero
, mbz
);
1008 /* Determines number of iterations of loop whose ending condition
1009 is IV0 < IV1. TYPE is the type of the iv. The number of
1010 iterations is stored to NITER. BNDS bounds the difference
1011 IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know
1012 that the exit must be taken eventually. */
1015 number_of_iterations_lt (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
1016 struct tree_niter_desc
*niter
,
1017 bool exit_must_be_taken
, bounds
*bnds
)
1019 tree niter_type
= unsigned_type_for (type
);
1020 tree delta
, step
, s
;
1023 if (integer_nonzerop (iv0
->step
))
1025 niter
->control
= *iv0
;
1026 niter
->cmp
= LT_EXPR
;
1027 niter
->bound
= iv1
->base
;
1031 niter
->control
= *iv1
;
1032 niter
->cmp
= GT_EXPR
;
1033 niter
->bound
= iv0
->base
;
1036 delta
= fold_build2 (MINUS_EXPR
, niter_type
,
1037 fold_convert (niter_type
, iv1
->base
),
1038 fold_convert (niter_type
, iv0
->base
));
1040 /* First handle the special case that the step is +-1. */
1041 if ((integer_onep (iv0
->step
) && integer_zerop (iv1
->step
))
1042 || (integer_all_onesp (iv1
->step
) && integer_zerop (iv0
->step
)))
1044 /* for (i = iv0->base; i < iv1->base; i++)
1048 for (i = iv1->base; i > iv0->base; i--).
1050 In both cases # of iterations is iv1->base - iv0->base, assuming that
1051 iv1->base >= iv0->base.
1053 First try to derive a lower bound on the value of
1054 iv1->base - iv0->base, computed in full precision. If the difference
1055 is nonnegative, we are done, otherwise we must record the
1058 if (mpz_sgn (bnds
->below
) < 0)
1059 niter
->may_be_zero
= fold_build2 (LT_EXPR
, boolean_type_node
,
1060 iv1
->base
, iv0
->base
);
1061 niter
->niter
= delta
;
1062 niter
->max
= mpz_get_double_int (niter_type
, bnds
->up
, false);
1066 if (integer_nonzerop (iv0
->step
))
1067 step
= fold_convert (niter_type
, iv0
->step
);
1069 step
= fold_convert (niter_type
,
1070 fold_build1 (NEGATE_EXPR
, type
, iv1
->step
));
1072 /* If we can determine the final value of the control iv exactly, we can
1073 transform the condition to != comparison. In particular, this will be
1074 the case if DELTA is constant. */
1075 if (number_of_iterations_lt_to_ne (type
, iv0
, iv1
, niter
, &delta
, step
,
1076 exit_must_be_taken
, bnds
))
1080 zps
.base
= build_int_cst (niter_type
, 0);
1082 /* number_of_iterations_lt_to_ne will add assumptions that ensure that
1083 zps does not overflow. */
1084 zps
.no_overflow
= true;
1086 return number_of_iterations_ne (type
, &zps
, delta
, niter
, true, bnds
);
1089 /* Make sure that the control iv does not overflow. */
1090 if (!assert_no_overflow_lt (type
, iv0
, iv1
, niter
, step
))
1093 /* We determine the number of iterations as (delta + step - 1) / step. For
1094 this to work, we must know that iv1->base >= iv0->base - step + 1,
1095 otherwise the loop does not roll. */
1096 assert_loop_rolls_lt (type
, iv0
, iv1
, niter
, bnds
);
1098 s
= fold_build2 (MINUS_EXPR
, niter_type
,
1099 step
, build_int_cst (niter_type
, 1));
1100 delta
= fold_build2 (PLUS_EXPR
, niter_type
, delta
, s
);
1101 niter
->niter
= fold_build2 (FLOOR_DIV_EXPR
, niter_type
, delta
, step
);
1105 mpz_set_double_int (mstep
, tree_to_double_int (step
), true);
1106 mpz_add (tmp
, bnds
->up
, mstep
);
1107 mpz_sub_ui (tmp
, tmp
, 1);
1108 mpz_fdiv_q (tmp
, tmp
, mstep
);
1109 niter
->max
= mpz_get_double_int (niter_type
, tmp
, false);
1116 /* Determines number of iterations of loop whose ending condition
1117 is IV0 <= IV1. TYPE is the type of the iv. The number of
1118 iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if
1119 we know that this condition must eventually become false (we derived this
1120 earlier, and possibly set NITER->assumptions to make sure this
1121 is the case). BNDS bounds the difference IV1->base - IV0->base. */
1124 number_of_iterations_le (tree type
, affine_iv
*iv0
, affine_iv
*iv1
,
1125 struct tree_niter_desc
*niter
, bool exit_must_be_taken
,
1130 if (POINTER_TYPE_P (type
))
1133 /* Say that IV0 is the control variable. Then IV0 <= IV1 iff
1134 IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest
1135 value of the type. This we must know anyway, since if it is
1136 equal to this value, the loop rolls forever. We do not check
1137 this condition for pointer type ivs, as the code cannot rely on
1138 the object to that the pointer points being placed at the end of
1139 the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is
1140 not defined for pointers). */
1142 if (!exit_must_be_taken
&& !POINTER_TYPE_P (type
))
1144 if (integer_nonzerop (iv0
->step
))
1145 assumption
= fold_build2 (NE_EXPR
, boolean_type_node
,
1146 iv1
->base
, TYPE_MAX_VALUE (type
));
1148 assumption
= fold_build2 (NE_EXPR
, boolean_type_node
,
1149 iv0
->base
, TYPE_MIN_VALUE (type
));
1151 if (integer_zerop (assumption
))
1153 if (!integer_nonzerop (assumption
))
1154 niter
->assumptions
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1155 niter
->assumptions
, assumption
);
1158 if (integer_nonzerop (iv0
->step
))
1160 if (POINTER_TYPE_P (type
))
1161 iv1
->base
= fold_build_pointer_plus_hwi (iv1
->base
, 1);
1163 iv1
->base
= fold_build2 (PLUS_EXPR
, type1
, iv1
->base
,
1164 build_int_cst (type1
, 1));
1166 else if (POINTER_TYPE_P (type
))
1167 iv0
->base
= fold_build_pointer_plus_hwi (iv0
->base
, -1);
1169 iv0
->base
= fold_build2 (MINUS_EXPR
, type1
,
1170 iv0
->base
, build_int_cst (type1
, 1));
1172 bounds_add (bnds
, double_int_one
, type1
);
1174 return number_of_iterations_lt (type
, iv0
, iv1
, niter
, exit_must_be_taken
,
1178 /* Dumps description of affine induction variable IV to FILE. */
1181 dump_affine_iv (FILE *file
, affine_iv
*iv
)
1183 if (!integer_zerop (iv
->step
))
1184 fprintf (file
, "[");
1186 print_generic_expr (dump_file
, iv
->base
, TDF_SLIM
);
1188 if (!integer_zerop (iv
->step
))
1190 fprintf (file
, ", + , ");
1191 print_generic_expr (dump_file
, iv
->step
, TDF_SLIM
);
1192 fprintf (file
, "]%s", iv
->no_overflow
? "(no_overflow)" : "");
1196 /* Determine the number of iterations according to condition (for staying
1197 inside loop) which compares two induction variables using comparison
1198 operator CODE. The induction variable on left side of the comparison
1199 is IV0, the right-hand side is IV1. Both induction variables must have
1200 type TYPE, which must be an integer or pointer type. The steps of the
1201 ivs must be constants (or NULL_TREE, which is interpreted as constant zero).
1203 LOOP is the loop whose number of iterations we are determining.
1205 ONLY_EXIT is true if we are sure this is the only way the loop could be
1206 exited (including possibly non-returning function calls, exceptions, etc.)
1207 -- in this case we can use the information whether the control induction
1208 variables can overflow or not in a more efficient way.
1210 if EVERY_ITERATION is true, we know the test is executed on every iteration.
1212 The results (number of iterations and assumptions as described in
1213 comments at struct tree_niter_desc in tree-flow.h) are stored to NITER.
