1 /* Scalar evolution detector.
2 Copyright (C) 2003-2016 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <s.pop@laposte.net>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
24 This pass analyzes the evolution of scalar variables in loop
25 structures. The algorithm is based on the SSA representation,
26 and on the loop hierarchy tree. This algorithm is not based on
27 the notion of versions of a variable, as it was the case for the
28 previous implementations of the scalar evolution algorithm, but
29 it assumes that each defined name is unique.
31 The notation used in this file is called "chains of recurrences",
32 and has been proposed by Eugene Zima, Robert Van Engelen, and
33 others for describing induction variables in programs. For example
34 "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0
35 when entering in the loop_1 and has a step 2 in this loop, in other
36 words "for (b = 0; b < N; b+=2);". Note that the coefficients of
37 this chain of recurrence (or chrec [shrek]) can contain the name of
38 other variables, in which case they are called parametric chrecs.
39 For example, "b -> {a, +, 2}_1" means that the initial value of "b"
40 is the value of "a". In most of the cases these parametric chrecs
41 are fully instantiated before their use because symbolic names can
42 hide some difficult cases such as self-references described later
43 (see the Fibonacci example).
45 A short sketch of the algorithm is:
47 Given a scalar variable to be analyzed, follow the SSA edge to
50 - When the definition is a GIMPLE_ASSIGN: if the right hand side
51 (RHS) of the definition cannot be statically analyzed, the answer
52 of the analyzer is: "don't know".
53 Otherwise, for all the variables that are not yet analyzed in the
54 RHS, try to determine their evolution, and finally try to
55 evaluate the operation of the RHS that gives the evolution
56 function of the analyzed variable.
58 - When the definition is a condition-phi-node: determine the
59 evolution function for all the branches of the phi node, and
60 finally merge these evolutions (see chrec_merge).
62 - When the definition is a loop-phi-node: determine its initial
63 condition, that is the SSA edge defined in an outer loop, and
64 keep it symbolic. Then determine the SSA edges that are defined
65 in the body of the loop. Follow the inner edges until ending on
66 another loop-phi-node of the same analyzed loop. If the reached
67 loop-phi-node is not the starting loop-phi-node, then we keep
68 this definition under a symbolic form. If the reached
69 loop-phi-node is the same as the starting one, then we compute a
70 symbolic stride on the return path. The result is then the
71 symbolic chrec {initial_condition, +, symbolic_stride}_loop.
75 Example 1: Illustration of the basic algorithm.
81 | if (c > 10) exit_loop
84 Suppose that we want to know the number of iterations of the
85 loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We
86 ask the scalar evolution analyzer two questions: what's the
87 scalar evolution (scev) of "c", and what's the scev of "10". For
88 "10" the answer is "10" since it is a scalar constant. For the
89 scalar variable "c", it follows the SSA edge to its definition,
90 "c = b + 1", and then asks again what's the scev of "b".
91 Following the SSA edge, we end on a loop-phi-node "b = phi (a,
92 c)", where the initial condition is "a", and the inner loop edge
93 is "c". The initial condition is kept under a symbolic form (it
94 may be the case that the copy constant propagation has done its
95 work and we end with the constant "3" as one of the edges of the
96 loop-phi-node). The update edge is followed to the end of the
97 loop, and until reaching again the starting loop-phi-node: b -> c
98 -> b. At this point we have drawn a path from "b" to "b" from
99 which we compute the stride in the loop: in this example it is
100 "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now
101 that the scev for "b" is known, it is possible to compute the
102 scev for "c", that is "c -> {a + 1, +, 1}_1". In order to
103 determine the number of iterations in the loop_1, we have to
104 instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some
105 more analysis the scev {4, +, 1}_1, or in other words, this is
106 the function "f (x) = x + 4", where x is the iteration count of
107 the loop_1. Now we have to solve the inequality "x + 4 > 10",
108 and take the smallest iteration number for which the loop is
109 exited: x = 7. This loop runs from x = 0 to x = 7, and in total
110 there are 8 iterations. In terms of loop normalization, we have
111 created a variable that is implicitly defined, "x" or just "_1",
112 and all the other analyzed scalars of the loop are defined in
113 function of this variable:
119 or in terms of a C program:
122 | for (x = 0; x <= 7; x++)
128 Example 2a: Illustration of the algorithm on nested loops.
139 For analyzing the scalar evolution of "a", the algorithm follows
140 the SSA edge into the loop's body: "a -> b". "b" is an inner
141 loop-phi-node, and its analysis as in Example 1, gives:
146 Following the SSA edge for the initial condition, we end on "c = a
147 + 2", and then on the starting loop-phi-node "a". From this point,
148 the loop stride is computed: back on "c = a + 2" we get a "+2" in
149 the loop_1, then on the loop-phi-node "b" we compute the overall
150 effect of the inner loop that is "b = c + 30", and we get a "+30"
151 in the loop_1. That means that the overall stride in loop_1 is
152 equal to "+32", and the result is:
157 Example 2b: Multivariate chains of recurrences.
170 Analyzing the access function of array A with
171 instantiate_parameters (loop_1, "j + k"), we obtain the
172 instantiation and the analysis of the scalar variables "j" and "k"
173 in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end
174 value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is
175 {0, +, 1}_1. To obtain the evolution function in loop_3 and
176 instantiate the scalar variables up to loop_1, one has to use:
177 instantiate_scev (block_before_loop (loop_1), loop_3, "j + k").
178 The result of this call is {{0, +, 1}_1, +, 1}_2.
180 Example 3: Higher degree polynomials.
194 instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1
195 instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
197 Example 4: Lucas, Fibonacci, or mixers in general.
209 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
210 following semantics: during the first iteration of the loop_1, the
211 variable contains the value 1, and then it contains the value "c".
212 Note that this syntax is close to the syntax of the loop-phi-node:
213 "a -> (1, c)_1" vs. "a = phi (1, c)".
215 The symbolic chrec representation contains all the semantics of the
216 original code. What is more difficult is to use this information.
218 Example 5: Flip-flops, or exchangers.
230 Based on these symbolic chrecs, it is possible to refine this
231 information into the more precise PERIODIC_CHRECs:
236 This transformation is not yet implemented.
240 You can find a more detailed description of the algorithm in:
241 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
242 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
243 this is a preliminary report and some of the details of the
244 algorithm have changed. I'm working on a research report that
245 updates the description of the algorithms to reflect the design
246 choices used in this implementation.
248 A set of slides show a high level overview of the algorithm and run
249 an example through the scalar evolution analyzer:
250 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
252 The slides that I have presented at the GCC Summit'04 are available
253 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
258 #include "coretypes.h"
264 #include "gimple-pretty-print.h"
265 #include "fold-const.h"
266 #include "gimplify.h"
267 #include "gimple-iterator.h"
268 #include "gimplify-me.h"
269 #include "tree-cfg.h"
270 #include "tree-ssa-loop-ivopts.h"
271 #include "tree-ssa-loop-manip.h"
272 #include "tree-ssa-loop-niter.h"
273 #include "tree-ssa-loop.h"
274 #include "tree-ssa.h"
276 #include "tree-chrec.h"
277 #include "tree-affine.h"
278 #include "tree-scalar-evolution.h"
279 #include "dumpfile.h"
281 #include "tree-ssa-propagate.h"
282 #include "gimple-fold.h"
284 static tree
analyze_scalar_evolution_1 (struct loop
*, tree
, tree
);
285 static tree
analyze_scalar_evolution_for_address_of (struct loop
*loop
,
288 /* The cached information about an SSA name with version NAME_VERSION,
289 claiming that below basic block with index INSTANTIATED_BELOW, the
290 value of the SSA name can be expressed as CHREC. */
292 struct GTY((for_user
)) scev_info_str
{
293 unsigned int name_version
;
294 int instantiated_below
;
298 /* Counters for the scev database. */
299 static unsigned nb_set_scev
= 0;
300 static unsigned nb_get_scev
= 0;
302 /* The following trees are unique elements. Thus the comparison of
303 another element to these elements should be done on the pointer to
304 these trees, and not on their value. */
306 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
307 tree chrec_not_analyzed_yet
;
309 /* Reserved to the cases where the analyzer has detected an
310 undecidable property at compile time. */
311 tree chrec_dont_know
;
313 /* When the analyzer has detected that a property will never
314 happen, then it qualifies it with chrec_known. */
317 struct scev_info_hasher
: ggc_ptr_hash
<scev_info_str
>
319 static hashval_t
hash (scev_info_str
*i
);
320 static bool equal (const scev_info_str
*a
, const scev_info_str
*b
);
323 static GTY (()) hash_table
<scev_info_hasher
> *scalar_evolution_info
;
326 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
328 static inline struct scev_info_str
*
329 new_scev_info_str (basic_block instantiated_below
, tree var
)
331 struct scev_info_str
*res
;
333 res
= ggc_alloc
<scev_info_str
> ();
334 res
->name_version
= SSA_NAME_VERSION (var
);
335 res
->chrec
= chrec_not_analyzed_yet
;
336 res
->instantiated_below
= instantiated_below
->index
;
341 /* Computes a hash function for database element ELT. */
344 scev_info_hasher::hash (scev_info_str
*elt
)
346 return elt
->name_version
^ elt
->instantiated_below
;
349 /* Compares database elements E1 and E2. */
352 scev_info_hasher::equal (const scev_info_str
*elt1
, const scev_info_str
*elt2
)
354 return (elt1
->name_version
== elt2
->name_version
355 && elt1
->instantiated_below
== elt2
->instantiated_below
);
358 /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
359 A first query on VAR returns chrec_not_analyzed_yet. */
362 find_var_scev_info (basic_block instantiated_below
, tree var
)
364 struct scev_info_str
*res
;
365 struct scev_info_str tmp
;
367 tmp
.name_version
= SSA_NAME_VERSION (var
);
368 tmp
.instantiated_below
= instantiated_below
->index
;
369 scev_info_str
**slot
= scalar_evolution_info
->find_slot (&tmp
, INSERT
);
372 *slot
= new_scev_info_str (instantiated_below
, var
);
378 /* Return true when CHREC contains symbolic names defined in
382 chrec_contains_symbols_defined_in_loop (const_tree chrec
, unsigned loop_nb
)
386 if (chrec
== NULL_TREE
)
389 if (is_gimple_min_invariant (chrec
))
392 if (TREE_CODE (chrec
) == SSA_NAME
)
395 loop_p def_loop
, loop
;
397 if (SSA_NAME_IS_DEFAULT_DEF (chrec
))
400 def
= SSA_NAME_DEF_STMT (chrec
);
401 def_loop
= loop_containing_stmt (def
);
402 loop
= get_loop (cfun
, loop_nb
);
404 if (def_loop
== NULL
)
407 if (loop
== def_loop
|| flow_loop_nested_p (loop
, def_loop
))
413 n
= TREE_OPERAND_LENGTH (chrec
);
414 for (i
= 0; i
< n
; i
++)
415 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec
, i
),
421 /* Return true when PHI is a loop-phi-node. */
424 loop_phi_node_p (gimple
*phi
)
426 /* The implementation of this function is based on the following
427 property: "all the loop-phi-nodes of a loop are contained in the
428 loop's header basic block". */
430 return loop_containing_stmt (phi
)->header
== gimple_bb (phi
);
433 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
434 In general, in the case of multivariate evolutions we want to get
435 the evolution in different loops. LOOP specifies the level for
436 which to get the evolution.
440 | for (j = 0; j < 100; j++)
442 | for (k = 0; k < 100; k++)
444 | i = k + j; - Here the value of i is a function of j, k.
446 | ... = i - Here the value of i is a function of j.
448 | ... = i - Here the value of i is a scalar.
