1 /* Scalar evolution detector.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Sebastian Pop <s.pop@laposte.net>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
25 This pass analyzes the evolution of scalar variables in loop
26 structures. The algorithm is based on the SSA representation,
27 and on the loop hierarchy tree. This algorithm is not based on
28 the notion of versions of a variable, as it was the case for the
29 previous implementations of the scalar evolution algorithm, but
30 it assumes that each defined name is unique.
32 The notation used in this file is called "chains of recurrences",
33 and has been proposed by Eugene Zima, Robert Van Engelen, and
34 others for describing induction variables in programs. For example
35 "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0
36 when entering in the loop_1 and has a step 2 in this loop, in other
37 words "for (b = 0; b < N; b+=2);". Note that the coefficients of
38 this chain of recurrence (or chrec [shrek]) can contain the name of
39 other variables, in which case they are called parametric chrecs.
40 For example, "b -> {a, +, 2}_1" means that the initial value of "b"
41 is the value of "a". In most of the cases these parametric chrecs
42 are fully instantiated before their use because symbolic names can
43 hide some difficult cases such as self-references described later
44 (see the Fibonacci example).
46 A short sketch of the algorithm is:
48 Given a scalar variable to be analyzed, follow the SSA edge to
51 - When the definition is a GIMPLE_ASSIGN: if the right hand side
52 (RHS) of the definition cannot be statically analyzed, the answer
53 of the analyzer is: "don't know".
54 Otherwise, for all the variables that are not yet analyzed in the
55 RHS, try to determine their evolution, and finally try to
56 evaluate the operation of the RHS that gives the evolution
57 function of the analyzed variable.
59 - When the definition is a condition-phi-node: determine the
60 evolution function for all the branches of the phi node, and
61 finally merge these evolutions (see chrec_merge).
63 - When the definition is a loop-phi-node: determine its initial
64 condition, that is the SSA edge defined in an outer loop, and
65 keep it symbolic. Then determine the SSA edges that are defined
66 in the body of the loop. Follow the inner edges until ending on
67 another loop-phi-node of the same analyzed loop. If the reached
68 loop-phi-node is not the starting loop-phi-node, then we keep
69 this definition under a symbolic form. If the reached
70 loop-phi-node is the same as the starting one, then we compute a
71 symbolic stride on the return path. The result is then the
72 symbolic chrec {initial_condition, +, symbolic_stride}_loop.
76 Example 1: Illustration of the basic algorithm.
82 | if (c > 10) exit_loop
85 Suppose that we want to know the number of iterations of the
86 loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We
87 ask the scalar evolution analyzer two questions: what's the
88 scalar evolution (scev) of "c", and what's the scev of "10". For
89 "10" the answer is "10" since it is a scalar constant. For the
90 scalar variable "c", it follows the SSA edge to its definition,
91 "c = b + 1", and then asks again what's the scev of "b".
92 Following the SSA edge, we end on a loop-phi-node "b = phi (a,
93 c)", where the initial condition is "a", and the inner loop edge
94 is "c". The initial condition is kept under a symbolic form (it
95 may be the case that the copy constant propagation has done its
96 work and we end with the constant "3" as one of the edges of the
97 loop-phi-node). The update edge is followed to the end of the
98 loop, and until reaching again the starting loop-phi-node: b -> c
99 -> b. At this point we have drawn a path from "b" to "b" from
100 which we compute the stride in the loop: in this example it is
101 "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now
102 that the scev for "b" is known, it is possible to compute the
103 scev for "c", that is "c -> {a + 1, +, 1}_1". In order to
104 determine the number of iterations in the loop_1, we have to
105 instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some
106 more analysis the scev {4, +, 1}_1, or in other words, this is
107 the function "f (x) = x + 4", where x is the iteration count of
108 the loop_1. Now we have to solve the inequality "x + 4 > 10",
109 and take the smallest iteration number for which the loop is
110 exited: x = 7. This loop runs from x = 0 to x = 7, and in total
111 there are 8 iterations. In terms of loop normalization, we have
112 created a variable that is implicitly defined, "x" or just "_1",
113 and all the other analyzed scalars of the loop are defined in
114 function of this variable:
120 or in terms of a C program:
123 | for (x = 0; x <= 7; x++)
129 Example 2a: Illustration of the algorithm on nested loops.
140 For analyzing the scalar evolution of "a", the algorithm follows
141 the SSA edge into the loop's body: "a -> b". "b" is an inner
142 loop-phi-node, and its analysis as in Example 1, gives:
147 Following the SSA edge for the initial condition, we end on "c = a
148 + 2", and then on the starting loop-phi-node "a". From this point,
149 the loop stride is computed: back on "c = a + 2" we get a "+2" in
150 the loop_1, then on the loop-phi-node "b" we compute the overall
151 effect of the inner loop that is "b = c + 30", and we get a "+30"
152 in the loop_1. That means that the overall stride in loop_1 is
153 equal to "+32", and the result is:
158 Example 2b: Multivariate chains of recurrences.
171 Analyzing the access function of array A with
172 instantiate_parameters (loop_1, "j + k"), we obtain the
173 instantiation and the analysis of the scalar variables "j" and "k"
174 in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end
175 value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is
176 {0, +, 1}_1. To obtain the evolution function in loop_3 and
177 instantiate the scalar variables up to loop_1, one has to use:
178 instantiate_scev (block_before_loop (loop_1), loop_3, "j + k").
179 The result of this call is {{0, +, 1}_1, +, 1}_2.
181 Example 3: Higher degree polynomials.
195 instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1
196 instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
198 Example 4: Lucas, Fibonacci, or mixers in general.
210 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
211 following semantics: during the first iteration of the loop_1, the
212 variable contains the value 1, and then it contains the value "c".
213 Note that this syntax is close to the syntax of the loop-phi-node:
214 "a -> (1, c)_1" vs. "a = phi (1, c)".
216 The symbolic chrec representation contains all the semantics of the
217 original code. What is more difficult is to use this information.
219 Example 5: Flip-flops, or exchangers.
231 Based on these symbolic chrecs, it is possible to refine this
232 information into the more precise PERIODIC_CHRECs:
237 This transformation is not yet implemented.
241 You can find a more detailed description of the algorithm in:
242 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
243 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
244 this is a preliminary report and some of the details of the
245 algorithm have changed. I'm working on a research report that
246 updates the description of the algorithms to reflect the design
247 choices used in this implementation.
249 A set of slides show a high level overview of the algorithm and run
250 an example through the scalar evolution analyzer:
251 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
253 The slides that I have presented at the GCC Summit'04 are available
254 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
259 #include "coretypes.h"
260 #include "gimple-pretty-print.h"
261 #include "tree-flow.h"
263 #include "tree-chrec.h"
264 #include "tree-scalar-evolution.h"
265 #include "dumpfile.h"
268 static tree
analyze_scalar_evolution_1 (struct loop
*, tree
, tree
);
269 static tree
analyze_scalar_evolution_for_address_of (struct loop
*loop
,
272 /* The cached information about an SSA name VAR, claiming that below
273 basic block INSTANTIATED_BELOW, the value of VAR can be expressed
276 struct GTY(()) scev_info_str
{
277 basic_block instantiated_below
;
282 /* Counters for the scev database. */
283 static unsigned nb_set_scev
= 0;
284 static unsigned nb_get_scev
= 0;
286 /* The following trees are unique elements. Thus the comparison of
287 another element to these elements should be done on the pointer to
288 these trees, and not on their value. */
290 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
291 tree chrec_not_analyzed_yet
;
293 /* Reserved to the cases where the analyzer has detected an
294 undecidable property at compile time. */
295 tree chrec_dont_know
;
297 /* When the analyzer has detected that a property will never
298 happen, then it qualifies it with chrec_known. */
301 static GTY ((param_is (struct scev_info_str
))) htab_t scalar_evolution_info
;
304 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
306 static inline struct scev_info_str
*
307 new_scev_info_str (basic_block instantiated_below
, tree var
)
309 struct scev_info_str
*res
;
311 res
= ggc_alloc_scev_info_str ();
313 res
->chrec
= chrec_not_analyzed_yet
;
314 res
->instantiated_below
= instantiated_below
;
319 /* Computes a hash function for database element ELT. */
322 hash_scev_info (const void *elt
)
324 return SSA_NAME_VERSION (((const struct scev_info_str
*) elt
)->var
);
327 /* Compares database elements E1 and E2. */
330 eq_scev_info (const void *e1
, const void *e2
)
332 const struct scev_info_str
*elt1
= (const struct scev_info_str
*) e1
;
333 const struct scev_info_str
*elt2
= (const struct scev_info_str
*) e2
;
335 return (elt1
->var
== elt2
->var
336 && elt1
->instantiated_below
== elt2
->instantiated_below
);
339 /* Deletes database element E. */
342 del_scev_info (void *e
)
347 /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
348 A first query on VAR returns chrec_not_analyzed_yet. */
351 find_var_scev_info (basic_block instantiated_below
, tree var
)
353 struct scev_info_str
*res
;
354 struct scev_info_str tmp
;
358 tmp
.instantiated_below
= instantiated_below
;
359 slot
= htab_find_slot (scalar_evolution_info
, &tmp
, INSERT
);
362 *slot
= new_scev_info_str (instantiated_below
, var
);
363 res
= (struct scev_info_str
*) *slot
;
368 /* Return true when CHREC contains symbolic names defined in
372 chrec_contains_symbols_defined_in_loop (const_tree chrec
, unsigned loop_nb
)
376 if (chrec
== NULL_TREE
)
379 if (is_gimple_min_invariant (chrec
))
382 if (TREE_CODE (chrec
) == SSA_NAME
)
385 loop_p def_loop
, loop
;
387 if (SSA_NAME_IS_DEFAULT_DEF (chrec
))
390 def
= SSA_NAME_DEF_STMT (chrec
);
391 def_loop
= loop_containing_stmt (def
);
392 loop
= get_loop (loop_nb
);
394 if (def_loop
== NULL
)
397 if (loop
== def_loop
|| flow_loop_nested_p (loop
, def_loop
))
403 n
= TREE_OPERAND_LENGTH (chrec
);
404 for (i
= 0; i
< n
; i
++)
405 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec
, i
),
411 /* Return true when PHI is a loop-phi-node. */
414 loop_phi_node_p (gimple phi
)
416 /* The implementation of this function is based on the following
417 property: "all the loop-phi-nodes of a loop are contained in the
418 loop's header basic block". */
420 return loop_containing_stmt (phi
)->header
== gimple_bb (phi
);
423 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
424 In general, in the case of multivariate evolutions we want to get
425 the evolution in different loops. LOOP specifies the level for
426 which to get the evolution.
430 | for (j = 0; j < 100; j++)
432 | for (k = 0; k < 100; k++)
434 | i = k + j; - Here the value of i is a function of j, k.
436 | ... = i - Here the value of i is a function of j.
438 | ... = i - Here the value of i is a scalar.
