1 /* Scalar evolution detector.
2 Copyright (C) 2003-2015 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 /* Given a loop-phi-node, return the initial conditions of the
1526 variable on entry of the loop. When the CCP has propagated
1527 constants into the loop-phi-node, the initial condition is
1528 instantiated, otherwise the initial condition is kept symbolic.
1529 This analyzer does not analyze the evolution outside the current
1530 loop, and leaves this task to the on-demand tree reconstructor. */
1533 analyze_initial_condition (gphi
*loop_phi_node
)
1536 tree init_cond
= chrec_not_analyzed_yet
;
1537 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1539 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1541 fprintf (dump_file
, "(analyze_initial_condition \n");
1542 fprintf (dump_file
, " (loop_phi_node = \n");
1543 print_gimple_stmt (dump_file
, loop_phi_node
, 0, 0);
1544 fprintf (dump_file
, ")\n");
1547 n
= gimple_phi_num_args (loop_phi_node
);
1548 for (i
= 0; i
< n
; i
++)
1550 tree branch
= PHI_ARG_DEF (loop_phi_node
, i
);
1551 basic_block bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1553 /* When the branch is oriented to the loop's body, it does
1554 not contribute to the initial condition. */
1555 if (flow_bb_inside_loop_p (loop
, bb
))
1558 if (init_cond
== chrec_not_analyzed_yet
)
1564 if (TREE_CODE (branch
) == SSA_NAME
)
1566 init_cond
= chrec_dont_know
;
1570 init_cond
= chrec_merge (init_cond
, branch
);
1573 /* Ooops -- a loop without an entry??? */
1574 if (init_cond
== chrec_not_analyzed_yet
)
1575 init_cond
= chrec_dont_know
;
1577 /* During early loop unrolling we do not have fully constant propagated IL.
1578 Handle degenerate PHIs here to not miss important unrollings. */
1579 if (TREE_CODE (init_cond
) == SSA_NAME
)
1581 gimple
*def
= SSA_NAME_DEF_STMT (init_cond
);
1582 if (gphi
*phi
= dyn_cast
<gphi
*> (def
))
1584 tree res
= degenerate_phi_result (phi
);
1585 if (res
!= NULL_TREE
1586 /* Only allow invariants here, otherwise we may break
1587 loop-closed SSA form. */
1588 && is_gimple_min_invariant (res
))
1593 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1595 fprintf (dump_file
, " (init_cond = ");
1596 print_generic_expr (dump_file
, init_cond
, 0);
1597 fprintf (dump_file
, "))\n");
1603 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1606 interpret_loop_phi (struct loop
*loop
, gphi
*loop_phi_node
)
1609 struct loop
*phi_loop
= loop_containing_stmt (loop_phi_node
);
1612 if (phi_loop
!= loop
)
1614 struct loop
*subloop
;
1615 tree evolution_fn
= analyze_scalar_evolution
1616 (phi_loop
, PHI_RESULT (loop_phi_node
));
1618 /* Dive one level deeper. */
1619 subloop
= superloop_at_depth (phi_loop
, loop_depth (loop
) + 1);
1621 /* Interpret the subloop. */
1622 res
= compute_overall_effect_of_inner_loop (subloop
, evolution_fn
);
1626 /* Otherwise really interpret the loop phi. */
1627 init_cond
= analyze_initial_condition (loop_phi_node
);
1628 res
= analyze_evolution_in_loop (loop_phi_node
, init_cond
);
1630 /* Verify we maintained the correct initial condition throughout
1631 possible conversions in the SSA chain. */
1632 if (res
!= chrec_dont_know
)
1634 tree new_init
= res
;
1635 if (CONVERT_EXPR_P (res
)
1636 && TREE_CODE (TREE_OPERAND (res
, 0)) == POLYNOMIAL_CHREC
)
1637 new_init
= fold_convert (TREE_TYPE (res
),
1638 CHREC_LEFT (TREE_OPERAND (res
, 0)));
1639 else if (TREE_CODE (res
) == POLYNOMIAL_CHREC
)
1640 new_init
= CHREC_LEFT (res
);
1641 STRIP_USELESS_TYPE_CONVERSION (new_init
);
1642 if (TREE_CODE (new_init
) == POLYNOMIAL_CHREC
1643 || !operand_equal_p (init_cond
, new_init
, 0))
1644 return chrec_dont_know
;
1650 /* This function merges the branches of a condition-phi-node,
1651 contained in the outermost loop, and whose arguments are already
1655 interpret_condition_phi (struct loop
*loop
, gphi
*condition_phi
)
1657 int i
, n
= gimple_phi_num_args (condition_phi
);
1658 tree res
= chrec_not_analyzed_yet
;
1660 for (i
= 0; i
< n
; i
++)
1664 if (backedge_phi_arg_p (condition_phi
, i
))
1666 res
= chrec_dont_know
;
1670 branch_chrec
= analyze_scalar_evolution
1671 (loop
, PHI_ARG_DEF (condition_phi
, i
));
1673 res
= chrec_merge (res
, branch_chrec
);
1679 /* Interpret the operation RHS1 OP RHS2. If we didn't
1680 analyze this node before, follow the definitions until ending
1681 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1682 return path, this function propagates evolutions (ala constant copy
1683 propagation). OPND1 is not a GIMPLE expression because we could
1684 analyze the effect of an inner loop: see interpret_loop_phi. */
1687 interpret_rhs_expr (struct loop
*loop
, gimple
*at_stmt
,
1688 tree type
, tree rhs1
, enum tree_code code
, tree rhs2
)
1690 tree res
, chrec1
, chrec2
;
1693 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1695 if (is_gimple_min_invariant (rhs1
))
1696 return chrec_convert (type
, rhs1
, at_stmt
);
1698 if (code
== SSA_NAME
)
1699 return chrec_convert (type
, analyze_scalar_evolution (loop
, rhs1
),
1702 if (code
== ASSERT_EXPR
)
1704 rhs1
= ASSERT_EXPR_VAR (rhs1
);
1705 return chrec_convert (type
, analyze_scalar_evolution (loop
, rhs1
),
1713 if (TREE_CODE (TREE_OPERAND (rhs1
, 0)) == MEM_REF
1714 || handled_component_p (TREE_OPERAND (rhs1
, 0)))
1717 HOST_WIDE_INT bitsize
, bitpos
;
1718 int unsignedp
, reversep
;
1724 base
= get_inner_reference (TREE_OPERAND (rhs1
, 0),
1725 &bitsize
, &bitpos
, &offset
, &mode
,
1726 &unsignedp
, &reversep
, &volatilep
,
1729 if (TREE_CODE (base
) == MEM_REF
)
1731 rhs2
= TREE_OPERAND (base
, 1);
1732 rhs1
= TREE_OPERAND (base
, 0);
1734 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1735 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1736 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1737 chrec2
= chrec_convert (TREE_TYPE (rhs2
), chrec2
, at_stmt
);
1738 chrec1
= instantiate_parameters (loop
, chrec1
);
1739 chrec2
= instantiate_parameters (loop
, chrec2
);
1740 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1744 chrec1
= analyze_scalar_evolution_for_address_of (loop
, base
);
1745 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1749 if (offset
!= NULL_TREE
)
1751 chrec2
= analyze_scalar_evolution (loop
, offset
);
1752 chrec2
= chrec_convert (TREE_TYPE (offset
), chrec2
, at_stmt
);
1753 chrec2
= instantiate_parameters (loop
, chrec2
);
1754 res
= chrec_fold_plus (type
, res
, chrec2
);
1759 gcc_assert ((bitpos
% BITS_PER_UNIT
) == 0);
1761 unitpos
= size_int (bitpos
/ BITS_PER_UNIT
);
1762 chrec3
= analyze_scalar_evolution (loop
, unitpos
);
1763 chrec3
= chrec_convert (TREE_TYPE (unitpos
), chrec3
, at_stmt
);
1764 chrec3
= instantiate_parameters (loop
, chrec3
);
1765 res
= chrec_fold_plus (type
, res
, chrec3
);
1769 res
= chrec_dont_know
;
1772 case POINTER_PLUS_EXPR
:
1773 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1774 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1775 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1776 chrec2
= chrec_convert (TREE_TYPE (rhs2
), chrec2
, at_stmt
);
1777 chrec1
= instantiate_parameters (loop
, chrec1
);
1778 chrec2
= instantiate_parameters (loop
, chrec2
);
1779 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1783 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1784 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1785 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1786 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1787 chrec1
= instantiate_parameters (loop
, chrec1
);
1788 chrec2
= instantiate_parameters (loop
, chrec2
);
1789 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1793 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1794 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1795 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1796 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1797 chrec1
= instantiate_parameters (loop
, chrec1
);
1798 chrec2
= instantiate_parameters (loop
, chrec2
);
1799 res
= chrec_fold_minus (type
, chrec1
, chrec2
);
1803 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1804 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1805 /* TYPE may be integer, real or complex, so use fold_convert. */
1806 chrec1
= instantiate_parameters (loop
, chrec1
);
1807 res
= chrec_fold_multiply (type
, chrec1
,
1808 fold_convert (type
, integer_minus_one_node
));
1812 /* Handle ~X as -1 - X. */
1813 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1814 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1815 chrec1
= instantiate_parameters (loop
, chrec1
);
1816 res
= chrec_fold_minus (type
,
1817 fold_convert (type
, integer_minus_one_node
),
1822 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1823 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1824 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1825 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1826 chrec1
= instantiate_parameters (loop
, chrec1
);
1827 chrec2
= instantiate_parameters (loop
, chrec2
);
1828 res
= chrec_fold_multiply (type
, chrec1
, chrec2
);
1833 /* Handle A<<B as A * (1<<B). */
1834 tree uns
= unsigned_type_for (type
);
1835 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1836 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1837 chrec1
= chrec_convert (uns
, chrec1
, at_stmt
);
1838 chrec1
= instantiate_parameters (loop
, chrec1
);
1839 chrec2
= instantiate_parameters (loop
, chrec2
);
1841 tree one
= build_int_cst (uns
, 1);
1842 chrec2
= fold_build2 (LSHIFT_EXPR
, uns
, one
, chrec2
);
1843 res
= chrec_fold_multiply (uns
, chrec1
, chrec2
);
1844 res
= chrec_convert (type
, res
, at_stmt
);
1849 /* In case we have a truncation of a widened operation that in
1850 the truncated type has undefined overflow behavior analyze
1851 the operation done in an unsigned type of the same precision
1852 as the final truncation. We cannot derive a scalar evolution
1853 for the widened operation but for the truncated result. */
1854 if (TREE_CODE (type
) == INTEGER_TYPE
1855 && TREE_CODE (TREE_TYPE (rhs1
)) == INTEGER_TYPE
1856 && TYPE_PRECISION (type
) < TYPE_PRECISION (TREE_TYPE (rhs1
))
1857 && TYPE_OVERFLOW_UNDEFINED (type
)
1858 && TREE_CODE (rhs1
) == SSA_NAME
1859 && (def
= SSA_NAME_DEF_STMT (rhs1
))
1860 && is_gimple_assign (def
)
1861 && TREE_CODE_CLASS (gimple_assign_rhs_code (def
)) == tcc_binary
1862 && TREE_CODE (gimple_assign_rhs2 (def
)) == INTEGER_CST
)
1864 tree utype
= unsigned_type_for (type
);
1865 chrec1
= interpret_rhs_expr (loop
, at_stmt
, utype
,
1866 gimple_assign_rhs1 (def
),
1867 gimple_assign_rhs_code (def
),
1868 gimple_assign_rhs2 (def
));
1871 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1872 res
= chrec_convert (type
, chrec1
, at_stmt
);
1876 res
= chrec_dont_know
;
1883 /* Interpret the expression EXPR. */
1886 interpret_expr (struct loop
*loop
, gimple
*at_stmt
, tree expr
)
1888 enum tree_code code
;
1889 tree type
= TREE_TYPE (expr
), op0
, op1
;
1891 if (automatically_generated_chrec_p (expr
))
1894 if (TREE_CODE (expr
) == POLYNOMIAL_CHREC
1895 || get_gimple_rhs_class (TREE_CODE (expr
)) == GIMPLE_TERNARY_RHS
)
1896 return chrec_dont_know
;
1898 extract_ops_from_tree (expr
, &code
, &op0
, &op1
);
1900 return interpret_rhs_expr (loop
, at_stmt
, type
,
1904 /* Interpret the rhs of the assignment STMT. */
1907 interpret_gimple_assign (struct loop
*loop
, gimple
*stmt
)
1909 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
1910 enum tree_code code
= gimple_assign_rhs_code (stmt
);
1912 return interpret_rhs_expr (loop
, stmt
, type
,
1913 gimple_assign_rhs1 (stmt
), code
,
1914 gimple_assign_rhs2 (stmt
));
1919 /* This section contains all the entry points:
1920 - number_of_iterations_in_loop,
1921 - analyze_scalar_evolution,
1922 - instantiate_parameters.
1925 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1926 common ancestor of DEF_LOOP and USE_LOOP. */
1929 compute_scalar_evolution_in_loop (struct loop
*wrto_loop
,
1930 struct loop
*def_loop
,
1936 if (def_loop
== wrto_loop
)
1939 def_loop
= superloop_at_depth (def_loop
, loop_depth (wrto_loop
) + 1);
1940 res
= compute_overall_effect_of_inner_loop (def_loop
, ev
);
1942 if (no_evolution_in_loop_p (res
, wrto_loop
->num
, &val
) && val
)
1945 return analyze_scalar_evolution_1 (wrto_loop
, res
, chrec_not_analyzed_yet
);
1948 /* Helper recursive function. */
1951 analyze_scalar_evolution_1 (struct loop
*loop
, tree var
, tree res
)
1953 tree type
= TREE_TYPE (var
);
1956 struct loop
*def_loop
;
1958 if (loop
== NULL
|| TREE_CODE (type
) == VECTOR_TYPE
)
1959 return chrec_dont_know
;
1961 if (TREE_CODE (var
) != SSA_NAME
)
1962 return interpret_expr (loop
, NULL
, var
);
1964 def
= SSA_NAME_DEF_STMT (var
);
1965 bb
= gimple_bb (def
);
1966 def_loop
= bb
? bb
->loop_father
: NULL
;
1969 || !flow_bb_inside_loop_p (loop
, bb
))
1971 /* Keep the symbolic form. */
1976 if (res
!= chrec_not_analyzed_yet
)
1978 if (loop
!= bb
->loop_father
)
1979 res
= compute_scalar_evolution_in_loop
1980 (find_common_loop (loop
, bb
->loop_father
), bb
->loop_father
, res
);
1985 if (loop
!= def_loop
)
1987 res
= analyze_scalar_evolution_1 (def_loop
, var
, chrec_not_analyzed_yet
);
1988 res
= compute_scalar_evolution_in_loop (loop
, def_loop
, res
);
1993 switch (gimple_code (def
))
1996 res
= interpret_gimple_assign (loop
, def
);
2000 if (loop_phi_node_p (def
))
2001 res
= interpret_loop_phi (loop
, as_a
<gphi
*> (def
));
2003 res
= interpret_condition_phi (loop
, as_a
<gphi
*> (def
));
2007 res
= chrec_dont_know
;
2013 /* Keep the symbolic form. */
2014 if (res
== chrec_dont_know
)
2017 if (loop
== def_loop
)
2018 set_scalar_evolution (block_before_loop (loop
), var
, res
);