1214 Returns false if it fails to determine number of iterations, true if it
1215 was determined (possibly with some assumptions). */
1218 number_of_iterations_cond (struct loop
*loop
,
1219 tree type
, affine_iv
*iv0
, enum tree_code code
,
1220 affine_iv
*iv1
, struct tree_niter_desc
*niter
,
1221 bool only_exit
, bool every_iteration
)
1223 bool exit_must_be_taken
= false, ret
;
1226 /* If the test is not executed every iteration, wrapping may make the test
1228 TODO: the overflow case can be still used as unreliable estimate of upper
1229 bound. But we have no API to pass it down to number of iterations code
1230 and, at present, it will not use it anyway. */
1231 if (!every_iteration
1232 && (!iv0
->no_overflow
|| !iv1
->no_overflow
1233 || code
== NE_EXPR
|| code
== EQ_EXPR
))
1236 /* The meaning of these assumptions is this:
1238 then the rest of information does not have to be valid
1239 if may_be_zero then the loop does not roll, even if
1241 niter
->assumptions
= boolean_true_node
;
1242 niter
->may_be_zero
= boolean_false_node
;
1243 niter
->niter
= NULL_TREE
;
1244 niter
->max
= double_int_zero
;
1246 niter
->bound
= NULL_TREE
;
1247 niter
->cmp
= ERROR_MARK
;
1249 /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that
1250 the control variable is on lhs. */
1251 if (code
== GE_EXPR
|| code
== GT_EXPR
1252 || (code
== NE_EXPR
&& integer_zerop (iv0
->step
)))
1255 code
= swap_tree_comparison (code
);
1258 if (POINTER_TYPE_P (type
))
1260 /* Comparison of pointers is undefined unless both iv0 and iv1 point
1261 to the same object. If they do, the control variable cannot wrap
1262 (as wrap around the bounds of memory will never return a pointer
1263 that would be guaranteed to point to the same object, even if we
1264 avoid undefined behavior by casting to size_t and back). */
1265 iv0
->no_overflow
= true;
1266 iv1
->no_overflow
= true;
1269 /* If the control induction variable does not overflow and the only exit
1270 from the loop is the one that we analyze, we know it must be taken
1274 if (!integer_zerop (iv0
->step
) && iv0
->no_overflow
)
1275 exit_must_be_taken
= true;
1276 else if (!integer_zerop (iv1
->step
) && iv1
->no_overflow
)
1277 exit_must_be_taken
= true;
1280 /* We can handle the case when neither of the sides of the comparison is
1281 invariant, provided that the test is NE_EXPR. This rarely occurs in
1282 practice, but it is simple enough to manage. */
1283 if (!integer_zerop (iv0
->step
) && !integer_zerop (iv1
->step
))
1285 tree step_type
= POINTER_TYPE_P (type
) ? sizetype
: type
;
1286 if (code
!= NE_EXPR
)
1289 iv0
->step
= fold_binary_to_constant (MINUS_EXPR
, step_type
,
1290 iv0
->step
, iv1
->step
);
1291 iv0
->no_overflow
= false;
1292 iv1
->step
= build_int_cst (step_type
, 0);
1293 iv1
->no_overflow
= true;
1296 /* If the result of the comparison is a constant, the loop is weird. More
1297 precise handling would be possible, but the situation is not common enough
1298 to waste time on it. */
1299 if (integer_zerop (iv0
->step
) && integer_zerop (iv1
->step
))
1302 /* Ignore loops of while (i-- < 10) type. */
1303 if (code
!= NE_EXPR
)
1305 if (iv0
->step
&& tree_int_cst_sign_bit (iv0
->step
))
1308 if (!integer_zerop (iv1
->step
) && !tree_int_cst_sign_bit (iv1
->step
))
1312 /* If the loop exits immediately, there is nothing to do. */
1313 if (integer_zerop (fold_build2 (code
, boolean_type_node
, iv0
->base
, iv1
->base
)))
1315 niter
->niter
= build_int_cst (unsigned_type_for (type
), 0);
1316 niter
->max
= double_int_zero
;
1320 /* OK, now we know we have a senseful loop. Handle several cases, depending
1321 on what comparison operator is used. */
1322 bound_difference (loop
, iv1
->base
, iv0
->base
, &bnds
);
1324 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1327 "Analyzing # of iterations of loop %d\n", loop
->num
);
1329 fprintf (dump_file
, " exit condition ");
1330 dump_affine_iv (dump_file
, iv0
);
1331 fprintf (dump_file
, " %s ",
1332 code
== NE_EXPR
? "!="
1333 : code
== LT_EXPR
? "<"
1335 dump_affine_iv (dump_file
, iv1
);
1336 fprintf (dump_file
, "\n");
1338 fprintf (dump_file
, " bounds on difference of bases: ");
1339 mpz_out_str (dump_file
, 10, bnds
.below
);
1340 fprintf (dump_file
, " ... ");
1341 mpz_out_str (dump_file
, 10, bnds
.up
);
1342 fprintf (dump_file
, "\n");
1348 gcc_assert (integer_zerop (iv1
->step
));
1349 ret
= number_of_iterations_ne (type
, iv0
, iv1
->base
, niter
,
1350 exit_must_be_taken
, &bnds
);
1354 ret
= number_of_iterations_lt (type
, iv0
, iv1
, niter
, exit_must_be_taken
,
1359 ret
= number_of_iterations_le (type
, iv0
, iv1
, niter
, exit_must_be_taken
,
1367 mpz_clear (bnds
.up
);
1368 mpz_clear (bnds
.below
);
1370 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1374 fprintf (dump_file
, " result:\n");
1375 if (!integer_nonzerop (niter
->assumptions
))
1377 fprintf (dump_file
, " under assumptions ");
1378 print_generic_expr (dump_file
, niter
->assumptions
, TDF_SLIM
);
1379 fprintf (dump_file
, "\n");
1382 if (!integer_zerop (niter
->may_be_zero
))
1384 fprintf (dump_file
, " zero if ");
1385 print_generic_expr (dump_file
, niter
->may_be_zero
, TDF_SLIM
);
1386 fprintf (dump_file
, "\n");
1389 fprintf (dump_file
, " # of iterations ");
1390 print_generic_expr (dump_file
, niter
->niter
, TDF_SLIM
);
1391 fprintf (dump_file
, ", bounded by ");
1392 dump_double_int (dump_file
, niter
->max
, true);
1393 fprintf (dump_file
, "\n");
1396 fprintf (dump_file
, " failed\n\n");
1401 /* Substitute NEW for OLD in EXPR and fold the result. */
1404 simplify_replace_tree (tree expr
, tree old
, tree new_tree
)
1407 tree ret
= NULL_TREE
, e
, se
;
1412 /* Do not bother to replace constants. */
1413 if (CONSTANT_CLASS_P (old
))
1417 || operand_equal_p (expr
, old
, 0))
1418 return unshare_expr (new_tree
);
1423 n
= TREE_OPERAND_LENGTH (expr
);
1424 for (i
= 0; i
< n
; i
++)
1426 e
= TREE_OPERAND (expr
, i
);
1427 se
= simplify_replace_tree (e
, old
, new_tree
);
1432 ret
= copy_node (expr
);
1434 TREE_OPERAND (ret
, i
) = se
;
1437 return (ret
? fold (ret
) : expr
);
1440 /* Expand definitions of ssa names in EXPR as long as they are simple
1441 enough, and return the new expression. */
1444 expand_simple_operations (tree expr
)
1447 tree ret
= NULL_TREE
, e
, ee
, e1
;
1448 enum tree_code code
;
1451 if (expr
== NULL_TREE
)
1454 if (is_gimple_min_invariant (expr
))
1457 code
= TREE_CODE (expr
);
1458 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code
)))
1460 n
= TREE_OPERAND_LENGTH (expr
);
1461 for (i
= 0; i
< n
; i
++)
1463 e
= TREE_OPERAND (expr
, i
);
1464 ee
= expand_simple_operations (e
);
1469 ret
= copy_node (expr
);
1471 TREE_OPERAND (ret
, i
) = ee
;
1477 fold_defer_overflow_warnings ();
1479 fold_undefer_and_ignore_overflow_warnings ();
1483 if (TREE_CODE (expr
) != SSA_NAME
)
1486 stmt
= SSA_NAME_DEF_STMT (expr
);
1487 if (gimple_code (stmt
) == GIMPLE_PHI
)
1489 basic_block src
, dest
;
1491 if (gimple_phi_num_args (stmt
) != 1)
1493 e
= PHI_ARG_DEF (stmt
, 0);
1495 /* Avoid propagating through loop exit phi nodes, which
1496 could break loop-closed SSA form restrictions. */
1497 dest
= gimple_bb (stmt
);
1498 src
= single_pred (dest
);
1499 if (TREE_CODE (e
) == SSA_NAME
1500 && src
->loop_father
!= dest
->loop_father
)
1503 return expand_simple_operations (e
);
1505 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
1508 e
= gimple_assign_rhs1 (stmt
);
1509 code
= gimple_assign_rhs_code (stmt
);
1510 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1512 if (is_gimple_min_invariant (e
))
1515 if (code
== SSA_NAME
)
1516 return expand_simple_operations (e
);
1524 /* Casts are simple. */
1525 ee
= expand_simple_operations (e
);
1526 return fold_build1 (code
, TREE_TYPE (expr
), ee
);
1530 case POINTER_PLUS_EXPR
:
1531 /* And increments and decrements by a constant are simple. */
1532 e1
= gimple_assign_rhs2 (stmt
);
1533 if (!is_gimple_min_invariant (e1
))
1536 ee
= expand_simple_operations (e
);
1537 return fold_build2 (code
, TREE_TYPE (expr
), ee
, e1
);
1544 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1545 expression (or EXPR unchanged, if no simplification was possible). */
1548 tree_simplify_using_condition_1 (tree cond
, tree expr
)
1551 tree e
, te
, e0
, e1
, e2
, notcond
;
1552 enum tree_code code
= TREE_CODE (expr
);
1554 if (code
== INTEGER_CST
)
1557 if (code
== TRUTH_OR_EXPR
1558 || code
== TRUTH_AND_EXPR
1559 || code
== COND_EXPR
)
1563 e0
= tree_simplify_using_condition_1 (cond
, TREE_OPERAND (expr
, 0));
1564 if (TREE_OPERAND (expr
, 0) != e0
)
1567 e1
= tree_simplify_using_condition_1 (cond
, TREE_OPERAND (expr
, 1));
1568 if (TREE_OPERAND (expr
, 1) != e1
)
1571 if (code
== COND_EXPR
)
1573 e2
= tree_simplify_using_condition_1 (cond
, TREE_OPERAND (expr
, 2));
1574 if (TREE_OPERAND (expr
, 2) != e2
)
1582 if (code
== COND_EXPR
)
1583 expr
= fold_build3 (code
, boolean_type_node
, e0
, e1
, e2
);
1585 expr
= fold_build2 (code
, boolean_type_node
, e0
, e1
);
1591 /* In case COND is equality, we may be able to simplify EXPR by copy/constant
1592 propagation, and vice versa. Fold does not handle this, since it is
1593 considered too expensive. */
1594 if (TREE_CODE (cond
) == EQ_EXPR
)
1596 e0
= TREE_OPERAND (cond
, 0);
1597 e1
= TREE_OPERAND (cond
, 1);
1599 /* We know that e0 == e1. Check whether we cannot simplify expr
1601 e
= simplify_replace_tree (expr
, e0
, e1
);
1602 if (integer_zerop (e
) || integer_nonzerop (e
))
1605 e
= simplify_replace_tree (expr
, e1
, e0
);
1606 if (integer_zerop (e
) || integer_nonzerop (e
))
1609 if (TREE_CODE (expr
) == EQ_EXPR
)
1611 e0
= TREE_OPERAND (expr
, 0);
1612 e1
= TREE_OPERAND (expr
, 1);
1614 /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */
1615 e
= simplify_replace_tree (cond
, e0
, e1
);
1616 if (integer_zerop (e
))
1618 e
= simplify_replace_tree (cond
, e1
, e0
);
1619 if (integer_zerop (e
))
1622 if (TREE_CODE (expr
) == NE_EXPR
)
1624 e0
= TREE_OPERAND (expr
, 0);
1625 e1
= TREE_OPERAND (expr
, 1);
1627 /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */
1628 e
= simplify_replace_tree (cond
, e0
, e1
);
1629 if (integer_zerop (e
))
1630 return boolean_true_node
;
1631 e
= simplify_replace_tree (cond
, e1
, e0
);
1632 if (integer_zerop (e
))
1633 return boolean_true_node
;
1636 te
= expand_simple_operations (expr
);
1638 /* Check whether COND ==> EXPR. */
1639 notcond
= invert_truthvalue (cond
);
1640 e
= fold_binary (TRUTH_OR_EXPR
, boolean_type_node
, notcond
, te
);
1641 if (e
&& integer_nonzerop (e
))
1644 /* Check whether COND ==> not EXPR. */
1645 e
= fold_binary (TRUTH_AND_EXPR
, boolean_type_node
, cond
, te
);
1646 if (e
&& integer_zerop (e
))
1652 /* Tries to simplify EXPR using the condition COND. Returns the simplified
1653 expression (or EXPR unchanged, if no simplification was possible).