444 | i_1 = phi (i_0, i_2)
448 This loop has the same effect as:
449 LOOP_1 has the same effect as:
453 The overall effect of the loop, "i_0 + 20" in the previous example,
454 is obtained by passing in the parameters: LOOP = 1,
455 EVOLUTION_FN = {i_0, +, 2}_1.
459 compute_overall_effect_of_inner_loop (struct loop
*loop
, tree evolution_fn
)
463 if (evolution_fn
== chrec_dont_know
)
464 return chrec_dont_know
;
466 else if (TREE_CODE (evolution_fn
) == POLYNOMIAL_CHREC
)
468 struct loop
*inner_loop
= get_chrec_loop (evolution_fn
);
470 if (inner_loop
== loop
471 || flow_loop_nested_p (loop
, inner_loop
))
473 tree nb_iter
= number_of_latch_executions (inner_loop
);
475 if (nb_iter
== chrec_dont_know
)
476 return chrec_dont_know
;
481 /* evolution_fn is the evolution function in LOOP. Get
482 its value in the nb_iter-th iteration. */
483 res
= chrec_apply (inner_loop
->num
, evolution_fn
, nb_iter
);
485 if (chrec_contains_symbols_defined_in_loop (res
, loop
->num
))
486 res
= instantiate_parameters (loop
, res
);
488 /* Continue the computation until ending on a parent of LOOP. */
489 return compute_overall_effect_of_inner_loop (loop
, res
);
496 /* If the evolution function is an invariant, there is nothing to do. */
497 else if (no_evolution_in_loop_p (evolution_fn
, loop
->num
, &val
) && val
)
501 return chrec_dont_know
;
504 /* Associate CHREC to SCALAR. */
507 set_scalar_evolution (basic_block instantiated_below
, tree scalar
, tree chrec
)
511 if (TREE_CODE (scalar
) != SSA_NAME
)
514 scalar_info
= find_var_scev_info (instantiated_below
, scalar
);
518 if (dump_flags
& TDF_SCEV
)
520 fprintf (dump_file
, "(set_scalar_evolution \n");
521 fprintf (dump_file
, " instantiated_below = %d \n",
522 instantiated_below
->index
);
523 fprintf (dump_file
, " (scalar = ");
524 print_generic_expr (dump_file
, scalar
, 0);
525 fprintf (dump_file
, ")\n (scalar_evolution = ");
526 print_generic_expr (dump_file
, chrec
, 0);
527 fprintf (dump_file
, "))\n");
529 if (dump_flags
& TDF_STATS
)
533 *scalar_info
= chrec
;
536 /* Retrieve the chrec associated to SCALAR instantiated below
537 INSTANTIATED_BELOW block. */
540 get_scalar_evolution (basic_block instantiated_below
, tree scalar
)
546 if (dump_flags
& TDF_SCEV
)
548 fprintf (dump_file
, "(get_scalar_evolution \n");
549 fprintf (dump_file
, " (scalar = ");
550 print_generic_expr (dump_file
, scalar
, 0);
551 fprintf (dump_file
, ")\n");
553 if (dump_flags
& TDF_STATS
)
557 switch (TREE_CODE (scalar
))
560 res
= *find_var_scev_info (instantiated_below
, scalar
);
570 res
= chrec_not_analyzed_yet
;
574 if (dump_file
&& (dump_flags
& TDF_SCEV
))
576 fprintf (dump_file
, " (scalar_evolution = ");
577 print_generic_expr (dump_file
, res
, 0);
578 fprintf (dump_file
, "))\n");
584 /* Helper function for add_to_evolution. Returns the evolution
585 function for an assignment of the form "a = b + c", where "a" and
586 "b" are on the strongly connected component. CHREC_BEFORE is the
587 information that we already have collected up to this point.
588 TO_ADD is the evolution of "c".
590 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
591 evolution the expression TO_ADD, otherwise construct an evolution
592 part for this loop. */
595 add_to_evolution_1 (unsigned loop_nb
, tree chrec_before
, tree to_add
,
598 tree type
, left
, right
;
599 struct loop
*loop
= get_loop (loop_nb
), *chloop
;
601 switch (TREE_CODE (chrec_before
))
603 case POLYNOMIAL_CHREC
:
604 chloop
= get_chrec_loop (chrec_before
);
606 || flow_loop_nested_p (chloop
, loop
))
610 type
= chrec_type (chrec_before
);
612 /* When there is no evolution part in this loop, build it. */
617 right
= SCALAR_FLOAT_TYPE_P (type
)
618 ? build_real (type
, dconst0
)
619 : build_int_cst (type
, 0);
623 var
= CHREC_VARIABLE (chrec_before
);
624 left
= CHREC_LEFT (chrec_before
);
625 right
= CHREC_RIGHT (chrec_before
);
628 to_add
= chrec_convert (type
, to_add
, at_stmt
);
629 right
= chrec_convert_rhs (type
, right
, at_stmt
);
630 right
= chrec_fold_plus (chrec_type (right
), right
, to_add
);
631 return build_polynomial_chrec (var
, left
, right
);
635 gcc_assert (flow_loop_nested_p (loop
, chloop
));
637 /* Search the evolution in LOOP_NB. */
638 left
= add_to_evolution_1 (loop_nb
, CHREC_LEFT (chrec_before
),
640 right
= CHREC_RIGHT (chrec_before
);
641 right
= chrec_convert_rhs (chrec_type (left
), right
, at_stmt
);
642 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before
),
647 /* These nodes do not depend on a loop. */
648 if (chrec_before
== chrec_dont_know
)
649 return chrec_dont_know
;
652 right
= chrec_convert_rhs (chrec_type (left
), to_add
, at_stmt
);
653 return build_polynomial_chrec (loop_nb
, left
, right
);
657 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
660 Description (provided for completeness, for those who read code in
661 a plane, and for my poor 62 bytes brain that would have forgotten
662 all this in the next two or three months):
664 The algorithm of translation of programs from the SSA representation
665 into the chrecs syntax is based on a pattern matching. After having
666 reconstructed the overall tree expression for a loop, there are only
667 two cases that can arise:
669 1. a = loop-phi (init, a + expr)
670 2. a = loop-phi (init, expr)
672 where EXPR is either a scalar constant with respect to the analyzed
673 loop (this is a degree 0 polynomial), or an expression containing
674 other loop-phi definitions (these are higher degree polynomials).
681 | a = phi (init, a + 5)
688 | a = phi (inita, 2 * b + 3)
689 | b = phi (initb, b + 1)
692 For the first case, the semantics of the SSA representation is:
694 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
696 that is, there is a loop index "x" that determines the scalar value
697 of the variable during the loop execution. During the first
698 iteration, the value is that of the initial condition INIT, while
699 during the subsequent iterations, it is the sum of the initial
700 condition with the sum of all the values of EXPR from the initial
701 iteration to the before last considered iteration.
703 For the second case, the semantics of the SSA program is:
705 | a (x) = init, if x = 0;
706 | expr (x - 1), otherwise.
708 The second case corresponds to the PEELED_CHREC, whose syntax is
709 close to the syntax of a loop-phi-node:
711 | phi (init, expr) vs. (init, expr)_x
713 The proof of the translation algorithm for the first case is a
714 proof by structural induction based on the degree of EXPR.
717 When EXPR is a constant with respect to the analyzed loop, or in
718 other words when EXPR is a polynomial of degree 0, the evolution of
719 the variable A in the loop is an affine function with an initial
720 condition INIT, and a step EXPR. In order to show this, we start
721 from the semantics of the SSA representation:
723 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
725 and since "expr (j)" is a constant with respect to "j",
727 f (x) = init + x * expr
729 Finally, based on the semantics of the pure sum chrecs, by
730 identification we get the corresponding chrecs syntax:
732 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
733 f (x) -> {init, +, expr}_x
736 Suppose that EXPR is a polynomial of degree N with respect to the
737 analyzed loop_x for which we have already determined that it is
738 written under the chrecs syntax:
740 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
742 We start from the semantics of the SSA program:
744 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
746 | f (x) = init + \sum_{j = 0}^{x - 1}
747 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
749 | f (x) = init + \sum_{j = 0}^{x - 1}
750 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
752 | f (x) = init + \sum_{k = 0}^{n - 1}
753 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
755 | f (x) = init + \sum_{k = 0}^{n - 1}
756 | (b_k * \binom{x}{k + 1})
758 | f (x) = init + b_0 * \binom{x}{1} + ...
759 | + b_{n-1} * \binom{x}{n}
761 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
762 | + b_{n-1} * \binom{x}{n}
765 And finally from the definition of the chrecs syntax, we identify:
766 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
768 This shows the mechanism that stands behind the add_to_evolution
769 function. An important point is that the use of symbolic
770 parameters avoids the need of an analysis schedule.
777 | a = phi (inita, a + 2 + b)
778 | b = phi (initb, b + 1)
781 When analyzing "a", the algorithm keeps "b" symbolically:
783 | a -> {inita, +, 2 + b}_1
785 Then, after instantiation, the analyzer ends on the evolution:
787 | a -> {inita, +, 2 + initb, +, 1}_1
792 add_to_evolution (unsigned loop_nb
, tree chrec_before
, enum tree_code code
,
793 tree to_add
, gimple at_stmt
)
795 tree type
= chrec_type (to_add
);
796 tree res
= NULL_TREE
;
798 if (to_add
== NULL_TREE
)
801 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
802 instantiated at this point. */
803 if (TREE_CODE (to_add
) == POLYNOMIAL_CHREC
)
804 /* This should not happen. */
805 return chrec_dont_know
;
807 if (dump_file
&& (dump_flags
& TDF_SCEV
))
809 fprintf (dump_file
, "(add_to_evolution \n");
810 fprintf (dump_file
, " (loop_nb = %d)\n", loop_nb
);
811 fprintf (dump_file
, " (chrec_before = ");
812 print_generic_expr (dump_file
, chrec_before
, 0);
813 fprintf (dump_file
, ")\n (to_add = ");
814 print_generic_expr (dump_file
, to_add
, 0);
815 fprintf (dump_file
, ")\n");
818 if (code
== MINUS_EXPR
)
819 to_add
= chrec_fold_multiply (type
, to_add
, SCALAR_FLOAT_TYPE_P (type
)
820 ? build_real (type
, dconstm1
)
821 : build_int_cst_type (type
, -1));
823 res
= add_to_evolution_1 (loop_nb
, chrec_before
, to_add
, at_stmt
);
825 if (dump_file
&& (dump_flags
& TDF_SCEV
))
827 fprintf (dump_file
, " (res = ");
828 print_generic_expr (dump_file
, res
, 0);
829 fprintf (dump_file
, "))\n");
837 /* This section selects the loops that will be good candidates for the
838 scalar evolution analysis. For the moment, greedily select all the
839 loop nests we could analyze. */
841 /* For a loop with a single exit edge, return the COND_EXPR that
842 guards the exit edge. If the expression is too difficult to
843 analyze, then give up. */
846 get_loop_exit_condition (const struct loop
*loop
)
849 edge exit_edge
= single_exit (loop
);
851 if (dump_file
&& (dump_flags
& TDF_SCEV
))
852 fprintf (dump_file
, "(get_loop_exit_condition \n ");
858 stmt
= last_stmt (exit_edge
->src
);
859 if (gimple_code (stmt
) == GIMPLE_COND
)
863 if (dump_file
&& (dump_flags
& TDF_SCEV
))
865 print_gimple_stmt (dump_file
, res
, 0, 0);
866 fprintf (dump_file
, ")\n");
872 /* Recursively determine and enqueue the exit conditions for a loop. */
875 get_exit_conditions_rec (struct loop
*loop
,
876 VEC(gimple
,heap
) **exit_conditions
)
881 /* Recurse on the inner loops, then on the next (sibling) loops. */
882 get_exit_conditions_rec (loop
->inner
, exit_conditions
);
883 get_exit_conditions_rec (loop
->next
, exit_conditions
);
885 if (single_exit (loop
))
887 gimple loop_condition
= get_loop_exit_condition (loop
);
890 VEC_safe_push (gimple
, heap
, *exit_conditions
, loop_condition
);
894 /* Select the candidate loop nests for the analysis. This function
895 initializes the EXIT_CONDITIONS array. */
898 select_loops_exit_conditions (VEC(gimple
,heap
) **exit_conditions
)
900 struct loop
*function_body
= current_loops
->tree_root
;
902 get_exit_conditions_rec (function_body
->inner
, exit_conditions
);
906 /* Depth first search algorithm. */
908 typedef enum t_bool
{
915 static t_bool
follow_ssa_edge (struct loop
*loop
, gimple
, gimple
, tree
*, int);
917 /* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
918 Return true if the strongly connected component has been found. */
921 follow_ssa_edge_binary (struct loop
*loop
, gimple at_stmt
,
922 tree type
, tree rhs0
, enum tree_code code
, tree rhs1
,
923 gimple halting_phi
, tree
*evolution_of_loop
, int limit
)
925 t_bool res
= t_false
;
930 case POINTER_PLUS_EXPR
:
932 if (TREE_CODE (rhs0
) == SSA_NAME
)
934 if (TREE_CODE (rhs1
) == SSA_NAME
)
936 /* Match an assignment under the form:
939 /* We want only assignments of form "name + name" contribute to
940 LIMIT, as the other cases do not necessarily contribute to
941 the complexity of the expression. */
944 evol
= *evolution_of_loop
;
945 res
= follow_ssa_edge
946 (loop
, SSA_NAME_DEF_STMT (rhs0
), halting_phi
, &evol
, limit
);
949 *evolution_of_loop
= add_to_evolution
951 chrec_convert (type
, evol
, at_stmt
),
952 code
, rhs1
, at_stmt
);
954 else if (res
== t_false
)
956 res
= follow_ssa_edge
957 (loop
, SSA_NAME_DEF_STMT (rhs1
), halting_phi
,
958 evolution_of_loop
, limit
);
961 *evolution_of_loop
= add_to_evolution
963 chrec_convert (type
, *evolution_of_loop
, at_stmt
),
964 code
, rhs0
, at_stmt
);
966 else if (res
== t_dont_know
)
967 *evolution_of_loop
= chrec_dont_know
;
970 else if (res
== t_dont_know
)
971 *evolution_of_loop
= chrec_dont_know
;
976 /* Match an assignment under the form:
978 res
= follow_ssa_edge
979 (loop
, SSA_NAME_DEF_STMT (rhs0
), halting_phi
,
980 evolution_of_loop
, limit
);
982 *evolution_of_loop
= add_to_evolution
983 (loop
->num
, chrec_convert (type
, *evolution_of_loop
,
985 code
, rhs1
, at_stmt
);
987 else if (res
== t_dont_know
)
988 *evolution_of_loop
= chrec_dont_know
;
992 else if (TREE_CODE (rhs1
) == SSA_NAME
)
994 /* Match an assignment under the form:
996 res
= follow_ssa_edge
997 (loop
, SSA_NAME_DEF_STMT (rhs1
), halting_phi
,
998 evolution_of_loop
, limit
);
1000 *evolution_of_loop
= add_to_evolution
1001 (loop
->num
, chrec_convert (type
, *evolution_of_loop
,
1003 code
, rhs0
, at_stmt
);
1005 else if (res
== t_dont_know
)
1006 *evolution_of_loop
= chrec_dont_know
;
1010 /* Otherwise, match an assignment under the form:
1012 /* And there is nothing to do. */
1017 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1018 if (TREE_CODE (rhs0
) == SSA_NAME
)
1020 /* Match an assignment under the form:
1023 /* We want only assignments of form "name - name" contribute to
1024 LIMIT, as the other cases do not necessarily contribute to
1025 the complexity of the expression. */
1026 if (TREE_CODE (rhs1
) == SSA_NAME
)
1029 res
= follow_ssa_edge (loop
, SSA_NAME_DEF_STMT (rhs0
), halting_phi
,
1030 evolution_of_loop
, limit
);
1032 *evolution_of_loop
= add_to_evolution
1033 (loop
->num
, chrec_convert (type
, *evolution_of_loop
, at_stmt
),
1034 MINUS_EXPR
, rhs1
, at_stmt
);
1036 else if (res
== t_dont_know
)
1037 *evolution_of_loop
= chrec_dont_know
;
1040 /* Otherwise, match an assignment under the form:
1042 /* And there is nothing to do. */
1053 /* Follow the ssa edge into the expression EXPR.