454 | i_1 = phi (i_0, i_2)
458 This loop has the same effect as:
459 LOOP_1 has the same effect as:
463 The overall effect of the loop, "i_0 + 20" in the previous example,
464 is obtained by passing in the parameters: LOOP = 1,
465 EVOLUTION_FN = {i_0, +, 2}_1.
469 compute_overall_effect_of_inner_loop (struct loop
*loop
, tree evolution_fn
)
473 if (evolution_fn
== chrec_dont_know
)
474 return chrec_dont_know
;
476 else if (TREE_CODE (evolution_fn
) == POLYNOMIAL_CHREC
)
478 struct loop
*inner_loop
= get_chrec_loop (evolution_fn
);
480 if (inner_loop
== loop
481 || flow_loop_nested_p (loop
, inner_loop
))
483 tree nb_iter
= number_of_latch_executions (inner_loop
);
485 if (nb_iter
== chrec_dont_know
)
486 return chrec_dont_know
;
491 /* evolution_fn is the evolution function in LOOP. Get
492 its value in the nb_iter-th iteration. */
493 res
= chrec_apply (inner_loop
->num
, evolution_fn
, nb_iter
);
495 if (chrec_contains_symbols_defined_in_loop (res
, loop
->num
))
496 res
= instantiate_parameters (loop
, res
);
498 /* Continue the computation until ending on a parent of LOOP. */
499 return compute_overall_effect_of_inner_loop (loop
, res
);
506 /* If the evolution function is an invariant, there is nothing to do. */
507 else if (no_evolution_in_loop_p (evolution_fn
, loop
->num
, &val
) && val
)
511 return chrec_dont_know
;
514 /* Associate CHREC to SCALAR. */
517 set_scalar_evolution (basic_block instantiated_below
, tree scalar
, tree chrec
)
521 if (TREE_CODE (scalar
) != SSA_NAME
)
524 scalar_info
= find_var_scev_info (instantiated_below
, scalar
);
528 if (dump_flags
& TDF_SCEV
)
530 fprintf (dump_file
, "(set_scalar_evolution \n");
531 fprintf (dump_file
, " instantiated_below = %d \n",
532 instantiated_below
->index
);
533 fprintf (dump_file
, " (scalar = ");
534 print_generic_expr (dump_file
, scalar
, 0);
535 fprintf (dump_file
, ")\n (scalar_evolution = ");
536 print_generic_expr (dump_file
, chrec
, 0);
537 fprintf (dump_file
, "))\n");
539 if (dump_flags
& TDF_STATS
)
543 *scalar_info
= chrec
;
546 /* Retrieve the chrec associated to SCALAR instantiated below
547 INSTANTIATED_BELOW block. */
550 get_scalar_evolution (basic_block instantiated_below
, tree scalar
)
556 if (dump_flags
& TDF_SCEV
)
558 fprintf (dump_file
, "(get_scalar_evolution \n");
559 fprintf (dump_file
, " (scalar = ");
560 print_generic_expr (dump_file
, scalar
, 0);
561 fprintf (dump_file
, ")\n");
563 if (dump_flags
& TDF_STATS
)
567 switch (TREE_CODE (scalar
))
570 res
= *find_var_scev_info (instantiated_below
, scalar
);
580 res
= chrec_not_analyzed_yet
;
584 if (dump_file
&& (dump_flags
& TDF_SCEV
))
586 fprintf (dump_file
, " (scalar_evolution = ");
587 print_generic_expr (dump_file
, res
, 0);
588 fprintf (dump_file
, "))\n");
594 /* Helper function for add_to_evolution. Returns the evolution
595 function for an assignment of the form "a = b + c", where "a" and
596 "b" are on the strongly connected component. CHREC_BEFORE is the
597 information that we already have collected up to this point.
598 TO_ADD is the evolution of "c".
600 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
601 evolution the expression TO_ADD, otherwise construct an evolution
602 part for this loop. */
605 add_to_evolution_1 (unsigned loop_nb
, tree chrec_before
, tree to_add
,
608 tree type
, left
, right
;
609 struct loop
*loop
= get_loop (cfun
, loop_nb
), *chloop
;
611 switch (TREE_CODE (chrec_before
))
613 case POLYNOMIAL_CHREC
:
614 chloop
= get_chrec_loop (chrec_before
);
616 || flow_loop_nested_p (chloop
, loop
))
620 type
= chrec_type (chrec_before
);
622 /* When there is no evolution part in this loop, build it. */
627 right
= SCALAR_FLOAT_TYPE_P (type
)
628 ? build_real (type
, dconst0
)
629 : build_int_cst (type
, 0);
633 var
= CHREC_VARIABLE (chrec_before
);
634 left
= CHREC_LEFT (chrec_before
);
635 right
= CHREC_RIGHT (chrec_before
);
638 to_add
= chrec_convert (type
, to_add
, at_stmt
);
639 right
= chrec_convert_rhs (type
, right
, at_stmt
);
640 right
= chrec_fold_plus (chrec_type (right
), right
, to_add
);
641 return build_polynomial_chrec (var
, left
, right
);
645 gcc_assert (flow_loop_nested_p (loop
, chloop
));
647 /* Search the evolution in LOOP_NB. */
648 left
= add_to_evolution_1 (loop_nb
, CHREC_LEFT (chrec_before
),
650 right
= CHREC_RIGHT (chrec_before
);
651 right
= chrec_convert_rhs (chrec_type (left
), right
, at_stmt
);
652 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before
),
657 /* These nodes do not depend on a loop. */
658 if (chrec_before
== chrec_dont_know
)
659 return chrec_dont_know
;
662 right
= chrec_convert_rhs (chrec_type (left
), to_add
, at_stmt
);
663 return build_polynomial_chrec (loop_nb
, left
, right
);
667 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
670 Description (provided for completeness, for those who read code in
671 a plane, and for my poor 62 bytes brain that would have forgotten
672 all this in the next two or three months):
674 The algorithm of translation of programs from the SSA representation
675 into the chrecs syntax is based on a pattern matching. After having
676 reconstructed the overall tree expression for a loop, there are only
677 two cases that can arise:
679 1. a = loop-phi (init, a + expr)
680 2. a = loop-phi (init, expr)
682 where EXPR is either a scalar constant with respect to the analyzed
683 loop (this is a degree 0 polynomial), or an expression containing
684 other loop-phi definitions (these are higher degree polynomials).
691 | a = phi (init, a + 5)
698 | a = phi (inita, 2 * b + 3)
699 | b = phi (initb, b + 1)
702 For the first case, the semantics of the SSA representation is:
704 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
706 that is, there is a loop index "x" that determines the scalar value
707 of the variable during the loop execution. During the first
708 iteration, the value is that of the initial condition INIT, while
709 during the subsequent iterations, it is the sum of the initial
710 condition with the sum of all the values of EXPR from the initial
711 iteration to the before last considered iteration.
713 For the second case, the semantics of the SSA program is:
715 | a (x) = init, if x = 0;
716 | expr (x - 1), otherwise.
718 The second case corresponds to the PEELED_CHREC, whose syntax is
719 close to the syntax of a loop-phi-node:
721 | phi (init, expr) vs. (init, expr)_x
723 The proof of the translation algorithm for the first case is a
724 proof by structural induction based on the degree of EXPR.
727 When EXPR is a constant with respect to the analyzed loop, or in
728 other words when EXPR is a polynomial of degree 0, the evolution of
729 the variable A in the loop is an affine function with an initial
730 condition INIT, and a step EXPR. In order to show this, we start
731 from the semantics of the SSA representation:
733 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
735 and since "expr (j)" is a constant with respect to "j",
737 f (x) = init + x * expr
739 Finally, based on the semantics of the pure sum chrecs, by
740 identification we get the corresponding chrecs syntax:
742 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
743 f (x) -> {init, +, expr}_x
746 Suppose that EXPR is a polynomial of degree N with respect to the
747 analyzed loop_x for which we have already determined that it is
748 written under the chrecs syntax:
750 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
752 We start from the semantics of the SSA program:
754 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
756 | f (x) = init + \sum_{j = 0}^{x - 1}
757 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
759 | f (x) = init + \sum_{j = 0}^{x - 1}
760 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
762 | f (x) = init + \sum_{k = 0}^{n - 1}
763 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
765 | f (x) = init + \sum_{k = 0}^{n - 1}
766 | (b_k * \binom{x}{k + 1})
768 | f (x) = init + b_0 * \binom{x}{1} + ...
769 | + b_{n-1} * \binom{x}{n}
771 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
772 | + b_{n-1} * \binom{x}{n}
775 And finally from the definition of the chrecs syntax, we identify:
776 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
778 This shows the mechanism that stands behind the add_to_evolution
779 function. An important point is that the use of symbolic
780 parameters avoids the need of an analysis schedule.
787 | a = phi (inita, a + 2 + b)
788 | b = phi (initb, b + 1)
791 When analyzing "a", the algorithm keeps "b" symbolically:
793 | a -> {inita, +, 2 + b}_1
795 Then, after instantiation, the analyzer ends on the evolution:
797 | a -> {inita, +, 2 + initb, +, 1}_1
802 add_to_evolution (unsigned loop_nb
, tree chrec_before
, enum tree_code code
,
803 tree to_add
, gimple
*at_stmt
)
805 tree type
= chrec_type (to_add
);
806 tree res
= NULL_TREE
;
808 if (to_add
== NULL_TREE
)
811 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
812 instantiated at this point. */
813 if (TREE_CODE (to_add
) == POLYNOMIAL_CHREC
)
814 /* This should not happen. */
815 return chrec_dont_know
;
817 if (dump_file
&& (dump_flags
& TDF_SCEV
))
819 fprintf (dump_file
, "(add_to_evolution \n");
820 fprintf (dump_file
, " (loop_nb = %d)\n", loop_nb
);
821 fprintf (dump_file
, " (chrec_before = ");
822 print_generic_expr (dump_file
, chrec_before
, 0);
823 fprintf (dump_file
, ")\n (to_add = ");
824 print_generic_expr (dump_file
, to_add
, 0);
825 fprintf (dump_file
, ")\n");
828 if (code
== MINUS_EXPR
)
829 to_add
= chrec_fold_multiply (type
, to_add
, SCALAR_FLOAT_TYPE_P (type
)
830 ? build_real (type
, dconstm1
)
831 : build_int_cst_type (type
, -1));
833 res
= add_to_evolution_1 (loop_nb
, chrec_before
, to_add
, at_stmt
);
835 if (dump_file
&& (dump_flags
& TDF_SCEV
))
837 fprintf (dump_file
, " (res = ");
838 print_generic_expr (dump_file
, res
, 0);
839 fprintf (dump_file
, "))\n");
847 /* This section selects the loops that will be good candidates for the
848 scalar evolution analysis. For the moment, greedily select all the
849 loop nests we could analyze. */
851 /* For a loop with a single exit edge, return the COND_EXPR that
852 guards the exit edge. If the expression is too difficult to
853 analyze, then give up. */
856 get_loop_exit_condition (const struct loop
*loop
)
859 edge exit_edge
= single_exit (loop
);
861 if (dump_file
&& (dump_flags
& TDF_SCEV
))
862 fprintf (dump_file
, "(get_loop_exit_condition \n ");
868 stmt
= last_stmt (exit_edge
->src
);
869 if (gcond
*cond_stmt
= dyn_cast
<gcond
*> (stmt
))
873 if (dump_file
&& (dump_flags
& TDF_SCEV
))
875 print_gimple_stmt (dump_file
, res
, 0, 0);
876 fprintf (dump_file
, ")\n");
883 /* Depth first search algorithm. */
892 static t_bool
follow_ssa_edge (struct loop
*loop
, gimple
*, gphi
*,
895 /* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
896 Return true if the strongly connected component has been found. */
899 follow_ssa_edge_binary (struct loop
*loop
, gimple
*at_stmt
,
900 tree type
, tree rhs0
, enum tree_code code
, tree rhs1
,
901 gphi
*halting_phi
, tree
*evolution_of_loop
,
904 t_bool res
= t_false
;
909 case POINTER_PLUS_EXPR
:
911 if (TREE_CODE (rhs0
) == SSA_NAME
)
913 if (TREE_CODE (rhs1
) == SSA_NAME
)
915 /* Match an assignment under the form:
918 /* We want only assignments of form "name + name" contribute to
919 LIMIT, as the other cases do not necessarily contribute to
920 the complexity of the expression. */
923 evol
= *evolution_of_loop
;
924 evol
= add_to_evolution
926 chrec_convert (type
, evol
, at_stmt
),
927 code
, rhs1
, at_stmt
);
928 res
= follow_ssa_edge
929 (loop
, SSA_NAME_DEF_STMT (rhs0
), halting_phi
, &evol
, limit
);
931 *evolution_of_loop
= evol
;
932 else if (res
== t_false
)
934 *evolution_of_loop
= add_to_evolution
936 chrec_convert (type
, *evolution_of_loop
, at_stmt
),
937 code
, rhs0
, at_stmt
);
938 res
= follow_ssa_edge
939 (loop
, SSA_NAME_DEF_STMT (rhs1
), halting_phi
,
940 evolution_of_loop
, limit
);
943 else if (res
== t_dont_know
)
944 *evolution_of_loop
= chrec_dont_know
;
947 else if (res
== t_dont_know
)
948 *evolution_of_loop
= chrec_dont_know
;
953 /* Match an assignment under the form:
955 *evolution_of_loop
= add_to_evolution
956 (loop
->num
, chrec_convert (type
, *evolution_of_loop
,
958 code
, rhs1
, at_stmt
);
959 res
= follow_ssa_edge
960 (loop
, SSA_NAME_DEF_STMT (rhs0
), halting_phi
,
961 evolution_of_loop
, limit
);
964 else if (res
== t_dont_know
)
965 *evolution_of_loop
= chrec_dont_know
;
969 else if (TREE_CODE (rhs1
) == SSA_NAME
)
971 /* Match an assignment under the form:
973 *evolution_of_loop
= add_to_evolution
974 (loop
->num
, chrec_convert (type
, *evolution_of_loop
,
976 code
, rhs0
, at_stmt
);
977 res
= follow_ssa_edge
978 (loop
, SSA_NAME_DEF_STMT (rhs1
), halting_phi
,
979 evolution_of_loop
, limit
);
982 else if (res
== t_dont_know
)
983 *evolution_of_loop
= chrec_dont_know
;
987 /* Otherwise, match an assignment under the form:
989 /* And there is nothing to do. */
994 /* This case is under the form "opnd0 = rhs0 - rhs1". */
995 if (TREE_CODE (rhs0
) == SSA_NAME
)
997 /* Match an assignment under the form:
1000 /* We want only assignments of form "name - name" contribute to
1001 LIMIT, as the other cases do not necessarily contribute to
1002 the complexity of the expression. */
1003 if (TREE_CODE (rhs1
) == SSA_NAME
)
1006 *evolution_of_loop
= add_to_evolution
1007 (loop
->num
, chrec_convert (type
, *evolution_of_loop
, at_stmt
),
1008 MINUS_EXPR
, rhs1
, at_stmt
);
1009 res
= follow_ssa_edge (loop
, SSA_NAME_DEF_STMT (rhs0
), halting_phi
,
1010 evolution_of_loop
, limit
);
1013 else if (res
== t_dont_know
)
1014 *evolution_of_loop
= chrec_dont_know
;
1017 /* Otherwise, match an assignment under the form:
1019 /* And there is nothing to do. */
1030 /* Follow the ssa edge into the expression EXPR.