1654 Wrapper around tree_simplify_using_condition_1 that ensures that chains
1655 of simple operations in definitions of ssa names in COND are expanded,
1656 so that things like casts or incrementing the value of the bound before
1657 the loop do not cause us to fail. */
1660 tree_simplify_using_condition (tree cond
, tree expr
)
1662 cond
= expand_simple_operations (cond
);
1664 return tree_simplify_using_condition_1 (cond
, expr
);
1667 /* Tries to simplify EXPR using the conditions on entry to LOOP.
1668 Returns the simplified expression (or EXPR unchanged, if no
1669 simplification was possible).*/
1672 simplify_using_initial_conditions (struct loop
*loop
, tree expr
)
1680 if (TREE_CODE (expr
) == INTEGER_CST
)
1683 /* Limit walking the dominators to avoid quadraticness in
1684 the number of BBs times the number of loops in degenerate
1686 for (bb
= loop
->header
;
1687 bb
!= ENTRY_BLOCK_PTR
&& cnt
< MAX_DOMINATORS_TO_WALK
;
1688 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
1690 if (!single_pred_p (bb
))
1692 e
= single_pred_edge (bb
);
1694 if (!(e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)))
1697 stmt
= last_stmt (e
->src
);
1698 cond
= fold_build2 (gimple_cond_code (stmt
),
1700 gimple_cond_lhs (stmt
),
1701 gimple_cond_rhs (stmt
));
1702 if (e
->flags
& EDGE_FALSE_VALUE
)
1703 cond
= invert_truthvalue (cond
);
1704 expr
= tree_simplify_using_condition (cond
, expr
);
1711 /* Tries to simplify EXPR using the evolutions of the loop invariants
1712 in the superloops of LOOP. Returns the simplified expression
1713 (or EXPR unchanged, if no simplification was possible). */
1716 simplify_using_outer_evolutions (struct loop
*loop
, tree expr
)
1718 enum tree_code code
= TREE_CODE (expr
);
1722 if (is_gimple_min_invariant (expr
))
1725 if (code
== TRUTH_OR_EXPR
1726 || code
== TRUTH_AND_EXPR
1727 || code
== COND_EXPR
)
1731 e0
= simplify_using_outer_evolutions (loop
, TREE_OPERAND (expr
, 0));
1732 if (TREE_OPERAND (expr
, 0) != e0
)
1735 e1
= simplify_using_outer_evolutions (loop
, TREE_OPERAND (expr
, 1));
1736 if (TREE_OPERAND (expr
, 1) != e1
)
1739 if (code
== COND_EXPR
)
1741 e2
= simplify_using_outer_evolutions (loop
, TREE_OPERAND (expr
, 2));
1742 if (TREE_OPERAND (expr
, 2) != e2
)
1750 if (code
== COND_EXPR
)
1751 expr
= fold_build3 (code
, boolean_type_node
, e0
, e1
, e2
);
1753 expr
= fold_build2 (code
, boolean_type_node
, e0
, e1
);
1759 e
= instantiate_parameters (loop
, expr
);
1760 if (is_gimple_min_invariant (e
))
1766 /* Returns true if EXIT is the only possible exit from LOOP. */
1769 loop_only_exit_p (const struct loop
*loop
, const_edge exit
)
1772 gimple_stmt_iterator bsi
;
1776 if (exit
!= single_exit (loop
))
1779 body
= get_loop_body (loop
);
1780 for (i
= 0; i
< loop
->num_nodes
; i
++)
1782 for (bsi
= gsi_start_bb (body
[i
]); !gsi_end_p (bsi
); gsi_next (&bsi
))
1784 call
= gsi_stmt (bsi
);
1785 if (gimple_code (call
) != GIMPLE_CALL
)
1788 if (gimple_has_side_effects (call
))
1800 /* Stores description of number of iterations of LOOP derived from
1801 EXIT (an exit edge of the LOOP) in NITER. Returns true if some
1802 useful information could be derived (and fields of NITER has
1803 meaning described in comments at struct tree_niter_desc
1804 declaration), false otherwise. If WARN is true and
1805 -Wunsafe-loop-optimizations was given, warn if the optimizer is going to use
1806 potentially unsafe assumptions.
1807 When EVERY_ITERATION is true, only tests that are known to be executed
1808 every iteration are considered (i.e. only test that alone bounds the loop).
1812 number_of_iterations_exit (struct loop
*loop
, edge exit
,
1813 struct tree_niter_desc
*niter
,
1814 bool warn
, bool every_iteration
)
1819 enum tree_code code
;
1823 safe
= dominated_by_p (CDI_DOMINATORS
, loop
->latch
, exit
->src
);
1825 if (every_iteration
&& !safe
)
1828 niter
->assumptions
= boolean_false_node
;
1829 stmt
= last_stmt (exit
->src
);
1830 if (!stmt
|| gimple_code (stmt
) != GIMPLE_COND
)
1833 /* We want the condition for staying inside loop. */
1834 code
= gimple_cond_code (stmt
);
1835 if (exit
->flags
& EDGE_TRUE_VALUE
)
1836 code
= invert_tree_comparison (code
, false);
1851 op0
= gimple_cond_lhs (stmt
);
1852 op1
= gimple_cond_rhs (stmt
);
1853 type
= TREE_TYPE (op0
);
1855 if (TREE_CODE (type
) != INTEGER_TYPE
1856 && !POINTER_TYPE_P (type
))
1859 if (!simple_iv (loop
, loop_containing_stmt (stmt
), op0
, &iv0
, false))
1861 if (!simple_iv (loop
, loop_containing_stmt (stmt
), op1
, &iv1
, false))
1864 /* We don't want to see undefined signed overflow warnings while
1865 computing the number of iterations. */
1866 fold_defer_overflow_warnings ();
1868 iv0
.base
= expand_simple_operations (iv0
.base
);
1869 iv1
.base
= expand_simple_operations (iv1
.base
);
1870 if (!number_of_iterations_cond (loop
, type
, &iv0
, code
, &iv1
, niter
,
1871 loop_only_exit_p (loop
, exit
), safe
))
1873 fold_undefer_and_ignore_overflow_warnings ();
1879 niter
->assumptions
= simplify_using_outer_evolutions (loop
,
1880 niter
->assumptions
);
1881 niter
->may_be_zero
= simplify_using_outer_evolutions (loop
,
1882 niter
->may_be_zero
);
1883 niter
->niter
= simplify_using_outer_evolutions (loop
, niter
->niter
);
1887 = simplify_using_initial_conditions (loop
,
1888 niter
->assumptions
);
1890 = simplify_using_initial_conditions (loop
,
1891 niter
->may_be_zero
);
1893 fold_undefer_and_ignore_overflow_warnings ();
1895 /* If NITER has simplified into a constant, update MAX. */
1896 if (TREE_CODE (niter
->niter
) == INTEGER_CST
)
1897 niter
->max
= tree_to_double_int (niter
->niter
);
1899 if (integer_onep (niter
->assumptions
))
1902 /* With -funsafe-loop-optimizations we assume that nothing bad can happen.
1903 But if we can prove that there is overflow or some other source of weird
1904 behavior, ignore the loop even with -funsafe-loop-optimizations. */
1905 if (integer_zerop (niter
->assumptions
) || !single_exit (loop
))
1908 if (flag_unsafe_loop_optimizations
)
1909 niter
->assumptions
= boolean_true_node
;
1913 const char *wording
;
1914 location_t loc
= gimple_location (stmt
);
1916 /* We can provide a more specific warning if one of the operator is
1917 constant and the other advances by +1 or -1. */
1918 if (!integer_zerop (iv1
.step
)
1919 ? (integer_zerop (iv0
.step
)
1920 && (integer_onep (iv1
.step
) || integer_all_onesp (iv1
.step
)))
1921 : (integer_onep (iv0
.step
) || integer_all_onesp (iv0
.step
)))
1923 flag_unsafe_loop_optimizations
1924 ? N_("assuming that the loop is not infinite")
1925 : N_("cannot optimize possibly infinite loops");
1928 flag_unsafe_loop_optimizations
1929 ? N_("assuming that the loop counter does not overflow")
1930 : N_("cannot optimize loop, the loop counter may overflow");
1932 warning_at ((LOCATION_LINE (loc
) > 0) ? loc
: input_location
,
1933 OPT_Wunsafe_loop_optimizations
, "%s", gettext (wording
));
1936 return flag_unsafe_loop_optimizations
;
1939 /* Try to determine the number of iterations of LOOP. If we succeed,
1940 expression giving number of iterations is returned and *EXIT is
1941 set to the edge from that the information is obtained. Otherwise
1942 chrec_dont_know is returned. */
1945 find_loop_niter (struct loop
*loop
, edge
*exit
)
1948 vec
<edge
> exits
= get_loop_exit_edges (loop
);
1950 tree niter
= NULL_TREE
, aniter
;
1951 struct tree_niter_desc desc
;
1954 FOR_EACH_VEC_ELT (exits
, i
, ex
)
1956 if (!number_of_iterations_exit (loop
, ex
, &desc
, false))
1959 if (integer_nonzerop (desc
.may_be_zero
))
1961 /* We exit in the first iteration through this exit.
1962 We won't find anything better. */
1963 niter
= build_int_cst (unsigned_type_node
, 0);
1968 if (!integer_zerop (desc
.may_be_zero
))
1971 aniter
= desc
.niter
;
1975 /* Nothing recorded yet. */
1981 /* Prefer constants, the lower the better. */
1982 if (TREE_CODE (aniter
) != INTEGER_CST
)
1985 if (TREE_CODE (niter
) != INTEGER_CST
)
1992 if (tree_int_cst_lt (aniter
, niter
))
2001 return niter
? niter
: chrec_dont_know
;
2004 /* Return true if loop is known to have bounded number of iterations. */
2007 finite_loop_p (struct loop
*loop
)
2012 if (flag_unsafe_loop_optimizations
)
2014 flags
= flags_from_decl_or_type (current_function_decl
);
2015 if ((flags
& (ECF_CONST
|ECF_PURE
)) && !(flags
& ECF_LOOPING_CONST_OR_PURE
))
2017 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2018 fprintf (dump_file
, "Found loop %i to be finite: it is within pure or const function.\n",
2023 if (loop
->any_upper_bound
2024 || max_loop_iterations (loop
, &nit
))
2026 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2027 fprintf (dump_file
, "Found loop %i to be finite: upper bound found.\n",
2036 Analysis of a number of iterations of a loop by a brute-force evaluation.
2040 /* Bound on the number of iterations we try to evaluate. */
2042 #define MAX_ITERATIONS_TO_TRACK \
2043 ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK))
2045 /* Returns the loop phi node of LOOP such that ssa name X is derived from its
2046 result by a chain of operations such that all but exactly one of their
2047 operands are constants. */
2050 chain_of_csts_start (struct loop
*loop
, tree x
)
2052 gimple stmt
= SSA_NAME_DEF_STMT (x
);
2054 basic_block bb
= gimple_bb (stmt
);
2055 enum tree_code code
;
2058 || !flow_bb_inside_loop_p (loop
, bb
))
2061 if (gimple_code (stmt
) == GIMPLE_PHI
)
2063 if (bb
== loop
->header
)
2069 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
2072 code
= gimple_assign_rhs_code (stmt
);
2073 if (gimple_references_memory_p (stmt
)
2074 || TREE_CODE_CLASS (code
) == tcc_reference
2075 || (code
== ADDR_EXPR
2076 && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt
))))
2079 use
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_USE
);
2080 if (use
== NULL_TREE
)
2083 return chain_of_csts_start (loop
, use
);
2086 /* Determines whether the expression X is derived from a result of a phi node
2087 in header of LOOP such that
2089 * the derivation of X consists only from operations with constants
2090 * the initial value of the phi node is constant
2091 * the value of the phi node in the next iteration can be derived from the
2092 value in the current iteration by a chain of operations with constants.