1054 Return true if the strongly connected component has been found. */
1057 follow_ssa_edge_expr (struct loop
*loop
, gimple at_stmt
, tree expr
,
1058 gimple halting_phi
, tree
*evolution_of_loop
, int limit
)
1060 enum tree_code code
= TREE_CODE (expr
);
1061 tree type
= TREE_TYPE (expr
), rhs0
, rhs1
;
1064 /* The EXPR is one of the following cases:
1068 - a POINTER_PLUS_EXPR,
1071 - other cases are not yet handled. */
1076 /* This assignment is under the form "a_1 = (cast) rhs. */
1077 res
= follow_ssa_edge_expr (loop
, at_stmt
, TREE_OPERAND (expr
, 0),
1078 halting_phi
, evolution_of_loop
, limit
);
1079 *evolution_of_loop
= chrec_convert (type
, *evolution_of_loop
, at_stmt
);
1083 /* This assignment is under the form "a_1 = 7". */
1088 /* This assignment is under the form: "a_1 = b_2". */
1089 res
= follow_ssa_edge
1090 (loop
, SSA_NAME_DEF_STMT (expr
), halting_phi
, evolution_of_loop
, limit
);
1093 case POINTER_PLUS_EXPR
:
1096 /* This case is under the form "rhs0 +- rhs1". */
1097 rhs0
= TREE_OPERAND (expr
, 0);
1098 rhs1
= TREE_OPERAND (expr
, 1);
1099 type
= TREE_TYPE (rhs0
);
1100 STRIP_USELESS_TYPE_CONVERSION (rhs0
);
1101 STRIP_USELESS_TYPE_CONVERSION (rhs1
);
1102 res
= follow_ssa_edge_binary (loop
, at_stmt
, type
, rhs0
, code
, rhs1
,
1103 halting_phi
, evolution_of_loop
, limit
);
1107 /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */
1108 if (TREE_CODE (TREE_OPERAND (expr
, 0)) == MEM_REF
)
1110 expr
= TREE_OPERAND (expr
, 0);
1111 rhs0
= TREE_OPERAND (expr
, 0);
1112 rhs1
= TREE_OPERAND (expr
, 1);
1113 type
= TREE_TYPE (rhs0
);
1114 STRIP_USELESS_TYPE_CONVERSION (rhs0
);
1115 STRIP_USELESS_TYPE_CONVERSION (rhs1
);
1116 res
= follow_ssa_edge_binary (loop
, at_stmt
, type
,
1117 rhs0
, POINTER_PLUS_EXPR
, rhs1
,
1118 halting_phi
, evolution_of_loop
, limit
);
1125 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1126 It must be handled as a copy assignment of the form a_1 = a_2. */
1127 rhs0
= ASSERT_EXPR_VAR (expr
);
1128 if (TREE_CODE (rhs0
) == SSA_NAME
)
1129 res
= follow_ssa_edge (loop
, SSA_NAME_DEF_STMT (rhs0
),
1130 halting_phi
, evolution_of_loop
, limit
);
1143 /* Follow the ssa edge into the right hand side of an assignment STMT.
1144 Return true if the strongly connected component has been found. */
1147 follow_ssa_edge_in_rhs (struct loop
*loop
, gimple stmt
,
1148 gimple halting_phi
, tree
*evolution_of_loop
, int limit
)
1150 enum tree_code code
= gimple_assign_rhs_code (stmt
);
1151 tree type
= gimple_expr_type (stmt
), rhs1
, rhs2
;
1157 /* This assignment is under the form "a_1 = (cast) rhs. */
1158 res
= follow_ssa_edge_expr (loop
, stmt
, gimple_assign_rhs1 (stmt
),
1159 halting_phi
, evolution_of_loop
, limit
);
1160 *evolution_of_loop
= chrec_convert (type
, *evolution_of_loop
, stmt
);
1163 case POINTER_PLUS_EXPR
:
1166 rhs1
= gimple_assign_rhs1 (stmt
);
1167 rhs2
= gimple_assign_rhs2 (stmt
);
1168 type
= TREE_TYPE (rhs1
);
1169 res
= follow_ssa_edge_binary (loop
, stmt
, type
, rhs1
, code
, rhs2
,
1170 halting_phi
, evolution_of_loop
, limit
);
1174 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1175 res
= follow_ssa_edge_expr (loop
, stmt
, gimple_assign_rhs1 (stmt
),
1176 halting_phi
, evolution_of_loop
, limit
);
1185 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1188 backedge_phi_arg_p (gimple phi
, int i
)
1190 const_edge e
= gimple_phi_arg_edge (phi
, i
);
1192 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1193 about updating it anywhere, and this should work as well most of the
1195 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
1201 /* Helper function for one branch of the condition-phi-node. Return
1202 true if the strongly connected component has been found following
1205 static inline t_bool
1206 follow_ssa_edge_in_condition_phi_branch (int i
,
1208 gimple condition_phi
,
1210 tree
*evolution_of_branch
,
1211 tree init_cond
, int limit
)
1213 tree branch
= PHI_ARG_DEF (condition_phi
, i
);
1214 *evolution_of_branch
= chrec_dont_know
;
1216 /* Do not follow back edges (they must belong to an irreducible loop, which
1217 we really do not want to worry about). */
1218 if (backedge_phi_arg_p (condition_phi
, i
))
1221 if (TREE_CODE (branch
) == SSA_NAME
)
1223 *evolution_of_branch
= init_cond
;
1224 return follow_ssa_edge (loop
, SSA_NAME_DEF_STMT (branch
), halting_phi
,
1225 evolution_of_branch
, limit
);
1228 /* This case occurs when one of the condition branches sets
1229 the variable to a constant: i.e. a phi-node like
1230 "a_2 = PHI <a_7(5), 2(6)>;".
1232 FIXME: This case have to be refined correctly:
1233 in some cases it is possible to say something better than
1234 chrec_dont_know, for example using a wrap-around notation. */
1238 /* This function merges the branches of a condition-phi-node in a
1242 follow_ssa_edge_in_condition_phi (struct loop
*loop
,
1243 gimple condition_phi
,
1245 tree
*evolution_of_loop
, int limit
)
1248 tree init
= *evolution_of_loop
;
1249 tree evolution_of_branch
;
1250 t_bool res
= follow_ssa_edge_in_condition_phi_branch (0, loop
, condition_phi
,
1252 &evolution_of_branch
,
1254 if (res
== t_false
|| res
== t_dont_know
)
1257 *evolution_of_loop
= evolution_of_branch
;
1259 n
= gimple_phi_num_args (condition_phi
);
1260 for (i
= 1; i
< n
; i
++)
1262 /* Quickly give up when the evolution of one of the branches is
1264 if (*evolution_of_loop
== chrec_dont_know
)
1267 /* Increase the limit by the PHI argument number to avoid exponential
1268 time and memory complexity. */
1269 res
= follow_ssa_edge_in_condition_phi_branch (i
, loop
, condition_phi
,
1271 &evolution_of_branch
,
1273 if (res
== t_false
|| res
== t_dont_know
)
1276 *evolution_of_loop
= chrec_merge (*evolution_of_loop
,
1277 evolution_of_branch
);
1283 /* Follow an SSA edge in an inner loop. It computes the overall
1284 effect of the loop, and following the symbolic initial conditions,
1285 it follows the edges in the parent loop. The inner loop is
1286 considered as a single statement. */
1289 follow_ssa_edge_inner_loop_phi (struct loop
*outer_loop
,
1290 gimple loop_phi_node
,
1292 tree
*evolution_of_loop
, int limit
)
1294 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1295 tree ev
= analyze_scalar_evolution (loop
, PHI_RESULT (loop_phi_node
));
1297 /* Sometimes, the inner loop is too difficult to analyze, and the
1298 result of the analysis is a symbolic parameter. */
1299 if (ev
== PHI_RESULT (loop_phi_node
))
1301 t_bool res
= t_false
;
1302 int i
, n
= gimple_phi_num_args (loop_phi_node
);
1304 for (i
= 0; i
< n
; i
++)
1306 tree arg
= PHI_ARG_DEF (loop_phi_node
, i
);
1309 /* Follow the edges that exit the inner loop. */
1310 bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1311 if (!flow_bb_inside_loop_p (loop
, bb
))
1312 res
= follow_ssa_edge_expr (outer_loop
, loop_phi_node
,
1314 evolution_of_loop
, limit
);
1319 /* If the path crosses this loop-phi, give up. */
1321 *evolution_of_loop
= chrec_dont_know
;
1326 /* Otherwise, compute the overall effect of the inner loop. */
1327 ev
= compute_overall_effect_of_inner_loop (loop
, ev
);
1328 return follow_ssa_edge_expr (outer_loop
, loop_phi_node
, ev
, halting_phi
,
1329 evolution_of_loop
, limit
);
1332 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1333 path that is analyzed on the return walk. */
1336 follow_ssa_edge (struct loop
*loop
, gimple def
, gimple halting_phi
,
1337 tree
*evolution_of_loop
, int limit
)
1339 struct loop
*def_loop
;
1341 if (gimple_nop_p (def
))
1344 /* Give up if the path is longer than the MAX that we allow. */
1345 if (limit
> PARAM_VALUE (PARAM_SCEV_MAX_EXPR_COMPLEXITY
))
1348 def_loop
= loop_containing_stmt (def
);
1350 switch (gimple_code (def
))
1353 if (!loop_phi_node_p (def
))
1354 /* DEF is a condition-phi-node. Follow the branches, and
1355 record their evolutions. Finally, merge the collected
1356 information and set the approximation to the main
1358 return follow_ssa_edge_in_condition_phi
1359 (loop
, def
, halting_phi
, evolution_of_loop
, limit
);
1361 /* When the analyzed phi is the halting_phi, the
1362 depth-first search is over: we have found a path from
1363 the halting_phi to itself in the loop. */
1364 if (def
== halting_phi
)
1367 /* Otherwise, the evolution of the HALTING_PHI depends
1368 on the evolution of another loop-phi-node, i.e. the
1369 evolution function is a higher degree polynomial. */
1370 if (def_loop
== loop
)
1374 if (flow_loop_nested_p (loop
, def_loop
))
1375 return follow_ssa_edge_inner_loop_phi
1376 (loop
, def
, halting_phi
, evolution_of_loop
, limit
+ 1);
1382 return follow_ssa_edge_in_rhs (loop
, def
, halting_phi
,
1383 evolution_of_loop
, limit
);
1386 /* At this level of abstraction, the program is just a set
1387 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
1388 other node to be handled. */
1395 /* Given a LOOP_PHI_NODE, this function determines the evolution
1396 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1399 analyze_evolution_in_loop (gimple loop_phi_node
,
1402 int i
, n
= gimple_phi_num_args (loop_phi_node
);
1403 tree evolution_function
= chrec_not_analyzed_yet
;
1404 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1407 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1409 fprintf (dump_file
, "(analyze_evolution_in_loop \n");
1410 fprintf (dump_file
, " (loop_phi_node = ");
1411 print_gimple_stmt (dump_file
, loop_phi_node
, 0, 0);
1412 fprintf (dump_file
, ")\n");
1415 for (i
= 0; i
< n
; i
++)
1417 tree arg
= PHI_ARG_DEF (loop_phi_node
, i
);
1422 /* Select the edges that enter the loop body. */
1423 bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1424 if (!flow_bb_inside_loop_p (loop
, bb
))
1427 if (TREE_CODE (arg
) == SSA_NAME
)
1431 ssa_chain
= SSA_NAME_DEF_STMT (arg
);
1433 /* Pass in the initial condition to the follow edge function. */
1435 res
= follow_ssa_edge (loop
, ssa_chain
, loop_phi_node
, &ev_fn
, 0);
1437 /* If ev_fn has no evolution in the inner loop, and the
1438 init_cond is not equal to ev_fn, then we have an
1439 ambiguity between two possible values, as we cannot know
1440 the number of iterations at this point. */
1441 if (TREE_CODE (ev_fn
) != POLYNOMIAL_CHREC
1442 && no_evolution_in_loop_p (ev_fn
, loop
->num
, &val
) && val
1443 && !operand_equal_p (init_cond
, ev_fn
, 0))
1444 ev_fn
= chrec_dont_know
;
1449 /* When it is impossible to go back on the same
1450 loop_phi_node by following the ssa edges, the
1451 evolution is represented by a peeled chrec, i.e. the
1452 first iteration, EV_FN has the value INIT_COND, then
1453 all the other iterations it has the value of ARG.