1031 Return true if the strongly connected component has been found. */
1034 follow_ssa_edge_expr (struct loop
*loop
, gimple
*at_stmt
, tree expr
,
1035 gphi
*halting_phi
, tree
*evolution_of_loop
,
1038 enum tree_code code
= TREE_CODE (expr
);
1039 tree type
= TREE_TYPE (expr
), rhs0
, rhs1
;
1042 /* The EXPR is one of the following cases:
1046 - a POINTER_PLUS_EXPR,
1049 - other cases are not yet handled. */
1054 /* This assignment is under the form "a_1 = (cast) rhs. */
1055 res
= follow_ssa_edge_expr (loop
, at_stmt
, TREE_OPERAND (expr
, 0),
1056 halting_phi
, evolution_of_loop
, limit
);
1057 *evolution_of_loop
= chrec_convert (type
, *evolution_of_loop
, at_stmt
);
1061 /* This assignment is under the form "a_1 = 7". */
1066 /* This assignment is under the form: "a_1 = b_2". */
1067 res
= follow_ssa_edge
1068 (loop
, SSA_NAME_DEF_STMT (expr
), halting_phi
, evolution_of_loop
, limit
);
1071 case POINTER_PLUS_EXPR
:
1074 /* This case is under the form "rhs0 +- rhs1". */
1075 rhs0
= TREE_OPERAND (expr
, 0);
1076 rhs1
= TREE_OPERAND (expr
, 1);
1077 type
= TREE_TYPE (rhs0
);
1078 STRIP_USELESS_TYPE_CONVERSION (rhs0
);
1079 STRIP_USELESS_TYPE_CONVERSION (rhs1
);
1080 res
= follow_ssa_edge_binary (loop
, at_stmt
, type
, rhs0
, code
, rhs1
,
1081 halting_phi
, evolution_of_loop
, limit
);
1085 /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */
1086 if (TREE_CODE (TREE_OPERAND (expr
, 0)) == MEM_REF
)
1088 expr
= TREE_OPERAND (expr
, 0);
1089 rhs0
= TREE_OPERAND (expr
, 0);
1090 rhs1
= TREE_OPERAND (expr
, 1);
1091 type
= TREE_TYPE (rhs0
);
1092 STRIP_USELESS_TYPE_CONVERSION (rhs0
);
1093 STRIP_USELESS_TYPE_CONVERSION (rhs1
);
1094 res
= follow_ssa_edge_binary (loop
, at_stmt
, type
,
1095 rhs0
, POINTER_PLUS_EXPR
, rhs1
,
1096 halting_phi
, evolution_of_loop
, limit
);
1103 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1104 It must be handled as a copy assignment of the form a_1 = a_2. */
1105 rhs0
= ASSERT_EXPR_VAR (expr
);
1106 if (TREE_CODE (rhs0
) == SSA_NAME
)
1107 res
= follow_ssa_edge (loop
, SSA_NAME_DEF_STMT (rhs0
),
1108 halting_phi
, evolution_of_loop
, limit
);
1121 /* Follow the ssa edge into the right hand side of an assignment STMT.
1122 Return true if the strongly connected component has been found. */
1125 follow_ssa_edge_in_rhs (struct loop
*loop
, gimple
*stmt
,
1126 gphi
*halting_phi
, tree
*evolution_of_loop
,
1129 enum tree_code code
= gimple_assign_rhs_code (stmt
);
1130 tree type
= gimple_expr_type (stmt
), rhs1
, rhs2
;
1136 /* This assignment is under the form "a_1 = (cast) rhs. */
1137 res
= follow_ssa_edge_expr (loop
, stmt
, gimple_assign_rhs1 (stmt
),
1138 halting_phi
, evolution_of_loop
, limit
);
1139 *evolution_of_loop
= chrec_convert (type
, *evolution_of_loop
, stmt
);
1142 case POINTER_PLUS_EXPR
:
1145 rhs1
= gimple_assign_rhs1 (stmt
);
1146 rhs2
= gimple_assign_rhs2 (stmt
);
1147 type
= TREE_TYPE (rhs1
);
1148 res
= follow_ssa_edge_binary (loop
, stmt
, type
, rhs1
, code
, rhs2
,
1149 halting_phi
, evolution_of_loop
, limit
);
1153 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1154 res
= follow_ssa_edge_expr (loop
, stmt
, gimple_assign_rhs1 (stmt
),
1155 halting_phi
, evolution_of_loop
, limit
);
1164 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1167 backedge_phi_arg_p (gphi
*phi
, int i
)
1169 const_edge e
= gimple_phi_arg_edge (phi
, i
);
1171 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1172 about updating it anywhere, and this should work as well most of the
1174 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
1180 /* Helper function for one branch of the condition-phi-node. Return
1181 true if the strongly connected component has been found following
1184 static inline t_bool
1185 follow_ssa_edge_in_condition_phi_branch (int i
,
1187 gphi
*condition_phi
,
1189 tree
*evolution_of_branch
,
1190 tree init_cond
, int limit
)
1192 tree branch
= PHI_ARG_DEF (condition_phi
, i
);
1193 *evolution_of_branch
= chrec_dont_know
;
1195 /* Do not follow back edges (they must belong to an irreducible loop, which
1196 we really do not want to worry about). */
1197 if (backedge_phi_arg_p (condition_phi
, i
))
1200 if (TREE_CODE (branch
) == SSA_NAME
)
1202 *evolution_of_branch
= init_cond
;
1203 return follow_ssa_edge (loop
, SSA_NAME_DEF_STMT (branch
), halting_phi
,
1204 evolution_of_branch
, limit
);
1207 /* This case occurs when one of the condition branches sets
1208 the variable to a constant: i.e. a phi-node like
1209 "a_2 = PHI <a_7(5), 2(6)>;".
1211 FIXME: This case have to be refined correctly:
1212 in some cases it is possible to say something better than
1213 chrec_dont_know, for example using a wrap-around notation. */
1217 /* This function merges the branches of a condition-phi-node in a
1221 follow_ssa_edge_in_condition_phi (struct loop
*loop
,
1222 gphi
*condition_phi
,
1224 tree
*evolution_of_loop
, int limit
)
1227 tree init
= *evolution_of_loop
;
1228 tree evolution_of_branch
;
1229 t_bool res
= follow_ssa_edge_in_condition_phi_branch (0, loop
, condition_phi
,
1231 &evolution_of_branch
,
1233 if (res
== t_false
|| res
== t_dont_know
)
1236 *evolution_of_loop
= evolution_of_branch
;
1238 n
= gimple_phi_num_args (condition_phi
);
1239 for (i
= 1; i
< n
; i
++)
1241 /* Quickly give up when the evolution of one of the branches is
1243 if (*evolution_of_loop
== chrec_dont_know
)
1246 /* Increase the limit by the PHI argument number to avoid exponential
1247 time and memory complexity. */
1248 res
= follow_ssa_edge_in_condition_phi_branch (i
, loop
, condition_phi
,
1250 &evolution_of_branch
,
1252 if (res
== t_false
|| res
== t_dont_know
)
1255 *evolution_of_loop
= chrec_merge (*evolution_of_loop
,
1256 evolution_of_branch
);
1262 /* Follow an SSA edge in an inner loop. It computes the overall
1263 effect of the loop, and following the symbolic initial conditions,
1264 it follows the edges in the parent loop. The inner loop is
1265 considered as a single statement. */
1268 follow_ssa_edge_inner_loop_phi (struct loop
*outer_loop
,
1269 gphi
*loop_phi_node
,
1271 tree
*evolution_of_loop
, int limit
)
1273 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1274 tree ev
= analyze_scalar_evolution (loop
, PHI_RESULT (loop_phi_node
));
1276 /* Sometimes, the inner loop is too difficult to analyze, and the
1277 result of the analysis is a symbolic parameter. */
1278 if (ev
== PHI_RESULT (loop_phi_node
))
1280 t_bool res
= t_false
;
1281 int i
, n
= gimple_phi_num_args (loop_phi_node
);
1283 for (i
= 0; i
< n
; i
++)
1285 tree arg
= PHI_ARG_DEF (loop_phi_node
, i
);
1288 /* Follow the edges that exit the inner loop. */
1289 bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1290 if (!flow_bb_inside_loop_p (loop
, bb
))
1291 res
= follow_ssa_edge_expr (outer_loop
, loop_phi_node
,
1293 evolution_of_loop
, limit
);
1298 /* If the path crosses this loop-phi, give up. */
1300 *evolution_of_loop
= chrec_dont_know
;
1305 /* Otherwise, compute the overall effect of the inner loop. */
1306 ev
= compute_overall_effect_of_inner_loop (loop
, ev
);
1307 return follow_ssa_edge_expr (outer_loop
, loop_phi_node
, ev
, halting_phi
,
1308 evolution_of_loop
, limit
);
1311 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1312 path that is analyzed on the return walk. */
1315 follow_ssa_edge (struct loop
*loop
, gimple
*def
, gphi
*halting_phi
,
1316 tree
*evolution_of_loop
, int limit
)
1318 struct loop
*def_loop
;
1320 if (gimple_nop_p (def
))
1323 /* Give up if the path is longer than the MAX that we allow. */
1324 if (limit
> PARAM_VALUE (PARAM_SCEV_MAX_EXPR_COMPLEXITY
))
1327 def_loop
= loop_containing_stmt (def
);
1329 switch (gimple_code (def
))
1332 if (!loop_phi_node_p (def
))
1333 /* DEF is a condition-phi-node. Follow the branches, and
1334 record their evolutions. Finally, merge the collected
1335 information and set the approximation to the main
1337 return follow_ssa_edge_in_condition_phi
1338 (loop
, as_a
<gphi
*> (def
), halting_phi
, evolution_of_loop
,
1341 /* When the analyzed phi is the halting_phi, the
1342 depth-first search is over: we have found a path from
1343 the halting_phi to itself in the loop. */
1344 if (def
== halting_phi
)
1347 /* Otherwise, the evolution of the HALTING_PHI depends
1348 on the evolution of another loop-phi-node, i.e. the
1349 evolution function is a higher degree polynomial. */
1350 if (def_loop
== loop
)
1354 if (flow_loop_nested_p (loop
, def_loop
))
1355 return follow_ssa_edge_inner_loop_phi
1356 (loop
, as_a
<gphi
*> (def
), halting_phi
, evolution_of_loop
,
1363 return follow_ssa_edge_in_rhs (loop
, def
, halting_phi
,
1364 evolution_of_loop
, limit
);
1367 /* At this level of abstraction, the program is just a set
1368 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
1369 other node to be handled. */
1375 /* Simplify PEELED_CHREC represented by (init_cond, arg) in LOOP.
1376 Handle below case and return the corresponding POLYNOMIAL_CHREC:
1378 # i_17 = PHI <i_13(5), 0(3)>
1379 # _20 = PHI <_5(5), start_4(D)(3)>
1382 _5 = start_4(D) + i_13;
1384 Though variable _20 appears as a PEELED_CHREC in the form of
1385 (start_4, _5)_LOOP, it's a POLYNOMIAL_CHREC like {start_4, 1}_LOOP.