2094 If such phi node exists, it is returned, otherwise NULL is returned. */
2097 get_base_for (struct loop
*loop
, tree x
)
2102 if (is_gimple_min_invariant (x
))
2105 phi
= chain_of_csts_start (loop
, x
);
2109 init
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
2110 next
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (loop
));
2112 if (TREE_CODE (next
) != SSA_NAME
)
2115 if (!is_gimple_min_invariant (init
))
2118 if (chain_of_csts_start (loop
, next
) != phi
)
2124 /* Given an expression X, then
2126 * if X is NULL_TREE, we return the constant BASE.
2127 * otherwise X is a SSA name, whose value in the considered loop is derived
2128 by a chain of operations with constant from a result of a phi node in
2129 the header of the loop. Then we return value of X when the value of the
2130 result of this phi node is given by the constant BASE. */
2133 get_val_for (tree x
, tree base
)
2137 gcc_assert (is_gimple_min_invariant (base
));
2142 stmt
= SSA_NAME_DEF_STMT (x
);
2143 if (gimple_code (stmt
) == GIMPLE_PHI
)
2146 gcc_assert (is_gimple_assign (stmt
));
2148 /* STMT must be either an assignment of a single SSA name or an
2149 expression involving an SSA name and a constant. Try to fold that
2150 expression using the value for the SSA name. */
2151 if (gimple_assign_ssa_name_copy_p (stmt
))
2152 return get_val_for (gimple_assign_rhs1 (stmt
), base
);
2153 else if (gimple_assign_rhs_class (stmt
) == GIMPLE_UNARY_RHS
2154 && TREE_CODE (gimple_assign_rhs1 (stmt
)) == SSA_NAME
)
2156 return fold_build1 (gimple_assign_rhs_code (stmt
),
2157 gimple_expr_type (stmt
),
2158 get_val_for (gimple_assign_rhs1 (stmt
), base
));
2160 else if (gimple_assign_rhs_class (stmt
) == GIMPLE_BINARY_RHS
)
2162 tree rhs1
= gimple_assign_rhs1 (stmt
);
2163 tree rhs2
= gimple_assign_rhs2 (stmt
);
2164 if (TREE_CODE (rhs1
) == SSA_NAME
)
2165 rhs1
= get_val_for (rhs1
, base
);
2166 else if (TREE_CODE (rhs2
) == SSA_NAME
)
2167 rhs2
= get_val_for (rhs2
, base
);
2170 return fold_build2 (gimple_assign_rhs_code (stmt
),
2171 gimple_expr_type (stmt
), rhs1
, rhs2
);
2178 /* Tries to count the number of iterations of LOOP till it exits by EXIT
2179 by brute force -- i.e. by determining the value of the operands of the
2180 condition at EXIT in first few iterations of the loop (assuming that
2181 these values are constant) and determining the first one in that the
2182 condition is not satisfied. Returns the constant giving the number
2183 of the iterations of LOOP if successful, chrec_dont_know otherwise. */
2186 loop_niter_by_eval (struct loop
*loop
, edge exit
)
2189 tree op
[2], val
[2], next
[2], aval
[2];
2194 cond
= last_stmt (exit
->src
);
2195 if (!cond
|| gimple_code (cond
) != GIMPLE_COND
)
2196 return chrec_dont_know
;
2198 cmp
= gimple_cond_code (cond
);
2199 if (exit
->flags
& EDGE_TRUE_VALUE
)
2200 cmp
= invert_tree_comparison (cmp
, false);
2210 op
[0] = gimple_cond_lhs (cond
);
2211 op
[1] = gimple_cond_rhs (cond
);
2215 return chrec_dont_know
;
2218 for (j
= 0; j
< 2; j
++)
2220 if (is_gimple_min_invariant (op
[j
]))
2223 next
[j
] = NULL_TREE
;
2228 phi
= get_base_for (loop
, op
[j
]);
2230 return chrec_dont_know
;
2231 val
[j
] = PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
2232 next
[j
] = PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (loop
));
2236 /* Don't issue signed overflow warnings. */
2237 fold_defer_overflow_warnings ();
2239 for (i
= 0; i
< MAX_ITERATIONS_TO_TRACK
; i
++)
2241 for (j
= 0; j
< 2; j
++)
2242 aval
[j
] = get_val_for (op
[j
], val
[j
]);
2244 acnd
= fold_binary (cmp
, boolean_type_node
, aval
[0], aval
[1]);
2245 if (acnd
&& integer_zerop (acnd
))
2247 fold_undefer_and_ignore_overflow_warnings ();
2248 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2250 "Proved that loop %d iterates %d times using brute force.\n",
2252 return build_int_cst (unsigned_type_node
, i
);
2255 for (j
= 0; j
< 2; j
++)
2257 val
[j
] = get_val_for (next
[j
], val
[j
]);
2258 if (!is_gimple_min_invariant (val
[j
]))
2260 fold_undefer_and_ignore_overflow_warnings ();
2261 return chrec_dont_know
;
2266 fold_undefer_and_ignore_overflow_warnings ();
2268 return chrec_dont_know
;
2271 /* Finds the exit of the LOOP by that the loop exits after a constant
2272 number of iterations and stores the exit edge to *EXIT. The constant
2273 giving the number of iterations of LOOP is returned. The number of
2274 iterations is determined using loop_niter_by_eval (i.e. by brute force
2275 evaluation). If we are unable to find the exit for that loop_niter_by_eval
2276 determines the number of iterations, chrec_dont_know is returned. */
2279 find_loop_niter_by_eval (struct loop
*loop
, edge
*exit
)
2282 vec
<edge
> exits
= get_loop_exit_edges (loop
);
2284 tree niter
= NULL_TREE
, aniter
;
2288 /* Loops with multiple exits are expensive to handle and less important. */
2289 if (!flag_expensive_optimizations
2290 && exits
.length () > 1)
2293 return chrec_dont_know
;
2296 FOR_EACH_VEC_ELT (exits
, i
, ex
)
2298 if (!just_once_each_iteration_p (loop
, ex
->src
))
2301 aniter
= loop_niter_by_eval (loop
, ex
);
2302 if (chrec_contains_undetermined (aniter
))
2306 && !tree_int_cst_lt (aniter
, niter
))
2314 return niter
? niter
: chrec_dont_know
;
2319 Analysis of upper bounds on number of iterations of a loop.
2323 static double_int
derive_constant_upper_bound_ops (tree
, tree
,
2324 enum tree_code
, tree
);
2326 /* Returns a constant upper bound on the value of the right-hand side of
2327 an assignment statement STMT. */
2330 derive_constant_upper_bound_assign (gimple stmt
)
2332 enum tree_code code
= gimple_assign_rhs_code (stmt
);
2333 tree op0
= gimple_assign_rhs1 (stmt
);
2334 tree op1
= gimple_assign_rhs2 (stmt
);
2336 return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt
)),
2340 /* Returns a constant upper bound on the value of expression VAL. VAL
2341 is considered to be unsigned. If its type is signed, its value must
2345 derive_constant_upper_bound (tree val
)
2347 enum tree_code code
;
2350 extract_ops_from_tree (val
, &code
, &op0
, &op1
);
2351 return derive_constant_upper_bound_ops (TREE_TYPE (val
), op0
, code
, op1
);
2354 /* Returns a constant upper bound on the value of expression OP0 CODE OP1,
2355 whose type is TYPE. The expression is considered to be unsigned. If
2356 its type is signed, its value must be nonnegative. */
2359 derive_constant_upper_bound_ops (tree type
, tree op0
,
2360 enum tree_code code
, tree op1
)
2363 double_int bnd
, max
, mmax
, cst
;
2366 if (INTEGRAL_TYPE_P (type
))
2367 maxt
= TYPE_MAX_VALUE (type
);
2369 maxt
= upper_bound_in_type (type
, type
);
2371 max
= tree_to_double_int (maxt
);
2376 return tree_to_double_int (op0
);
2379 subtype
= TREE_TYPE (op0
);
2380 if (!TYPE_UNSIGNED (subtype
)
2381 /* If TYPE is also signed, the fact that VAL is nonnegative implies
2382 that OP0 is nonnegative. */
2383 && TYPE_UNSIGNED (type
)
2384 && !tree_expr_nonnegative_p (op0
))
2386 /* If we cannot prove that the casted expression is nonnegative,
2387 we cannot establish more useful upper bound than the precision
2388 of the type gives us. */
2392 /* We now know that op0 is an nonnegative value. Try deriving an upper
2394 bnd
= derive_constant_upper_bound (op0
);
2396 /* If the bound does not fit in TYPE, max. value of TYPE could be
2404 case POINTER_PLUS_EXPR
:
2406 if (TREE_CODE (op1
) != INTEGER_CST
2407 || !tree_expr_nonnegative_p (op0
))
2410 /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to
2411 choose the most logical way how to treat this constant regardless
2412 of the signedness of the type. */
2413 cst
= tree_to_double_int (op1
);
2414 cst
= cst
.sext (TYPE_PRECISION (type
));
2415 if (code
!= MINUS_EXPR
)
2418 bnd
= derive_constant_upper_bound (op0
);
2420 if (cst
.is_negative ())
2423 /* Avoid CST == 0x80000... */
2424 if (cst
.is_negative ())
2427 /* OP0 + CST. We need to check that
2428 BND <= MAX (type) - CST. */
2438 /* OP0 - CST, where CST >= 0.