1454 For the moment, PEELED_CHREC nodes are not built. */
1456 ev_fn
= chrec_dont_know
;
1458 /* When there are multiple back edges of the loop (which in fact never
1459 happens currently, but nevertheless), merge their evolutions. */
1460 evolution_function
= chrec_merge (evolution_function
, ev_fn
);
1463 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1465 fprintf (dump_file
, " (evolution_function = ");
1466 print_generic_expr (dump_file
, evolution_function
, 0);
1467 fprintf (dump_file
, "))\n");
1470 return evolution_function
;
1473 /* Given a loop-phi-node, return the initial conditions of the
1474 variable on entry of the loop. When the CCP has propagated
1475 constants into the loop-phi-node, the initial condition is
1476 instantiated, otherwise the initial condition is kept symbolic.
1477 This analyzer does not analyze the evolution outside the current
1478 loop, and leaves this task to the on-demand tree reconstructor. */
1481 analyze_initial_condition (gimple loop_phi_node
)
1484 tree init_cond
= chrec_not_analyzed_yet
;
1485 struct loop
*loop
= loop_containing_stmt (loop_phi_node
);
1487 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1489 fprintf (dump_file
, "(analyze_initial_condition \n");
1490 fprintf (dump_file
, " (loop_phi_node = \n");
1491 print_gimple_stmt (dump_file
, loop_phi_node
, 0, 0);
1492 fprintf (dump_file
, ")\n");
1495 n
= gimple_phi_num_args (loop_phi_node
);
1496 for (i
= 0; i
< n
; i
++)
1498 tree branch
= PHI_ARG_DEF (loop_phi_node
, i
);
1499 basic_block bb
= gimple_phi_arg_edge (loop_phi_node
, i
)->src
;
1501 /* When the branch is oriented to the loop's body, it does
1502 not contribute to the initial condition. */
1503 if (flow_bb_inside_loop_p (loop
, bb
))
1506 if (init_cond
== chrec_not_analyzed_yet
)
1512 if (TREE_CODE (branch
) == SSA_NAME
)
1514 init_cond
= chrec_dont_know
;
1518 init_cond
= chrec_merge (init_cond
, branch
);
1521 /* Ooops -- a loop without an entry??? */
1522 if (init_cond
== chrec_not_analyzed_yet
)
1523 init_cond
= chrec_dont_know
;
1525 /* During early loop unrolling we do not have fully constant propagated IL.
1526 Handle degenerate PHIs here to not miss important unrollings. */
1527 if (TREE_CODE (init_cond
) == SSA_NAME
)
1529 gimple def
= SSA_NAME_DEF_STMT (init_cond
);
1531 if (gimple_code (def
) == GIMPLE_PHI
1532 && (res
= degenerate_phi_result (def
)) != NULL_TREE
1533 /* Only allow invariants here, otherwise we may break
1534 loop-closed SSA form. */
1535 && is_gimple_min_invariant (res
))
1539 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1541 fprintf (dump_file
, " (init_cond = ");
1542 print_generic_expr (dump_file
, init_cond
, 0);
1543 fprintf (dump_file
, "))\n");
1549 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1552 interpret_loop_phi (struct loop
*loop
, gimple loop_phi_node
)
1555 struct loop
*phi_loop
= loop_containing_stmt (loop_phi_node
);
1558 if (phi_loop
!= loop
)
1560 struct loop
*subloop
;
1561 tree evolution_fn
= analyze_scalar_evolution
1562 (phi_loop
, PHI_RESULT (loop_phi_node
));
1564 /* Dive one level deeper. */
1565 subloop
= superloop_at_depth (phi_loop
, loop_depth (loop
) + 1);
1567 /* Interpret the subloop. */
1568 res
= compute_overall_effect_of_inner_loop (subloop
, evolution_fn
);
1572 /* Otherwise really interpret the loop phi. */
1573 init_cond
= analyze_initial_condition (loop_phi_node
);
1574 res
= analyze_evolution_in_loop (loop_phi_node
, init_cond
);
1576 /* Verify we maintained the correct initial condition throughout
1577 possible conversions in the SSA chain. */
1578 if (res
!= chrec_dont_know
)
1580 tree new_init
= res
;
1581 if (CONVERT_EXPR_P (res
)
1582 && TREE_CODE (TREE_OPERAND (res
, 0)) == POLYNOMIAL_CHREC
)
1583 new_init
= fold_convert (TREE_TYPE (res
),
1584 CHREC_LEFT (TREE_OPERAND (res
, 0)));
1585 else if (TREE_CODE (res
) == POLYNOMIAL_CHREC
)
1586 new_init
= CHREC_LEFT (res
);
1587 STRIP_USELESS_TYPE_CONVERSION (new_init
);
1588 if (TREE_CODE (new_init
) == POLYNOMIAL_CHREC
1589 || !operand_equal_p (init_cond
, new_init
, 0))
1590 return chrec_dont_know
;
1596 /* This function merges the branches of a condition-phi-node,
1597 contained in the outermost loop, and whose arguments are already
1601 interpret_condition_phi (struct loop
*loop
, gimple condition_phi
)
1603 int i
, n
= gimple_phi_num_args (condition_phi
);
1604 tree res
= chrec_not_analyzed_yet
;
1606 for (i
= 0; i
< n
; i
++)
1610 if (backedge_phi_arg_p (condition_phi
, i
))
1612 res
= chrec_dont_know
;
1616 branch_chrec
= analyze_scalar_evolution
1617 (loop
, PHI_ARG_DEF (condition_phi
, i
));
1619 res
= chrec_merge (res
, branch_chrec
);
1625 /* Interpret the operation RHS1 OP RHS2. If we didn't
1626 analyze this node before, follow the definitions until ending
1627 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1628 return path, this function propagates evolutions (ala constant copy
1629 propagation). OPND1 is not a GIMPLE expression because we could
1630 analyze the effect of an inner loop: see interpret_loop_phi. */
1633 interpret_rhs_expr (struct loop
*loop
, gimple at_stmt
,
1634 tree type
, tree rhs1
, enum tree_code code
, tree rhs2
)
1636 tree res
, chrec1
, chrec2
;
1639 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1641 if (is_gimple_min_invariant (rhs1
))
1642 return chrec_convert (type
, rhs1
, at_stmt
);
1644 if (code
== SSA_NAME
)
1645 return chrec_convert (type
, analyze_scalar_evolution (loop
, rhs1
),
1648 if (code
== ASSERT_EXPR
)
1650 rhs1
= ASSERT_EXPR_VAR (rhs1
);
1651 return chrec_convert (type
, analyze_scalar_evolution (loop
, rhs1
),
1659 if (TREE_CODE (TREE_OPERAND (rhs1
, 0)) == MEM_REF
1660 || handled_component_p (TREE_OPERAND (rhs1
, 0)))
1662 enum machine_mode mode
;
1663 HOST_WIDE_INT bitsize
, bitpos
;
1670 base
= get_inner_reference (TREE_OPERAND (rhs1
, 0),
1671 &bitsize
, &bitpos
, &offset
,
1672 &mode
, &unsignedp
, &volatilep
, false);
1674 if (TREE_CODE (base
) == MEM_REF
)
1676 rhs2
= TREE_OPERAND (base
, 1);
1677 rhs1
= TREE_OPERAND (base
, 0);
1679 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1680 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1681 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1682 chrec2
= chrec_convert (TREE_TYPE (rhs2
), chrec2
, at_stmt
);
1683 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1687 chrec1
= analyze_scalar_evolution_for_address_of (loop
, base
);
1688 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1692 if (offset
!= NULL_TREE
)
1694 chrec2
= analyze_scalar_evolution (loop
, offset
);
1695 chrec2
= chrec_convert (TREE_TYPE (offset
), chrec2
, at_stmt
);
1696 res
= chrec_fold_plus (type
, res
, chrec2
);
1701 gcc_assert ((bitpos
% BITS_PER_UNIT
) == 0);
1703 unitpos
= size_int (bitpos
/ BITS_PER_UNIT
);
1704 chrec3
= analyze_scalar_evolution (loop
, unitpos
);
1705 chrec3
= chrec_convert (TREE_TYPE (unitpos
), chrec3
, at_stmt
);
1706 res
= chrec_fold_plus (type
, res
, chrec3
);
1710 res
= chrec_dont_know
;
1713 case POINTER_PLUS_EXPR
:
1714 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1715 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1716 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1717 chrec2
= chrec_convert (TREE_TYPE (rhs2
), chrec2
, at_stmt
);
1718 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1722 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1723 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1724 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1725 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1726 res
= chrec_fold_plus (type
, chrec1
, chrec2
);
1730 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1731 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1732 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1733 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1734 res
= chrec_fold_minus (type
, chrec1
, chrec2
);
1738 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1739 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1740 /* TYPE may be integer, real or complex, so use fold_convert. */
1741 res
= chrec_fold_multiply (type
, chrec1
,
1742 fold_convert (type
, integer_minus_one_node
));
1746 /* Handle ~X as -1 - X. */
1747 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1748 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1749 res
= chrec_fold_minus (type
,
1750 fold_convert (type
, integer_minus_one_node
),
1755 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1756 chrec2
= analyze_scalar_evolution (loop
, rhs2
);
1757 chrec1
= chrec_convert (type
, chrec1
, at_stmt
);
1758 chrec2
= chrec_convert (type
, chrec2
, at_stmt
);
1759 res
= chrec_fold_multiply (type
, chrec1
, chrec2
);
1763 /* In case we have a truncation of a widened operation that in
1764 the truncated type has undefined overflow behavior analyze
1765 the operation done in an unsigned type of the same precision
1766 as the final truncation. We cannot derive a scalar evolution
1767 for the widened operation but for the truncated result. */
1768 if (TREE_CODE (type
) == INTEGER_TYPE
1769 && TREE_CODE (TREE_TYPE (rhs1
)) == INTEGER_TYPE
1770 && TYPE_PRECISION (type
) < TYPE_PRECISION (TREE_TYPE (rhs1
))
1771 && TYPE_OVERFLOW_UNDEFINED (type
)
1772 && TREE_CODE (rhs1
) == SSA_NAME
1773 && (def
= SSA_NAME_DEF_STMT (rhs1
))
1774 && is_gimple_assign (def
)
1775 && TREE_CODE_CLASS (gimple_assign_rhs_code (def
)) == tcc_binary
1776 && TREE_CODE (gimple_assign_rhs2 (def
)) == INTEGER_CST
)
1778 tree utype
= unsigned_type_for (type
);
1779 chrec1
= interpret_rhs_expr (loop
, at_stmt
, utype
,
1780 gimple_assign_rhs1 (def
),
1781 gimple_assign_rhs_code (def
),
1782 gimple_assign_rhs2 (def
));
1785 chrec1
= analyze_scalar_evolution (loop
, rhs1
);
1786 res
= chrec_convert (type
, chrec1
, at_stmt
);
1790 res
= chrec_dont_know
;
1797 /* Interpret the expression EXPR. */
1800 interpret_expr (struct loop
*loop
, gimple at_stmt
, tree expr
)
1802 enum tree_code code
;
1803 tree type
= TREE_TYPE (expr
), op0
, op1
;
1805 if (automatically_generated_chrec_p (expr
))
1808 if (TREE_CODE (expr
) == POLYNOMIAL_CHREC
1809 || get_gimple_rhs_class (TREE_CODE (expr
)) == GIMPLE_TERNARY_RHS
)
1810 return chrec_dont_know
;
1812 extract_ops_from_tree (expr
, &code
, &op0
, &op1
);
1814 return interpret_rhs_expr (loop
, at_stmt
, type
,
1818 /* Interpret the rhs of the assignment STMT. */
1821 interpret_gimple_assign (struct loop
*loop
, gimple stmt
)
1823 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
1824 enum tree_code code
= gimple_assign_rhs_code (stmt
);
1826 return interpret_rhs_expr (loop
, stmt
, type
,
1827 gimple_assign_rhs1 (stmt
), code
,
1828 gimple_assign_rhs2 (stmt
));
1833 /* This section contains all the entry points:
1834 - number_of_iterations_in_loop,
1835 - analyze_scalar_evolution,
1836 - instantiate_parameters.