1390 simplify_peeled_chrec (struct loop
*loop
, tree arg
, tree init_cond
)
1392 aff_tree aff1
, aff2
;
1393 tree ev
, left
, right
, type
, step_val
;
1394 hash_map
<tree
, name_expansion
*> *peeled_chrec_map
= NULL
;
1396 ev
= instantiate_parameters (loop
, analyze_scalar_evolution (loop
, arg
));
1397 if (ev
== NULL_TREE
|| TREE_CODE (ev
) != POLYNOMIAL_CHREC
)
1398 return chrec_dont_know
;
1400 left
= CHREC_LEFT (ev
);
1401 right
= CHREC_RIGHT (ev
);
1402 type
= TREE_TYPE (left
);
1403 step_val
= chrec_fold_plus (type
, init_cond
, right
);
1405 /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP
1406 if "left" equals to "init + right". */
1407 if (operand_equal_p (left
, step_val
, 0))
1409 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1410 fprintf (dump_file
, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n");
1412 return build_polynomial_chrec (loop
->num
, init_cond
, right
);
1415 /* Try harder to check if they are equal. */
1416 tree_to_aff_combination_expand (left
, type
, &aff1
, &peeled_chrec_map
);
1417 tree_to_aff_combination_expand (step_val
, type
, &aff2
, &peeled_chrec_map
);
1418 free_affine_expand_cache (&peeled_chrec_map
);
1419 aff_combination_scale (&aff2
, -1);
1420 aff_combination_add (&aff1
, &aff2
);
1422 /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP
1423 if "left" equals to "init + right". */
1424 if (aff_combination_zero_p (&aff1
))
1426 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1427 fprintf (dump_file
, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n");
1429 return build_polynomial_chrec (loop
->num
, init_cond
, right
);
1431 return chrec_dont_know
;
1434 /* Given a LOOP_PHI_NODE, this function determines the evolution
1435 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1438 analyze_evolution_in_loop (gphi
*loop_phi_node
,
1441 int i
, n
= gimple_phi_num_args (loop_phi_node
);
1442 tree evolution_function
= chrec_not_analyzed_yet
;
1443 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1445 static bool simplify_peeled_chrec_p
= true;
1447 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1449 fprintf (dump_file
, "(analyze_evolution_in_loop \n");
1450 fprintf (dump_file
, " (loop_phi_node = ");
1451 print_gimple_stmt (dump_file
, loop_phi_node
, 0, 0);
1452 fprintf (dump_file
, ")\n");
1455 for (i
= 0; i
< n
; i
++)
1457 tree arg
= PHI_ARG_DEF (loop_phi_node
, i
);
1462 /* Select the edges that enter the loop body. */
1463 bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1464 if (!flow_bb_inside_loop_p (loop
, bb
))
1467 if (TREE_CODE (arg
) == SSA_NAME
)
1471 ssa_chain
= SSA_NAME_DEF_STMT (arg
);
1473 /* Pass in the initial condition to the follow edge function. */
1475 res
= follow_ssa_edge (loop
, ssa_chain
, loop_phi_node
, &ev_fn
, 0);
1477 /* If ev_fn has no evolution in the inner loop, and the
1478 init_cond is not equal to ev_fn, then we have an
1479 ambiguity between two possible values, as we cannot know
1480 the number of iterations at this point. */
1481 if (TREE_CODE (ev_fn
) != POLYNOMIAL_CHREC
1482 && no_evolution_in_loop_p (ev_fn
, loop
->num
, &val
) && val
1483 && !operand_equal_p (init_cond
, ev_fn
, 0))
1484 ev_fn
= chrec_dont_know
;
1489 /* When it is impossible to go back on the same
1490 loop_phi_node by following the ssa edges, the
1491 evolution is represented by a peeled chrec, i.e. the
1492 first iteration, EV_FN has the value INIT_COND, then
1493 all the other iterations it has the value of ARG.
1494 For the moment, PEELED_CHREC nodes are not built. */
1497 ev_fn
= chrec_dont_know
;
1498 /* Try to recognize POLYNOMIAL_CHREC which appears in
1499 the form of PEELED_CHREC, but guard the process with
1500 a bool variable to keep the analyzer from infinite
1501 recurrence for real PEELED_RECs. */
1502 if (simplify_peeled_chrec_p
&& TREE_CODE (arg
) == SSA_NAME
)
1504 simplify_peeled_chrec_p
= false;
1505 ev_fn
= simplify_peeled_chrec (loop
, arg
, init_cond
);
1506 simplify_peeled_chrec_p
= true;
1510 /* When there are multiple back edges of the loop (which in fact never
1511 happens currently, but nevertheless), merge their evolutions. */
1512 evolution_function
= chrec_merge (evolution_function
, ev_fn
);
1515 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1517 fprintf (dump_file
, " (evolution_function = ");
1518 print_generic_expr (dump_file
, evolution_function
, 0);
1519 fprintf (dump_file
, "))\n");
1522 return evolution_function
;
1525 /* Looks to see if VAR is a copy of a constant (via straightforward assignments
1526 or degenerate phi's). If so, returns the constant; else, returns VAR. */
1529 follow_copies_to_constant (tree var
)
1532 while (TREE_CODE (res
) == SSA_NAME
)
1534 gimple
*def
= SSA_NAME_DEF_STMT (res
);
1535 if (gphi
*phi
= dyn_cast
<gphi
*> (def
))
1537 if (tree rhs
= degenerate_phi_result (phi
))
1542 else if (gimple_assign_single_p (def
))
1543 /* Will exit loop if not an SSA_NAME. */
1544 res
= gimple_assign_rhs1 (def
);
1548 if (CONSTANT_CLASS_P (res
))
1553 /* Given a loop-phi-node, return the initial conditions of the
1554 variable on entry of the loop. When the CCP has propagated
1555 constants into the loop-phi-node, the initial condition is
1556 instantiated, otherwise the initial condition is kept symbolic.
1557 This analyzer does not analyze the evolution outside the current
1558 loop, and leaves this task to the on-demand tree reconstructor. */
1561 analyze_initial_condition (gphi
*loop_phi_node
)
1564 tree init_cond
= chrec_not_analyzed_yet
;
1565 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1567 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1569 fprintf (dump_file
, "(analyze_initial_condition \n");
1570 fprintf (dump_file
, " (loop_phi_node = \n");
1571 print_gimple_stmt (dump_file
, loop_phi_node
, 0, 0);
1572 fprintf (dump_file
, ")\n");
1575 n
= gimple_phi_num_args (loop_phi_node
);
1576 for (i
= 0; i
< n
; i
++)
1578 tree branch
= PHI_ARG_DEF (loop_phi_node
, i
);
1579 basic_block bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1581 /* When the branch is oriented to the loop's body, it does
1582 not contribute to the initial condition. */
1583 if (flow_bb_inside_loop_p (loop
, bb
))
1586 if (init_cond
== chrec_not_analyzed_yet
)
1592 if (TREE_CODE (branch
) == SSA_NAME
)
1594 init_cond
= chrec_dont_know
;
1598 init_cond
= chrec_merge (init_cond
, branch
);
1601 /* Ooops -- a loop without an entry??? */
1602 if (init_cond
== chrec_not_analyzed_yet
)
1603 init_cond
= chrec_dont_know
;
1605 /* We may not have fully constant propagated IL. Handle degenerate PHIs here
1606 to not miss important early loop unrollings. */
1607 init_cond
= follow_copies_to_constant (init_cond
);
1609 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1611 fprintf (dump_file
, " (init_cond = ");
1612 print_generic_expr (dump_file
, init_cond
, 0);
1613 fprintf (dump_file
, "))\n");
1619 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1622 interpret_loop_phi (struct loop
*loop
, gphi
*loop_phi_node
)
1625 struct loop
*phi_loop
= loop_containing_stmt (loop_phi_node
);
1628 if (phi_loop
!= loop
)
1630 struct loop
*subloop
;
1631 tree evolution_fn
= analyze_scalar_evolution
1632 (phi_loop
, PHI_RESULT (loop_phi_node
));
1634 /* Dive one level deeper. */
1635 subloop
= superloop_at_depth (phi_loop
, loop_depth (loop
) + 1);
1637 /* Interpret the subloop. */
1638 res
= compute_overall_effect_of_inner_loop (subloop
, evolution_fn
);
1642 /* Otherwise really interpret the loop phi. */
1643 init_cond
= analyze_initial_condition (loop_phi_node
);
1644 res
= analyze_evolution_in_loop (loop_phi_node
, init_cond
);
1646 /* Verify we maintained the correct initial condition throughout
1647 possible conversions in the SSA chain. */
1648 if (res
!= chrec_dont_know
)
1650 tree new_init
= res
;
1651 if (CONVERT_EXPR_P (res
)
1652 && TREE_CODE (TREE_OPERAND (res
, 0)) == POLYNOMIAL_CHREC
)
1653 new_init
= fold_convert (TREE_TYPE (res
),
1654 CHREC_LEFT (TREE_OPERAND (res
, 0)));
1655 else if (TREE_CODE (res
) == POLYNOMIAL_CHREC
)
1656 new_init
= CHREC_LEFT (res
);
1657 STRIP_USELESS_TYPE_CONVERSION (new_init
);
1658 if (TREE_CODE (new_init
) == POLYNOMIAL_CHREC
1659 || !operand_equal_p (init_cond
, new_init
, 0))
1660 return chrec_dont_know
;
1666 /* This function merges the branches of a condition-phi-node,
1667 contained in the outermost loop, and whose arguments are already
1671 interpret_condition_phi (struct loop
*loop
, gphi
*condition_phi
)
1673 int i
, n
= gimple_phi_num_args (condition_phi
);
1674 tree res
= chrec_not_analyzed_yet
;
1676 for (i
= 0; i
< n
; i
++)
1680 if (backedge_phi_arg_p (condition_phi
, i
))
1682 res
= chrec_dont_know
;
1686 branch_chrec
= analyze_scalar_evolution
1687 (loop
, PHI_ARG_DEF (condition_phi
, i
));
1689 res
= chrec_merge (res
, branch_chrec
);
1695 /* Interpret the operation RHS1 OP RHS2. If we didn't
1696 analyze this node before, follow the definitions until ending
1697 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1698 return path, this function propagates evolutions (ala constant copy
1699 propagation). OPND1 is not a GIMPLE expression because we could
1700 analyze the effect of an inner loop: see interpret_loop_phi. */
1703 interpret_rhs_expr (struct loop
*loop
, gimple
*at_stmt
,
1704 tree type
, tree rhs1
, enum tree_code code
, tree rhs2
)
1706 tree res
, chrec1
, chrec2
;
1709 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1711 if (is_gimple_min_invariant (rhs1
))
1712 return chrec_convert (type
, rhs1
, at_stmt
);
1714 if (code
== SSA_NAME
)
1715 return chrec_convert (type
, analyze_scalar_evolution (loop
, rhs1
),
1718 if (code
== ASSERT_EXPR
)
1720 rhs1
= ASSERT_EXPR_VAR (rhs1
);
1721 return chrec_convert (type
, analyze_scalar_evolution (loop
, rhs1
),
1729 if (TREE_CODE (TREE_OPERAND (rhs1
, 0)) == MEM_REF
1730 || handled_component_p (TREE_OPERAND (rhs1
, 0)))
1733 HOST_WIDE_INT bitsize
, bitpos
;
1734 int unsignedp
, reversep
;
1740 base
= get_inner_reference (TREE_OPERAND (rhs1
, 0),
1741 &bitsize
, &bitpos
, &offset
, &mode
,
1742 &unsignedp
, &reversep
, &volatilep
,
1745 if (TREE_CODE (base
) == MEM_REF
)
1747 rhs2
= TREE_OPERAND (base
, 1);
1748 rhs1
= TREE_OPERAND (base
, 0);
1750 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1751 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1752 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1753 chrec2
= chrec_convert (TREE_TYPE (rhs2
), chrec2
, at_stmt
);
1754 chrec1
= instantiate_parameters (loop
, chrec1
);
1755 chrec2
= instantiate_parameters (loop
, chrec2
);
1756 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1760 chrec1
= analyze_scalar_evolution_for_address_of (loop
, base
);
1761 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1765 if (offset
!= NULL_TREE
)
1767 chrec2
= analyze_scalar_evolution (loop
, offset
);
1768 chrec2
= chrec_convert (TREE_TYPE (offset
), chrec2
, at_stmt
);
1769 chrec2
= instantiate_parameters (loop
, chrec2
);
1770 res
= chrec_fold_plus (type
, res
, chrec2
);
1775 gcc_assert ((bitpos
% BITS_PER_UNIT
) == 0);
1777 unitpos
= size_int (bitpos
/ BITS_PER_UNIT
);
1778 chrec3
= analyze_scalar_evolution (loop
, unitpos
);
1779 chrec3
= chrec_convert (TREE_TYPE (unitpos
), chrec3
, at_stmt
);
1780 chrec3
= instantiate_parameters (loop
, chrec3
);
1781 res
= chrec_fold_plus (type
, res
, chrec3
);
1785 res
= chrec_dont_know
;
1788 case POINTER_PLUS_EXPR
:
1789 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1790 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1791 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1792 chrec2
= chrec_convert (TREE_TYPE (rhs2
), chrec2
, at_stmt
);
1793 chrec1
= instantiate_parameters (loop
, chrec1
);
1794 chrec2
= instantiate_parameters (loop
, chrec2
);
1795 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1799 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1800 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1801 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1802 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1803 chrec1
= instantiate_parameters (loop
, chrec1
);
1804 chrec2
= instantiate_parameters (loop
, chrec2
);
1805 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1809 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1810 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1811 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1812 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1813 chrec1
= instantiate_parameters (loop
, chrec1
);
1814 chrec2
= instantiate_parameters (loop
, chrec2
);
1815 res
= chrec_fold_minus (type
, chrec1
, chrec2