2440 If TYPE is signed, we have already verified that OP0 >= 0, and we
2441 know that the result is nonnegative. This implies that
2444 If TYPE is unsigned, we must additionally know that OP0 >= CST,
2445 otherwise the operation underflows.
2448 /* This should only happen if the type is unsigned; however, for
2449 buggy programs that use overflowing signed arithmetics even with
2450 -fno-wrapv, this condition may also be true for signed values. */
2454 if (TYPE_UNSIGNED (type
))
2456 tree tem
= fold_binary (GE_EXPR
, boolean_type_node
, op0
,
2457 double_int_to_tree (type
, cst
));
2458 if (!tem
|| integer_nonzerop (tem
))
2467 case FLOOR_DIV_EXPR
:
2468 case EXACT_DIV_EXPR
:
2469 if (TREE_CODE (op1
) != INTEGER_CST
2470 || tree_int_cst_sign_bit (op1
))
2473 bnd
= derive_constant_upper_bound (op0
);
2474 return bnd
.udiv (tree_to_double_int (op1
), FLOOR_DIV_EXPR
);
2477 if (TREE_CODE (op1
) != INTEGER_CST
2478 || tree_int_cst_sign_bit (op1
))
2480 return tree_to_double_int (op1
);
2483 stmt
= SSA_NAME_DEF_STMT (op0
);
2484 if (gimple_code (stmt
) != GIMPLE_ASSIGN
2485 || gimple_assign_lhs (stmt
) != op0
)
2487 return derive_constant_upper_bound_assign (stmt
);
2494 /* Records that every statement in LOOP is executed I_BOUND times.
2495 REALISTIC is true if I_BOUND is expected to be close to the real number
2496 of iterations. UPPER is true if we are sure the loop iterates at most
2500 record_niter_bound (struct loop
*loop
, double_int i_bound
, bool realistic
,
2503 /* Update the bounds only when there is no previous estimation, or when the
2504 current estimation is smaller. */
2506 && (!loop
->any_upper_bound
2507 || i_bound
.ult (loop
->nb_iterations_upper_bound
)))
2509 loop
->any_upper_bound
= true;
2510 loop
->nb_iterations_upper_bound
= i_bound
;
2513 && (!loop
->any_estimate
2514 || i_bound
.ult (loop
->nb_iterations_estimate
)))
2516 loop
->any_estimate
= true;
2517 loop
->nb_iterations_estimate
= i_bound
;
2520 /* If an upper bound is smaller than the realistic estimate of the
2521 number of iterations, use the upper bound instead. */
2522 if (loop
->any_upper_bound
2523 && loop
->any_estimate
2524 && loop
->nb_iterations_upper_bound
.ult (loop
->nb_iterations_estimate
))
2525 loop
->nb_iterations_estimate
= loop
->nb_iterations_upper_bound
;
2528 /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT
2529 is true if the loop is exited immediately after STMT, and this exit
2530 is taken at last when the STMT is executed BOUND + 1 times.
2531 REALISTIC is true if BOUND is expected to be close to the real number
2532 of iterations. UPPER is true if we are sure the loop iterates at most
2533 BOUND times. I_BOUND is an unsigned double_int upper estimate on BOUND. */
2536 record_estimate (struct loop
*loop
, tree bound
, double_int i_bound
,
2537 gimple at_stmt
, bool is_exit
, bool realistic
, bool upper
)
2541 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2543 fprintf (dump_file
, "Statement %s", is_exit
? "(exit)" : "");
2544 print_gimple_stmt (dump_file
, at_stmt
, 0, TDF_SLIM
);
2545 fprintf (dump_file
, " is %sexecuted at most ",
2546 upper
? "" : "probably ");
2547 print_generic_expr (dump_file
, bound
, TDF_SLIM
);
2548 fprintf (dump_file
, " (bounded by ");
2549 dump_double_int (dump_file
, i_bound
, true);
2550 fprintf (dump_file
, ") + 1 times in loop %d.\n", loop
->num
);
2553 /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the
2554 real number of iterations. */
2555 if (TREE_CODE (bound
) != INTEGER_CST
)
2558 gcc_checking_assert (i_bound
== tree_to_double_int (bound
));
2559 if (!upper
&& !realistic
)
2562 /* If we have a guaranteed upper bound, record it in the appropriate
2566 struct nb_iter_bound
*elt
= ggc_alloc_nb_iter_bound ();
2568 elt
->bound
= i_bound
;
2569 elt
->stmt
= at_stmt
;
2570 elt
->is_exit
= is_exit
;
2571 elt
->next
= loop
->bounds
;
2575 /* If statement is executed on every path to the loop latch, we can directly
2576 infer the upper bound on the # of iterations of the loop. */
2577 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, gimple_bb (at_stmt
)))
2580 /* Update the number of iteration estimates according to the bound.
2581 If at_stmt is an exit then the loop latch is executed at most BOUND times,
2582 otherwise it can be executed BOUND + 1 times. We will lower the estimate
2583 later if such statement must be executed on last iteration */
2585 delta
= double_int_zero
;
2587 delta
= double_int_one
;
2590 /* If an overflow occurred, ignore the result. */
2591 if (i_bound
.ult (delta
))
2594 record_niter_bound (loop
, i_bound
, realistic
, upper
);
2597 /* Record the estimate on number of iterations of LOOP based on the fact that
2598 the induction variable BASE + STEP * i evaluated in STMT does not wrap and
2599 its values belong to the range <LOW, HIGH>. REALISTIC is true if the
2600 estimated number of iterations is expected to be close to the real one.
2601 UPPER is true if we are sure the induction variable does not wrap. */
2604 record_nonwrapping_iv (struct loop
*loop
, tree base
, tree step
, gimple stmt
,
2605 tree low
, tree high
, bool realistic
, bool upper
)
2607 tree niter_bound
, extreme
, delta
;
2608 tree type
= TREE_TYPE (base
), unsigned_type
;
2611 if (TREE_CODE (step
) != INTEGER_CST
|| integer_zerop (step
))
2614 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2616 fprintf (dump_file
, "Induction variable (");
2617 print_generic_expr (dump_file
, TREE_TYPE (base
), TDF_SLIM
);
2618 fprintf (dump_file
, ") ");
2619 print_generic_expr (dump_file
, base
, TDF_SLIM
);
2620 fprintf (dump_file
, " + ");
2621 print_generic_expr (dump_file
, step
, TDF_SLIM
);
2622 fprintf (dump_file
, " * iteration does not wrap in statement ");
2623 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
2624 fprintf (dump_file
, " in loop %d.\n", loop
->num
);
2627 unsigned_type
= unsigned_type_for (type
);
2628 base
= fold_convert (unsigned_type
, base
);
2629 step
= fold_convert (unsigned_type
, step
);
2631 if (tree_int_cst_sign_bit (step
))
2633 extreme
= fold_convert (unsigned_type
, low
);
2634 if (TREE_CODE (base
) != INTEGER_CST
)
2635 base
= fold_convert (unsigned_type
, high
);
2636 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, base
, extreme
);
2637 step
= fold_build1 (NEGATE_EXPR
, unsigned_type
, step
);
2641 extreme
= fold_convert (unsigned_type
, high
);
2642 if (TREE_CODE (base
) != INTEGER_CST
)
2643 base
= fold_convert (unsigned_type
, low
);
2644 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, extreme
, base
);
2647 /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value
2648 would get out of the range. */
2649 niter_bound
= fold_build2 (FLOOR_DIV_EXPR
, unsigned_type
, delta
, step
);
2650 max
= derive_constant_upper_bound (niter_bound
);
2651 record_estimate (loop
, niter_bound
, max
, stmt
, false, realistic
, upper
);
2654 /* Determine information about number of iterations a LOOP from the index
2655 IDX of a data reference accessed in STMT. RELIABLE is true if STMT is
2656 guaranteed to be executed in every iteration of LOOP. Callback for
2666 idx_infer_loop_bounds (tree base
, tree
*idx
, void *dta
)
2668 struct ilb_data
*data
= (struct ilb_data
*) dta
;
2669 tree ev
, init
, step
;
2670 tree low
, high
, type
, next
;
2671 bool sign
, upper
= true, at_end
= false;
2672 struct loop
*loop
= data
->loop
;
2673 bool reliable
= true;
2675 if (TREE_CODE (base
) != ARRAY_REF
)
2678 /* For arrays at the end of the structure, we are not guaranteed that they
2679 do not really extend over their declared size. However, for arrays of
2680 size greater than one, this is unlikely to be intended. */
2681 if (array_at_struct_end_p (base
))
2687 struct loop
*dloop
= loop_containing_stmt (data
->stmt
);
2691 ev
= analyze_scalar_evolution (dloop
, *idx
);
2692 ev
= instantiate_parameters (loop
, ev
);
2693 init
= initial_condition (ev
);
2694 step
= evolution_part_in_loop_num (ev
, loop
->num
);
2698 || TREE_CODE (step
) != INTEGER_CST
2699 || integer_zerop (step
)
2700 || tree_contains_chrecs (init
, NULL
)
2701 || chrec_contains_symbols_defined_in_loop (init
, loop
->num
))
2704 low
= array_ref_low_bound (base
);
2705 high
= array_ref_up_bound (base
);
2707 /* The case of nonconstant bounds could be handled, but it would be
2709 if (TREE_CODE (low
) != INTEGER_CST
2711 || TREE_CODE (high
) != INTEGER_CST
)
2713 sign
= tree_int_cst_sign_bit (step
);
2714 type
= TREE_TYPE (step
);
2716 /* The array of length 1 at the end of a structure most likely extends
2717 beyond its bounds. */
2719 && operand_equal_p (low
, high
, 0))
2722 /* In case the relevant bound of the array does not fit in type, or
2723 it does, but bound + step (in type) still belongs into the range of the
2724 array, the index may wrap and still stay within the range of the array
2725 (consider e.g. if the array is indexed by the full range of
2728 To make things simpler, we require both bounds to fit into type, although
2729 there are cases where this would not be strictly necessary. */
2730 if (!int_fits_type_p (high
, type
)
2731 || !int_fits_type_p (low
, type
))
2733 low
= fold_convert (type
, low
);
2734 high
= fold_convert (type
, high
);
2737 next
= fold_binary (PLUS_EXPR
, type
, low
, step
);
2739 next
= fold_binary (PLUS_EXPR
, type
, high
, step
);
2741 if (tree_int_cst_compare (low
, next
) <= 0
2742 && tree_int_cst_compare (next
, high
) <= 0)