2023 /* Analyzes and returns the scalar evolution of the ssa_name VAR in
2024 LOOP. LOOP is the loop in which the variable is used.
2026 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
2027 pointer to the statement that uses this variable, in order to
2028 determine the evolution function of the variable, use the following
2031 loop_p loop = loop_containing_stmt (stmt);
2032 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
2033 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
2037 analyze_scalar_evolution (struct loop
*loop
, tree var
)
2041 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2043 fprintf (dump_file
, "(analyze_scalar_evolution \n");
2044 fprintf (dump_file
, " (loop_nb = %d)\n", loop
->num
);
2045 fprintf (dump_file
, " (scalar = ");
2046 print_generic_expr (dump_file
, var
, 0);
2047 fprintf (dump_file
, ")\n");
2050 res
= get_scalar_evolution (block_before_loop (loop
), var
);
2051 res
= analyze_scalar_evolution_1 (loop
, var
, res
);
2053 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2054 fprintf (dump_file
, ")\n");
2059 /* Analyzes and returns the scalar evolution of VAR address in LOOP. */
2062 analyze_scalar_evolution_for_address_of (struct loop
*loop
, tree var
)
2064 return analyze_scalar_evolution (loop
, build_fold_addr_expr (var
));
2067 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
2068 WRTO_LOOP (which should be a superloop of USE_LOOP)
2070 FOLDED_CASTS is set to true if resolve_mixers used
2071 chrec_convert_aggressive (TODO -- not really, we are way too conservative
2072 at the moment in order to keep things simple).
2074 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
2077 for (i = 0; i < 100; i++) -- loop 1
2079 for (j = 0; j < 100; j++) -- loop 2
2086 for (t = 0; t < 100; t++) -- loop 3
2093 Both k1 and k2 are invariants in loop3, thus
2094 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
2095 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
2097 As they are invariant, it does not matter whether we consider their
2098 usage in loop 3 or loop 2, hence
2099 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
2100 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
2101 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
2102 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
2104 Similarly for their evolutions with respect to loop 1. The values of K2
2105 in the use in loop 2 vary independently on loop 1, thus we cannot express
2106 the evolution with respect to loop 1:
2107 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2108 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2109 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2110 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2112 The value of k2 in the use in loop 1 is known, though:
2113 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2114 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2118 analyze_scalar_evolution_in_loop (struct loop
*wrto_loop
, struct loop
*use_loop
,
2119 tree version
, bool *folded_casts
)
2122 tree ev
= version
, tmp
;
2124 /* We cannot just do
2126 tmp = analyze_scalar_evolution (use_loop, version);
2127 ev = resolve_mixers (wrto_loop, tmp, folded_casts);
2129 as resolve_mixers would query the scalar evolution with respect to
2130 wrto_loop. For example, in the situation described in the function
2131 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2134 analyze_scalar_evolution (use_loop, version) = k2
2136 and resolve_mixers (loop1, k2, folded_casts) finds that the value of
2137 k2 in loop 1 is 100, which is a wrong result, since we are interested
2138 in the value in loop 3.
2140 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2141 each time checking that there is no evolution in the inner loop. */
2144 *folded_casts
= false;
2147 tmp
= analyze_scalar_evolution (use_loop
, ev
);
2148 ev
= resolve_mixers (use_loop
, tmp
, folded_casts
);
2150 if (use_loop
== wrto_loop
)
2153 /* If the value of the use changes in the inner loop, we cannot express
2154 its value in the outer loop (we might try to return interval chrec,
2155 but we do not have a user for it anyway) */
2156 if (!no_evolution_in_loop_p (ev
, use_loop
->num
, &val
)
2158 return chrec_dont_know
;
2160 use_loop
= loop_outer (use_loop
);
2165 /* Hashtable helpers for a temporary hash-table used when
2166 instantiating a CHREC or resolving mixers. For this use
2167 instantiated_below is always the same. */
2169 struct instantiate_cache_type
2172 vec
<scev_info_str
> entries
;
2174 instantiate_cache_type () : map (NULL
), entries (vNULL
) {}
2175 ~instantiate_cache_type ();
2176 tree
get (unsigned slot
) { return entries
[slot
].chrec
; }
2177 void set (unsigned slot
, tree chrec
) { entries
[slot
].chrec
= chrec
; }
2180 instantiate_cache_type::~instantiate_cache_type ()
2189 /* Cache to avoid infinite recursion when instantiating an SSA name.
2190 Live during the outermost instantiate_scev or resolve_mixers call. */
2191 static instantiate_cache_type
*global_cache
;
2193 /* Computes a hash function for database element ELT. */
2195 static inline hashval_t
2196 hash_idx_scev_info (const void *elt_
)
2198 unsigned idx
= ((size_t) elt_
) - 2;
2199 return scev_info_hasher::hash (&global_cache
->entries
[idx
]);
2202 /* Compares database elements E1 and E2. */
2205 eq_idx_scev_info (const void *e1
, const void *e2
)
2207 unsigned idx1
= ((size_t) e1
) - 2;
2208 return scev_info_hasher::equal (&global_cache
->entries
[idx1
],
2209 (const scev_info_str
*) e2
);
2212 /* Returns from CACHE the slot number of the cached chrec for NAME. */
2215 get_instantiated_value_entry (instantiate_cache_type
&cache
,
2216 tree name
, basic_block instantiate_below
)
2220 cache
.map
= htab_create (10, hash_idx_scev_info
, eq_idx_scev_info
, NULL
);
2221 cache
.entries
.create (10);
2225 e
.name_version
= SSA_NAME_VERSION (name
);
2226 e
.instantiated_below
= instantiate_below
->index
;
2227 void **slot
= htab_find_slot_with_hash (cache
.map
, &e
,
2228 scev_info_hasher::hash (&e
), INSERT
);
2231 e
.chrec
= chrec_not_analyzed_yet
;
2232 *slot
= (void *)(size_t)(cache
.entries
.length () + 2);
2233 cache
.entries
.safe_push (e
);
2236 return ((size_t)*slot
) - 2;
2240 /* Return the closed_loop_phi node for VAR. If there is none, return
2244 loop_closed_phi_def (tree var
)
2251 if (var
== NULL_TREE
2252 || TREE_CODE (var
) != SSA_NAME
)
2255 loop
= loop_containing_stmt (SSA_NAME_DEF_STMT (var
));
2256 exit
= single_exit (loop
);
2260 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); gsi_next (&psi
))
2263 if (PHI_ARG_DEF_FROM_EDGE (phi
, exit
) == var
)
2264 return PHI_RESULT (phi
);
2270 static tree
instantiate_scev_r (basic_block
, struct loop
*, struct loop
*,
2273 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2274 and EVOLUTION_LOOP, that were left under a symbolic form.
2276 CHREC is an SSA_NAME to be instantiated.
2278 CACHE is the cache of already instantiated values.
2280 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2281 conversions that may wrap in signed/pointer type are folded, as long
2282 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2283 then we don't do such fold.