);
1819 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1820 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1821 /* TYPE may be integer, real or complex, so use fold_convert. */
1822 chrec1
= instantiate_parameters (loop
, chrec1
);
1823 res
= chrec_fold_multiply (type
, chrec1
,
1824 fold_convert (type
, integer_minus_one_node
));
1828 /* Handle ~X as -1 - X. */
1829 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1830 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1831 chrec1
= instantiate_parameters (loop
, chrec1
);
1832 res
= chrec_fold_minus (type
,
1833 fold_convert (type
, integer_minus_one_node
),
1838 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1839 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1840 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1841 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1842 chrec1
= instantiate_parameters (loop
, chrec1
);
1843 chrec2
= instantiate_parameters (loop
, chrec2
);
1844 res
= chrec_fold_multiply (type
, chrec1
, chrec2
);
1849 /* Handle A<<B as A * (1<<B). */
1850 tree uns
= unsigned_type_for (type
);
1851 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1852 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1853 chrec1
= chrec_convert (uns
, chrec1
, at_stmt
);
1854 chrec1
= instantiate_parameters (loop
, chrec1
);
1855 chrec2
= instantiate_parameters (loop
, chrec2
);
1857 tree one
= build_int_cst (uns
, 1);
1858 chrec2
= fold_build2 (LSHIFT_EXPR
, uns
, one
, chrec2
);
1859 res
= chrec_fold_multiply (uns
, chrec1
, chrec2
);
1860 res
= chrec_convert (type
, res
, at_stmt
);
1865 /* In case we have a truncation of a widened operation that in
1866 the truncated type has undefined overflow behavior analyze
1867 the operation done in an unsigned type of the same precision
1868 as the final truncation. We cannot derive a scalar evolution
1869 for the widened operation but for the truncated result. */
1870 if (TREE_CODE (type
) == INTEGER_TYPE
1871 && TREE_CODE (TREE_TYPE (rhs1
)) == INTEGER_TYPE
1872 && TYPE_PRECISION (type
) < TYPE_PRECISION (TREE_TYPE (rhs1
))
1873 && TYPE_OVERFLOW_UNDEFINED (type
)
1874 && TREE_CODE (rhs1
) == SSA_NAME
1875 && (def
= SSA_NAME_DEF_STMT (rhs1
))
1876 && is_gimple_assign (def
)
1877 && TREE_CODE_CLASS (gimple_assign_rhs_code (def
)) == tcc_binary
1878 && TREE_CODE (gimple_assign_rhs2 (def
)) == INTEGER_CST
)
1880 tree utype
= unsigned_type_for (type
);
1881 chrec1
= interpret_rhs_expr (loop
, at_stmt
, utype
,
1882 gimple_assign_rhs1 (def
),
1883 gimple_assign_rhs_code (def
),
1884 gimple_assign_rhs2 (def
));
1887 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1888 res
= chrec_convert (type
, chrec1
, at_stmt
);
1892 res
= chrec_dont_know
;
1899 /* Interpret the expression EXPR. */
1902 interpret_expr (struct loop
*loop
, gimple
*at_stmt
, tree expr
)
1904 enum tree_code code
;
1905 tree type
= TREE_TYPE (expr
), op0
, op1
;
1907 if (automatically_generated_chrec_p (expr
))
1910 if (TREE_CODE (expr
) == POLYNOMIAL_CHREC
1911 || get_gimple_rhs_class (TREE_CODE (expr
)) == GIMPLE_TERNARY_RHS
)
1912 return chrec_dont_know
;
1914 extract_ops_from_tree (expr
, &code
, &op0
, &op1
);
1916 return interpret_rhs_expr (loop
, at_stmt
, type
,
1920 /* Interpret the rhs of the assignment STMT. */
1923 interpret_gimple_assign (struct loop
*loop
, gimple
*stmt
)
1925 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
1926 enum tree_code code
= gimple_assign_rhs_code (stmt
);
1928 return interpret_rhs_expr (loop
, stmt
, type
,
1929 gimple_assign_rhs1 (stmt
), code
,
1930 gimple_assign_rhs2 (stmt
));
1935 /* This section contains all the entry points:
1936 - number_of_iterations_in_loop,
1937 - analyze_scalar_evolution,
1938 - instantiate_parameters.
1941 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1942 common ancestor of DEF_LOOP and USE_LOOP. */
1945 compute_scalar_evolution_in_loop (struct loop
*wrto_loop
,
1946 struct loop
*def_loop
,
1952 if (def_loop
== wrto_loop
)
1955 def_loop
= superloop_at_depth (def_loop
, loop_depth (wrto_loop
) + 1);
1956 res
= compute_overall_effect_of_inner_loop (def_loop
, ev
);
1958 if (no_evolution_in_loop_p (res
, wrto_loop
->num
, &val
) && val
)
1961 return analyze_scalar_evolution_1 (wrto_loop
, res
, chrec_not_analyzed_yet
);
1964 /* Helper recursive function. */
1967 analyze_scalar_evolution_1 (struct loop
*loop
, tree var
, tree res
)
1969 tree type
= TREE_TYPE (var
);
1972 struct loop
*def_loop
;
1974 if (loop
== NULL
|| TREE_CODE (type
) == VECTOR_TYPE
)
1975 return chrec_dont_know
;
1977 if (TREE_CODE (var
) != SSA_NAME
)
1978 return interpret_expr (loop
, NULL
, var
);
1980 def
= SSA_NAME_DEF_STMT (var
);
1981 bb
= gimple_bb (def
);
1982 def_loop
= bb
? bb
->loop_father
: NULL
;
1985 || !flow_bb_inside_loop_p (loop
, bb
))
1987 /* Keep symbolic form, but look through obvious copies for constants. */
1988 res
= follow_copies_to_constant (var
);
1992 if (res
!= chrec_not_analyzed_yet
)
1994 if (loop
!= bb
->loop_father
)
1995 res
= compute_scalar_evolution_in_loop
1996 (find_common_loop (loop
, bb
->loop_father
), bb
->loop_father
, res
);
2001 if (loop
!= def_loop
)
2003 res
= analyze_scalar_evolution_1 (def_loop
, var
, chrec_not_analyzed_yet
);
2004 res
= compute_scalar_evolution_in_loop (loop
, def_loop
, res
);
2009 switch (gimple_code (def
))
2012 res
= interpret_gimple_assign (loop
, def
);
2016 if (loop_phi_node_p (def
))
2017 res
= interpret_loop_phi (loop
, as_a
<gphi
*> (def
));
2019 res
= interpret_condition_phi (loop
, as_a
<gphi
*> (def
));
2023 res
= chrec_dont_know
;
2029 /* Keep the symbolic form. */
2030 if (res
== chrec_dont_know
)
2033 if (loop
== def_loop
)
2034 set_scalar_evolution (block_before_loop (loop
), var
, res
);
2039 /* Analyzes and returns the scalar evolution of the ssa_name VAR in
2040 LOOP. LOOP is the loop in which the variable is used.
2042 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
2043 pointer to the statement that uses this variable, in order to
2044 determine the evolution function of the variable, use the following
2047 loop_p loop = loop_containing_stmt (stmt);
2048 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
2049 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
2053 analyze_scalar_evolution (struct loop
*loop
, tree var
)
2057 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2059 fprintf (dump_file
, "(analyze_scalar_evolution \n");
2060 fprintf (dump_file
, " (loop_nb = %d)\n", loop
->num
);
2061 fprintf (dump_file
, " (scalar = ");
2062 print_generic_expr (dump_file
, var
, 0);
2063 fprintf (dump_file
, ")\n");
2066 res
= get_scalar_evolution (block_before_loop (loop
), var
);
2067 res
= analyze_scalar_evolution_1 (loop
, var
, res
);
2069 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2070 fprintf (dump_file
, ")\n");
2075 /* Analyzes and returns the scalar evolution of VAR address in LOOP. */
2078 analyze_scalar_evolution_for_address_of (struct loop
*loop
, tree var
)
2080 return analyze_scalar_evolution (loop
, build_fold_addr_expr (var
));
2083 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
2084 WRTO_LOOP (which should be a superloop of USE_LOOP)
2086 FOLDED_CASTS is set to true if resolve_mixers used
2087 chrec_convert_aggressive (TODO -- not really, we are way too conservative
2088 at the moment in order to keep things simple).
2090 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
2093 for (i = 0; i < 100; i++) -- loop 1
2095 for (j = 0; j < 100; j++) -- loop 2
2102 for (t = 0; t < 100; t++) -- loop 3
2109 Both k1 and k2 are invariants in loop3, thus
2110 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
2111 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
2113 As they are invariant, it does not matter whether we consider their
2114 usage in loop 3 or loop 2, hence
2115 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
2116 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
2117 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
2118 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
2120 Similarly for their evolutions with respect to loop 1. The values of K2
2121 in the use in loop 2 vary independently on loop 1, thus we cannot express
2122 the evolution with respect to loop 1:
2123 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2124 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2125 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2126 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2128 The value of k2 in the use in loop 1 is known, though:
2129 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2130 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2134 analyze_scalar_evolution_in_loop (struct loop
*wrto_loop
, struct loop
*use_loop
,
2135 tree version
, bool *folded_casts
)
2138 tree ev
= version
, tmp
;
2140 /* We cannot just do
2142 tmp = analyze_scalar_evolution (use_loop, version);
2143 ev = resolve_mixers (wrto_loop, tmp, folded_casts);
2145 as resolve_mixers would query the scalar evolution with respect to
2146 wrto_loop. For example, in the situation described in the function
2147 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2150 analyze_scalar_evolution (use_loop, version) = k2
2152 and resolve_mixers (loop1, k2, folded_casts) finds that the value of
2153 k2 in loop 1 is 100, which is a wrong result, since we are interested
2154 in the value in loop 3.
2156 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2157 each time checking that there is no evolution in the inner loop. */
2160 *folded_casts
= false;
2163 tmp
= analyze_scalar_evolution (use_loop
, ev
);
2164 ev
= resolve_mixers (use_loop
, tmp
, folded_casts
);
2166 if (use_loop
== wrto_loop
)
2169 /* If the value of the use changes in the inner loop, we cannot express
2170 its value in the outer loop (we might try to return interval chrec,
2171 but we do not have a user for it anyway) */
2172 if (!no_evolution_in_loop_p (ev
, use_loop
->num
, &val
)
2174 return chrec_dont_know
;
2176 use_loop
= loop_outer (use_loop
);
2181 /* Hashtable helpers for a temporary hash-table used when
2182 instantiating a CHREC or resolving mixers. For this use
2183 instantiated_below is always the same. */
2185 struct instantiate_cache_type
2188 vec
<scev_info_str
> entries
;
2190 instantiate_cache_type () : map (NULL
), entries (vNULL
) {}
2191 ~instantiate_cache_type ();
2192 tree
get (unsigned slot
) { return entries
[slot
].chrec
; }
2193 void set (unsigned slot
, tree chrec
) { entries
[slot
].chrec
= chrec
; }
2196 instantiate_cache_type::~instantiate_cache_type ()
2205 /* Cache to avoid infinite recursion when instantiating an SSA name.
2206 Live during the outermost instantiate_scev or resolve_mixers call. */
2207 static instantiate_cache_type
*global_cache
;
2209 /* Computes a hash function for database element ELT. */
2211 static inline hashval_t
2212 hash_idx_scev_info (const void *elt_
)
2214 unsigned idx
= ((size_t) elt_
) - 2;
2215 return scev_info_hasher::hash (&global_cache
->entries
[idx
]);
2218 /* Compares database elements E1 and E2. */
2221 eq_idx_scev_info (const void *e1
, const void *e2
)
2223 unsigned idx1
= ((size_t) e1
) - 2;
2224 return scev_info_hasher::equal (&global_cache
->entries
[idx1
],
2225 (const scev_info_str
*) e2
);
2228 /* Returns from CACHE the slot number of the cached chrec for NAME. */
2231 get_instantiated_value_entry (instantiate_cache_type
&cache
,
2232 tree name
, basic_block instantiate_below
)
2236 cache
.map
= htab_create (10, hash_idx_scev_info
, eq_idx_scev_info
, NULL
);
2237 cache
.entries
.create (10);
2241 e
.name_version
= SSA_NAME_VERSION (name
);
2242 e
.instantiated_below
= instantiate_below
->index
;
2243 void **slot
= htab_find_slot_with_hash (cache
.map
, &e
,
2244 scev_info_hasher::hash (&e
), INSERT
);
2247 e
.chrec
= chrec_not_analyzed_yet
;
2248 *slot
= (void *)(size_t)(cache
.entries
.length () + 2);
2249 cache
.entries
.safe_push (e
);
2252 return ((size_t)*slot
) - 2;
2256 /* Return the closed_loop_phi node for VAR. If there is none, return
2260 loop_closed_phi_def (tree var
)
2267 if (var
== NULL_TREE
2268 || TREE_CODE (var
) != SSA_NAME
)
2271 loop
= loop_containing_stmt (SSA_NAME_DEF_STMT (var
));
2272 exit
= single_exit (loop
);
2276 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); gsi_next (&psi
))
2279 if (PHI_ARG_DEF_FROM_EDGE (phi
, exit
) == var
)
2280 return PHI_RESULT (phi
);
2286 static tree
instantiate_scev_r (basic_block
, struct loop
*, struct loop
*,
2289 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2290 and EVOLUTION_LOOP, that were left under a symbolic form.