2745 /* If access is not executed on every iteration, we must ensure that overlow may
2746 not make the access valid later. */
2747 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, gimple_bb (data
->stmt
))
2748 && scev_probably_wraps_p (initial_condition_in_loop_num (ev
, loop
->num
),
2749 step
, data
->stmt
, loop
, true))
2752 record_nonwrapping_iv (loop
, init
, step
, data
->stmt
, low
, high
, reliable
, upper
);
2756 /* Determine information about number of iterations a LOOP from the bounds
2757 of arrays in the data reference REF accessed in STMT. RELIABLE is true if
2758 STMT is guaranteed to be executed in every iteration of LOOP.*/
2761 infer_loop_bounds_from_ref (struct loop
*loop
, gimple stmt
, tree ref
)
2763 struct ilb_data data
;
2767 for_each_index (&ref
, idx_infer_loop_bounds
, &data
);
2770 /* Determine information about number of iterations of a LOOP from the way
2771 arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be
2772 executed in every iteration of LOOP. */
2775 infer_loop_bounds_from_array (struct loop
*loop
, gimple stmt
)
2777 if (is_gimple_assign (stmt
))
2779 tree op0
= gimple_assign_lhs (stmt
);
2780 tree op1
= gimple_assign_rhs1 (stmt
);
2782 /* For each memory access, analyze its access function
2783 and record a bound on the loop iteration domain. */
2784 if (REFERENCE_CLASS_P (op0
))
2785 infer_loop_bounds_from_ref (loop
, stmt
, op0
);
2787 if (REFERENCE_CLASS_P (op1
))
2788 infer_loop_bounds_from_ref (loop
, stmt
, op1
);
2790 else if (is_gimple_call (stmt
))
2793 unsigned i
, n
= gimple_call_num_args (stmt
);
2795 lhs
= gimple_call_lhs (stmt
);
2796 if (lhs
&& REFERENCE_CLASS_P (lhs
))
2797 infer_loop_bounds_from_ref (loop
, stmt
, lhs
);
2799 for (i
= 0; i
< n
; i
++)
2801 arg
= gimple_call_arg (stmt
, i
);
2802 if (REFERENCE_CLASS_P (arg
))
2803 infer_loop_bounds_from_ref (loop
, stmt
, arg
);
2808 /* Determine information about number of iterations of a LOOP from the fact
2809 that pointer arithmetics in STMT does not overflow. */
2812 infer_loop_bounds_from_pointer_arith (struct loop
*loop
, gimple stmt
)
2814 tree def
, base
, step
, scev
, type
, low
, high
;
2817 if (!is_gimple_assign (stmt
)
2818 || gimple_assign_rhs_code (stmt
) != POINTER_PLUS_EXPR
)
2821 def
= gimple_assign_lhs (stmt
);
2822 if (TREE_CODE (def
) != SSA_NAME
)
2825 type
= TREE_TYPE (def
);
2826 if (!nowrap_type_p (type
))
2829 ptr
= gimple_assign_rhs1 (stmt
);
2830 if (!expr_invariant_in_loop_p (loop
, ptr
))
2833 var
= gimple_assign_rhs2 (stmt
);
2834 if (TYPE_PRECISION (type
) != TYPE_PRECISION (TREE_TYPE (var
)))
2837 scev
= instantiate_parameters (loop
, analyze_scalar_evolution (loop
, def
));
2838 if (chrec_contains_undetermined (scev
))
2841 base
= initial_condition_in_loop_num (scev
, loop
->num
);
2842 step
= evolution_part_in_loop_num (scev
, loop
->num
);
2845 || TREE_CODE (step
) != INTEGER_CST
2846 || tree_contains_chrecs (base
, NULL
)
2847 || chrec_contains_symbols_defined_in_loop (base
, loop
->num
))
2850 low
= lower_bound_in_type (type
, type
);
2851 high
= upper_bound_in_type (type
, type
);
2853 /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot
2854 produce a NULL pointer. The contrary would mean NULL points to an object,
2855 while NULL is supposed to compare unequal with the address of all objects.
2856 Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a
2857 NULL pointer since that would mean wrapping, which we assume here not to
2858 happen. So, we can exclude NULL from the valid range of pointer
2860 if (flag_delete_null_pointer_checks
&& int_cst_value (low
) == 0)
2861 low
= build_int_cstu (TREE_TYPE (low
), TYPE_ALIGN_UNIT (TREE_TYPE (type
)));
2863 record_nonwrapping_iv (loop
, base
, step
, stmt
, low
, high
, false, true);
2866 /* Determine information about number of iterations of a LOOP from the fact
2867 that signed arithmetics in STMT does not overflow. */
2870 infer_loop_bounds_from_signedness (struct loop
*loop
, gimple stmt
)
2872 tree def
, base
, step
, scev
, type
, low
, high
;
2874 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
2877 def
= gimple_assign_lhs (stmt
);
2879 if (TREE_CODE (def
) != SSA_NAME
)
2882 type
= TREE_TYPE (def
);
2883 if (!INTEGRAL_TYPE_P (type
)
2884 || !TYPE_OVERFLOW_UNDEFINED (type
))
2887 scev
= instantiate_parameters (loop
, analyze_scalar_evolution (loop
, def
));
2888 if (chrec_contains_undetermined (scev
))
2891 base
= initial_condition_in_loop_num (scev
, loop
->num
);
2892 step
= evolution_part_in_loop_num (scev
, loop
->num
);
2895 || TREE_CODE (step
) != INTEGER_CST
2896 || tree_contains_chrecs (base
, NULL
)
2897 || chrec_contains_symbols_defined_in_loop (base
, loop
->num
))
2900 low
= lower_bound_in_type (type
, type
);
2901 high
= upper_bound_in_type (type
, type
);
2903 record_nonwrapping_iv (loop
, base
, step
, stmt
, low
, high
, false, true);
2906 /* The following analyzers are extracting informations on the bounds
2907 of LOOP from the following undefined behaviors:
2909 - data references should not access elements over the statically
2912 - signed variables should not overflow when flag_wrapv is not set.
2916 infer_loop_bounds_from_undefined (struct loop
*loop
)
2920 gimple_stmt_iterator bsi
;
2924 bbs
= get_loop_body (loop
);
2926 for (i
= 0; i
< loop
->num_nodes
; i
++)
2930 /* If BB is not executed in each iteration of the loop, we cannot
2931 use the operations in it to infer reliable upper bound on the
2932 # of iterations of the loop. However, we can use it as a guess.
2933 Reliable guesses come only from array bounds. */
2934 reliable
= dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
);
2936 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
2938 gimple stmt
= gsi_stmt (bsi
);
2940 infer_loop_bounds_from_array (loop
, stmt
);
2944 infer_loop_bounds_from_signedness (loop
, stmt
);
2945 infer_loop_bounds_from_pointer_arith (loop
, stmt
);
2954 /* Converts VAL to double_int. */
2957 gcov_type_to_double_int (gcov_type val
)
2961 ret
.low
= (unsigned HOST_WIDE_INT
) val
;
2962 /* If HOST_BITS_PER_WIDE_INT == HOST_BITS_PER_WIDEST_INT, avoid shifting by
2963 the size of type. */
2964 val
>>= HOST_BITS_PER_WIDE_INT
- 1;
2966 ret
.high
= (unsigned HOST_WIDE_INT
) val
;
2971 /* Compare double ints, callback for qsort. */
2974 double_int_cmp (const void *p1
, const void *p2
)
2976 const double_int
*d1
= (const double_int
*)p1
;
2977 const double_int
*d2
= (const double_int
*)p2
;
2985 /* Return index of BOUND in BOUNDS array sorted in increasing order.
2986 Lookup by binary search. */
2989 bound_index (vec
<double_int
> bounds
, double_int bound
)
2991 unsigned int end
= bounds
.length ();
2992 unsigned int begin
= 0;
2994 /* Find a matching index by means of a binary search. */
2995 while (begin
!= end
)
2997 unsigned int middle
= (begin
+ end
) / 2;
2998 double_int index
= bounds
[middle
];
3002 else if (index
.ult (bound
))
3010 /* Used to hold vector of queues of basic blocks bellow. */
3011 typedef vec
<basic_block
> bb_queue
;
3013 /* We recorded loop bounds only for statements dominating loop latch (and thus
3014 executed each loop iteration). If there are any bounds on statements not
3015 dominating the loop latch we can improve the estimate by walking the loop
3016 body and seeing if every path from loop header to loop latch contains
3017 some bounded statement. */
3020 discover_iteration_bound_by_body_walk (struct loop
*loop
)
3022 pointer_map_t
*bb_bounds
;
3023 struct nb_iter_bound
*elt
;
3024 vec
<double_int
> bounds
= vNULL
;
3025 vec
<bb_queue
> queues
= vNULL
;
3026 bb_queue queue
= bb_queue();
3027 ptrdiff_t queue_index
;
3028 ptrdiff_t latch_index
= 0;
3029 pointer_map_t
*block_priority
;
3031 /* Discover what bounds may interest us. */
3032 for (elt
= loop
->bounds
; elt
; elt
= elt
->next
)
3034 double_int bound
= elt
->bound
;
3036 /* Exit terminates loop at given iteration, while non-exits produce undefined
3037 effect on the next iteration. */
3040 bound
+= double_int_one
;
3041 /* If an overflow occurred, ignore the result. */
3042 if (bound
.is_zero ())
3046 if (!loop
->any_upper_bound
3047 || bound
.ult (loop
->nb_iterations_upper_bound
))
3048 bounds
.safe_push (bound
);
3051 /* Exit early if there is nothing to do. */
3052 if (!bounds
.exists ())
3055 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3056 fprintf (dump_file
, " Trying to walk loop body to reduce the bound.\n");
3058 /* Sort the bounds in decreasing order. */
3059 qsort (bounds
.address (), bounds
.length (),
3060 sizeof (double_int
), double_int_cmp
);
3062 /* For every basic block record the lowest bound that is guaranteed to
3063 terminate the loop. */
3065 bb_bounds
= pointer_map_create ();
3066 for (elt
= loop
->bounds
; elt
; elt
= elt
->next
)
3068 double_int bound
= elt
->bound
;
3071 bound
+= double_int_one
;
3072 /* If an overflow occurred, ignore the result. */
3073 if (bound
.is_zero ())
3077 if (!loop
->any_upper_bound
3078 || bound
.ult (loop
->nb_iterations_upper_bound
))
3080 ptrdiff_t index
= bound_index (bounds
, bound
);
3081 void **entry
= pointer_map_contains (bb_bounds
,
3082 gimple_bb (elt
->stmt
));
3084 *pointer_map_insert (bb_bounds
,
3085 gimple_bb (elt
->stmt
)) = (void *)index
;
3086 else if ((ptrdiff_t)*entry
> index
)
3087 *entry
= (void *)index
;
3091 block_priority
= pointer_map_create ();
3093 /* Perform shortest path discovery loop->header ... loop->latch.