2285 SIZE_EXPR is used for computing the size of the expression to be
2286 instantiated, and to stop if it exceeds some limit. */
2289 instantiate_scev_name (basic_block instantiate_below
,
2290 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2292 bool *fold_conversions
,
2296 struct loop
*def_loop
;
2297 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (chrec
));
2299 /* A parameter (or loop invariant and we do not want to include
2300 evolutions in outer loops), nothing to do. */
2302 || loop_depth (def_bb
->loop_father
) == 0
2303 || dominated_by_p (CDI_DOMINATORS
, instantiate_below
, def_bb
))
2306 /* We cache the value of instantiated variable to avoid exponential
2307 time complexity due to reevaluations. We also store the convenient
2308 value in the cache in order to prevent infinite recursion -- we do
2309 not want to instantiate the SSA_NAME if it is in a mixer
2310 structure. This is used for avoiding the instantiation of
2311 recursively defined functions, such as:
2313 | a_2 -> {0, +, 1, +, a_2}_1 */
2315 unsigned si
= get_instantiated_value_entry (*global_cache
,
2316 chrec
, instantiate_below
);
2317 if (global_cache
->get (si
) != chrec_not_analyzed_yet
)
2318 return global_cache
->get (si
);
2320 /* On recursion return chrec_dont_know. */
2321 global_cache
->set (si
, chrec_dont_know
);
2323 def_loop
= find_common_loop (evolution_loop
, def_bb
->loop_father
);
2325 /* If the analysis yields a parametric chrec, instantiate the
2327 res
= analyze_scalar_evolution (def_loop
, chrec
);
2329 /* Don't instantiate default definitions. */
2330 if (TREE_CODE (res
) == SSA_NAME
2331 && SSA_NAME_IS_DEFAULT_DEF (res
))
2334 /* Don't instantiate loop-closed-ssa phi nodes. */
2335 else if (TREE_CODE (res
) == SSA_NAME
2336 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res
)))
2337 > loop_depth (def_loop
))
2340 res
= loop_closed_phi_def (chrec
);
2344 /* When there is no loop_closed_phi_def, it means that the
2345 variable is not used after the loop: try to still compute the
2346 value of the variable when exiting the loop. */
2347 if (res
== NULL_TREE
)
2349 loop_p loop
= loop_containing_stmt (SSA_NAME_DEF_STMT (chrec
));
2350 res
= analyze_scalar_evolution (loop
, chrec
);
2351 res
= compute_overall_effect_of_inner_loop (loop
, res
);
2352 res
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2354 fold_conversions
, size_expr
);
2356 else if (!dominated_by_p (CDI_DOMINATORS
, instantiate_below
,
2357 gimple_bb (SSA_NAME_DEF_STMT (res
))))
2358 res
= chrec_dont_know
;
2361 else if (res
!= chrec_dont_know
)
2364 && def_bb
->loop_father
!= inner_loop
2365 && !flow_loop_nested_p (def_bb
->loop_father
, inner_loop
))
2366 /* ??? We could try to compute the overall effect of the loop here. */
2367 res
= chrec_dont_know
;
2369 res
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2371 fold_conversions
, size_expr
);
2374 /* Store the correct value to the cache. */
2375 global_cache
->set (si
, res
);
2379 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2380 and EVOLUTION_LOOP, that were left under a symbolic form.
2382 CHREC is a polynomial chain of recurrence to be instantiated.
2384 CACHE is the cache of already instantiated values.
2386 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2387 conversions that may wrap in signed/pointer type are folded, as long
2388 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2389 then we don't do such fold.
2391 SIZE_EXPR is used for computing the size of the expression to be
2392 instantiated, and to stop if it exceeds some limit. */
2395 instantiate_scev_poly (basic_block instantiate_below
,
2396 struct loop
*evolution_loop
, struct loop
*,
2397 tree chrec
, bool *fold_conversions
, int size_expr
)
2400 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2401 get_chrec_loop (chrec
),
2402 CHREC_LEFT (chrec
), fold_conversions
,
2404 if (op0
== chrec_dont_know
)
2405 return chrec_dont_know
;
2407 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2408 get_chrec_loop (chrec
),
2409 CHREC_RIGHT (chrec
), fold_conversions
,
2411 if (op1
== chrec_dont_know
)
2412 return chrec_dont_know
;
2414 if (CHREC_LEFT (chrec
) != op0
2415 || CHREC_RIGHT (chrec
) != op1
)
2417 op1
= chrec_convert_rhs (chrec_type (op0
), op1
, NULL
);
2418 chrec
= build_polynomial_chrec (CHREC_VARIABLE (chrec
), op0
, op1
);
2424 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2425 and EVOLUTION_LOOP, that were left under a symbolic form.
2427 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
2429 CACHE is the cache of already instantiated values.
2431 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2432 conversions that may wrap in signed/pointer type are folded, as long
2433 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2434 then we don't do such fold.
2436 SIZE_EXPR is used for computing the size of the expression to be
2437 instantiated, and to stop if it exceeds some limit. */
2440 instantiate_scev_binary (basic_block instantiate_below
,
2441 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2442 tree chrec
, enum tree_code code
,
2443 tree type
, tree c0
, tree c1
,
2444 bool *fold_conversions
, int size_expr
)
2447 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
, inner_loop
,
2448 c0
, fold_conversions
, size_expr
);
2449 if (op0
== chrec_dont_know
)
2450 return chrec_dont_know
;
2452 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
, inner_loop
,
2453 c1
, fold_conversions
, size_expr
);
2454 if (op1
== chrec_dont_know
)
2455 return chrec_dont_know
;
2460 op0
= chrec_convert (type
, op0
, NULL
);
2461 op1
= chrec_convert_rhs (type
, op1
, NULL
);
2465 case POINTER_PLUS_EXPR
:
2467 return chrec_fold_plus (type
, op0
, op1
);
2470 return chrec_fold_minus (type
, op0
, op1
);
2473 return chrec_fold_multiply (type
, op0
, op1
);
2480 return chrec
? chrec
: fold_build2 (code
, type
, c0
, c1
);
2483 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2484 and EVOLUTION_LOOP, that were left under a symbolic form.
2486 "CHREC" is an array reference to be instantiated.
2488 CACHE is the cache of already instantiated values.
2490 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2491 conversions that may wrap in signed/pointer type are folded, as long
2492 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2493 then we don't do such fold.
2495 SIZE_EXPR is used for computing the size of the expression to be
2496 instantiated, and to stop if it exceeds some limit. */
2499 instantiate_array_ref (basic_block instantiate_below
,
2500 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2501 tree chrec
, bool *fold_conversions
, int size_expr
)
2504 tree index
= TREE_OPERAND (chrec
, 1);
2505 tree op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2507 fold_conversions
, size_expr
);
2509 if (op1
== chrec_dont_know
)
2510 return chrec_dont_know
;
2512 if (chrec
&& op1
== index
)
2515 res
= unshare_expr (chrec
);
2516 TREE_OPERAND (res
, 1) = op1
;
2520 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2521 and EVOLUTION_LOOP, that were left under a symbolic form.
2523 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2526 CACHE is the cache of already instantiated values.
2528 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2529 conversions that may wrap in signed/pointer type are folded, as long
2530 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2531 then we don't do such fold.
2533 SIZE_EXPR is used for computing the size of the expression to be
2534 instantiated, and to stop if it exceeds some limit. */
2537 instantiate_scev_convert (basic_block instantiate_below
,
2538 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2539 tree chrec
, tree type
, tree op
,
2540 bool *fold_conversions
, int size_expr
)
2542 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2544 fold_conversions
, size_expr
);
2546 if (op0
== chrec_dont_know
)
2547 return chrec_dont_know
;
2549 if (fold_conversions
)
2551 tree tmp
= chrec_convert_aggressive (type
, op0
, fold_conversions
);
2555 /* If we used chrec_convert_aggressive, we can no longer assume that
2556 signed chrecs do not overflow, as chrec_convert does, so avoid
2557 calling it in that case. */
2558 if (*fold_conversions
)
2560 if (chrec
&& op0
== op
)
2563 return fold_convert (type
, op0
);
2567 return chrec_convert (type
, op0
, NULL
);
2570 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2571 and EVOLUTION_LOOP, that were left under a symbolic form.