1839 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1840 common ancestor of DEF_LOOP and USE_LOOP. */
1843 compute_scalar_evolution_in_loop (struct loop
*wrto_loop
,
1844 struct loop
*def_loop
,
1850 if (def_loop
== wrto_loop
)
1853 def_loop
= superloop_at_depth (def_loop
, loop_depth (wrto_loop
) + 1);
1854 res
= compute_overall_effect_of_inner_loop (def_loop
, ev
);
1856 if (no_evolution_in_loop_p (res
, wrto_loop
->num
, &val
) && val
)
1859 return analyze_scalar_evolution_1 (wrto_loop
, res
, chrec_not_analyzed_yet
);
1862 /* Helper recursive function. */
1865 analyze_scalar_evolution_1 (struct loop
*loop
, tree var
, tree res
)
1867 tree type
= TREE_TYPE (var
);
1870 struct loop
*def_loop
;
1872 if (loop
== NULL
|| TREE_CODE (type
) == VECTOR_TYPE
)
1873 return chrec_dont_know
;
1875 if (TREE_CODE (var
) != SSA_NAME
)
1876 return interpret_expr (loop
, NULL
, var
);
1878 def
= SSA_NAME_DEF_STMT (var
);
1879 bb
= gimple_bb (def
);
1880 def_loop
= bb
? bb
->loop_father
: NULL
;
1883 || !flow_bb_inside_loop_p (loop
, bb
))
1885 /* Keep the symbolic form. */
1890 if (res
!= chrec_not_analyzed_yet
)
1892 if (loop
!= bb
->loop_father
)
1893 res
= compute_scalar_evolution_in_loop
1894 (find_common_loop (loop
, bb
->loop_father
), bb
->loop_father
, res
);
1899 if (loop
!= def_loop
)
1901 res
= analyze_scalar_evolution_1 (def_loop
, var
, chrec_not_analyzed_yet
);
1902 res
= compute_scalar_evolution_in_loop (loop
, def_loop
, res
);
1907 switch (gimple_code (def
))
1910 res
= interpret_gimple_assign (loop
, def
);
1914 if (loop_phi_node_p (def
))
1915 res
= interpret_loop_phi (loop
, def
);
1917 res
= interpret_condition_phi (loop
, def
);
1921 res
= chrec_dont_know
;
1927 /* Keep the symbolic form. */
1928 if (res
== chrec_dont_know
)
1931 if (loop
== def_loop
)
1932 set_scalar_evolution (block_before_loop (loop
), var
, res
);
1937 /* Analyzes and returns the scalar evolution of the ssa_name VAR in
1938 LOOP. LOOP is the loop in which the variable is used.
1940 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
1941 pointer to the statement that uses this variable, in order to
1942 determine the evolution function of the variable, use the following
1945 loop_p loop = loop_containing_stmt (stmt);
1946 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
1947 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
1951 analyze_scalar_evolution (struct loop
*loop
, tree var
)
1955 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1957 fprintf (dump_file
, "(analyze_scalar_evolution \n");
1958 fprintf (dump_file
, " (loop_nb = %d)\n", loop
->num
);
1959 fprintf (dump_file
, " (scalar = ");
1960 print_generic_expr (dump_file
, var
, 0);
1961 fprintf (dump_file
, ")\n");
1964 res
= get_scalar_evolution (block_before_loop (loop
), var
);
1965 res
= analyze_scalar_evolution_1 (loop
, var
, res
);
1967 if (dump_file
&& (dump_flags
& TDF_SCEV
))
1968 fprintf (dump_file
, ")\n");
1973 /* Analyzes and returns the scalar evolution of VAR address in LOOP. */
1976 analyze_scalar_evolution_for_address_of (struct loop
*loop
, tree var
)
1978 return analyze_scalar_evolution (loop
, build_fold_addr_expr (var
));
1981 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
1982 WRTO_LOOP (which should be a superloop of USE_LOOP)
1984 FOLDED_CASTS is set to true if resolve_mixers used
1985 chrec_convert_aggressive (TODO -- not really, we are way too conservative
1986 at the moment in order to keep things simple).
1988 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
1991 for (i = 0; i < 100; i++) -- loop 1
1993 for (j = 0; j < 100; j++) -- loop 2
2000 for (t = 0; t < 100; t++) -- loop 3
2007 Both k1 and k2 are invariants in loop3, thus
2008 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
2009 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
2011 As they are invariant, it does not matter whether we consider their
2012 usage in loop 3 or loop 2, hence
2013 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
2014 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
2015 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
2016 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
2018 Similarly for their evolutions with respect to loop 1. The values of K2
2019 in the use in loop 2 vary independently on loop 1, thus we cannot express
2020 the evolution with respect to loop 1:
2021 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2022 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2023 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2024 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2026 The value of k2 in the use in loop 1 is known, though:
2027 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2028 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2032 analyze_scalar_evolution_in_loop (struct loop
*wrto_loop
, struct loop
*use_loop
,
2033 tree version
, bool *folded_casts
)
2036 tree ev
= version
, tmp
;
2038 /* We cannot just do
2040 tmp = analyze_scalar_evolution (use_loop, version);
2041 ev = resolve_mixers (wrto_loop, tmp);
2043 as resolve_mixers would query the scalar evolution with respect to
2044 wrto_loop. For example, in the situation described in the function
2045 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2048 analyze_scalar_evolution (use_loop, version) = k2
2050 and resolve_mixers (loop1, k2) finds that the value of k2 in loop 1
2051 is 100, which is a wrong result, since we are interested in the
2054 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2055 each time checking that there is no evolution in the inner loop. */
2058 *folded_casts
= false;
2061 tmp
= analyze_scalar_evolution (use_loop
, ev
);
2062 ev
= resolve_mixers (use_loop
, tmp
);
2064 if (folded_casts
&& tmp
!= ev
)
2065 *folded_casts
= true;
2067 if (use_loop
== wrto_loop
)
2070 /* If the value of the use changes in the inner loop, we cannot express
2071 its value in the outer loop (we might try to return interval chrec,
2072 but we do not have a user for it anyway) */
2073 if (!no_evolution_in_loop_p (ev
, use_loop
->num
, &val
)
2075 return chrec_dont_know
;
2077 use_loop
= loop_outer (use_loop
);
2081 /* Returns from CACHE the value for VERSION instantiated below
2082 INSTANTIATED_BELOW block. */
2085 get_instantiated_value (htab_t cache
, basic_block instantiated_below
,
2088 struct scev_info_str
*info
, pattern
;
2090 pattern
.var
= version
;
2091 pattern
.instantiated_below
= instantiated_below
;
2092 info
= (struct scev_info_str
*) htab_find (cache
, &pattern
);
2100 /* Sets in CACHE the value of VERSION instantiated below basic block
2101 INSTANTIATED_BELOW to VAL. */
2104 set_instantiated_value (htab_t cache
, basic_block instantiated_below
,
2105 tree version
, tree val
)
2107 struct scev_info_str
*info
, pattern
;
2110 pattern
.var
= version
;
2111 pattern
.instantiated_below
= instantiated_below
;
2112 slot
= htab_find_slot (cache
, &pattern
, INSERT
);
2115 *slot
= new_scev_info_str (instantiated_below
, version
);
2116 info
= (struct scev_info_str
*) *slot
;
2120 /* Return the closed_loop_phi node for VAR. If there is none, return
2124 loop_closed_phi_def (tree var
)
2129 gimple_stmt_iterator psi
;
2131 if (var
== NULL_TREE
2132 || TREE_CODE (var
) != SSA_NAME
)
2135 loop
= loop_containing_stmt (SSA_NAME_DEF_STMT (var
));
2136 exit
= single_exit (loop
);
2140 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); gsi_next (&psi
))
2142 phi
= gsi_stmt (psi
);
2143 if (PHI_ARG_DEF_FROM_EDGE (phi
, exit
) == var
)
2144 return PHI_RESULT (phi
);
2150 static tree
instantiate_scev_r (basic_block
, struct loop
*, tree
, bool,
2153 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2154 and EVOLUTION_LOOP, that were left under a symbolic form.
2156 CHREC is an SSA_NAME to be instantiated.
2158 CACHE is the cache of already instantiated values.
2160 FOLD_CONVERSIONS should be set to true when the conversions that
2161 may wrap in signed/pointer type are folded, as long as the value of
2162 the chrec is preserved.
2164 SIZE_EXPR is used for computing the size of the expression to be
2165 instantiated, and to stop if it exceeds some limit. */
2168 instantiate_scev_name (basic_block instantiate_below
,
2169 struct loop
*evolution_loop
, tree chrec
,
2170 bool fold_conversions
, htab_t cache
, int size_expr
)
2173 struct loop
*def_loop
;
2174 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (chrec
));
2176 /* A parameter (or loop invariant and we do not want to include
2177 evolutions in outer loops), nothing to do. */
2179 || loop_depth (def_bb
->loop_father
) == 0
2180 || dominated_by_p (CDI_DOMINATORS
, instantiate_below
, def_bb
))
2183 /* We cache the value of instantiated variable to avoid exponential
2184 time complexity due to reevaluations. We also store the convenient
2185 value in the cache in order to prevent infinite recursion -- we do
2186 not want to instantiate the SSA_NAME if it is in a mixer
2187 structure. This is used for avoiding the instantiation of
2188 recursively defined functions, such as:
2190 | a_2 -> {0, +, 1, +, a_2}_1 */
2192 res
= get_instantiated_value (cache
, instantiate_below
, chrec
);
2196 res
= chrec_dont_know
;
2197 set_instantiated_value (cache
, instantiate_below
, chrec
, res
);
2199 def_loop
= find_common_loop (evolution_loop
, def_bb
->loop_father
);
2201 /* If the analysis yields a parametric chrec, instantiate the
2203 res
= analyze_scalar_evolution (def_loop
, chrec
);
2205 /* Don't instantiate default definitions. */
2206 if (TREE_CODE (res
) == SSA_NAME
2207 && SSA_NAME_IS_DEFAULT_DEF (res
))
2210 /* Don't instantiate loop-closed-ssa phi nodes. */
2211 else if (TREE_CODE (res
) == SSA_NAME
2212 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res
)))
2213 > loop_depth (def_loop
))
2216 res
= loop_closed_phi_def (chrec
);
2220 /* When there is no loop_closed_phi_def, it means that the
2221 variable is not used after the loop: try to still compute the
2222 value of the variable when exiting the loop. */
2223 if (res
== NULL_TREE
)
2225 loop_p loop
= loop_containing_stmt (SSA_NAME_DEF_STMT (chrec
));
2226 res
= analyze_scalar_evolution (loop
, chrec
);
2227 res
= compute_overall_effect_of_inner_loop (loop
, res
);
2228 res
= instantiate_scev_r (instantiate_below
, evolution_loop
, res
,
2229 fold_conversions
, cache
, size_expr
);
2231 else if (!dominated_by_p (CDI_DOMINATORS
, instantiate_below
,
2232 gimple_bb (SSA_NAME_DEF_STMT (res
))))
2233 res
= chrec_dont_know
;
2236 else if (res
!= chrec_dont_know
)
2237 res
= instantiate_scev_r (instantiate_below
, evolution_loop
, res
,
2238 fold_conversions
, cache
, size_expr
);
2240 /* Store the correct value to the cache. */
2241 set_instantiated_value (cache
, instantiate_below
, chrec
, res
);
2245 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2246 and EVOLUTION_LOOP, that were left under a symbolic form.