3095 The "distance" is given by the smallest loop bound of basic block
3096 present in the path and we look for path with largest smallest bound
3099 To avoid the need for fibonaci heap on double ints we simply compress
3100 double ints into indexes to BOUNDS array and then represent the queue
3101 as arrays of queues for every index.
3102 Index of BOUNDS.length() means that the execution of given BB has
3103 no bounds determined.
3105 VISITED is a pointer map translating basic block into smallest index
3106 it was inserted into the priority queue with. */
3109 /* Start walk in loop header with index set to infinite bound. */
3110 queue_index
= bounds
.length ();
3111 queues
.safe_grow_cleared (queue_index
+ 1);
3112 queue
.safe_push (loop
->header
);
3113 queues
[queue_index
] = queue
;
3114 *pointer_map_insert (block_priority
, loop
->header
) = (void *)queue_index
;
3116 for (; queue_index
>= 0; queue_index
--)
3118 if (latch_index
< queue_index
)
3120 while (queues
[queue_index
].length ())
3123 ptrdiff_t bound_index
= queue_index
;
3128 queue
= queues
[queue_index
];
3131 /* OK, we later inserted the BB with lower priority, skip it. */
3132 if ((ptrdiff_t)*pointer_map_contains (block_priority
, bb
) > queue_index
)
3135 /* See if we can improve the bound. */
3136 entry
= pointer_map_contains (bb_bounds
, bb
);
3137 if (entry
&& (ptrdiff_t)*entry
< bound_index
)
3138 bound_index
= (ptrdiff_t)*entry
;
3140 /* Insert succesors into the queue, watch for latch edge
3141 and record greatest index we saw. */
3142 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3144 bool insert
= false;
3147 if (loop_exit_edge_p (loop
, e
))
3150 if (e
== loop_latch_edge (loop
)
3151 && latch_index
< bound_index
)
3152 latch_index
= bound_index
;
3153 else if (!(entry
= pointer_map_contains (block_priority
, e
->dest
)))
3156 *pointer_map_insert (block_priority
, e
->dest
) = (void *)bound_index
;
3158 else if ((ptrdiff_t)*entry
< bound_index
)
3161 *entry
= (void *)bound_index
;
3166 bb_queue queue2
= queues
[bound_index
];
3167 queue2
.safe_push (e
->dest
);
3168 queues
[bound_index
] = queue2
;
3174 queues
[queue_index
].release ();
3177 gcc_assert (latch_index
>= 0);
3178 if ((unsigned)latch_index
< bounds
.length ())
3180 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3182 fprintf (dump_file
, "Found better loop bound ");
3183 dump_double_int (dump_file
, bounds
[latch_index
], true);
3184 fprintf (dump_file
, "\n");
3186 record_niter_bound (loop
, bounds
[latch_index
], false, true);
3190 pointer_map_destroy (bb_bounds
);
3191 pointer_map_destroy (block_priority
);
3194 /* See if every path cross the loop goes through a statement that is known
3195 to not execute at the last iteration. In that case we can decrese iteration
3199 maybe_lower_iteration_bound (struct loop
*loop
)
3201 pointer_set_t
*not_executed_last_iteration
= NULL
;
3202 struct nb_iter_bound
*elt
;
3203 bool found_exit
= false;
3204 vec
<basic_block
> queue
= vNULL
;
3207 /* Collect all statements with interesting (i.e. lower than
3208 nb_iterations_upper_bound) bound on them.
3210 TODO: Due to the way record_estimate choose estimates to store, the bounds
3211 will be always nb_iterations_upper_bound-1. We can change this to record
3212 also statements not dominating the loop latch and update the walk bellow
3213 to the shortest path algorthm. */
3214 for (elt
= loop
->bounds
; elt
; elt
= elt
->next
)
3217 && elt
->bound
.ult (loop
->nb_iterations_upper_bound
))
3219 if (!not_executed_last_iteration
)
3220 not_executed_last_iteration
= pointer_set_create ();
3221 pointer_set_insert (not_executed_last_iteration
, elt
->stmt
);
3224 if (!not_executed_last_iteration
)
3227 /* Start DFS walk in the loop header and see if we can reach the
3228 loop latch or any of the exits (including statements with side
3229 effects that may terminate the loop otherwise) without visiting
3230 any of the statements known to have undefined effect on the last
3232 queue
.safe_push (loop
->header
);
3233 visited
= BITMAP_ALLOC (NULL
);
3234 bitmap_set_bit (visited
, loop
->header
->index
);
3239 basic_block bb
= queue
.pop ();
3240 gimple_stmt_iterator gsi
;
3241 bool stmt_found
= false;
3243 /* Loop for possible exits and statements bounding the execution. */
3244 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
3246 gimple stmt
= gsi_stmt (gsi
);
3247 if (pointer_set_contains (not_executed_last_iteration
, stmt
))
3252 if (gimple_has_side_effects (stmt
))
3261 /* If no bounding statement is found, continue the walk. */
3267 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3269 if (loop_exit_edge_p (loop
, e
)
3270 || e
== loop_latch_edge (loop
))
3275 if (bitmap_set_bit (visited
, e
->dest
->index
))
3276 queue
.safe_push (e
->dest
);
3280 while (queue
.length () && !found_exit
);
3282 /* If every path through the loop reach bounding statement before exit,
3283 then we know the last iteration of the loop will have undefined effect
3284 and we can decrease number of iterations. */
3288 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3289 fprintf (dump_file
, "Reducing loop iteration estimate by 1; "
3290 "undefined statement must be executed at the last iteration.\n");
3291 record_niter_bound (loop
, loop
->nb_iterations_upper_bound
- double_int_one
,
3294 BITMAP_FREE (visited
);
3298 /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P
3299 is true also use estimates derived from undefined behavior. */
3302 estimate_numbers_of_iterations_loop (struct loop
*loop
)
3307 struct tree_niter_desc niter_desc
;
3312 /* Give up if we already have tried to compute an estimation. */
3313 if (loop
->estimate_state
!= EST_NOT_COMPUTED
)
3316 loop
->estimate_state
= EST_AVAILABLE
;
3317 /* Force estimate compuation but leave any existing upper bound in place. */
3318 loop
->any_estimate
= false;
3320 exits
= get_loop_exit_edges (loop
);
3321 likely_exit
= single_likely_exit (loop
);
3322 FOR_EACH_VEC_ELT (exits
, i
, ex
)
3324 if (!number_of_iterations_exit (loop
, ex
, &niter_desc
, false, false))
3327 niter
= niter_desc
.niter
;
3328 type
= TREE_TYPE (niter
);
3329 if (TREE_CODE (niter_desc
.may_be_zero
) != INTEGER_CST
)
3330 niter
= build3 (COND_EXPR
, type
, niter_desc
.may_be_zero
,
3331 build_int_cst (type
, 0),
3333 record_estimate (loop
, niter
, niter_desc
.max
,
3334 last_stmt (ex
->src
),
3335 true, ex
== likely_exit
, true);
3339 if (flag_aggressive_loop_optimizations
)
3340 infer_loop_bounds_from_undefined (loop
);
3342 discover_iteration_bound_by_body_walk (loop
);
3344 maybe_lower_iteration_bound (loop
);
3346 /* If we have a measured profile, use it to estimate the number of
3348 if (loop
->header
->count
!= 0)
3350 gcov_type nit
= expected_loop_iterations_unbounded (loop
) + 1;
3351 bound
= gcov_type_to_double_int (nit
);
3352 record_niter_bound (loop
, bound
, true, false);
3356 /* Sets NIT to the estimated number of executions of the latch of the
3357 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
3358 large as the number of iterations. If we have no reliable estimate,
3359 the function returns false, otherwise returns true. */
3362 estimated_loop_iterations (struct loop
*loop
, double_int
*nit
)
3364 /* When SCEV information is available, try to update loop iterations
3365 estimate. Otherwise just return whatever we recorded earlier. */
3366 if (scev_initialized_p ())
3367 estimate_numbers_of_iterations_loop (loop
);
3369 /* Even if the bound is not recorded, possibly we can derrive one from
3371 if (!loop
->any_estimate
)
3373 if (loop
->header
->count
)
3375 *nit
= gcov_type_to_double_int
3376 (expected_loop_iterations_unbounded (loop
) + 1);
3382 *nit
= loop
->nb_iterations_estimate
;
3386 /* Sets NIT to an upper bound for the maximum number of executions of the
3387 latch of the LOOP. If we have no reliable estimate, the function returns
3388 false, otherwise returns true. */
3391 max_loop_iterations (struct loop
*loop
, double_int
*nit
)
3393 /* When SCEV information is available, try to update loop iterations
3394 estimate. Otherwise just return whatever we recorded earlier. */
3395 if (scev_initialized_p ())
3396 estimate_numbers_of_iterations_loop (loop
);
3397 if (!loop
->any_upper_bound
)
3400 *nit
= loop
->nb_iterations_upper_bound
;
3404 /* Similar to estimated_loop_iterations, but returns the estimate only
3405 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3406 on the number of iterations of LOOP could not be derived, returns -1. */
3409 estimated_loop_iterations_int (struct loop
*loop
)
3412 HOST_WIDE_INT hwi_nit
;
3414 if (!estimated_loop_iterations (loop
, &nit
))
3417 if (!nit
.fits_shwi ())
3419 hwi_nit
= nit
.to_shwi ();
3421 return hwi_nit
< 0 ? -1 : hwi_nit
;
3424 /* Similar to max_loop_iterations, but returns the estimate only
3425 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
3426 on the number of iterations of LOOP could not be derived, returns -1. */
3429 max_loop_iterations_int (struct loop
*loop
)
3432 HOST_WIDE_INT hwi_nit
;
3434 if (!max_loop_iterations (loop
, &nit
))
3437 if (!nit
.fits_shwi ())
3439 hwi_nit
= nit
.to_shwi ();
3441 return hwi_nit
< 0 ? -1 : hwi_nit
;
3444 /* Returns an upper bound on the number of executions of statements
3445 in the LOOP. For statements before the loop exit, this exceeds
3446 the number of execution of the latch by one. */
3449 max_stmt_executions_int (struct loop
*loop
)
3451 HOST_WIDE_INT nit
= max_loop_iterations_int (loop
);
3457 snit
= (HOST_WIDE_INT
) ((unsigned HOST_WIDE_INT
) nit
+ 1);
3459 /* If the computation overflows, return -1. */
3460 return snit
< 0 ? -1 : snit
;
3463 /* Returns an estimate for the number of executions of statements
3464 in the LOOP. For statements before the loop exit, this exceeds
3465 the number of execution of the latch by one. */
3468 estimated_stmt_executions_int (struct loop
*loop
)
3470 HOST_WIDE_INT nit
= estimated_loop_iterations_int (loop
);
3476 snit
= (HOST_WIDE_INT
) ((unsigned HOST_WIDE_INT
) nit
+ 1);
3478 /* If the computation overflows, return -1. */
3479 return snit
< 0 ? -1 : snit
;
3482 /* Sets NIT to the estimated maximum number of executions of the latch of the
3483 LOOP, plus one. If we have no reliable estimate, the function returns
3484 false, otherwise returns true. */
3487 max_stmt_executions (struct loop
*loop
, double_int
*nit
)
3489 double_int nit_minus_one
;
3491 if (!max_loop_iterations (loop
, nit
))
3494 nit_minus_one
= *nit
;
3496 *nit
+= double_int_one
;
3498 return (*nit
).ugt (nit_minus_one
);
3501 /* Sets NIT to the estimated number of executions of the latch of the
3502 LOOP, plus one. If we have no reliable estimate, the function returns
3503 false, otherwise returns true. */
3506 estimated_stmt_executions (struct loop
*loop
, double_int
*nit
)
3508 double_int nit_minus_one
;
3510 if (!estimated_loop_iterations (loop
, nit
))
3513 nit_minus_one
= *nit
;
3515 *nit
+= double_int_one
;
3517 return (*nit
).ugt (nit_minus_one
);
3520 /* Records estimates on numbers of iterations of loops. */
3523 estimate_numbers_of_iterations (void)
3528 /* We don't want to issue signed overflow warnings while getting
3529 loop iteration estimates. */
3530 fold_defer_overflow_warnings ();
3532 FOR_EACH_LOOP (li
, loop
, 0)
3534 estimate_numbers_of_iterations_loop (loop
);
3537 fold_undefer_and_ignore_overflow_warnings ();
3540 /* Returns true if statement S1 dominates statement S2. */
3543 stmt_dominates_stmt_p (gimple s1
, gimple s2
)
3545 basic_block bb1
= gimple_bb (s1
), bb2
= gimple_bb (s2
);
3553 gimple_stmt_iterator bsi
;
3555 if (gimple_code (s2
) == GIMPLE_PHI
)
3558 if (gimple_code (s1
) == GIMPLE_PHI
)
3561 for (bsi
= gsi_start_bb (bb1
); gsi_stmt (bsi
) != s2
; gsi_next (&bsi
))
3562 if (gsi_stmt (bsi
) == s1
)
3568 return dominated_by_p (CDI_DOMINATORS
, bb2
, bb1
);