2573 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
2574 Handle ~X as -1 - X.
2575 Handle -X as -1 * X.
2577 CACHE is the cache of already instantiated values.
2579 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2580 conversions that may wrap in signed/pointer type are folded, as long
2581 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2582 then we don't do such fold.
2584 SIZE_EXPR is used for computing the size of the expression to be
2585 instantiated, and to stop if it exceeds some limit. */
2588 instantiate_scev_not (basic_block instantiate_below
,
2589 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2591 enum tree_code code
, tree type
, tree op
,
2592 bool *fold_conversions
, int size_expr
)
2594 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2596 fold_conversions
, size_expr
);
2598 if (op0
== chrec_dont_know
)
2599 return chrec_dont_know
;
2603 op0
= chrec_convert (type
, op0
, NULL
);
2608 return chrec_fold_minus
2609 (type
, fold_convert (type
, integer_minus_one_node
), op0
);
2612 return chrec_fold_multiply
2613 (type
, fold_convert (type
, integer_minus_one_node
), op0
);
2620 return chrec
? chrec
: fold_build1 (code
, type
, op0
);
2623 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2624 and EVOLUTION_LOOP, that were left under a symbolic form.
2626 CHREC is an expression with 3 operands to be instantiated.
2628 CACHE is the cache of already instantiated values.
2630 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2631 conversions that may wrap in signed/pointer type are folded, as long
2632 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2633 then we don't do such fold.
2635 SIZE_EXPR is used for computing the size of the expression to be
2636 instantiated, and to stop if it exceeds some limit. */
2639 instantiate_scev_3 (basic_block instantiate_below
,
2640 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2642 bool *fold_conversions
, int size_expr
)
2645 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2646 inner_loop
, TREE_OPERAND (chrec
, 0),
2647 fold_conversions
, size_expr
);
2648 if (op0
== chrec_dont_know
)
2649 return chrec_dont_know
;
2651 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2652 inner_loop
, TREE_OPERAND (chrec
, 1),
2653 fold_conversions
, size_expr
);
2654 if (op1
== chrec_dont_know
)
2655 return chrec_dont_know
;
2657 op2
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2658 inner_loop
, TREE_OPERAND (chrec
, 2),
2659 fold_conversions
, size_expr
);
2660 if (op2
== chrec_dont_know
)
2661 return chrec_dont_know
;
2663 if (op0
== TREE_OPERAND (chrec
, 0)
2664 && op1
== TREE_OPERAND (chrec
, 1)
2665 && op2
== TREE_OPERAND (chrec
, 2))
2668 return fold_build3 (TREE_CODE (chrec
),
2669 TREE_TYPE (chrec
), op0
, op1
, op2
);
2672 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2673 and EVOLUTION_LOOP, that were left under a symbolic form.
2675 CHREC is an expression with 2 operands to be instantiated.
2677 CACHE is the cache of already instantiated values.
2679 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2680 conversions that may wrap in signed/pointer type are folded, as long
2681 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2682 then we don't do such fold.
2684 SIZE_EXPR is used for computing the size of the expression to be
2685 instantiated, and to stop if it exceeds some limit. */
2688 instantiate_scev_2 (basic_block instantiate_below
,
2689 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2691 bool *fold_conversions
, int size_expr
)
2694 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2695 inner_loop
, TREE_OPERAND (chrec
, 0),
2696 fold_conversions
, size_expr
);
2697 if (op0
== chrec_dont_know
)
2698 return chrec_dont_know
;
2700 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2701 inner_loop
, TREE_OPERAND (chrec
, 1),
2702 fold_conversions
, size_expr
);
2703 if (op1
== chrec_dont_know
)
2704 return chrec_dont_know
;
2706 if (op0
== TREE_OPERAND (chrec
, 0)
2707 && op1
== TREE_OPERAND (chrec
, 1))
2710 return fold_build2 (TREE_CODE (chrec
), TREE_TYPE (chrec
), op0
, op1
);
2713 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2714 and EVOLUTION_LOOP, that were left under a symbolic form.
2716 CHREC is an expression with 2 operands to be instantiated.
2718 CACHE is the cache of already instantiated values.
2720 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2721 conversions that may wrap in signed/pointer type are folded, as long
2722 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2723 then we don't do such fold.
2725 SIZE_EXPR is used for computing the size of the expression to be
2726 instantiated, and to stop if it exceeds some limit. */
2729 instantiate_scev_1 (basic_block instantiate_below
,
2730 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2732 bool *fold_conversions
, int size_expr
)
2734 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2735 inner_loop
, TREE_OPERAND (chrec
, 0),
2736 fold_conversions
, size_expr
);
2738 if (op0
== chrec_dont_know
)
2739 return chrec_dont_know
;
2741 if (op0
== TREE_OPERAND (chrec
, 0))
2744 return fold_build1 (TREE_CODE (chrec
), TREE_TYPE (chrec
), op0
);
2747 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2748 and EVOLUTION_LOOP, that were left under a symbolic form.
2750 CHREC is the scalar evolution to instantiate.
2752 CACHE is the cache of already instantiated values.
2754 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2755 conversions that may wrap in signed/pointer type are folded, as long
2756 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2757 then we don't do such fold.
2759 SIZE_EXPR is used for computing the size of the expression to be
2760 instantiated, and to stop if it exceeds some limit. */
2763 instantiate_scev_r (basic_block instantiate_below
,
2764 struct loop
*evolution_loop
, struct loop
*inner_loop
,
2766 bool *fold_conversions
, int size_expr
)
2768 /* Give up if the expression is larger than the MAX that we allow. */
2769 if (size_expr
++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE
))
2770 return chrec_dont_know
;
2772 if (chrec
== NULL_TREE
2773 || automatically_generated_chrec_p (chrec
)
2774 || is_gimple_min_invariant (chrec
))
2777 switch (TREE_CODE (chrec
))
2780 return instantiate_scev_name (instantiate_below
, evolution_loop
,
2782 fold_conversions
, size_expr
);
2784 case POLYNOMIAL_CHREC
:
2785 return instantiate_scev_poly (instantiate_below
, evolution_loop
,
2787 fold_conversions
, size_expr
);
2789 case POINTER_PLUS_EXPR
:
2793 return instantiate_scev_binary (instantiate_below
, evolution_loop
,
2795 TREE_CODE (chrec
), chrec_type (chrec
),
2796 TREE_OPERAND (chrec
, 0),
2797 TREE_OPERAND (chrec
, 1),
2798 fold_conversions
, size_expr
);
2801 return instantiate_scev_convert (instantiate_below
, evolution_loop
,
2803 TREE_TYPE (chrec
), TREE_OPERAND (chrec
, 0),
2804 fold_conversions
, size_expr
);
2808 return instantiate_scev_not (instantiate_below
, evolution_loop
,
2810 TREE_CODE (chrec
), TREE_TYPE (chrec
),
2811 TREE_OPERAND (chrec
, 0),
2812 fold_conversions
, size_expr
);
2815 case SCEV_NOT_KNOWN
:
2816 return chrec_dont_know
;
2822 return instantiate_array_ref (instantiate_below
, evolution_loop
,
2824 fold_conversions
, size_expr
);
2830 if (VL_EXP_CLASS_P (chrec
))
2831 return chrec_dont_know
;
2833 switch (TREE_CODE_LENGTH (TREE_CODE (chrec
)))
2836 return instantiate_scev_3 (instantiate_below
, evolution_loop
,
2838 fold_conversions
, size_expr
);
2841 return instantiate_scev_2 (instantiate_below
, evolution_loop
,
2843 fold_conversions
, size_expr
);
2846 return instantiate_scev_1 (instantiate_below
, evolution_loop
,
2848 fold_conversions
, size_expr
);
2857 /* Too complicated to handle. */
2858 return chrec_dont_know
;
2861 /* Analyze all the parameters of the chrec that were left under a
2862 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2863 recursive instantiation of parameters: a parameter is a variable
2864 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2865 a function parameter. */
2868 instantiate_scev (basic_block instantiate_below
, struct loop
*evolution_loop
,
2873 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2875 fprintf (dump_file
, "(instantiate_scev \n");
2876 fprintf (dump_file
, " (instantiate_below = %d)\n", instantiate_below
->index
);
2877 fprintf (dump_file
, " (evolution_loop = %d)\n", evolution_loop
->num
);
2878 fprintf (dump_file
, " (chrec = ");
2879 print_generic_expr (dump_file
, chrec
, 0);
2880 fprintf (dump_file
, ")\n");
2886 global_cache
= new instantiate_cache_type
;
2890 res
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2891 NULL
, chrec
, NULL
, 0);
2895 delete global_cache
;
2896 global_cache
= NULL
;
2899 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2901 fprintf (dump_file
, " (res = ");
2902 print_generic_expr (dump_file
, res
, 0);
2903 fprintf (dump_file
, "))\n");
2909 /* Similar to instantiate_parameters, but does not introduce the
2910 evolutions in outer loops for LOOP invariants in CHREC, and does not
2911 care about causing overflows, as long as they do not affect value
2912 of an expression. */
2915 resolve_mixers (struct loop
*loop
, tree chrec
, bool *folded_casts
)
2918 bool fold_conversions
= false;
2921 global_cache
= new instantiate_cache_type
;
2925 tree ret
= instantiate_scev_r (block_before_loop (loop
), loop
, NULL
,
2926 chrec
, &fold_conversions
, 0);
2928 if (folded_casts
&& !*folded_casts
)
2929 *folded_casts
= fold_conversions
;
2933 delete global_cache
;
2934 global_cache
= NULL
;
2940 /* Entry point for the analysis of the number of iterations pass.