2248 CHREC is a polynomial chain of recurrence to be instantiated.
2250 CACHE is the cache of already instantiated values.
2252 FOLD_CONVERSIONS should be set to true when the conversions that
2253 may wrap in signed/pointer type are folded, as long as the value of
2254 the chrec is preserved.
2256 SIZE_EXPR is used for computing the size of the expression to be
2257 instantiated, and to stop if it exceeds some limit. */
2260 instantiate_scev_poly (basic_block instantiate_below
,
2261 struct loop
*evolution_loop
, tree chrec
,
2262 bool fold_conversions
, htab_t cache
, int size_expr
)
2265 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2266 CHREC_LEFT (chrec
), fold_conversions
, cache
,
2268 if (op0
== chrec_dont_know
)
2269 return chrec_dont_know
;
2271 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2272 CHREC_RIGHT (chrec
), fold_conversions
, cache
,
2274 if (op1
== chrec_dont_know
)
2275 return chrec_dont_know
;
2277 if (CHREC_LEFT (chrec
) != op0
2278 || CHREC_RIGHT (chrec
) != op1
)
2280 unsigned var
= CHREC_VARIABLE (chrec
);
2282 /* When the instantiated stride or base has an evolution in an
2283 innermost loop, return chrec_dont_know, as this is not a
2284 valid SCEV representation. In the reduced testcase for
2285 PR40281 we would have {0, +, {1, +, 1}_2}_1 that has no
2287 if ((tree_is_chrec (op0
) && CHREC_VARIABLE (op0
) > var
)
2288 || (tree_is_chrec (op1
) && CHREC_VARIABLE (op1
) > var
))
2289 return chrec_dont_know
;
2291 op1
= chrec_convert_rhs (chrec_type (op0
), op1
, NULL
);
2292 chrec
= build_polynomial_chrec (var
, op0
, op1
);
2298 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2299 and EVOLUTION_LOOP, that were left under a symbolic form.
2301 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
2303 CACHE is the cache of already instantiated values.
2305 FOLD_CONVERSIONS should be set to true when the conversions that
2306 may wrap in signed/pointer type are folded, as long as the value of
2307 the chrec is preserved.
2309 SIZE_EXPR is used for computing the size of the expression to be
2310 instantiated, and to stop if it exceeds some limit. */
2313 instantiate_scev_binary (basic_block instantiate_below
,
2314 struct loop
*evolution_loop
, tree chrec
, enum tree_code code
,
2315 tree type
, tree c0
, tree c1
,
2316 bool fold_conversions
, htab_t cache
, int size_expr
)
2319 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2320 c0
, fold_conversions
, cache
,
2322 if (op0
== chrec_dont_know
)
2323 return chrec_dont_know
;
2325 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2326 c1
, fold_conversions
, cache
,
2328 if (op1
== chrec_dont_know
)
2329 return chrec_dont_know
;
2334 op0
= chrec_convert (type
, op0
, NULL
);
2335 op1
= chrec_convert_rhs (type
, op1
, NULL
);
2339 case POINTER_PLUS_EXPR
:
2341 return chrec_fold_plus (type
, op0
, op1
);
2344 return chrec_fold_minus (type
, op0
, op1
);
2347 return chrec_fold_multiply (type
, op0
, op1
);
2354 return chrec
? chrec
: fold_build2 (code
, type
, c0
, c1
);
2357 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2358 and EVOLUTION_LOOP, that were left under a symbolic form.
2360 "CHREC" is an array reference to be instantiated.
2362 CACHE is the cache of already instantiated values.
2364 FOLD_CONVERSIONS should be set to true when the conversions that
2365 may wrap in signed/pointer type are folded, as long as the value of
2366 the chrec is preserved.
2368 SIZE_EXPR is used for computing the size of the expression to be
2369 instantiated, and to stop if it exceeds some limit. */
2372 instantiate_array_ref (basic_block instantiate_below
,
2373 struct loop
*evolution_loop
, tree chrec
,
2374 bool fold_conversions
, htab_t cache
, int size_expr
)
2377 tree index
= TREE_OPERAND (chrec
, 1);
2378 tree op1
= instantiate_scev_r (instantiate_below
, evolution_loop
, index
,
2379 fold_conversions
, cache
, size_expr
);
2381 if (op1
== chrec_dont_know
)
2382 return chrec_dont_know
;
2384 if (chrec
&& op1
== index
)
2387 res
= unshare_expr (chrec
);
2388 TREE_OPERAND (res
, 1) = op1
;
2392 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2393 and EVOLUTION_LOOP, that were left under a symbolic form.
2395 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2398 CACHE is the cache of already instantiated values.
2400 FOLD_CONVERSIONS should be set to true when the conversions that
2401 may wrap in signed/pointer type are folded, as long as the value of
2402 the chrec is preserved.
2404 SIZE_EXPR is used for computing the size of the expression to be
2405 instantiated, and to stop if it exceeds some limit. */
2408 instantiate_scev_convert (basic_block instantiate_below
,
2409 struct loop
*evolution_loop
, tree chrec
,
2411 bool fold_conversions
, htab_t cache
, int size_expr
)
2413 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
, op
,
2414 fold_conversions
, cache
, size_expr
);
2416 if (op0
== chrec_dont_know
)
2417 return chrec_dont_know
;
2419 if (fold_conversions
)
2421 tree tmp
= chrec_convert_aggressive (type
, op0
);
2426 if (chrec
&& op0
== op
)
2429 /* If we used chrec_convert_aggressive, we can no longer assume that
2430 signed chrecs do not overflow, as chrec_convert does, so avoid
2431 calling it in that case. */
2432 if (fold_conversions
)
2433 return fold_convert (type
, op0
);
2435 return chrec_convert (type
, op0
, NULL
);
2438 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2439 and EVOLUTION_LOOP, that were left under a symbolic form.
2441 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
2442 Handle ~X as -1 - X.
2443 Handle -X as -1 * X.
2445 CACHE is the cache of already instantiated values.
2447 FOLD_CONVERSIONS should be set to true when the conversions that
2448 may wrap in signed/pointer type are folded, as long as the value of
2449 the chrec is preserved.
2451 SIZE_EXPR is used for computing the size of the expression to be
2452 instantiated, and to stop if it exceeds some limit. */
2455 instantiate_scev_not (basic_block instantiate_below
,
2456 struct loop
*evolution_loop
, tree chrec
,
2457 enum tree_code code
, tree type
, tree op
,
2458 bool fold_conversions
, htab_t cache
, int size_expr
)
2460 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
, op
,
2461 fold_conversions
, cache
, size_expr
);
2463 if (op0
== chrec_dont_know
)
2464 return chrec_dont_know
;
2468 op0
= chrec_convert (type
, op0
, NULL
);
2473 return chrec_fold_minus
2474 (type
, fold_convert (type
, integer_minus_one_node
), op0
);
2477 return chrec_fold_multiply
2478 (type
, fold_convert (type
, integer_minus_one_node
), op0
);
2485 return chrec
? chrec
: fold_build1 (code
, type
, op0
);
2488 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2489 and EVOLUTION_LOOP, that were left under a symbolic form.
2491 CHREC is an expression with 3 operands to be instantiated.
2493 CACHE is the cache of already instantiated values.
2495 FOLD_CONVERSIONS should be set to true when the conversions that
2496 may wrap in signed/pointer type are folded, as long as the value of
2497 the chrec is preserved.
2499 SIZE_EXPR is used for computing the size of the expression to be
2500 instantiated, and to stop if it exceeds some limit. */
2503 instantiate_scev_3 (basic_block instantiate_below
,
2504 struct loop
*evolution_loop
, tree chrec
,
2505 bool fold_conversions
, htab_t cache
, int size_expr
)
2508 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2509 TREE_OPERAND (chrec
, 0),
2510 fold_conversions
, cache
, size_expr
);
2511 if (op0
== chrec_dont_know
)
2512 return chrec_dont_know
;
2514 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2515 TREE_OPERAND (chrec
, 1),
2516 fold_conversions
, cache
, size_expr
);
2517 if (op1
== chrec_dont_know
)
2518 return chrec_dont_know
;
2520 op2
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2521 TREE_OPERAND (chrec
, 2),
2522 fold_conversions
, cache
, size_expr
);
2523 if (op2
== chrec_dont_know
)
2524 return chrec_dont_know
;
2526 if (op0
== TREE_OPERAND (chrec
, 0)
2527 && op1
== TREE_OPERAND (chrec
, 1)
2528 && op2
== TREE_OPERAND (chrec
, 2))
2531 return fold_build3 (TREE_CODE (chrec
),
2532 TREE_TYPE (chrec
), op0
, op1
, op2
);
2535 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2536 and EVOLUTION_LOOP, that were left under a symbolic form.
2538 CHREC is an expression with 2 operands to be instantiated.
2540 CACHE is the cache of already instantiated values.
2542 FOLD_CONVERSIONS should be set to true when the conversions that
2543 may wrap in signed/pointer type are folded, as long as the value of
2544 the chrec is preserved.
2546 SIZE_EXPR is used for computing the size of the expression to be
2547 instantiated, and to stop if it exceeds some limit. */
2550 instantiate_scev_2 (basic_block instantiate_below
,
2551 struct loop
*evolution_loop
, tree chrec
,
2552 bool fold_conversions
, htab_t cache
, int size_expr
)
2555 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2556 TREE_OPERAND (chrec
, 0),
2557 fold_conversions
, cache
, size_expr
);
2558 if (op0
== chrec_dont_know
)
2559 return chrec_dont_know
;
2561 op1
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2562 TREE_OPERAND (chrec
, 1),
2563 fold_conversions
, cache
, size_expr
);
2564 if (op1
== chrec_dont_know
)
2565 return chrec_dont_know
;
2567 if (op0
== TREE_OPERAND (chrec
, 0)
2568 && op1
== TREE_OPERAND (chrec
, 1))
2571 return fold_build2 (TREE_CODE (chrec
), TREE_TYPE (chrec
), op0
, op1
);
2574 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2575 and EVOLUTION_LOOP, that were left under a symbolic form.
2577 CHREC is an expression with 2 operands to be instantiated.
2579 CACHE is the cache of already instantiated values.
2581 FOLD_CONVERSIONS should be set to true when the conversions that
2582 may wrap in signed/pointer type are folded, as long as the value of
2583 the chrec is preserved.
2585 SIZE_EXPR is used for computing the size of the expression to be
2586 instantiated, and to stop if it exceeds some limit. */
2589 instantiate_scev_1 (basic_block instantiate_below
,
2590 struct loop
*evolution_loop
, tree chrec
,
2591 bool fold_conversions
, htab_t cache
, int size_expr
)
2593 tree op0
= instantiate_scev_r (instantiate_below
, evolution_loop
,
2594 TREE_OPERAND (chrec
, 0),
2595 fold_conversions
, cache
, size_expr
);
2597 if (op0
== chrec_dont_know
)
2598 return chrec_dont_know
;
2600 if (op0
== TREE_OPERAND (chrec
, 0))
2603 return fold_build1 (TREE_CODE (chrec
), TREE_TYPE (chrec
), op0
);
2606 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2607 and EVOLUTION_LOOP, that were left under a symbolic form.
2609 CHREC is the scalar evolution to instantiate.
2611 CACHE is the cache of already instantiated values.
2613 FOLD_CONVERSIONS should be set to true when the conversions that
2614 may wrap in signed/pointer type are folded, as long as the value of
2615 the chrec is preserved.