2292 CHREC is an SSA_NAME to be instantiated.
2294 CACHE is the cache of already instantiated values.
2296 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2297 conversions that may wrap in signed/pointer type are folded, as long
2298 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2299 then we don't do such fold.
2301 SIZE_EXPR is used for computing the size of the expression to be
2302 instantiated, and to stop if it exceeds some limit. */
2305 instantiate_scev_name (basic_block instantiate_below
,
2306 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2308 bool *fold_conversions
,
2312 struct loop
*def_loop
;
2313 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (chrec
));
2315 /* A parameter (or loop invariant and we do not want to include
2316 evolutions in outer loops), nothing to do. */
2318 || loop_depth (def_bb
->loop_father
) == 0
2319 || dominated_by_p (CDI_DOMINATORS
, instantiate_below
, def_bb
))
2322 /* We cache the value of instantiated variable to avoid exponential
2323 time complexity due to reevaluations. We also store the convenient
2324 value in the cache in order to prevent infinite recursion -- we do
2325 not want to instantiate the SSA_NAME if it is in a mixer
2326 structure. This is used for avoiding the instantiation of
2327 recursively defined functions, such as:
2329 | a_2 -> {0, +, 1, +, a_2}_1 */
2331 unsigned si
= get_instantiated_value_entry (*global_cache
,
2332 chrec
, instantiate_below
);
2333 if (global_cache
->get (si
) != chrec_not_analyzed_yet
)
2334 return global_cache
->get (si
);
2336 /* On recursion return chrec_dont_know. */
2337 global_cache
->set (si
, chrec_dont_know
);
2339 def_loop
= find_common_loop (evolution_loop
, def_bb
->loop_father
);
2341 /* If the analysis yields a parametric chrec, instantiate the
2343 res
= analyze_scalar_evolution (def_loop
, chrec
);
2345 /* Don't instantiate default definitions. */
2346 if (TREE_CODE (res
) == SSA_NAME
2347 && SSA_NAME_IS_DEFAULT_DEF (res
))
2350 /* Don't instantiate loop-closed-ssa phi nodes. */
2351 else if (TREE_CODE (res
) == SSA_NAME
2352 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res
)))
2353 > loop_depth (def_loop
))
2356 res
= loop_closed_phi_def (chrec
);
2360 /* When there is no loop_closed_phi_def, it means that the
2361 variable is not used after the loop: try to still compute the
2362 value of the variable when exiting the loop. */
2363 if (res
== NULL_TREE
)
2365 loop_p loop
= loop_containing_stmt (SSA_NAME_DEF_STMT (chrec
));
2366 res
= analyze_scalar_evolution (loop
, chrec
);
2367 res
= compute_overall_effect_of_inner_loop (loop
, res
);
2368 res
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2370 fold_conversions
, size_expr
);
2372 else if (!dominated_by_p (CDI_DOMINATORS
, instantiate_below
,
2373 gimple_bb (SSA_NAME_DEF_STMT (res
))))
2374 res
= chrec_dont_know
;
2377 else if (res
!= chrec_dont_know
)
2380 && def_bb
->loop_father
!= inner_loop
2381 && !flow_loop_nested_p (def_bb
->loop_father
, inner_loop
))
2382 /* ??? We could try to compute the overall effect of the loop here. */
2383 res
= chrec_dont_know
;
2385 res
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2387 fold_conversions
, size_expr
);
2390 /* Store the correct value to the cache. */
2391 global_cache
->set (si
, res
);
2395 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2396 and EVOLUTION_LOOP, that were left under a symbolic form.
2398 CHREC is a polynomial chain of recurrence to be instantiated.
2400 CACHE is the cache of already instantiated values.
2402 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2403 conversions that may wrap in signed/pointer type are folded, as long
2404 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2405 then we don't do such fold.
2407 SIZE_EXPR is used for computing the size of the expression to be
2408 instantiated, and to stop if it exceeds some limit. */
2411 instantiate_scev_poly (basic_block instantiate_below
,
2412 struct loop
*evolution_loop
, struct loop
*,
2413 tree chrec
, bool *fold_conversions
, int size_expr
)
2416 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2417 get_chrec_loop (chrec
),
2418 CHREC_LEFT (chrec
), fold_conversions
,
2420 if (op0
== chrec_dont_know
)
2421 return chrec_dont_know
;
2423 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2424 get_chrec_loop (chrec
),
2425 CHREC_RIGHT (chrec
), fold_conversions
,
2427 if (op1
== chrec_dont_know
)
2428 return chrec_dont_know
;
2430 if (CHREC_LEFT (chrec
) != op0
2431 || CHREC_RIGHT (chrec
) != op1
)
2433 op1
= chrec_convert_rhs (chrec_type (op0
), op1
, NULL
);
2434 chrec
= build_polynomial_chrec (CHREC_VARIABLE (chrec
), op0
, op1
);
2440 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2441 and EVOLUTION_LOOP, that were left under a symbolic form.
2443 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
2445 CACHE is the cache of already instantiated values.
2447 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2448 conversions that may wrap in signed/pointer type are folded, as long
2449 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2450 then we don't do such fold.
2452 SIZE_EXPR is used for computing the size of the expression to be
2453 instantiated, and to stop if it exceeds some limit. */
2456 instantiate_scev_binary (basic_block instantiate_below
,
2457 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2458 tree chrec
, enum tree_code code
,
2459 tree type
, tree c0
, tree c1
,
2460 bool *fold_conversions
, int size_expr
)
2463 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
, inner_loop
,
2464 c0
, fold_conversions
, size_expr
);
2465 if (op0
== chrec_dont_know
)
2466 return chrec_dont_know
;
2468 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
, inner_loop
,
2469 c1
, fold_conversions
, size_expr
);
2470 if (op1
== chrec_dont_know
)
2471 return chrec_dont_know
;
2476 op0
= chrec_convert (type
, op0
, NULL
);
2477 op1
= chrec_convert_rhs (type
, op1
, NULL
);
2481 case POINTER_PLUS_EXPR
:
2483 return chrec_fold_plus (type
, op0
, op1
);
2486 return chrec_fold_minus (type
, op0
, op1
);
2489 return chrec_fold_multiply (type
, op0
, op1
);
2496 return chrec
? chrec
: fold_build2 (code
, type
, c0
, c1
);
2499 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2500 and EVOLUTION_LOOP, that were left under a symbolic form.
2502 "CHREC" is an array reference to be instantiated.
2504 CACHE is the cache of already instantiated values.
2506 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2507 conversions that may wrap in signed/pointer type are folded, as long
2508 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2509 then we don't do such fold.
2511 SIZE_EXPR is used for computing the size of the expression to be
2512 instantiated, and to stop if it exceeds some limit. */
2515 instantiate_array_ref (basic_block instantiate_below
,
2516 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2517 tree chrec
, bool *fold_conversions
, int size_expr
)
2520 tree index
= TREE_OPERAND (chrec
, 1);
2521 tree op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2523 fold_conversions
, size_expr
);
2525 if (op1
== chrec_dont_know
)
2526 return chrec_dont_know
;
2528 if (chrec
&& op1
== index
)
2531 res
= unshare_expr (chrec
);
2532 TREE_OPERAND (res
, 1) = op1
;
2536 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2537 and EVOLUTION_LOOP, that were left under a symbolic form.
2539 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2542 CACHE is the cache of already instantiated values.
2544 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2545 conversions that may wrap in signed/pointer type are folded, as long
2546 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2547 then we don't do such fold.
2549 SIZE_EXPR is used for computing the size of the expression to be
2550 instantiated, and to stop if it exceeds some limit. */
2553 instantiate_scev_convert (basic_block instantiate_below
,
2554 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2555 tree chrec
, tree type
, tree op
,
2556 bool *fold_conversions
, int size_expr
)
2558 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2560 fold_conversions
, size_expr
);
2562 if (op0
== chrec_dont_know
)
2563 return chrec_dont_know
;
2565 if (fold_conversions
)
2567 tree tmp
= chrec_convert_aggressive (type
, op0
, fold_conversions
);
2571 /* If we used chrec_convert_aggressive, we can no longer assume that
2572 signed chrecs do not overflow, as chrec_convert does, so avoid
2573 calling it in that case. */
2574 if (*fold_conversions
)
2576 if (chrec
&& op0
== op
)
2579 return fold_convert (type
, op0
);
2583 return chrec_convert (type
, op0
, NULL
);
2586 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2587 and EVOLUTION_LOOP, that were left under a symbolic form.
2589 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
2590 Handle ~X as -1 - X.
2591 Handle -X as -1 * X.
2593 CACHE is the cache of already instantiated values.
2595 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2596 conversions that may wrap in signed/pointer type are folded, as long
2597 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2598 then we don't do such fold.
2600 SIZE_EXPR is used for computing the size of the expression to be
2601 instantiated, and to stop if it exceeds some limit. */
2604 instantiate_scev_not (basic_block instantiate_below
,
2605 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2607 enum tree_code code
, tree type
, tree op
,
2608 bool *fold_conversions
, int size_expr
)
2610 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2612 fold_conversions
, size_expr
);
2614 if (op0
== chrec_dont_know
)
2615 return chrec_dont_know
;
2619 op0
= chrec_convert (type
, op0
, NULL
);
2624 return chrec_fold_minus
2625 (type
, fold_convert (type
, integer_minus_one_node
), op0
);
2628 return chrec_fold_multiply
2629 (type
, fold_convert (type
, integer_minus_one_node
), op0
);
2636 return chrec
? chrec
: fold_build1 (code
, type
, op0
);
2639 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2640 and EVOLUTION_LOOP, that were left under a symbolic form.
2642 CHREC is an expression with 3 operands to be instantiated.
2644 CACHE is the cache of already instantiated values.
2646 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2647 conversions that may wrap in signed/pointer type are folded, as long
2648 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2649 then we don't do such fold.
2651 SIZE_EXPR is used for computing the size of the expression to be
2652 instantiated, and to stop if it exceeds some limit. */
2655 instantiate_scev_3 (basic_block instantiate_below
,
2656 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2658 bool *fold_conversions
, int size_expr
)
2661 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2662 inner_loop
, TREE_OPERAND (chrec
, 0),
2663 fold_conversions
, size_expr
);
2664 if (op0
== chrec_dont_know
)
2665 return chrec_dont_know
;
2667 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2668 inner_loop
, TREE_OPERAND (chrec
, 1),
2669 fold_conversions
, size_expr
);
2670 if (op1
== chrec_dont_know
)
2671 return chrec_dont_know
;
2673 op2
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2674 inner_loop
, TREE_OPERAND (chrec
, 2),
2675 fold_conversions
, size_expr
);
2676 if (op2
== chrec_dont_know
)
2677 return chrec_dont_know
;
2679 if (op0
== TREE_OPERAND (chrec
, 0)
2680 && op1
== TREE_OPERAND (chrec
, 1)
2681 && op2
== TREE_OPERAND (chrec
, 2))
2684 return fold_build3 (TREE_CODE (chrec
),
2685 TREE_TYPE (chrec
), op0
, op1
, op2
);
2688 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2689 and EVOLUTION_LOOP, that were left under a symbolic form.
2691 CHREC is an expression with 2 operands to be instantiated.
2693 CACHE is the cache of already instantiated values.
2695 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2696 conversions that may wrap in signed/pointer type are folded, as long
2697 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2698 then we don't do such fold.
2700 SIZE_EXPR is used for computing the size of the expression to be
2701 instantiated, and to stop if it exceeds some limit. */
2704 instantiate_scev_2 (basic_block instantiate_below
,
2705 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2707 bool *fold_conversions
, int size_expr
)
2710 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2711 inner_loop
, TREE_OPERAND (chrec
, 0),
2712 fold_conversions
, size_expr
);
2713 if (op0
== chrec_dont_know
)
2714 return chrec_dont_know
;
2716 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2717 inner_loop
, TREE_OPERAND (chrec
, 1),
2718 fold_conversions
, size_expr
);
2719 if (op1
== chrec_dont_know
)
2720 return chrec_dont_know
;
2722 if (op0
== TREE_OPERAND (chrec
, 0)
2723 && op1
== TREE_OPERAND (chrec
, 1))
2726 return fold_build2 (TREE_CODE (chrec
), TREE_TYPE (chrec
), op0
, op1
);
2729 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2730 and EVOLUTION_LOOP, that were left under a symbolic form.
2732 CHREC is an expression with 2 operands to be instantiated.
2734 CACHE is the cache of already instantiated values.
2736 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2737 conversions that may wrap in signed/pointer type are folded, as long
2738 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2739 then we don't do such fold.