2941 This function tries to safely approximate the number of iterations
2942 the loop will run. When this property is not decidable at compile
2943 time, the result is chrec_dont_know. Otherwise the result is a
2944 scalar or a symbolic parameter. When the number of iterations may
2945 be equal to zero and the property cannot be determined at compile
2946 time, the result is a COND_EXPR that represents in a symbolic form
2947 the conditions under which the number of iterations is not zero.
2949 Example of analysis: suppose that the loop has an exit condition:
2951 "if (b > 49) goto end_loop;"
2953 and that in a previous analysis we have determined that the
2954 variable 'b' has an evolution function:
2956 "EF = {23, +, 5}_2".
2958 When we evaluate the function at the point 5, i.e. the value of the
2959 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2960 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2961 the loop body has been executed 6 times. */
2964 number_of_latch_executions (struct loop
*loop
)
2967 struct tree_niter_desc niter_desc
;
2971 /* Determine whether the number of iterations in loop has already
2973 res
= loop
->nb_iterations
;
2977 may_be_zero
= NULL_TREE
;
2979 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2980 fprintf (dump_file
, "(number_of_iterations_in_loop = \n");
2982 res
= chrec_dont_know
;
2983 exit
= single_exit (loop
);
2985 if (exit
&& number_of_iterations_exit (loop
, exit
, &niter_desc
, false))
2987 may_be_zero
= niter_desc
.may_be_zero
;
2988 res
= niter_desc
.niter
;
2991 if (res
== chrec_dont_know
2993 || integer_zerop (may_be_zero
))
2995 else if (integer_nonzerop (may_be_zero
))
2996 res
= build_int_cst (TREE_TYPE (res
), 0);
2998 else if (COMPARISON_CLASS_P (may_be_zero
))
2999 res
= fold_build3 (COND_EXPR
, TREE_TYPE (res
), may_be_zero
,
3000 build_int_cst (TREE_TYPE (res
), 0), res
);
3002 res
= chrec_dont_know
;
3004 if (dump_file
&& (dump_flags
& TDF_SCEV
))
3006 fprintf (dump_file
, " (set_nb_iterations_in_loop = ");
3007 print_generic_expr (dump_file
, res
, 0);
3008 fprintf (dump_file
, "))\n");
3011 loop
->nb_iterations
= res
;
3016 /* Counters for the stats. */
3022 unsigned nb_affine_multivar
;
3023 unsigned nb_higher_poly
;
3024 unsigned nb_chrec_dont_know
;
3025 unsigned nb_undetermined
;
3028 /* Reset the counters. */
3031 reset_chrecs_counters (struct chrec_stats
*stats
)
3033 stats
->nb_chrecs
= 0;
3034 stats
->nb_affine
= 0;
3035 stats
->nb_affine_multivar
= 0;
3036 stats
->nb_higher_poly
= 0;
3037 stats
->nb_chrec_dont_know
= 0;
3038 stats
->nb_undetermined
= 0;
3041 /* Dump the contents of a CHREC_STATS structure. */
3044 dump_chrecs_stats (FILE *file
, struct chrec_stats
*stats
)
3046 fprintf (file
, "\n(\n");
3047 fprintf (file
, "-----------------------------------------\n");
3048 fprintf (file
, "%d\taffine univariate chrecs\n", stats
->nb_affine
);
3049 fprintf (file
, "%d\taffine multivariate chrecs\n", stats
->nb_affine_multivar
);
3050 fprintf (file
, "%d\tdegree greater than 2 polynomials\n",
3051 stats
->nb_higher_poly
);
3052 fprintf (file
, "%d\tchrec_dont_know chrecs\n", stats
->nb_chrec_dont_know
);
3053 fprintf (file
, "-----------------------------------------\n");
3054 fprintf (file
, "%d\ttotal chrecs\n", stats
->nb_chrecs
);
3055 fprintf (file
, "%d\twith undetermined coefficients\n",
3056 stats
->nb_undetermined
);
3057 fprintf (file
, "-----------------------------------------\n");
3058 fprintf (file
, "%d\tchrecs in the scev database\n",
3059 (int) scalar_evolution_info
->elements ());
3060 fprintf (file
, "%d\tsets in the scev database\n", nb_set_scev
);
3061 fprintf (file
, "%d\tgets in the scev database\n", nb_get_scev
);
3062 fprintf (file
, "-----------------------------------------\n");
3063 fprintf (file
, ")\n\n");
3066 /* Gather statistics about CHREC. */
3069 gather_chrec_stats (tree chrec
, struct chrec_stats
*stats
)
3071 if (dump_file
&& (dump_flags
& TDF_STATS
))
3073 fprintf (dump_file
, "(classify_chrec ");
3074 print_generic_expr (dump_file
, chrec
, 0);
3075 fprintf (dump_file
, "\n");
3080 if (chrec
== NULL_TREE
)
3082 stats
->nb_undetermined
++;
3086 switch (TREE_CODE (chrec
))
3088 case POLYNOMIAL_CHREC
:
3089 if (evolution_function_is_affine_p (chrec
))
3091 if (dump_file
&& (dump_flags
& TDF_STATS
))
3092 fprintf (dump_file
, " affine_univariate\n");
3095 else if (evolution_function_is_affine_multivariate_p (chrec
, 0))
3097 if (dump_file
&& (dump_flags
& TDF_STATS
))
3098 fprintf (dump_file
, " affine_multivariate\n");
3099 stats
->nb_affine_multivar
++;
3103 if (dump_file
&& (dump_flags
& TDF_STATS
))
3104 fprintf (dump_file
, " higher_degree_polynomial\n");
3105 stats
->nb_higher_poly
++;
3114 if (chrec_contains_undetermined (chrec
))
3116 if (dump_file
&& (dump_flags
& TDF_STATS
))
3117 fprintf (dump_file
, " undetermined\n");
3118 stats
->nb_undetermined
++;
3121 if (dump_file
&& (dump_flags
& TDF_STATS
))
3122 fprintf (dump_file
, ")\n");
3125 /* Classify the chrecs of the whole database. */
3128 gather_stats_on_scev_database (void)
3130 struct chrec_stats stats
;
3135 reset_chrecs_counters (&stats
);
3137 hash_table
<scev_info_hasher
>::iterator iter
;
3139 FOR_EACH_HASH_TABLE_ELEMENT (*scalar_evolution_info
, elt
, scev_info_str
*,
3141 gather_chrec_stats (elt
->chrec
, &stats
);
3143 dump_chrecs_stats (dump_file
, &stats
);
3151 initialize_scalar_evolutions_analyzer (void)
3153 /* The elements below are unique. */
3154 if (chrec_dont_know
== NULL_TREE
)
3156 chrec_not_analyzed_yet
= NULL_TREE
;
3157 chrec_dont_know
= make_node (SCEV_NOT_KNOWN
);
3158 chrec_known
= make_node (SCEV_KNOWN