2617 SIZE_EXPR is used for computing the size of the expression to be
2618 instantiated, and to stop if it exceeds some limit. */
2621 instantiate_scev_r (basic_block instantiate_below
,
2622 struct loop
*evolution_loop
, tree chrec
,
2623 bool fold_conversions
, htab_t cache
, int size_expr
)
2625 /* Give up if the expression is larger than the MAX that we allow. */
2626 if (size_expr
++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE
))
2627 return chrec_dont_know
;
2629 if (chrec
== NULL_TREE
2630 || automatically_generated_chrec_p (chrec
)
2631 || is_gimple_min_invariant (chrec
))
2634 switch (TREE_CODE (chrec
))
2637 return instantiate_scev_name (instantiate_below
, evolution_loop
, chrec
,
2638 fold_conversions
, cache
, size_expr
);
2640 case POLYNOMIAL_CHREC
:
2641 return instantiate_scev_poly (instantiate_below
, evolution_loop
, chrec
,
2642 fold_conversions
, cache
, size_expr
);
2644 case POINTER_PLUS_EXPR
:
2648 return instantiate_scev_binary (instantiate_below
, evolution_loop
, chrec
,
2649 TREE_CODE (chrec
), chrec_type (chrec
),
2650 TREE_OPERAND (chrec
, 0),
2651 TREE_OPERAND (chrec
, 1),
2652 fold_conversions
, cache
, size_expr
);
2655 return instantiate_scev_convert (instantiate_below
, evolution_loop
, chrec
,
2656 TREE_TYPE (chrec
), TREE_OPERAND (chrec
, 0),
2657 fold_conversions
, cache
, size_expr
);
2661 return instantiate_scev_not (instantiate_below
, evolution_loop
, chrec
,
2662 TREE_CODE (chrec
), TREE_TYPE (chrec
),
2663 TREE_OPERAND (chrec
, 0),
2664 fold_conversions
, cache
, size_expr
);
2667 case SCEV_NOT_KNOWN
:
2668 return chrec_dont_know
;
2674 return instantiate_array_ref (instantiate_below
, evolution_loop
, chrec
,
2675 fold_conversions
, cache
, size_expr
);
2681 if (VL_EXP_CLASS_P (chrec
))
2682 return chrec_dont_know
;
2684 switch (TREE_CODE_LENGTH (TREE_CODE (chrec
)))
2687 return instantiate_scev_3 (instantiate_below
, evolution_loop
, chrec
,
2688 fold_conversions
, cache
, size_expr
);
2691 return instantiate_scev_2 (instantiate_below
, evolution_loop
, chrec
,
2692 fold_conversions
, cache
, size_expr
);
2695 return instantiate_scev_1 (instantiate_below
, evolution_loop
, chrec
,
2696 fold_conversions
, cache
, size_expr
);
2705 /* Too complicated to handle. */
2706 return chrec_dont_know
;
2709 /* Analyze all the parameters of the chrec that were left under a
2710 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2711 recursive instantiation of parameters: a parameter is a variable
2712 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2713 a function parameter. */
2716 instantiate_scev (basic_block instantiate_below
, struct loop
*evolution_loop
,
2720 htab_t cache
= htab_create (10, hash_scev_info
, eq_scev_info
, del_scev_info
);
2722 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2724 fprintf (dump_file
, "(instantiate_scev \n");
2725 fprintf (dump_file
, " (instantiate_below = %d)\n", instantiate_below
->index
);
2726 fprintf (dump_file
, " (evolution_loop = %d)\n", evolution_loop
->num
);
2727 fprintf (dump_file
, " (chrec = ");
2728 print_generic_expr (dump_file
, chrec
, 0);
2729 fprintf (dump_file
, ")\n");
2732 res
= instantiate_scev_r (instantiate_below
, evolution_loop
, chrec
, false,
2735 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2737 fprintf (dump_file
, " (res = ");
2738 print_generic_expr (dump_file
, res
, 0);
2739 fprintf (dump_file
, "))\n");
2742 htab_delete (cache
);
2747 /* Similar to instantiate_parameters, but does not introduce the
2748 evolutions in outer loops for LOOP invariants in CHREC, and does not
2749 care about causing overflows, as long as they do not affect value
2750 of an expression. */
2753 resolve_mixers (struct loop
*loop
, tree chrec
)
2755 htab_t cache
= htab_create (10, hash_scev_info
, eq_scev_info
, del_scev_info
);
2756 tree ret
= instantiate_scev_r (block_before_loop (loop
), loop
, chrec
, true,
2758 htab_delete (cache
);
2762 /* Entry point for the analysis of the number of iterations pass.
2763 This function tries to safely approximate the number of iterations
2764 the loop will run. When this property is not decidable at compile
2765 time, the result is chrec_dont_know. Otherwise the result is a
2766 scalar or a symbolic parameter. When the number of iterations may
2767 be equal to zero and the property cannot be determined at compile
2768 time, the result is a COND_EXPR that represents in a symbolic form
2769 the conditions under which the number of iterations is not zero.
2771 Example of analysis: suppose that the loop has an exit condition:
2773 "if (b > 49) goto end_loop;"
2775 and that in a previous analysis we have determined that the
2776 variable 'b' has an evolution function:
2778 "EF = {23, +, 5}_2".
2780 When we evaluate the function at the point 5, i.e. the value of the
2781 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2782 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2783 the loop body has been executed 6 times. */
2786 number_of_latch_executions (struct loop
*loop
)
2789 struct tree_niter_desc niter_desc
;
2793 /* Determine whether the number of iterations in loop has already
2795 res
= loop
->nb_iterations
;
2799 may_be_zero
= NULL_TREE
;
2801 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2802 fprintf (dump_file
, "(number_of_iterations_in_loop = \n");
2804 res
= chrec_dont_know
;
2805 exit
= single_exit (loop
);
2807 if (exit
&& number_of_iterations_exit (loop
, exit
, &niter_desc
, false))
2809 may_be_zero
= niter_desc
.may_be_zero
;
2810 res
= niter_desc
.niter
;
2813 if (res
== chrec_dont_know
2815 || integer_zerop (may_be_zero
))
2817 else if (integer_nonzerop (may_be_zero
))
2818 res
= build_int_cst (TREE_TYPE (res
), 0);
2820 else if (COMPARISON_CLASS_P (may_be_zero
))
2821 res
= fold_build3 (COND_EXPR
, TREE_TYPE (res
), may_be_zero
,
2822 build_int_cst (TREE_TYPE (res
), 0), res
);
2824 res
= chrec_dont_know
;
2826 if (dump_file
&& (dump_flags
& TDF_SCEV
))
2828 fprintf (dump_file
, " (set_nb_iterations_in_loop = ");
2829 print_generic_expr (dump_file
, res
, 0);
2830 fprintf (dump_file
, "))\n");
2833 loop
->nb_iterations
= res
;
2837 /* Returns the number of executions of the exit condition of LOOP,
2838 i.e., the number by one higher than number_of_latch_executions.
2839 Note that unlike number_of_latch_executions, this number does
2840 not necessarily fit in the unsigned variant of the type of
2841 the control variable -- if the number of iterations is a constant,
2842 we return chrec_dont_know if adding one to number_of_latch_executions
2843 overflows; however, in case the number of iterations is symbolic
2844 expression, the caller is responsible for dealing with this
2845 the possible overflow. */
2848 number_of_exit_cond_executions (struct loop
*loop
)
2850 tree ret
= number_of_latch_executions (loop
);
2851 tree type
= chrec_type (ret
);
2853 if (chrec_contains_undetermined (ret
))
2856 ret
= chrec_fold_plus (type
, ret
, build_int_cst (type
, 1));
2857 if (TREE_CODE (ret
) == INTEGER_CST
2858 && TREE_OVERFLOW (ret
))
2859 return chrec_dont_know
;
2864 /* One of the drivers for testing the scalar evolutions analysis.
2865 This function computes the number of iterations for all the loops
2866 from the EXIT_CONDITIONS array. */
2869 number_of_iterations_for_all_loops (VEC(gimple
,heap
) **exit_conditions
)
2872 unsigned nb_chrec_dont_know_loops
= 0;
2873 unsigned nb_static_loops
= 0;
2876 FOR_EACH_VEC_ELT (gimple
, *exit_conditions
, i
, cond
)
2878 tree res
= number_of_latch_executions (loop_containing_stmt (cond
));
2879 if (chrec_contains_undetermined (res
))
2880 nb_chrec_dont_know_loops
++;
2887 fprintf (dump_file
, "\n(\n");
2888 fprintf (dump_file
, "-----------------------------------------\n");
2889 fprintf (dump_file
, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops
);
2890 fprintf (dump_file
, "%d\tnb_static_loops\n", nb_static_loops
);
2891 fprintf (dump_file
, "%d\tnb_total_loops\n", number_of_loops ());
2892 fprintf (dump_file
, "-----------------------------------------\n");
2893 fprintf (dump_file
, ")\n\n");
2895 print_loops (dump_file
, 3);
2901 /* Counters for the stats. */
2907 unsigned nb_affine_multivar
;
2908 unsigned nb_higher_poly
;
2909 unsigned nb_chrec_dont_know
;
2910 unsigned nb_undetermined
;
2913 /* Reset the counters. */
2916 reset_chrecs_counters (struct chrec_stats
*stats
)
2918 stats
->nb_chrecs
= 0;
2919 stats
->nb_affine
= 0;
2920 stats
->nb_affine_multivar
= 0;
2921 stats
->nb_higher_poly
= 0;
2922 stats
->nb_chrec_dont_know
= 0;
2923 stats
->nb_undetermined
= 0;
2926 /* Dump the contents of a CHREC_STATS structure. */
2929 dump_chrecs_stats (FILE *file
, struct chrec_stats
*stats
)
2931 fprintf (file
, "\n(\n");
2932 fprintf (file
, "-----------------------------------------\n");
2933 fprintf (file
, "%d\taffine univariate chrecs\n", stats
->nb_affine
);
2934 fprintf (file
, "%d\taffine multivariate chrecs\n", stats
->nb_affine_multivar
);
2935 fprintf (file
, "%d\tdegree greater than 2 polynomials\n",
2936 stats
->nb_higher_poly
);
2937 fprintf (file
, "%d\tchrec_dont_know chrecs\n", stats
->nb_chrec_dont_know
);
2938 fprintf (file
, "-----------------------------------------\n");
2939 fprintf (file
, "%d\ttotal chrecs\n", stats
->nb_chrecs
);
2940 fprintf (file
, "%d\twith undetermined coefficients\n",
2941 stats
->nb_undetermined
);
2942 fprintf (file
, "-----------------------------------------\n");
2943 fprintf (file
, "%d\tchrecs in the scev database\n",
2944 (int) htab_elements (scalar_evolution_info
));
2945 fprintf (file
, "%d\tsets in the scev database\n", nb_set_scev
);
2946 fprintf (file
, "%d\tgets in the scev database\n", nb_get_scev
);
2947 fprintf (file
, "-----------------------------------------\n");
2948 fprintf (file
, ")\n\n");
2951 /* Gather statistics about CHREC. */
2954 gather_chrec_stats (tree chrec
, struct chrec_stats
*stats
)
2956 if (dump_file
&& (dump_flags
& TDF_STATS
))
2958 fprintf (dump_file
, "(classify_chrec ");
2959 print_generic_expr (dump_file
, chrec
, 0);
2960 fprintf (dump_file
, "\n");
2965 if (chrec
== NULL_TREE
)
2967 stats
->nb_undetermined
++;
2971 switch (TREE_CODE (chrec
))
2973 case POLYNOMIAL_CHREC
:
2974 if (evolution_function_is_affine_p (chrec
))
2976 if (dump_file
&& (dump_flags
& TDF_STATS
))
2977 fprintf (dump_file
, " affine_univariate\n");
2980 else if (evolution_function_is_affine_multivariate_p (chrec
, 0))
2982 if (dump_file
&& (dump_flags
& TDF_STATS
))
2983 fprintf (dump_file
, " affine_multivariate\n");
2984 stats
->nb_affine_multivar
++;
2988 if (dump_file
&& (dump_flags
& TDF_STATS
))
2989 fprintf (dump_file
, " higher_degree_polynomial\n");
2990 stats
->nb_higher_poly
++;
2999 if (chrec_contains_undetermined (chrec
))
3001 if (dump_file
&& (dump_flags
& TDF_STATS
))
3002 fprintf (dump_file
, " undetermined\n");
3003 stats
->nb_undetermined
++;
3006 if (dump_file
&& (dump_flags
& TDF_STATS
))
3007 fprintf (dump_file
, ")\n");
3010 /* One of the drivers for testing the scalar evolutions analysis.