2741 SIZE_EXPR is used for computing the size of the expression to be
2742 instantiated, and to stop if it exceeds some limit. */
2745 instantiate_scev_1 (basic_block instantiate_below
,
2746 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2748 bool *fold_conversions
, int size_expr
)
2750 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2751 inner_loop
, TREE_OPERAND (chrec
, 0),
2752 fold_conversions
, size_expr
);
2754 if (op0
== chrec_dont_know
)
2755 return chrec_dont_know
;
2757 if (op0
== TREE_OPERAND (chrec
, 0))
2760 return fold_build1 (TREE_CODE (chrec
), TREE_TYPE (chrec
), op0
);
2763 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2764 and EVOLUTION_LOOP, that were left under a symbolic form.
2766 CHREC is the scalar evolution to instantiate.
2768 CACHE is the cache of already instantiated values.
2770 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2771 conversions that may wrap in signed/pointer type are folded, as long
2772 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2773 then we don't do such fold.
2775 SIZE_EXPR is used for computing the size of the expression to be
2776 instantiated, and to stop if it exceeds some limit. */
2779 instantiate_scev_r (basic_block instantiate_below
,
2780 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2782 bool *fold_conversions
, int size_expr
)
2784 /* Give up if the expression is larger than the MAX that we allow. */
2785 if (size_expr
++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE
))
2786 return chrec_dont_know
;
2788 if (chrec
== NULL_TREE
2789 || automatically_generated_chrec_p (chrec
)
2790 || is_gimple_min_invariant (chrec
))
2793 switch (TREE_CODE (chrec
))
2796 return instantiate_scev_name (instantiate_below
, evolution_loop
,
2798 fold_conversions
, size_expr
);
2800 case POLYNOMIAL_CHREC
:
2801 return instantiate_scev_poly (instantiate_below
, evolution_loop
,
2803 fold_conversions
, size_expr
);
2805 case POINTER_PLUS_EXPR
:
2809 return instantiate_scev_binary (instantiate_below
, evolution_loop
,
2811 TREE_CODE (chrec
), chrec_type (chrec
),
2812 TREE_OPERAND (chrec
, 0),
2813 TREE_OPERAND (chrec
, 1),
2814 fold_conversions
, size_expr
);
2817 return instantiate_scev_convert (instantiate_below
, evolution_loop
,
2819 TREE_TYPE (chrec
), TREE_OPERAND (chrec
, 0),
2820 fold_conversions
, size_expr
);
2824 return instantiate_scev_not (instantiate_below
, evolution_loop
,
2826 TREE_CODE (chrec
), TREE_TYPE (chrec
),
2827 TREE_OPERAND (chrec
, 0),
2828 fold_conversions
, size_expr
);
2831 case SCEV_NOT_KNOWN
:
2832 return chrec_dont_know
;
2838 return instantiate_array_ref (instantiate_below
, evolution_loop
,
2840 fold_conversions
, size_expr
);
2846 if (VL_EXP_CLASS_P (chrec
))
2847 return chrec_dont_know
;
2849 switch (TREE_CODE_LENGTH (TREE_CODE (chrec
)))
2852 return instantiate_scev_3 (instantiate_below
, evolution_loop
,
2854 fold_conversions
, size_expr
);
2857 return instantiate_scev_2 (instantiate_below
, evolution_loop
,
2859 fold_conversions
, size_expr
);
2862 return instantiate_scev_1 (instantiate_below
, evolution_loop
,
2864 fold_conversions
, size_expr
);
2873 /* Too complicated to handle. */
2874 return chrec_dont_know
;
2877 /* Analyze all the parameters of the chrec that were left under a
2878 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2879 recursive instantiation of parameters: a parameter is a variable
2880 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2881 a function parameter. */
2884 instantiate_scev (basic_block instantiate_below
, struct loop
*evolution_loop
,
2889 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2891 fprintf (dump_file
, "(instantiate_scev \n");
2892 fprintf (dump_file
, " (instantiate_below = %d)\n", instantiate_below
->index
);
2893 fprintf (dump_file
, " (evolution_loop = %d)\n", evolution_loop
->num
);
2894 fprintf (dump_file
, " (chrec = ");
2895 print_generic_expr (dump_file
, chrec
, 0);
2896 fprintf (dump_file
, ")\n");
2902 global_cache
= new instantiate_cache_type
;
2906 res
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2907 NULL
, chrec
, NULL
, 0);
2911 delete global_cache
;
2912 global_cache
= NULL
;
2915 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2917 fprintf (dump_file
, " (res = ");
2918 print_generic_expr (dump_file
, res
, 0);
2919 fprintf (dump_file
, "))\n");
2925 /* Similar to instantiate_parameters, but does not introduce the
2926 evolutions in outer loops for LOOP invariants in CHREC, and does not
2927 care about causing overflows, as long as they do not affect value
2928 of an expression. */
2931 resolve_mixers (struct loop
*loop
, tree chrec
, bool *folded_casts
)
2934 bool fold_conversions
= false;
2937 global_cache
= new instantiate_cache_type
;
2941 tree ret
= instantiate_scev_r (block_before_loop (loop
), loop
, NULL
,
2942 chrec
, &fold_conversions
, 0);
2944 if (folded_casts
&& !*folded_casts
)
2945 *folded_casts
= fold_conversions
;
2949 delete global_cache
;
2950 global_cache
= NULL
;
2956 /* Entry point for the analysis of the number of iterations pass.
2957 This function tries to safely approximate the number of iterations
2958 the loop will run. When this property is not decidable at compile
2959 time, the result is chrec_dont_know. Otherwise the result is a
2960 scalar or a symbolic parameter. When the number of iterations may
2961 be equal to zero and the property cannot be determined at compile
2962 time, the result is a COND_EXPR that represents in a symbolic form
2963 the conditions under which the number of iterations is not zero.
2965 Example of analysis: suppose that the loop has an exit condition:
2967 "if (b > 49) goto end_loop;"
2969 and that in a previous analysis we have determined that the
2970 variable 'b' has an evolution function:
2972 "EF = {23, +, 5}_2".
2974 When we evaluate the function at the point 5, i.e. the value of the
2975 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2976 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2977 the loop body has been executed 6 times. */
2980 number_of_latch_executions (struct loop
*loop
)
2983 struct tree_niter_desc niter_desc
;
2987 /* Determine whether the number of iterations in loop has already
2989 res
= loop
->nb_iterations
;
2993 may_be_zero
= NULL_TREE
;
2995 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2996 fprintf (dump_file
, "(number_of_iterations_in_loop = \n");
2998 res
= chrec_dont_know
;
2999 exit
= single_exit (loop
);
3001 if (exit
&& number_of_iterations_exit (loop
, exit
, &niter_desc
, false))
3003 may_be_zero
= niter_desc
.may_be_zero
;
3004 res
= niter_desc
.niter
;
3007 if (res
== chrec_dont_know
3009 || integer_zerop (may_be_zero
))
3011 else if (integer_nonzerop (may_be_zero
))
3012 res
= build_int_cst (TREE_TYPE (res
), 0);
3014 else if (COMPARISON_CLASS_P (may_be_zero
))
3015 res
= fold_build3 (COND_EXPR
, TREE_TYPE (res
), may_be_zero
,
3016 build_int_cst (TREE_TYPE (res
), 0), res
);
3018 res
= chrec_dont_know
;
3020 if (dump_file
&& (dump_flags
& TDF_SCEV
))
3022 fprintf (dump_file
, " (set_nb_iterations_in_loop = ");
3023 print_generic_expr (dump_file
, res
, 0);
3024 fprintf (dump_file
, "))\n");
3027 loop
->nb_iterations
= res
;
3032 /* Counters for the stats. */
3038 unsigned nb_affine_multivar
;
3039 unsigned nb_higher_poly
;
3040 unsigned nb_chrec_dont_know
;
3041 unsigned nb_undetermined
;
3044 /* Reset the counters. */
3047 reset_chrecs_counters (struct chrec_stats
*stats
)
3049 stats
->nb_chrecs
= 0;
3050 stats
->nb_affine
= 0;
3051 stats
->nb_affine_multivar
= 0;
3052 stats
->nb_higher_poly
= 0;
3053 stats
->nb_chrec_dont_know
= 0;
3054 stats
->nb_undetermined
= 0;
3057 /* Dump the contents of a CHREC_STATS structure. */
3060 dump_chrecs_stats (FILE *file
, struct chrec_stats
*stats
)
3062 fprintf (file
, "\n(\n");
3063 fprintf (file
, "-----------------------------------------\n");
3064 fprintf (file
, "%d\taffine univariate chrecs\n", stats
->nb_affine
);
3065 fprintf (file
, "%d\taffine multivariate chrecs\n", stats
->nb_affine_multivar
);
3066 fprintf (file
, "%d\tdegree greater than 2 polynomials\n",
3067 stats
->nb_higher_poly
);
3068 fprintf (file
, "%d\tchrec_dont_know chrecs\n", stats
->nb_chrec_dont_know
);
3069 fprintf (file
, "-----------------------------------------\n");
3070 fprintf (file
, "%d\ttotal chrecs\n", stats
->nb_chrecs
);
3071 fprintf (file
, "%d\twith undetermined coefficients\n",
3072 stats
->nb_undetermined
);
3073 fprintf (file
, "-----------------------------------------\n");
3074 fprintf (file
, "%d\tchrecs in the scev database\n",
3075 (int) scalar_evolution_info
->elements ());
3076 fprintf (file
, "%d\tsets in the scev database\n", nb_set_scev
);
3077 fprintf (file
, "%d\tgets in the scev database\n", nb_get_scev
);
3078 fprintf (file
, "-----------------------------------------\n");
3079 fprintf (file
, ")\n\n");
3082 /* Gather statistics about CHREC. */
3085 gather_chrec_stats (tree chrec
, struct chrec_stats
*stats
)
3087 if (dump_file
&& (dump_flags
& TDF_STATS
))
3089 fprintf (dump_file
, "(classify_chrec ");
3090 print_generic_expr (dump_file
, chrec
, 0);
3091 fprintf (dump_file
, "\n");
3096 if (chrec
== NULL_TREE
)
3098 stats
->nb_undetermined
++;
3102 switch (TREE_CODE (chrec
))
3104 case POLYNOMIAL_CHREC
:
3105 if (evolution_function_is_affine_p (chrec
))
3107 if (dump_file
&& (dump_flags
& TDF_STATS
))
3108 fprintf (dump_file
, " affine_univariate\n");
3111 else if (evolution_function_is_affine_multivariate_p (chrec
, 0))
3113 if (dump_file
&& (dump_flags
& TDF_STATS
))
3114 fprintf (dump_file
, " affine_multivariate\n");
3115 stats
->nb_affine_multivar
++;
3119 if (dump_file
&& (dump_flags
& TDF_STATS
))
3120 fprintf (dump_file
, " higher_degree_polynomial\n");
3121 stats
->nb_higher_poly
++;
3130 if (chrec_contains_undetermined (chrec
))
3132 if (dump_file
&& (dump_flags
& TDF_STATS
))
3133 fprintf (dump_file
, " undetermined\n");
3134 stats
->nb_undetermined
++;
3137 if (dump_file
&& (dump_flags
& TDF_STATS
))
3138 fprintf (dump_file
, ")\n");
3141 /* Classify the chrecs of the whole database. */
3144 gather_stats_on_scev_database (void)
3146 struct chrec_stats stats
;
3151 reset_chrecs_counters (&stats
);
3153 hash_table
<scev_info_hasher
>::iterator iter
;
3155 FOR_EACH_HASH_TABLE_ELEMENT (*scalar_evolution_info
, elt
, scev_info_str
*,
3157 gather_chrec_stats (elt
->chrec
, &stats
);
3159 dump_chrecs_stats (dump_file
, &stats
);
3167 initialize_scalar_evolutions_analyzer (void)
3169 /* The elements below are unique. */
3170 if (chrec_dont_know
== NULL_TREE
)
3172 chrec_not_analyzed_yet
= NULL_TREE
;
3173 chrec_dont_know
= make_node (SCEV_NOT_KNOWN
);
3174 chrec_known
= make_node (SCEV_KNOWN
);
3175 TREE_TYPE (chrec_dont_know
) = void_type_node
;
3176 TREE_TYPE (chrec_known
) = void_type_node
;
3180 /* Initialize the analysis of scalar evolutions for LOOPS. */
3183 scev_initialize (void)
3187 scalar_evolution_info
= hash_table
<scev_info_hasher
>::create_ggc (100);
3189 initialize_scalar_evolutions_analyzer ();
3191 FOR_EACH_LOOP (loop
, 0)
3193 loop
->nb_iterations
= NULL_TREE
;
3197 /* Return true if SCEV is initialized. */
3200 scev_initialized_p (void)
3202 return scalar_evolution_info
!= NULL
;
3205 /* Cleans up the information cached by the scalar evolutions analysis
3206 in the hash table. */
3209 scev_reset_htab (void)
3211 if (!scalar_evolution_info
)
3214 scalar_evolution_info
->empty ();
3217 /* Cleans up the information cached by the scalar evolutions analysis
3218 in the hash table and in the loop->nb_iterations. */
3227 FOR_EACH_LOOP (loop
, 0)
3229 loop
->nb_iterations
= NULL_TREE
;
3233 /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3234 respect to WRTO_LOOP and returns its base and step in IV if possible
3235 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3236 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3237 invariant in LOOP. Otherwise we require it to be an integer constant.