);
3159 TREE_TYPE (chrec_dont_know
) = void_type_node
;
3160 TREE_TYPE (chrec_known
) = void_type_node
;
3164 /* Initialize the analysis of scalar evolutions for LOOPS. */
3167 scev_initialize (void)
3171 scalar_evolution_info
= hash_table
<scev_info_hasher
>::create_ggc (100);
3173 initialize_scalar_evolutions_analyzer ();
3175 FOR_EACH_LOOP (loop
, 0)
3177 loop
->nb_iterations
= NULL_TREE
;
3181 /* Return true if SCEV is initialized. */
3184 scev_initialized_p (void)
3186 return scalar_evolution_info
!= NULL
;
3189 /* Cleans up the information cached by the scalar evolutions analysis
3190 in the hash table. */
3193 scev_reset_htab (void)
3195 if (!scalar_evolution_info
)
3198 scalar_evolution_info
->empty ();
3201 /* Cleans up the information cached by the scalar evolutions analysis
3202 in the hash table and in the loop->nb_iterations. */
3211 FOR_EACH_LOOP (loop
, 0)
3213 loop
->nb_iterations
= NULL_TREE
;
3217 /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3218 respect to WRTO_LOOP and returns its base and step in IV if possible
3219 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3220 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3221 invariant in LOOP. Otherwise we require it to be an integer constant.
3223 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3224 because it is computed in signed arithmetics). Consequently, adding an
3227 for (i = IV->base; ; i += IV->step)
3229 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3230 false for the type of the induction variable, or you can prove that i does
3231 not wrap by some other argument. Otherwise, this might introduce undefined
3234 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3236 must be used instead. */
3239 simple_iv (struct loop
*wrto_loop
, struct loop
*use_loop
, tree op
,
3240 affine_iv
*iv
, bool allow_nonconstant_step
)
3242 enum tree_code code
;
3243 tree type
, ev
, base
, e
, stop
;
3245 bool folded_casts
, overflow
;
3247 iv
->base
= NULL_TREE
;
3248 iv
->step
= NULL_TREE
;
3249 iv
->no_overflow
= false;
3251 type
= TREE_TYPE (op
);
3252 if (!POINTER_TYPE_P (type
)
3253 && !INTEGRAL_TYPE_P (type
))
3256 ev
= analyze_scalar_evolution_in_loop (wrto_loop
, use_loop
, op
,
3258 if (chrec_contains_undetermined (ev
)
3259 || chrec_contains_symbols_defined_in_loop (ev
, wrto_loop
->num
))
3262 if (tree_does_not_contain_chrecs (ev
))
3265 iv
->step
= build_int_cst (TREE_TYPE (ev
), 0);
3266 iv
->no_overflow
= true;
3270 if (TREE_CODE (ev
) != POLYNOMIAL_CHREC
3271 || CHREC_VARIABLE (ev
) != (unsigned) wrto_loop
->num
)
3274 iv
->step
= CHREC_RIGHT (ev
);
3275 if ((!allow_nonconstant_step
&& TREE_CODE (iv
->step
) != INTEGER_CST
)
3276 || tree_contains_chrecs (iv
->step
, NULL
))
3279 iv
->base
= CHREC_LEFT (ev
);
3280 if (tree_contains_chrecs (iv
->base
, NULL
))
3283 iv
->no_overflow
= (!folded_casts
&& ANY_INTEGRAL_TYPE_P (type
)
3284 && TYPE_OVERFLOW_UNDEFINED (type
));
3286 /* Try to simplify iv base:
3288 (signed T) ((unsigned T)base + step) ;; TREE_TYPE (base) == signed T
3289 == (signed T)(unsigned T)base + step
3292 If we can prove operation (base + step) doesn't overflow or underflow.
3293 Specifically, we try to prove below conditions are satisfied:
3295 base <= UPPER_BOUND (type) - step ;;step > 0
3296 base >= LOWER_BOUND (type) - step ;;step < 0
3298 This is done by proving the reverse conditions are false using loop's
3301 The is necessary to make loop niter, or iv overflow analysis easier
3304 int foo (int *a, signed char s, signed char l)
3307 for (i = s; i < l; i++)
3312 Note variable I is firstly converted to type unsigned char, incremented,
3313 then converted back to type signed char. */
3315 if (wrto_loop
->num
!= use_loop
->num
)
3318 if (!CONVERT_EXPR_P (iv
->base
) || TREE_CODE (iv
->step
) != INTEGER_CST
)
3321 type
= TREE_TYPE (iv
->base
);
3322 e
= TREE_OPERAND (iv
->base
, 0);
3323 if (TREE_CODE (e
) != PLUS_EXPR
3324 || TREE_CODE (TREE_OPERAND (e
, 1)) != INTEGER_CST
3325 || !tree_int_cst_equal (iv
->step
,
3326 fold_convert (type
, TREE_OPERAND (e
, 1))))
3328 e
= TREE_OPERAND (e
, 0);
3329 if (!CONVERT_EXPR_P (e
))
3331 base
= TREE_OPERAND (e
, 0);
3332 if (!useless_type_conversion_p (type
, TREE_TYPE (base
)))
3335 if (tree_int_cst_sign_bit (iv
->step
))
3338 extreme
= wi::min_value (type
);
3343 extreme
= wi::max_value (type
);
3346 extreme
= wi::sub (extreme
, iv
->step
, TYPE_SIGN (type
), &overflow
);
3349 e
= fold_build2 (code
, boolean_type_node
, base
,
3350 wide_int_to_tree (type
, extreme
));
3351 stop
= (TREE_CODE (base
) == SSA_NAME
) ? base
: NULL
;
3352 e
= simplify_using_initial_conditions (use_loop
, e
, stop
);
3353 if (!integer_zerop (e
))
3356 if (POINTER_TYPE_P (TREE_TYPE (base
)))
3357 code
= POINTER_PLUS_EXPR
;
3361 iv
->base
= fold_build2 (code
, TREE_TYPE (base
), base
, iv
->step
);
3365 /* Finalize the scalar evolution analysis. */
3368 scev_finalize (void)
3370 if (!scalar_evolution_info
)
3372 scalar_evolution_info
->empty ();
3373 scalar_evolution_info
= NULL
;
3376 /* Returns true if the expression EXPR is considered to be too expensive
3377 for scev_const_prop. */
3380 expression_expensive_p (tree expr
)
3382 enum tree_code code
;
3384 if (is_gimple_val (expr
))
3387 code
= TREE_CODE (expr
);
3388 if (code
== TRUNC_DIV_EXPR
3389 || code
== CEIL_DIV_EXPR
3390 || code
== FLOOR_DIV_EXPR
3391 || code
== ROUND_DIV_EXPR
3392 || code
== TRUNC_MOD_EXPR
3393 || code
== CEIL_MOD_EXPR
3394 || code
== FLOOR_MOD_EXPR
3395 || code
== ROUND_MOD_EXPR
3396 || code
== EXACT_DIV_EXPR
)
3398 /* Division by power of two is usually cheap, so we allow it.