3011 This function analyzes the scalar evolution of all the scalars
3012 defined as loop phi nodes in one of the loops from the
3013 EXIT_CONDITIONS array.
3015 TODO Optimization: A loop is in canonical form if it contains only
3016 a single scalar loop phi node. All the other scalars that have an
3017 evolution in the loop are rewritten in function of this single
3018 index. This allows the parallelization of the loop. */
3021 analyze_scalar_evolution_for_all_loop_phi_nodes (VEC(gimple
,heap
) **exit_conditions
)
3024 struct chrec_stats stats
;
3026 gimple_stmt_iterator psi
;
3028 reset_chrecs_counters (&stats
);
3030 FOR_EACH_VEC_ELT (gimple
, *exit_conditions
, i
, cond
)
3036 loop
= loop_containing_stmt (cond
);
3039 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); gsi_next (&psi
))
3041 phi
= gsi_stmt (psi
);
3042 if (!virtual_operand_p (PHI_RESULT (phi
)))
3044 chrec
= instantiate_parameters
3046 analyze_scalar_evolution (loop
, PHI_RESULT (phi
)));
3048 if (dump_file
&& (dump_flags
& TDF_STATS
))
3049 gather_chrec_stats (chrec
, &stats
);
3054 if (dump_file
&& (dump_flags
& TDF_STATS
))
3055 dump_chrecs_stats (dump_file
, &stats
);
3058 /* Callback for htab_traverse, gathers information on chrecs in the
3062 gather_stats_on_scev_database_1 (void **slot
, void *stats
)
3064 struct scev_info_str
*entry
= (struct scev_info_str
*) *slot
;
3066 gather_chrec_stats (entry
->chrec
, (struct chrec_stats
*) stats
);
3071 /* Classify the chrecs of the whole database. */
3074 gather_stats_on_scev_database (void)
3076 struct chrec_stats stats
;
3081 reset_chrecs_counters (&stats
);
3083 htab_traverse (scalar_evolution_info
, gather_stats_on_scev_database_1
,
3086 dump_chrecs_stats (dump_file
, &stats
);
3094 initialize_scalar_evolutions_analyzer (void)
3096 /* The elements below are unique. */
3097 if (chrec_dont_know
== NULL_TREE
)
3099 chrec_not_analyzed_yet
= NULL_TREE
;
3100 chrec_dont_know
= make_node (SCEV_NOT_KNOWN
);
3101 chrec_known
= make_node (SCEV_KNOWN
);
3102 TREE_TYPE (chrec_dont_know
) = void_type_node
;
3103 TREE_TYPE (chrec_known
) = void_type_node
;
3107 /* Initialize the analysis of scalar evolutions for LOOPS. */
3110 scev_initialize (void)
3116 scalar_evolution_info
= htab_create_ggc (100, hash_scev_info
, eq_scev_info
,
3119 initialize_scalar_evolutions_analyzer ();
3121 FOR_EACH_LOOP (li
, loop
, 0)
3123 loop
->nb_iterations
= NULL_TREE
;
3127 /* Cleans up the information cached by the scalar evolutions analysis
3128 in the hash table. */
3131 scev_reset_htab (void)
3133 if (!scalar_evolution_info
)
3136 htab_empty (scalar_evolution_info
);
3139 /* Cleans up the information cached by the scalar evolutions analysis
3140 in the hash table and in the loop->nb_iterations. */
3153 FOR_EACH_LOOP (li
, loop
, 0)
3155 loop
->nb_iterations
= NULL_TREE
;
3159 /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3160 respect to WRTO_LOOP and returns its base and step in IV if possible
3161 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3162 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3163 invariant in LOOP. Otherwise we require it to be an integer constant.
3165 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3166 because it is computed in signed arithmetics). Consequently, adding an
3169 for (i = IV->base; ; i += IV->step)
3171 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3172 false for the type of the induction variable, or you can prove that i does
3173 not wrap by some other argument. Otherwise, this might introduce undefined
3176 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3178 must be used instead. */
3181 simple_iv (struct loop
*wrto_loop
, struct loop
*use_loop
, tree op
,
3182 affine_iv
*iv
, bool allow_nonconstant_step
)
3187 iv
->base
= NULL_TREE
;
3188 iv
->step
= NULL_TREE
;
3189 iv
->no_overflow
= false;
3191 type
= TREE_TYPE (op
);
3192 if (!POINTER_TYPE_P (type
)
3193 && !INTEGRAL_TYPE_P (type
))
3196 ev
= analyze_scalar_evolution_in_loop (wrto_loop
, use_loop
, op
,
3198 if (chrec_contains_undetermined (ev
)
3199 || chrec_contains_symbols_defined_in_loop (ev
, wrto_loop
->num
))
3202 if (tree_does_not_contain_chrecs (ev
))
3205 iv
->step
= build_int_cst (TREE_TYPE (ev
), 0);
3206 iv
->no_overflow
= true;
3210 if (TREE_CODE (ev
) != POLYNOMIAL_CHREC
3211 || CHREC_VARIABLE (ev
) != (unsigned) wrto_loop
->num
)
3214 iv
->step
= CHREC_RIGHT (ev
);
3215 if ((!allow_nonconstant_step
&& TREE_CODE (iv
->step
) != INTEGER_CST
)
3216 || tree_contains_chrecs (iv
->step
, NULL
))
3219 iv
->base
= CHREC_LEFT (ev
);
3220 if (tree_contains_chrecs (iv
->base
, NULL
))
3223 iv
->no_overflow
= !folded_casts
&& TYPE_OVERFLOW_UNDEFINED (type
);
3228 /* Runs the analysis of scalar evolutions. */
3231 scev_analysis (void)
3233 VEC(gimple
,heap
) *exit_conditions
;
3235 exit_conditions
= VEC_alloc (gimple
, heap
, 37);
3236 select_loops_exit_conditions (&exit_conditions
);
3238 if (dump_file
&& (dump_flags
& TDF_STATS
))
3239 analyze_scalar_evolution_for_all_loop_phi_nodes (&exit_conditions
);
3241 number_of_iterations_for_all_loops (&exit_conditions
);
3242 VEC_free (gimple
, heap
, exit_conditions
);
3245 /* Finalize the scalar evolution analysis. */
3248 scev_finalize (void)
3250 if (!scalar_evolution_info
)
3252 htab_delete (scalar_evolution_info
);
3253 scalar_evolution_info
= NULL
;
3256 /* Returns true if the expression EXPR is considered to be too expensive
3257 for scev_const_prop. */
3260 expression_expensive_p (tree expr
)
3262 enum tree_code code
;
3264 if (is_gimple_val (expr
))
3267 code
= TREE_CODE (expr
);
3268 if (code
== TRUNC_DIV_EXPR
3269 || code
== CEIL_DIV_EXPR
3270 || code
== FLOOR_DIV_EXPR
3271 || code
== ROUND_DIV_EXPR
3272 || code
== TRUNC_MOD_EXPR
3273 || code
== CEIL_MOD_EXPR
3274 || code
== FLOOR_MOD_EXPR
3275 || code
== ROUND_MOD_EXPR
3276 || code
== EXACT_DIV_EXPR
)
3278 /* Division by power of two is usually cheap, so we allow it.
3279 Forbid anything else. */
3280 if (!integer_pow2p (TREE_OPERAND (expr
, 1)))
3284 switch (TREE_CODE_CLASS (code
))
3287 case tcc_comparison
:
3288 if (expression_expensive_p (TREE_OPERAND (expr
, 1)))
3293 return expression_expensive_p (TREE_OPERAND (expr
, 0));
3300 /* Replace ssa names for that scev can prove they are constant by the
3301 appropriate constants. Also perform final value replacement in loops,
3302 in case the replacement expressions are cheap.
3304 We only consider SSA names defined by phi nodes; rest is left to the
3305 ordinary constant propagation pass. */
3308 scev_const_prop (void)
3311 tree name
, type
, ev
;
3313 struct loop
*loop
, *ex_loop
;
3314 bitmap ssa_names_to_remove
= NULL
;
3317 gimple_stmt_iterator psi
;
3319 if (number_of_loops () <= 1)
3324 loop
= bb
->loop_father
;
3326 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); gsi_next (&psi
))
3328 phi
= gsi_stmt (psi
);
3329 name
= PHI_RESULT (phi
);
3331 if (virtual_operand_p (name
))
3334 type
= TREE_TYPE (name
);
3336 if (!POINTER_TYPE_P (type
)
3337 && !INTEGRAL_TYPE_P (type
))
3340 ev
= resolve_mixers (loop
, analyze_scalar_evolution (loop
, name
));
3341 if (!is_gimple_min_invariant (ev
)
3342 || !may_propagate_copy (name
, ev
))
3345 /* Replace the uses of the name. */
3347 replace_uses_by (name
, ev
);
3349 if (!ssa_names_to_remove
)
3350 ssa_names_to_remove
= BITMAP_ALLOC (NULL
);
3351 bitmap_set_bit (ssa_names_to_remove
, SSA_NAME_VERSION (name
));
3355 /* Remove the ssa names that were replaced by constants. We do not
3356 remove them directly in the previous cycle, since this
3357 invalidates scev cache. */
3358 if (ssa_names_to_remove
)
3362 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove
, 0, i
, bi
)
3364 gimple_stmt_iterator psi
;
3365 name
= ssa_name (i
);
3366 phi
= SSA_NAME_DEF_STMT (name
);
3368 gcc_assert (gimple_code (phi
) == GIMPLE_PHI
);
3369 psi
= gsi_for_stmt (phi
);
3370 remove_phi_node (&psi
, true);
3373 BITMAP_FREE (ssa_names_to_remove
);
3377 /* Now the regular final value replacement. */
3378 FOR_EACH_LOOP (li
, loop
, LI_FROM_INNERMOST
)
3381 tree def
, rslt
, niter
;
3382 gimple_stmt_iterator bsi
;
3384 /* If we do not know exact number of iterations of the loop, we cannot
3385 replace the final value. */
3386 exit
= single_exit (loop
);
3390 niter
= number_of_latch_executions (loop
);
3391 if (niter
== chrec_dont_know
)
3394 /* Ensure that it is possible to insert new statements somewhere. */
3395 if (!single_pred_p (exit
->dest
))
3396 split_loop_exit_edge (exit
);
3397 bsi
= gsi_after_labels (exit
->dest
);
3399 ex_loop
= superloop_at_depth (loop
,
3400 loop_depth (exit
->dest
->loop_father
) + 1);
3402 for (psi
= gsi_start_phis (exit
->dest
); !gsi_end_p (psi
); )
3404 phi
= gsi_stmt (psi
);
3405 rslt
= PHI_RESULT (phi
);
3406 def
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
3407 if (virtual_operand_p (def
))
3413 if (!POINTER_TYPE_P (TREE_TYPE (def
))
3414 && !INTEGRAL_TYPE_P (TREE_TYPE (def
)))
3420 def
= analyze_scalar_evolution_in_loop (ex_loop
, loop
, def
, NULL
);
3421 def
= compute_overall_effect_of_inner_loop (ex_loop
, def
);
3422 if (!tree_does_not_contain_chrecs (def
)
3423 || chrec_contains_symbols_defined_in_loop (def
, ex_loop
->num
)
3424 /* Moving the computation from the loop may prolong life range
3425 of some ssa names, which may cause problems if they appear
3426 on abnormal edges. */
3427 || contains_abnormal_ssa_name_p (def
)
3428 /* Do not emit expensive expressions. The rationale is that
3429 when someone writes a code like
3431 while (n > 45) n -= 45;
3433 he probably knows that n is not large, and does not want it
3434 to be turned into n %= 45. */
3435 || expression_expensive_p (def
))
3441 /* Eliminate the PHI node and replace it by a computation outside
3443 def
= unshare_expr (def
);
3444 remove_phi_node (&psi
, false);
3446 def
= force_gimple_operand_gsi (&bsi
, def
, false, NULL_TREE
,
3447 true, GSI_SAME_STMT
);
3448 ass
= gimple_build_assign (rslt
, def
);
3449 gsi_insert_before (&bsi
, ass
, GSI_SAME_STMT
);
3455 #include "gt-tree-scalar-evolution.h"