3239 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3240 because it is computed in signed arithmetics). Consequently, adding an
3243 for (i = IV->base; ; i += IV->step)
3245 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3246 false for the type of the induction variable, or you can prove that i does
3247 not wrap by some other argument. Otherwise, this might introduce undefined
3250 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3252 must be used instead. */
3255 simple_iv (struct loop
*wrto_loop
, struct loop
*use_loop
, tree op
,
3256 affine_iv
*iv
, bool allow_nonconstant_step
)
3258 enum tree_code code
;
3259 tree type
, ev
, base
, e
, stop
;
3261 bool folded_casts
, overflow
;
3263 iv
->base
= NULL_TREE
;
3264 iv
->step
= NULL_TREE
;
3265 iv
->no_overflow
= false;
3267 type
= TREE_TYPE (op
);
3268 if (!POINTER_TYPE_P (type
)
3269 && !INTEGRAL_TYPE_P (type
))
3272 ev
= analyze_scalar_evolution_in_loop (wrto_loop
, use_loop
, op
,
3274 if (chrec_contains_undetermined (ev
)
3275 || chrec_contains_symbols_defined_in_loop (ev
, wrto_loop
->num
))
3278 if (tree_does_not_contain_chrecs (ev
))
3281 iv
->step
= build_int_cst (TREE_TYPE (ev
), 0);
3282 iv
->no_overflow
= true;
3286 if (TREE_CODE (ev
) != POLYNOMIAL_CHREC
3287 || CHREC_VARIABLE (ev
) != (unsigned) wrto_loop
->num
)
3290 iv
->step
= CHREC_RIGHT (ev
);
3291 if ((!allow_nonconstant_step
&& TREE_CODE (iv
->step
) != INTEGER_CST
)
3292 || tree_contains_chrecs (iv
->step
, NULL
))
3295 iv
->base
= CHREC_LEFT (ev
);
3296 if (tree_contains_chrecs (iv
->base
, NULL
))
3299 iv
->no_overflow
= (!folded_casts
&& ANY_INTEGRAL_TYPE_P (type
)
3300 && TYPE_OVERFLOW_UNDEFINED (type
));
3302 /* Try to simplify iv base:
3304 (signed T) ((unsigned T)base + step) ;; TREE_TYPE (base) == signed T
3305 == (signed T)(unsigned T)base + step
3308 If we can prove operation (base + step) doesn't overflow or underflow.
3309 Specifically, we try to prove below conditions are satisfied:
3311 base <= UPPER_BOUND (type) - step ;;step > 0
3312 base >= LOWER_BOUND (type) - step ;;step < 0
3314 This is done by proving the reverse conditions are false using loop's
3317 The is necessary to make loop niter, or iv overflow analysis easier
3320 int foo (int *a, signed char s, signed char l)
3323 for (i = s; i < l; i++)
3328 Note variable I is firstly converted to type unsigned char, incremented,
3329 then converted back to type signed char. */
3331 if (wrto_loop
->num
!= use_loop
->num
)
3334 if (!CONVERT_EXPR_P (iv
->base
) || TREE_CODE (iv
->step
) != INTEGER_CST
)
3337 type
= TREE_TYPE (iv
->base
);
3338 e
= TREE_OPERAND (iv
->base
, 0);
3339 if (TREE_CODE (e
) != PLUS_EXPR
3340 || TREE_CODE (TREE_OPERAND (e
, 1)) != INTEGER_CST
3341 || !tree_int_cst_equal (iv
->step
,
3342 fold_convert (type
, TREE_OPERAND (e
, 1))))
3344 e
= TREE_OPERAND (e
, 0);
3345 if (!CONVERT_EXPR_P (e
))
3347 base
= TREE_OPERAND (e
, 0);
3348 if (!useless_type_conversion_p (type
, TREE_TYPE (base
)))
3351 if (tree_int_cst_sign_bit (iv
->step
))
3354 extreme
= wi::min_value (type
);
3359 extreme
= wi::max_value (type
);
3362 extreme
= wi::sub (extreme
, iv
->step
, TYPE_SIGN (type
), &overflow
);
3365 e
= fold_build2 (code
, boolean_type_node
, base
,
3366 wide_int_to_tree (type
, extreme
));
3367 stop
= (TREE_CODE (base
) == SSA_NAME
) ? base
: NULL
;
3368 e
= simplify_using_initial_conditions (use_loop
, e
, stop
);
3369 if (!integer_zerop (e
))
3372 if (POINTER_TYPE_P (TREE_TYPE (base
)))
3373 code
= POINTER_PLUS_EXPR
;
3377 iv
->base
= fold_build2 (code
, TREE_TYPE (base
), base
, iv
->step
);
3381 /* Finalize the scalar evolution analysis. */
3384 scev_finalize (void)
3386 if (!scalar_evolution_info
)
3388 scalar_evolution_info
->empty ();
3389 scalar_evolution_info
= NULL
;
3392 /* Returns true if the expression EXPR is considered to be too expensive
3393 for scev_const_prop. */
3396 expression_expensive_p (tree expr
)
3398 enum tree_code code
;
3400 if (is_gimple_val (expr
))
3403 code
= TREE_CODE (expr
);
3404 if (code
== TRUNC_DIV_EXPR
3405 || code
== CEIL_DIV_EXPR
3406 || code
== FLOOR_DIV_EXPR
3407 || code
== ROUND_DIV_EXPR
3408 || code
== TRUNC_MOD_EXPR
3409 || code
== CEIL_MOD_EXPR
3410 || code
== FLOOR_MOD_EXPR
3411 || code
== ROUND_MOD_EXPR
3412 || code
== EXACT_DIV_EXPR
)
3414 /* Division by power of two is usually cheap, so we allow it.
3415 Forbid anything else. */
3416 if (!integer_pow2p (TREE_OPERAND (expr
, 1)))
3420 switch (TREE_CODE_CLASS (code
))
3423 case tcc_comparison
:
3424 if (expression_expensive_p (TREE_OPERAND (expr
, 1)))
3429 return expression_expensive_p (TREE_OPERAND (expr
, 0));
3436 /* Do final value replacement for LOOP. */
3439 final_value_replacement_loop (struct loop
*loop
)
3441 /* If we do not know exact number of iterations of the loop, we cannot
3442 replace the final value. */
3443 edge exit
= single_exit (loop
);
3447 tree niter
= number_of_latch_executions (loop
);
3448 if (niter
== chrec_dont_know
)
3451 /* Ensure that it is possible to insert new statements somewhere. */
3452 if (!single_pred_p (exit
->dest
))
3453 split_loop_exit_edge (exit
);
3455 /* Set stmt insertion pointer. All stmts are inserted before this point. */
3456 gimple_stmt_iterator gsi
= gsi_after_labels (exit
->dest
);
3458 struct loop
*ex_loop
3459 = superloop_at_depth (loop
,
3460 loop_depth (exit
->dest
->loop_father
) + 1);
3463 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); )
3465 gphi
*phi
= psi
.phi ();
3466 tree rslt
= PHI_RESULT (phi
);
3467 tree def
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
3468 if (virtual_operand_p (def
))
3474 if (!POINTER_TYPE_P (TREE_TYPE (def
))
3475 && !INTEGRAL_TYPE_P (TREE_TYPE (def
)))
3482 def
= analyze_scalar_evolution_in_loop (ex_loop
, loop
, def
,
3484 def
= compute_overall_effect_of_inner_loop (ex_loop
, def
);
3485 if (!tree_does_not_contain_chrecs (def
)
3486 || chrec_contains_symbols_defined_in_loop (def
, ex_loop
->num
)
3487 /* Moving the computation from the loop may prolong life range
3488 of some ssa names, which may cause problems if they appear
3489 on abnormal edges. */
3490 || contains_abnormal_ssa_name_p (def
)
3491 /* Do not emit expensive expressions. The rationale is that
3492 when someone writes a code like
3494 while (n > 45) n -= 45;
3496 he probably knows that n is not large, and does not want it
3497 to be turned into n %= 45. */
3498 || expression_expensive_p (def
))
3500 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3502 fprintf (dump_file
, "not replacing:\n ");
3503 print_gimple_stmt (dump_file
, phi
, 0, 0);
3504 fprintf (dump_file
, "\n");
3510 /* Eliminate the PHI node and replace it by a computation outside
3514 fprintf (dump_file
, "\nfinal value replacement:\n ");
3515 print_gimple_stmt (dump_file
, phi
, 0, 0);
3516 fprintf (dump_file
, " with\n ");
3518 def
= unshare_expr (def
);
3519 remove_phi_node (&psi
, false);
3521 /* If def's type has undefined overflow and there were folded
3522 casts, rewrite all stmts added for def into arithmetics
3523 with defined overflow behavior. */
3524 if (folded_casts
&& ANY_INTEGRAL_TYPE_P (TREE_TYPE (def
))
3525 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (def
)))
3528 gimple_stmt_iterator gsi2
;
3529 def
= force_gimple_operand (def
, &stmts
, true, NULL_TREE
);
3530 gsi2
= gsi_start (stmts
);
3531 while (!gsi_end_p (gsi2
))
3533 gimple
*stmt
= gsi_stmt (gsi2
);
3534 gimple_stmt_iterator gsi3
= gsi2
;
3536 gsi_remove (&gsi3
, false);
3537 if (is_gimple_assign (stmt
)
3538 && arith_code_with_undefined_signed_overflow
3539 (gimple_assign_rhs_code (stmt
)))
3540 gsi_insert_seq_before (&gsi
,
3541 rewrite_to_defined_overflow (stmt
),
3544 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
3548 def
= force_gimple_operand_gsi (&gsi
, def
, false, NULL_TREE
,
3549 true, GSI_SAME_STMT
);
3551 gassign
*ass
= gimple_build_assign (rslt
, def
);
3552 gsi_insert_before (&gsi
, ass
, GSI_SAME_STMT
);
3555 print_gimple_stmt (dump_file
, ass
, 0, 0);
3556 fprintf (dump_file
, "\n");
3561 /* Replace ssa names for that scev can prove they are constant by the
3562 appropriate constants. Also perform final value replacement in loops,
3563 in case the replacement expressions are cheap.
3565 We only consider SSA names defined by phi nodes; rest is left to the
3566 ordinary constant propagation pass. */
3569 scev_const_prop (void)
3572 tree name
, type
, ev
;
3575 bitmap ssa_names_to_remove
= NULL
;
3579 if (number_of_loops (cfun
) <= 1)
3582 FOR_EACH_BB_FN (bb
, cfun
)
3584 loop
= bb
->loop_father
;
3586 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); gsi_next (&psi
))
3589 name
= PHI_RESULT (phi
);
3591 if (virtual_operand_p (name
))
3594 type
= TREE_TYPE (name
);
3596 if (!POINTER_TYPE_P (type
)
3597 && !INTEGRAL_TYPE_P (type
))
3600 ev
= resolve_mixers (loop
, analyze_scalar_evolution (loop
, name
),
3602 if (!is_gimple_min_invariant (ev
)
3603 || !may_propagate_copy (name
, ev
))
3606 /* Replace the uses of the name. */
3609 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3611 fprintf (dump_file
, "Replacing uses of: ");
3612 print_generic_expr (dump_file
, name
, 0);
3613 fprintf (dump_file
, " with: ");
3614 print_generic_expr (dump_file
, ev
, 0);
3615 fprintf (dump_file
, "\n");
3617 replace_uses_by (name
, ev
);
3620 if (!ssa_names_to_remove
)
3621 ssa_names_to_remove
= BITMAP_ALLOC (NULL
);
3622 bitmap_set_bit (ssa_names_to_remove
, SSA_NAME_VERSION (name
));
3626 /* Remove the ssa names that were replaced by constants. We do not
3627 remove them directly in the previous cycle, since this
3628 invalidates scev cache. */
3629 if (ssa_names_to_remove
)
3633 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove
, 0, i
, bi
)
3635 gimple_stmt_iterator psi
;
3636 name
= ssa_name (i
);
3637 phi
= as_a
<gphi
*> (SSA_NAME_DEF_STMT (name
));
3639 gcc_assert (gimple_code (phi
) == GIMPLE_PHI
);
3640 psi
= gsi_for_stmt (phi
);
3641 remove_phi_node (&psi
, true);
3644 BITMAP_FREE (ssa_names_to_remove
);
3648 /* Now the regular final value replacement. */
3649 FOR_EACH_LOOP (loop
, LI_FROM_INNERMOST
)
3650 final_value_replacement_loop (loop
);
3655 #include "gt-tree-scalar-evolution.h"