3571 /* Returns true when we can prove that the number of executions of
3572 STMT in the loop is at most NITER, according to the bound on
3573 the number of executions of the statement NITER_BOUND->stmt recorded in
3574 NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT.
3576 ??? This code can become quite a CPU hog - we can have many bounds,
3577 and large basic block forcing stmt_dominates_stmt_p to be queried
3578 many times on a large basic blocks, so the whole thing is O(n^2)
3579 for scev_probably_wraps_p invocation (that can be done n times).
3581 It would make more sense (and give better answers) to remember BB
3582 bounds computed by discover_iteration_bound_by_body_walk. */
3585 n_of_executions_at_most (gimple stmt
,
3586 struct nb_iter_bound
*niter_bound
,
3589 double_int bound
= niter_bound
->bound
;
3590 tree nit_type
= TREE_TYPE (niter
), e
;
3593 gcc_assert (TYPE_UNSIGNED (nit_type
));
3595 /* If the bound does not even fit into NIT_TYPE, it cannot tell us that
3596 the number of iterations is small. */
3597 if (!double_int_fits_to_tree_p (nit_type
, bound
))
3600 /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1
3601 times. This means that:
3603 -- if NITER_BOUND->is_exit is true, then everything after
3604 it at most NITER_BOUND->bound times.
3606 -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT
3607 is executed, then NITER_BOUND->stmt is executed as well in the same
3608 iteration then STMT is executed at most NITER_BOUND->bound + 1 times.
3610 If we can determine that NITER_BOUND->stmt is always executed
3611 after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times.
3612 We conclude that if both statements belong to the same
3613 basic block and STMT is before NITER_BOUND->stmt and there are no
3614 statements with side effects in between. */
3616 if (niter_bound
->is_exit
)
3618 if (stmt
== niter_bound
->stmt
3619 || !stmt_dominates_stmt_p (niter_bound
->stmt
, stmt
))
3625 if (!stmt_dominates_stmt_p (niter_bound
->stmt
, stmt
))
3627 gimple_stmt_iterator bsi
;
3628 if (gimple_bb (stmt
) != gimple_bb (niter_bound
->stmt
)
3629 || gimple_code (stmt
) == GIMPLE_PHI
3630 || gimple_code (niter_bound
->stmt
) == GIMPLE_PHI
)
3633 /* By stmt_dominates_stmt_p we already know that STMT appears
3634 before NITER_BOUND->STMT. Still need to test that the loop
3635 can not be terinated by a side effect in between. */
3636 for (bsi
= gsi_for_stmt (stmt
); gsi_stmt (bsi
) != niter_bound
->stmt
;
3638 if (gimple_has_side_effects (gsi_stmt (bsi
)))
3640 bound
+= double_int_one
;
3641 if (bound
.is_zero ()
3642 || !double_int_fits_to_tree_p (nit_type
, bound
))
3648 e
= fold_binary (cmp
, boolean_type_node
,
3649 niter
, double_int_to_tree (nit_type
, bound
));
3650 return e
&& integer_nonzerop (e
);
3653 /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */
3656 nowrap_type_p (tree type
)
3658 if (INTEGRAL_TYPE_P (type
)
3659 && TYPE_OVERFLOW_UNDEFINED (type
))
3662 if (POINTER_TYPE_P (type
))
3668 /* Return false only when the induction variable BASE + STEP * I is
3669 known to not overflow: i.e. when the number of iterations is small
3670 enough with respect to the step and initial condition in order to
3671 keep the evolution confined in TYPEs bounds. Return true when the
3672 iv is known to overflow or when the property is not computable.
3674 USE_OVERFLOW_SEMANTICS is true if this function should assume that
3675 the rules for overflow of the given language apply (e.g., that signed
3676 arithmetics in C does not overflow). */
3679 scev_probably_wraps_p (tree base
, tree step
,
3680 gimple at_stmt
, struct loop
*loop
,
3681 bool use_overflow_semantics
)
3683 tree delta
, step_abs
;
3684 tree unsigned_type
, valid_niter
;
3685 tree type
= TREE_TYPE (step
);
3688 struct nb_iter_bound
*bound
;
3690 /* FIXME: We really need something like
3691 http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html.
3693 We used to test for the following situation that frequently appears
3694 during address arithmetics:
3696 D.1621_13 = (long unsigned intD.4) D.1620_12;
3697 D.1622_14 = D.1621_13 * 8;
3698 D.1623_15 = (doubleD.29 *) D.1622_14;
3700 And derived that the sequence corresponding to D_14
3701 can be proved to not wrap because it is used for computing a
3702 memory access; however, this is not really the case -- for example,
3703 if D_12 = (unsigned char) [254,+,1], then D_14 has values
3704 2032, 2040, 0, 8, ..., but the code is still legal. */
3706 if (chrec_contains_undetermined (base
)
3707 || chrec_contains_undetermined (step
))
3710 if (integer_zerop (step
))
3713 /* If we can use the fact that signed and pointer arithmetics does not
3714 wrap, we are done. */
3715 if (use_overflow_semantics
&& nowrap_type_p (TREE_TYPE (base
)))
3718 /* To be able to use estimates on number of iterations of the loop,
3719 we must have an upper bound on the absolute value of the step. */
3720 if (TREE_CODE (step
) != INTEGER_CST
)
3723 /* Don't issue signed overflow warnings. */
3724 fold_defer_overflow_warnings ();
3726 /* Otherwise, compute the number of iterations before we reach the
3727 bound of the type, and verify that the loop is exited before this
3729 unsigned_type
= unsigned_type_for (type
);
3730 base
= fold_convert (unsigned_type
, base
);
3732 if (tree_int_cst_sign_bit (step
))
3734 tree extreme
= fold_convert (unsigned_type
,
3735 lower_bound_in_type (type
, type
));
3736 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, base
, extreme
);
3737 step_abs
= fold_build1 (NEGATE_EXPR
, unsigned_type
,
3738 fold_convert (unsigned_type
, step
));
3742 tree extreme
= fold_convert (unsigned_type
,
3743 upper_bound_in_type (type
, type
));
3744 delta
= fold_build2 (MINUS_EXPR
, unsigned_type
, extreme
, base
);
3745 step_abs
= fold_convert (unsigned_type
, step
);
3748 valid_niter
= fold_build2 (FLOOR_DIV_EXPR
, unsigned_type
, delta
, step_abs
);
3750 estimate_numbers_of_iterations_loop (loop
);
3752 if (max_loop_iterations (loop
, &niter
)
3753 && double_int_fits_to_tree_p (TREE_TYPE (valid_niter
), niter
)
3754 && (e
= fold_binary (GT_EXPR
, boolean_type_node
, valid_niter
,
3755 double_int_to_tree (TREE_TYPE (valid_niter
),
3757 && integer_nonzerop (e
))
3759 fold_undefer_and_ignore_overflow_warnings ();
3763 for (bound
= loop
->bounds
; bound
; bound
= bound
->next
)
3765 if (n_of_executions_at_most (at_stmt
, bound
, valid_niter
))
3767 fold_undefer_and_ignore_overflow_warnings ();
3772 fold_undefer_and_ignore_overflow_warnings ();
3774 /* At this point we still don't have a proof that the iv does not
3775 overflow: give up. */
3779 /* Frees the information on upper bounds on numbers of iterations of LOOP. */
3782 free_numbers_of_iterations_estimates_loop (struct loop
*loop
)
3784 struct nb_iter_bound
*bound
, *next
;
3786 loop
->nb_iterations
= NULL
;
3787 loop
->estimate_state
= EST_NOT_COMPUTED
;
3788 for (bound
= loop
->bounds
; bound
; bound
= next
)
3794 loop
->bounds
= NULL
;
3797 /* Frees the information on upper bounds on numbers of iterations of loops. */
3800 free_numbers_of_iterations_estimates (void)
3805 FOR_EACH_LOOP (li
, loop
, 0)
3807 free_numbers_of_iterations_estimates_loop (loop
);
3811 /* Substitute value VAL for ssa name NAME inside expressions held
3815 substitute_in_loop_info (struct loop
*loop
, tree name
, tree val
)
3817 loop
->nb_iterations
= simplify_replace_tree (loop
->nb_iterations
, name
, val
);