3399 Forbid anything else. */
3400 if (!integer_pow2p (TREE_OPERAND (expr
, 1)))
3404 switch (TREE_CODE_CLASS (code
))
3407 case tcc_comparison
:
3408 if (expression_expensive_p (TREE_OPERAND (expr
, 1)))
3413 return expression_expensive_p (TREE_OPERAND (expr
, 0));
3420 /* Do final value replacement for LOOP. */
3423 final_value_replacement_loop (struct loop
*loop
)
3425 /* If we do not know exact number of iterations of the loop, we cannot
3426 replace the final value. */
3427 edge exit
= single_exit (loop
);
3431 tree niter
= number_of_latch_executions (loop
);
3432 if (niter
== chrec_dont_know
)
3435 /* Ensure that it is possible to insert new statements somewhere. */
3436 if (!single_pred_p (exit
->dest
))
3437 split_loop_exit_edge (exit
);
3439 /* Set stmt insertion pointer. All stmts are inserted before this point. */
3440 gimple_stmt_iterator gsi
= gsi_after_labels (exit
->dest
);
3442 struct loop
*ex_loop
3443 = superloop_at_depth (loop
,
3444 loop_depth (exit
->dest
->loop_father
) + 1);
3447 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); )
3449 gphi
*phi
= psi
.phi ();
3450 tree rslt
= PHI_RESULT (phi
);
3451 tree def
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
3452 if (virtual_operand_p (def
))
3458 if (!POINTER_TYPE_P (TREE_TYPE (def
))
3459 && !INTEGRAL_TYPE_P (TREE_TYPE (def
)))
3466 def
= analyze_scalar_evolution_in_loop (ex_loop
, loop
, def
,
3468 def
= compute_overall_effect_of_inner_loop (ex_loop
, def
);
3469 if (!tree_does_not_contain_chrecs (def
)
3470 || chrec_contains_symbols_defined_in_loop (def
, ex_loop
->num
)
3471 /* Moving the computation from the loop may prolong life range
3472 of some ssa names, which may cause problems if they appear
3473 on abnormal edges. */
3474 || contains_abnormal_ssa_name_p (def
)
3475 /* Do not emit expensive expressions. The rationale is that
3476 when someone writes a code like
3478 while (n > 45) n -= 45;
3480 he probably knows that n is not large, and does not want it
3481 to be turned into n %= 45. */
3482 || expression_expensive_p (def
))
3484 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3486 fprintf (dump_file
, "not replacing:\n ");
3487 print_gimple_stmt (dump_file
, phi
, 0, 0);
3488 fprintf (dump_file
, "\n");
3494 /* Eliminate the PHI node and replace it by a computation outside
3498 fprintf (dump_file
, "\nfinal value replacement:\n ");
3499 print_gimple_stmt (dump_file
, phi
, 0, 0);
3500 fprintf (dump_file
, " with\n ");
3502 def
= unshare_expr (def
);
3503 remove_phi_node (&psi
, false);
3505 /* If def's type has undefined overflow and there were folded
3506 casts, rewrite all stmts added for def into arithmetics
3507 with defined overflow behavior. */
3508 if (folded_casts
&& ANY_INTEGRAL_TYPE_P (TREE_TYPE (def
))
3509 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (def
)))
3512 gimple_stmt_iterator gsi2
;
3513 def
= force_gimple_operand (def
, &stmts
, true, NULL_TREE
);
3514 gsi2
= gsi_start (stmts
);
3515 while (!gsi_end_p (gsi2
))
3517 gimple
*stmt
= gsi_stmt (gsi2
);
3518 gimple_stmt_iterator gsi3
= gsi2
;
3520 gsi_remove (&gsi3
, false);
3521 if (is_gimple_assign (stmt
)
3522 && arith_code_with_undefined_signed_overflow
3523 (gimple_assign_rhs_code (stmt
)))
3524 gsi_insert_seq_before (&gsi
,
3525 rewrite_to_defined_overflow (stmt
),
3528 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
3532 def
= force_gimple_operand_gsi (&gsi
, def
, false, NULL_TREE
,
3533 true, GSI_SAME_STMT
);
3535 gassign
*ass
= gimple_build_assign (rslt
, def
);
3536 gsi_insert_before (&gsi
, ass
, GSI_SAME_STMT
);
3539 print_gimple_stmt (dump_file
, ass
, 0, 0);
3540 fprintf (dump_file
, "\n");
3545 /* Replace ssa names for that scev can prove they are constant by the
3546 appropriate constants. Also perform final value replacement in loops,
3547 in case the replacement expressions are cheap.
3549 We only consider SSA names defined by phi nodes; rest is left to the
3550 ordinary constant propagation pass. */
3553 scev_const_prop (void)
3556 tree name
, type
, ev
;
3559 bitmap ssa_names_to_remove
= NULL
;
3563 if (number_of_loops (cfun
) <= 1)
3566 FOR_EACH_BB_FN (bb
, cfun
)
3568 loop
= bb
->loop_father
;
3570 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); gsi_next (&psi
))
3573 name
= PHI_RESULT (phi
);
3575 if (virtual_operand_p (name
))
3578 type
= TREE_TYPE (name
);
3580 if (!POINTER_TYPE_P (type
)
3581 && !INTEGRAL_TYPE_P (type
))
3584 ev
= resolve_mixers (loop
, analyze_scalar_evolution (loop
, name
),
3586 if (!is_gimple_min_invariant (ev
)
3587 || !may_propagate_copy (name
, ev
))
3590 /* Replace the uses of the name. */
3593 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3595 fprintf (dump_file
, "Replacing uses of: ");
3596 print_generic_expr (dump_file
, name
, 0);
3597 fprintf (dump_file
, " with: ");
3598 print_generic_expr (dump_file
, ev
, 0);
3599 fprintf (dump_file
, "\n");
3601 replace_uses_by (name
, ev
);
3604 if (!ssa_names_to_remove
)
3605 ssa_names_to_remove
= BITMAP_ALLOC (NULL
);
3606 bitmap_set_bit (ssa_names_to_remove
, SSA_NAME_VERSION (name
));
3610 /* Remove the ssa names that were replaced by constants. We do not
3611 remove them directly in the previous cycle, since this
3612 invalidates scev cache. */
3613 if (ssa_names_to_remove
)
3617 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove
, 0, i
, bi
)
3619 gimple_stmt_iterator psi
;
3620 name
= ssa_name (i
);
3621 phi
= as_a
<gphi
*> (SSA_NAME_DEF_STMT (name
));
3623 gcc_assert (gimple_code (phi
) == GIMPLE_PHI
);
3624 psi
= gsi_for_stmt (phi
);
3625 remove_phi_node (&psi
, true);
3628 BITMAP_FREE (ssa_names_to_remove
);
3632 /* Now the regular final value replacement. */
3633 FOR_EACH_LOOP (loop
, LI_FROM_INNERMOST
)
3634 final_value_replacement_loop (loop
);
3639 #include "gt-tree-scalar-evolution.h"